My thirteenth prediction was the 'beyond the hype' grab-bag — healthcare plus the crypto/blockchain trust angle. Closing the forecast, I want to grade it the way I'd grade a portfolio: by which bets paid and which merely held.

Prediction thirteen: healthcare attracts significant AI investment and broad adoption in diagnostics and personalized care; crypto markets stay volatile but AI-enhanced blockchain — for authenticity and trust — sees renewed interest. My favorite spin-offs were a pediatric/parenting 'one-stop' AI assistant and blockchain-AI synergy for verifying outputs like synthetic-media authenticity.

Healthcare: the loud, clear win

Healthcare AI in 2026 is exactly the surge I predicted, and then some. Around 80% of hospitals now use AI in at least one clinical or operational function. The FDA had cleared roughly 1,250-1,300 AI/ML-enabled medical devices by mid-2025, radiology dominant — AI is now the default 'second reader' in diagnostic imaging. The diagnostics, imaging, genomics, and personalized-medicine adoption I named is real and measured.

The regulatory tailwind exceeded my expectations. The US didn't just invest in healthcare AI — it cleared the runway. The FDA's January 6, 2026 guidance relaxed clinical-decision-support device requirements, letting many generative diagnostic tools reach clinics without full premarket sign-off. New federal models — CMS ACCESS, FDA TEMPO Pilot — let selected manufacturers deploy without premarket authorization while covered. The FDA even deployed agentic AI in its own workflows as of December 2025. I predicted significant healthcare investment and adoption. I underestimated how aggressively US policy would accelerate it — which ties directly back to my prediction-2 miss about the US going build-fast rather than precautionary.

My pediatric/parenting 'one-stop AI assistant' spin-off didn't emerge as a single dominant product, but the category — AI-driven personalized health guidance with clinician oversight, ambient AI as 'the shell around the visit' — is exactly where consumer and clinical health tech converged. Right category, no single breakout name. Partial credit on the spin-off, full credit on the structural call.

The blockchain-trust angle: quieter, and reframed as provenance

Here's the nuanced grade. I predicted AI-enhanced blockchain would find renewed interest specifically for authenticity and trust — verifying outputs, fighting synthetic-media fraud. The literal 'blockchain' framing got quieter in 2025-2026 as crypto did its usual volatility dance. But the underlying need I identified — cryptographic verification of authenticity in an AI-saturated world — became one of the most important problems in tech, just under different branding.

Content provenance and authenticity verification — SynthID, C2PA, and the disclosure mandates now rolling out across jurisdictions — are precisely the 'AI-enhanced trust and authenticity' use case I flagged, addressing exactly the synthetic-media-verification problem I named. Separately, the agent-to-agent payment rails (prediction 3) are where crypto's machine-commerce thesis actually found traction. So the trust-and-authenticity need I predicted is real and growing; it just expressed itself as provenance standards and machine-payment protocols more than as headline 'blockchain' products. I had the need right and the label slightly dated.

The next 12 months: healthcare agentic care and provenance-by-default

My forecast for May 2027: healthcare AI crosses from documentation and diagnostics into agentic care — autonomous AI agents drafting personalized interventions and care adjustments for physician review, not just flagging data. The FDA's own agentic deployment and the ACCESS/TEMPO frameworks are the leading indicators. The bottleneck shifts from 'can AI do it' to reimbursement and liability — who pays, and who's responsible when an agent's care suggestion is wrong. Watch the CPT codes and Medicare benefit-category fights; that's where healthcare AI's 2027 economics get decided.

On the trust side: provenance becomes default infrastructure. As AI-generated content saturates everything, verifiable authenticity stops being optional and becomes a baseline requirement — embedded in cameras, platforms, and enterprise content pipelines. The 'AI-enhanced trust' need I named in January 2025 matures into invisible plumbing, the way HTTPS did. Whether the verification layer runs on a blockchain or a simpler cryptographic standard matters less than that it exists everywhere.

Closing the scorecard

Thirteen predictions, seventeen months, one honest reckoning. The headline: I was substantially right on the structural calls — narrow-AI-not-AGI, agent swarms, the one-person unicorn, humanoids as niche labor, the safe-zone autonomous map, energy as the binding constraint, the oligopoly-and-correction setup, AI video adoption, the 90% automation, and the healthcare surge. Where I missed, I missed on mechanism (China's open-source vector, the test-time-compute shift) and on second-order human cost (the disappearing entry-level rung), not on direction.

The forecaster's lesson I'm taking into the next thirteen predictions: getting the direction right is the easy 70%. The hard 30% — and the part that separates a useful forecast from a lucky one — is the mechanism and the second-order consequence. That's where I'll push harder next time. Same beach. Same coast. Sharper pencil.

____________________________________________

_{Sources: original forecast (Jan 6 2025); FDA AI-device clearance data and Jan 6 2026 clinical-decision-support guidance; CMS ACCESS / FDA TEMPO frameworks; MobiHealthNews and Chief Healthcare Executive 2026 predictions; Bipartisan Policy Center on AI CPT codes; SynthID/C2PA provenance standards; agent-to-agent payment protocols.}

First, the chart everyone is misreading.

Take global GDP since 1990 and split it into three economies. The non-digital economy — steel, food, construction, everything your grandfather would recognize — grew from roughly $22 trillion to $98 trillion. A 4.4x, and most of that is inflation plus China industrializing. The digital economy went from about half a trillion to $20 trillion, now 17.3% of world GDP per IDCA's 2026 research. That's 40x. Same planet, same 35 years, ten times the growth multiple.

The third line is AI. Today it's maybe $1.3 trillion — roughly 1% of world GDP. On the chart it looks like a rounding error hugging the x-axis.

That is exactly what the internet line looked like in 1995.

The bubble question is the wrong question

Here is what the bubble callers get right: valuations are stretched, capex is enormous, and most enterprise pilots fail. Fine. All true in 1999 too.

Here is what they get wrong: they think the equity bubble and the value creation are the same event. They never are. The dot-com crash vaporized $5 trillion in market cap between 2000 and 2002. Cisco shareholders lost 80% and waited two decades to break even. And the infrastructure that capital paid for — the fiber, the data centers, the protocols — went on to carry a digital economy now worth $20 trillion a year. Every year. Forever.

The bubble was real.

The value was realer.

Those are two different charts, and confusing them is how you end up smugly correct in 2001 and structurally poor by 2010.

There's also a mechanical difference this cycle that the 1999 comparison ignores. The dark fiber laid in the telecom boom sat mostly unlit for years — capacity waiting for demand that hadn't arrived. GPU capacity today is the opposite problem: demand queues for supply. IDC clocked AI infrastructure spending at nearly $90 billion in Q4 2025 alone, 62% year-over-year growth, and projects $487 billion for 2026 with the market passing $1 trillion by 2029. The St. Louis Fed measured information-processing equipment contributing 0.90 percentage points to US real GDP growth in Q1 2025 — more than two standard deviations above its long-run average. You can argue about whether Nvidia is worth its multiple. You cannot argue with utilization.

Could specific equities crater? Absolutely. Will some of the circular vendor-financing deals end in tears? Some will. Does any of that stop the red line on my chart from doing in one decade what the blue line needed three to do? IDC models AI driving 3.5% of global GDP by 2030 with $19.9 trillion of cumulative impact. Even if they're off by half, it's the largest value migration of your working life.

Disclosure, because I do this inline and always will: I run DCXPS, we build modular intelligence factories on stranded power, and our first $45M unit commissions in January 2027. I am talking my book. The book is public; check the math yourself.

The founders went quiet and started typing

Now the part that should actually worry the bubble callers.

The loudest bubble arguments come from people whose relationship with AI is a stock ticker and a podcast feed. Meanwhile, the people with the longest track records of being early went silent about eighteen months ago and did something almost nobody predicted.

They started coding again.

Garry Tan runs Y Combinator — arguably the best seat in the world for pattern-matching what's next. He didn't write a memo about agents. He published his personal agentic toolchain and the receipts: three production services and forty-plus shipped features in sixty days, part-time, a self-measured ~810x increase over his 2013 coding pace. Andrej Karpathy, one of the defining engineers of the deep-learning era, put it in eleven words this year: "I don't think I've typed a line of code since December." The man who taught half the industry to program stopped typing — and became more productive.

I watch the same thing every week from my own seat. Fund partners asking for agent-tool licenses before their analysts do. Founders in their fifties who last shipped code when Nokia was a phone company, now running overnight agent sessions. A generation of operators who spent twenty years delegating implementation suddenly demanding to touch it again.

Why?

Because the abstraction penalty flipped.

For thirty years, the rational move for a capital allocator was to stay high in the stack: strategy up here, implementation down there, a management layer translating between. AI collapsed the translation layer. When one person with agents can do what previously required a team, the person who can operate the agents directly holds an information advantage over the person reading their summary. The founders re-learning the stack aren't nostalgic. They're refusing to price an asset class through an interpreter.

Klarna showed the corporate version: its AI assistant reportedly handled two-thirds of customer-service chats in its first month, work equivalent to roughly 700 full-time agents. Accenture showed the brutal version — 11,000 positions eliminated between June and August 2025 while revenue rose 7% to $69.7 billion and GenAI bookings jumped from $3 billion to $5.1 billion. Read that pair of numbers again. Revenue up, humans down, AI bookings nearly doubled. That's not a company betting on a bubble. That's a company repricing labor in public.

Follow the money into the ground

If this were a speculative mania, the money would concentrate in the most liquid, most flippable layer: model-company equity, AI-labeled SaaS, tokens of things. Some does. But look at where the largest checks actually land: Meta committing $600 billion to data centers and AI infrastructure through 2028. Microsoft's $80 billion year. Google's $85 billion. Nvidia and TSMC forming a $500 billion manufacturing partnership. Sovereigns in the Gulf posting the fastest regional infrastructure growth on earth in Q4 2025.

Concrete.

Transformers.

Cooling loops.

Twenty-year assets.

Speculators buy what they can sell by Friday.

This capital is buying things that take years to energize, in a grid environment where 2,300+ GW sits stuck in US interconnection queues and historically about 14% ever reaches commercial operation, where PJM capacity prices went from $28.92 to $329.17 per MW-day in one auction cycle — an 11x. Nobody pours patient capital into a decade-long power problem because a chatbot demo impressed them. They do it because they've concluded compute is the scarce input of the next economy and electricity is the scarce input of compute.

The model layer, meanwhile, is getting squeezed from both ends — commoditized from below by open-weight releases, and increasingly subject to government direction from above. When the middle of a stack compresses, value migrates to the layers that can't be copied with a download: the physical infrastructure underneath and the orchestration on top. The smart money read that memo early. It's why the checks go into the ground.

The wealth landscape, 2031

So play it forward five years. Three shifts, and they compound.

It will close the same way.

Quietly.

Around the time your company hires its first "VP of Agent Operations" and the skill becomes a job description instead of an edge.

Share

The uncomfortable close

In 2000, the correct call was: the stocks are overpriced and the internet is underestimated. Both. Simultaneously. Almost nobody could hold both thoughts, so one crowd got wiped out in the crash and the other missed the largest value creation event of the century.

Same test, new era.

Yes, parts of the AI trade will burn.

And the red line at the bottom of my chart — 1% of world GDP, growing faster than anything in economic history — will look like the blue line looks today: obvious, enormous, and owned by the people who stopped debating it in 2026 and learned to operate it instead.

The skeptics are writing think-pieces.

The believers are writing code.

In five years, check which group changed its wealth bracket.

Prediction twelve: as the hype matures into pragmatic deployment, AI automates up to 90% of certain job tasks — entry-level coding, customer interactions, document review — making traditional methods in those areas uncompetitive.

My favorite spin-off was 'AI-Powered Outsourcing': specialized AI outsourcing departments and services that drastically reduce the need for human labor on repetitive tasks, with industry-specific marketplaces of legal, finance, and HR modules.

The 90% automation call: landed

The specific tasks I named are exactly the ones AI now dominates.

Entry-level coding: AI coding assistants write or co-write the majority of new code at many firms, and the 2026 adoption stats (Stack Overflow, GitHub, DORA, McKinsey) confirm AI-assisted development is the norm, not the exception.

Document review: AI scribes and summarization tools produce the largest measured time savings in knowledge work, with healthcare documentation as the flagship case.

Customer interactions: AI handles the high-volume, repetitive tier as a default.

The 'traditional methods become uncompetitive' framing held — firms still doing this work manually are simply more expensive than firms that automated it.

AI outsourcing: real, and reshaping the BPO industry

My 'AI-Powered Outsourcing' concept — AI replacing the traditional outsourcing-department model for repetitive tasks — is materializing as the genuine disruption of the business-process-outsourcing industry.

The logic I laid out (hyper-personalized, end-to-end-integrated, industry-specific automation marketplaces) is the direction BPO is being forced to move.

The companies whose entire business was offshoring repetitive cognitive labor to lower-cost regions are now competing against AI that does the same work at near-zero marginal cost. That's the structural squeeze I described, arriving on schedule.

Here's where I connect it to prediction 5 — and where I refuse to sugarcoat.

The 90% automation of entry-level tasks isn't a neutral efficiency story. Entry-level coding, junior document review, first-tier customer work — these aren't just tasks. They're how humans entered professions and climbed.

When AI does 90% of them, the entry-level job doesn't get more interesting. It gets eliminated or compressed to a fraction of its former headcount.

I framed this in January 2025 as 'human talent reallocation' to higher-value roles — and for the people already established, that reallocation is real and good.

But the cohort that would have started in those entry-level seats faces a structural problem I underweighted: you can't reallocate to a senior role you never got to begin.

The 90% automation I correctly predicted is, viewed from the other side, the mechanism removing the bottom rung of the career ladder across multiple white-collar fields.

Both things are true at once. The efficiency gain is real and I called it. The generational discontinuity is real and I should have weighted it more heavily. A good forecaster names the second-order effect, not just the first.

The next 12 months: the BPO reckoning and the apprenticeship counter-move

My forecast for May 2027: the traditional BPO and outsourcing industry hits its first real contraction directly attributable to AI — not a productivity story, a headcount story.

Firms whose value proposition was cheap repetitive cognitive labor lose contracts to AI-native competitors, and the offshore-BPO employment model that absorbed millions of entry-level workers globally starts to shrink.

This is one of the most under-discussed labor consequences of the automation I predicted, and it lands hardest outside the headlines — in the economies built on outsourced back-office work.

The counter-move worth watching: the smartest firms turn the missing rung into a deliberate strategy.

If everyone else stops hiring juniors because AI does junior work, the firm that builds a real human-apprenticeship-alongside-AI pipeline will own the senior talent of 2030 while competitors discover their pipeline ran dry.

The scarce asset stops being the AI. It becomes the human who has been deliberately grown alongside it.

Grade: hit on the automation, hit on the AI-outsourcing disruption, with a mandatory honesty tax — I called the 90% and underweighted what removing the bottom 90% does to the people who used to live there. The prediction was right. The framing was too comfortable.

^{Sources: original forecast (Jan 6 2025); 2026 AI coding adoption stats (Stack Overflow, GitHub Octoverse, DORA, McKinsey); healthcare AI scribe time-savings data; BPO industry disruption analysis; the author's own prediction 5 on workforce evolution; entry-level labor-market data 2025-26.}

My favorite spin-off was the 'Personalized Echo Chamber' — AI-driven, human-approved 'internal TV stations' for corporations and 'aligned broadcast networks' for political entities, with explicit warnings about misinformation and reinforcement.

AI video became enterprise infrastructure — fast

By early 2026, AI video generation crossed from novelty to production tool.

Google's Veo 3 generating 4K clips up to 60 seconds with native audio. Sora 2 holding character and scene continuity across shots.

Runway, Kling, Synthesia, HeyGen — the latter shipping synthetic presenters in 175+ languages with native lip-sync. Studios are cutting production costs 70-90%. The dominant use cases are exactly the ones I named: corporate training, onboarding, product demos, marketing, internal communications.

Enterprises adopting AI for commercial and consumer-facing media — that was the prediction, and it's now the default workflow.

Crucially, the market positioned itself precisely along my 'humans keep strategy' line. The 2026 consensus on AI video tools is that they produce a usable first draft, and humans handle brand alignment, emotional resonance, and final approval.

The tools got reframed from 'type a prompt, get a finished video' (the 2024 fantasy) toward editability — storyboard, remix, recut, human-in-the-loop.

The Sora consumer collapse: my caveat, written in blood

Here's the part that delights the forecaster in me. OpenAI's Sora launched as a consumer app and exploded — over 1 million downloads in under five days in late 2025.

Then it cratered.

Downloads fell 32% month-on-month in December 2025, another 45% in January 2026. Cumulative in-app purchase revenue from launch to shutdown: about $2.1 million total, against operating costs estimated near $1 million per day.

OpenAI confirmed the consumer app shuts down April 26, 2026. The Disney partnership collapsed over unsustainable economics. Deepfakes and guardrail-bypassing ran rampant.

Why did the consumer toy fail while enterprise video thrived?

Because the consumer app removed exactly the thing I said had to stay: human strategic oversight.

Unsupervised, prompt-anything video generation produced slop, abuse, and IP nightmares.

The enterprise tools that kept humans on brand alignment and approval — Veo on Vertex AI, Synthesia avatars for corporate content — became durable businesses.

My prediction wasn't just that enterprises would adopt AI video. It was that the human-approval layer was load-bearing. Sora's consumer failure is the cleanest possible proof of that caveat.

The Personalized Echo Chamber: arriving, with the dangers I flagged

My 'Personalized Echo Chamber' concept — organizations running AI-powered internal broadcast networks, with explicit misinformation risk — is materializing as hyper-personalized video at scale and AI-generated corporate/political content.

The provenance and disclosure infrastructure I implied was necessary is now being mandated: most platforms and jurisdictions require AI-content disclosure, and standards like SynthID and C2PA are being deployed. I called both the capability and the need for transparency safeguards. Both arrived.

The next 12 months: the buyer splits three ways

My forecast for May 2027: AI video consolidates around the buyer, not the model — consumer creators in the Gemini/YouTube ecosystem, agencies and tool-builders on programmatic APIs, and enterprise marketing teams on stable, integrated, indemnified platforms like Veo on Vertex.

The 'best model' framing dies; the 'right tool for your buyer and your risk tolerance' framing wins. Benchmark wars between Sora, Veo, and the rest will keep missing the point: these are three different products for three different buyers.

Watch provenance become a purchasing requirement, not a nicety. As deepfake liability and IP exposure mount, enterprises will only buy AI video from vendors who indemnify them and embed verifiable provenance.

The 'human-approved, traceable' layer I described in January 2025 becomes a procurement checkbox. The vendors who treated it as optional get filtered out.

_{Sources: original forecast (Jan 6 2025); Veo 3 / Sora 2 capability reporting (2026); Constellation Research and digitalapplied on Sora consumer shutdown (Mar-Apr 2026); HeyGen/Synthesia enterprise avatar adoption; SynthID/C2PA provenance standards; AI-content disclosure mandates 2026.}

Prediction ten: runaway valuations could trigger a market correction; high-burn startups get forced to consolidate or fold; a handful of players dominate foundational models as an oligopoly; lean, cost-efficient, specialized vendors survive and thrive through any correction. My favorite spin-off was the 'AI Application Powerhouse' — an app-store-style meta-layer aggregating AI tools.

The oligopoly: confirmed and entrenched

The foundation-model oligopoly I predicted is now simply the structure of the industry. OpenAI, Anthropic, Google at the frontier; a tight cluster of hyperscalers controlling the compute they run on.

Anthropic crossed $30B annualized revenue by April 2026; OpenAI around $24B. The capital required to play at the frontier — OpenAI's reported $1.4 trillion in multi-year compute commitments — is itself the moat that keeps the oligopoly closed.

Exactly as predicted: a handful of players who can afford the massive models dominate and shape pricing.

The correction: every precondition present, trigger still pending

Here's the careful part. I predicted a correction was possible. As of May 2026, the bubble debate has gone fully mainstream — and the specific mechanism analysts now fixate on is the one I'd have flagged: circular financing.

$1.4T — OpenAI's reported multi-year compute commitment

~$13B — OpenAI 2025 revenue against that commitment

95% — of enterprises reporting zero measurable return on genAI investment (MIT Media Lab)

$5T+ — Nvidia valuation peak, October 2025

And the demand-side warning is just as loud: an MIT Media Lab report found that despite $30-40 billion in enterprise genAI investment, 95% of organizations reported zero measurable return. That's the gap between deployment and value that breaks investment theses. I'm not calling the exact week of a correction — anyone who does is guessing. But every structural precondition I named in January 2025 is now visibly in place, and the mainstream financial press is naming them too.

Cost-efficiency as survival: validated by DeepSeek

My claim that cost-efficiency would be the survival trait got its proof point in DeepSeek — frontier-ish performance at a fraction of the cost, which repriced the entire conversation about what a model 'should' cost to train and run. The lean, efficiency-focused players didn't just survive; they forced the giants to defend their economics. As I keep saying: when oil becomes a commodity, Exxon still makes money — but not at 21x revenue multiples. The model layer is commoditizing, and the premium is the thing at risk, not the market.

The 'AI Application Powerhouse': partially built

My app-store-meta-layer concept partially materialized — aggregators and platforms bundling AI tools emerged, and the 'one-stop AI platform' is a recognized category. But no single dominant 'AI app store' claimed the aggregator crown the way I floated. The meta-layer is fragmented, not consolidated. Partial credit on the spin-off, full credit on the structural prediction.

The next 12 months: the trigger, or the grind

My forecast for May 2027: either a visible correction hits — most likely triggered by one high-burn AI company failing to refinance and forcing a repricing through the circular-financing loops — or the market grinds higher on the argument that, unlike 1999, these firms have real revenue and real cash flow (Morgan Stanley notes corporate cash flow is triple its 1999 level).

Watch the refinancing calendar, not the stock charts. The AI startups with enormous burn must refinance within the next couple of years, and lenders will finally demand revenue proof.

The moment one can't, the circular loops around Nvidia, OpenAI, and CoreWeave start to unwind — and unwinding is faster than building. That's the trigger I'd watch through 2027.

_{Sources: original forecast (Jan 6 2025); Bloomberg AI circular-deals analysis (Mar 2026); INSEAD 'Are We in an AI Bubble?' (Feb 2026); GMO via Medium analysis (Mar 2026); MIT Media Lab enterprise-genAI return study; DeepSeek cost reporting; Anthropic/OpenAI revenue figures 2026; Morgan Stanley cash-flow comparison.}

Seventeen months later: CNBC confirms the administration and OpenAI have discussed a government equity stake, with talks tracking back over a year.

Sam Altman has reportedly been discussing it with officials since the second term began. A sovereign wealth fund stands chartered by executive order since February, waiting for shares to be ceded into it — returns to flow back to citizens as an “AI dividend.”

A White House meeting is reportedly being arranged with the leading labs to explore what deeper government involvement looks like. Bernie Sanders, from the other end of the spectrum entirely, is drafting a one-time 50% equity tax on the top labs, shares deposited into the fund with voting power and board seats.

Left hand, right hand, same pocket.

Frontier AI companies are no longer being treated as startups. They are being treated, in the now-standard phrasing, as strategic infrastructure.

Under the sodium light of a Prague June night, let me say the quiet part in mono: the state does not regulate what it can own.

And this week, three stories that look unrelated — the stake talks, the labs publishing evidence of models improving the models, and IPOs engineered to float almost nothing — snapped into one structure.

I’m going to name it before someone in a think tank does it worse.

The Receipts — graded in public, as always

What I got right, January 2025: the U.S. goes security-first on AI. The Scorecard graded it in May and the call held — federal policy reorganized around security primacy, the chip war became the industry’s central fact, build-fast doctrine justified in the language of national survival.

What I wrote in April, when Anthropic handed Mythos Preview to twelve corporations — Apple, Microsoft, Google, Nvidia, JPMorgan, the Pentagon’s whole cloud choir — and zero government agencies: whoever decides who gets access to Mythos-class capability is making a decision that shapes cybersecurity, defense, and economic competition for years.

A private company made that decision in a blog post. I asked, in print, how long Washington would tolerate reading about its own strategic posture in someone else’s changelog.

Answer: about eight weeks.

Now the miss, because misses get more words than hits in this letter. I predicted the state’s grip would arrive as regulation — export-control logic, licensing, the compliance perimeter. Wrong instrument.

The state skipped the rulebook and reached for the cap table. I also thought the squeeze would take a decade. It took eighteen months from prediction to confirmed negotiation. Direction: hit. Mechanism: miss. Timing: embarrassingly conservative.

When I’m wrong, I’m usually wrong about how fast.

The trigger: the lab published the chart

Why now? Because the recursion stopped being a thought experiment and started being telemetry — published by the lab itself.

Anthropic released a chart of lines of code merged per active engineer, expressed as a multiple of its pre-2025 average.

Read it like an ECG: 1.2× in Q1 2025. 1.5×. 1.9×. 2.5×. Then 5.8× in Q1 2026. Then — with Mythos Preview running internally — 8.0× in the current quarter. Flat line for four years, then a vertical.

On open-ended coding problems where the right answer is unclear, Claude’s reported success rate hit 76% — a 50-point jump in six months — and Anthropic says its engineers now rate the model’s code as comparable to human-written, expecting it to surpass that level within the year.

The same company published a blog this week titled When AI Builds Itself, describing the risks of AI progressing to where it can improve itself autonomously. Not a doomer Substack. The lab. In its own changelog voice.

Yoshua Bengio — a literal grandfather of the field — responded that if labs are truly approaching recursive self-improvement, “a coordinated, verifiable, and globally respected pause” may be the only responsible path, possible “if others follow Anthropic’s lead.” And as the Towards AI write-up dryly noted: almost nobody expects a real pause to happen.

Stack the rest on top. Fable runs nine-plus hours unattended; it doubled FrontierCode Diamond in one generation; it reportedly throttles itself, silently, on exactly one category of work — frontier-LLM development. The lab drew a border checkpoint around recursion before the government did.

When the machine can improve the machine, it stops being a product and becomes a cascade — and America has exactly one precedent for cascades. In 1946, the Atomic Energy Act nationalized fissionable material and invented a legal category called “born secret”: certain knowledge classified at the instant of its creation.

In 1946 America invented born secret. In 2026 it is drafting born sovereign.

Two doors into the state

Here is the new texture the week added: the two leading labs are entering the state through different doors.

OpenAI is taking the Treasury door. Equity talks since the start of the second term. The sovereign wealth fund. The AI dividend. Altman and Sanders reportedly hold different visions of what the government’s role should be — partnership versus expropriation — but notice what they no longer disagree about: that the government will hold a position. The negotiation is over terms, not over whether.

Anthropic is taking the Fort Meade door. It does not appear to be in the equity conversations at all. Instead, per an FT report, it has placed “forward-deployed” engineers inside the National Security Agency to help the agency operate the restricted Mythos model for cyber operations.

It is reportedly working more closely with the Department of War while absorbing pressure from a designation as a supply-chain risk.

And it has reportedly filed its IPO prospectus confidentially with the SEC while announcing Mythos access for the European Union plus 150 partners across 15 countries — capability distribution conducted with the cadence of foreign policy, not product marketing.

One lab is negotiating with the Treasury. The other is embedding with the security state. Different doors, same building.

Equity is just the most legible form of capture; engineers inside the NSA is the less legible, more binding one. Either way, the April question — who decides who gets Mythos-class capability? — now has its answer forming in real time: increasingly, not the lab alone.

The precedent stack

None of this is improvisation. 1946: the Atomic Energy Act — the cascade nationalized. 1962: COMSAT — Congress charters a public-private hybrid because satellites were too strategic for either sector alone. 2008: Treasury takes 79.9% of AIG over a weekend, because some balance sheets are load-bearing. 2025–26: a ~10% stake in Intel. Stakes in IBM, quantum, critical minerals.

A golden share in U.S. Steel. A 15% cut of Nvidia and AMD’s China revenue. The February sovereign-wealth-fund order — the empty vault, waiting.

And look at the mechanics being floated for the labs, because this is where my entire thesis takes a bow. The framework isn’t the government writing checks. It’s barter: equity in exchange for federal land for data centers, accelerated energy permits, grid access at scale, security clearances.

The currency Washington is offering isn’t dollars. It’s interconnection.

Two thousand three hundred gigawatts sit in U.S. grid queues; fourteen percent historically reach operation. The labs are drowning in capability and starving for power, and the state noticed that its most valuable chip hums at 60 hertz. Every reader who watched me bang on about energy as the binding constraint for two years: this is what the constraint looks like when it learns to negotiate.

The construct: the Sovereign Float

Now the part you came for — why “in the IPO there will not be so much shares to sell,” and why that’s the design, not a defect. I’m calling it the Sovereign Float: when a technology becomes a strategic asset, its IPO stops being a sale and becomes a pricing ceremony. Control never trades. Six mechanisms lock the door:

One — the state locks. Government stakes, golden shares, sovereign-fund positions don’t float. Every share ceded to the fund is permanently subtracted from the sellable pool. Sanders’ version subtracts half the company at birth.

Two — the screen shrinks the buyers. Once Washington sits on the cap table, the cap table becomes a security perimeter. Foreign ownership gets capped, reviewed, unwound — ask MGX, the Abu Dhabi fund in OpenAI’s March round, how comfortable that seat stays.

Three — strategic holders hoard. Microsoft holds OpenAI for access, not returns. Google and Amazon hold Anthropic the same way. Compute partners hold for allocation. None of it ever hits the tape.

Four — shares are currency, not product. The listed paper exists to acquire things and pay people. SpaceX filed an exclusive right to buy Cursor — $4B of ARR — for $60 billion in largely its own freshly minted stock. You don’t sell your ammunition.

Five — the float is the collateral. A public mark lets you borrow against the whole at the price of the sliver. SpaceX carries a $20B bridge into Friday; the float’s job is to print the number the refinancing rests on.

Six — the index is the buyer of last resort. This is the engine the others bolt onto. Nasdaq 100 fast-track entry rules have been rewritten so trillion-dollar debutantes reach index inclusion far faster than companies historically could. Once inside, passive funds and ETFs must buy — trillions in capital that doesn’t read prospectuses, flowing in, as the Towards AI piece put it, almost automatically.

Forced buyers meet engineered scarcity. Index inclusion then shields the stock from normal market discipline: the valuation stops measuring profitability and starts measuring belief. Their phrase for the end state is the right one — too strategic to fail.

Friday is Exhibit A, fully assembled.

SpaceX lists at a fixed $135 — no bookbuilding — 555.56 million Class A shares, ~$75 billion raised against a $1.77 trillion valuation, with a public float somewhere between 3 and 7 percent depending on whose definition you use.

Musk keeps >82% of voting power; per the S-1, the public buys economics, not influence. Thirty percent of the offering goes to retail — triple the norm — index trackers are projected to hold ~30% of the float within fifteen trading days, and a 366-day insider lockup meters future supply like a drip line. Morningstar’s fair value sits 55% below the offer. Doesn’t matter. Wrong instrument for measuring a ceremony.

And behind it, the chamber is loading. OpenAI: IPO “as soon as this year,” per CNBC — with a government stake negotiation running in parallel. Anthropic: prospectus reportedly already filed, confidentially, with the SEC.

Three sovereign floats in twelve months, each adding a tranche of un-sellable shares — state, strategic, security-screened — making each public sliver thinner than the last.

Monday I wrote that they’d started metering the tokens. The Sovereign Float is the same meter, moved up a layer. First they rationed the intelligence. Now they ration the title deeds.

And catch the closing loop, because it’s almost elegant: the official justification for the government stake is that ordinary citizens won’t benefit from the AI boom unless they hold meaningful equity — which is true, and which the 3% float makes structurally impossible. The construct locks the public out of ownership, then offers to hold ownership on the public’s behalf and mail out a dividend. They took away your seat at the cap table and are offering you a coupon.

Two meters, one direction

China embeds AI in its Five-Year Plan as a national priority on par with defense, ¥10 trillion of AI-linked value targeted by 2030. America answers with sovereign wealth funds, golden shares, equity-for-megawatts barter, engineers seconded to the NSA, and IPOs structured so ownership never actually changes hands. The open-source ecosystem stays vibrant and global — increasingly Chinese — while the U.S. consolidates into a handful of mega-labs entangled with state power and financial markets simultaneously. Two systems, one org chart, converging from opposite ends of the ideology aisle.

I grew up behind a curtain where the state decided which industries were strategic. I recognize the paperwork. The fonts are nicer now.

The Towards AI piece ends on the right question, so I’ll sharpen it rather than repeat it: once these companies are too strategic to fail, can democratic institutions, markets, or regulators still shape them at all? My January 2025 self predicted the state would treat AI as a security problem. I undershot. The state is treating AI as a rival state — and you don’t regulate rival states. You take equity in them, you station your people inside them, you control their power supply, and you make very sure their shares never really trade.

For everyone outside that perimeter — the founders, the operators, the sovereign clients already asking for vendors without a government on the cap table — the conclusion has only hardened. The frontier is being fenced at the model layer, the ownership layer, and now the personnel layer. The only first-class seat left in this economy is the one you own outright: your iron, your weights, your dispatch, your power.

They can float four percent of a heartbeat at $135 a share, and the index will be ordered to buy it.

They cannot float yours.

Prediction nine: rising power demands for large-scale AI intensify scrutiny from regulators, investors, and the public. Massive centralized data centers draw criticism for inefficiency. Decentralization — edge computing, renewable integration, asynchronous scheduling — emerges to distribute energy draw.

My favorite spin-off: 'Mobile AI Data Centers,' containerized compute clusters anchored on or near renewable power sites, harvesting clean or stranded energy at the source.

Energy became THE constraint. Not a constraint. The constraint.

The numbers that vindicated this prediction are almost violent. The IEA's April 2025 Energy and AI report put global data-center electricity consumption on track to exceed 1,000 terawatt-hours by end of 2026 — Japan's entire annual consumption. Bloom Energy's January 2026 report: US data-center demand nearly doubling from 80 GW in 2025 to 150 GW by 2028. Wood Mackenzie (May 2026): US capacity needs going from 24 GW to 100 GW by 2030. Five tech companies spent $400B+ on AI infrastructure in 2025; the IEA projects that rising another 75% in 2026.

And the scrutiny I predicted?

The PJM Interconnection capacity auction for the 2026-2027 delivery year cleared at $329.17 per megawatt-day, up from $28.92 the prior year — a 1,037% increase, with data-center growth explicitly named as a major driver.

Community opposition has blocked tens of billions in data-center projects. Transformer lead times hit 160 weeks. The grid became the bottleneck. Exactly as predicted: power-grid stability and the scramble for real estate near generation became the defining infrastructure fight.

The centralized-inefficiency critique: validated harder than I expected

I said centralized mega-facilities would draw criticism for inefficiency. What I underestimated is how the inefficiency would compound with the GPU density curve. Facilities designed for 20-25 kW racks are being asked to serve workloads needing 200-250 kW per rack.

The mismatch means a lot of capital is going into buildings that are structurally under-equipped before they open.

The centralized model isn't just drawing criticism for energy inefficiency — it's drawing it for architectural obsolescence, which is a sharper version of the point I made in January 2025.

Mobile/modular data centers: from 'my favorite spin-off' to my actual work

This is the part where prediction became vocation. My January 2025 'Mobile AI Data Centers' concept — containerized clusters at renewable or stranded-energy sites, modular, pay-as-you-scale, resilient through distribution — is now the core infrastructure thesis I spend my days on.

The logic only got stronger: if power is the binding constraint and the grid can't be rebuilt fast (20-40 year infrastructure lifespans against an 18-month compute-density cycle), then the rational move is to bring modular compute to where the power already is, rather than dragging power to where someone wanted to build a cathedral.

I'm not going to pretend the modular model has won — centralized hyperscale is still where the headline capex goes.

But every structural pressure I named in January 2025 pushes toward distribution, and the gap between what hyperscale promises and what the grid can deliver is the widest it's ever been.

The next 12 months: the stranded-asset reckoning begins

My forecast for May 2027: the first major hyperscale facility is publicly acknowledged as under-equipped for the GPUs it was meant to host — a stranded or stranded-ish asset, built for a density that frontier workloads have already outrun.

When that happens, the conversation flips from 'how fast can we build gigawatt campuses' to 'how do we avoid building obsolescence.'

Watch energy-sovereign and behind-the-meter compute go from fringe to portfolio strategy.

As capacity prices stay elevated and interconnection queues run 5+ years, compute that brings its own power stops being a niche play and becomes the only way to deploy on a reasonable timeline.

The constraint I named in January 2025 doesn't ease in the next year. It tightens, and the tightening is the opportunity.

Grade: hit, and the most consequential one I made — not because it was the hardest call, but because it was the one I was willing to bet on. Predicting the constraint is cheap. Reorganizing around it is the expensive part, and the only part that pays.

____

^{Sources: original forecast (Jan 6 2025); IEA Energy and AI report (Apr 2025); Bloom Energy infrastructure report (Jan 2026); Wood Mackenzie (May 2026); PJM 2026-27 capacity auction; Quanta Services ($2.4T grid opportunity); transformer lead-time and interconnection-queue data 2026.}

The first half is unambiguously right. By 2026 the frontier labs converged hard. GPT-5, Gemini 3, Claude, and a wall of capable open models all clustered in capability. DeepSeek proved you could match frontier performance at a fraction of the cost. The 'advantage of any one organization diminishes' thesis is now the bear case on every frontier lab's valuation — the commoditization of the model layer is the single biggest story in AI economics. I called the convergence. It arrived.

The new-architecture call: real, but still insurgent

Here's the honest part. I predicted new architectures would 'rise.' What actually happened is subtler: new architectures are rising, but the transformer refused to die, and the most important shift wasn't a new architecture at all — it was a new axis of scaling.

On the architecture front, I was directionally right. State-space models — Mamba, Mamba-2, Mamba-3 — went from research curiosity to serious contender, delivering linear-time inference and 5x throughput gains over transformers on long sequences. Hybrid models like AI21's Jamba (interleaving Mamba and attention, scaled to 398B parameters) showed the attention-free and hybrid approaches can scale. Reasoning models built on Mamba demonstrated you can get strong reasoning without pure attention. The 'architectural fatigue' I described — diminishing returns from just scaling transformers — is real and widely acknowledged.

But — and I have to say it — the transformer did not get dethroned. It got supplemented. The dominant 2025-2026 models are still attention-based or hybrid, not pure alternatives. My prediction implied a changing of the guard. What we got is a coalition government.

What I genuinely missed: test-time compute

The biggest architectural story of 2025-2026 wasn't a new network architecture. It was a new place to spend compute: inference time. Chain-of-thought reasoning, test-time compute scaling, agentic loops — the realization that you can substitute inference-time thinking for raw model size. I gestured at 'search-based intelligence' in January 2025, which was actually a pretty good intuition for where reasoning models went. But I filed it under 'new architectures' when it's really a new scaling paradigm layered on top of the old architecture. I had the right instinct shelved in the wrong category.

This matters as a forecasting lesson. I was pattern-matching to 'the architecture changes' because that's the legible story. The actual change was economic and operational: inference became the cost center and the battleground, which is why efficiency-first designs like Mamba matter — not because they're prettier, but because inference repeated billions of times is the business model itself.

The next 12 months: efficiency stops being optional

Watch inference cost become the headline metric. Training cost is paid once; inference is paid forever. As reasoning models spend ever more compute at inference, the company with the most efficient inference architecture wins on margin even with an equally smart model. That's the structural reason I keep harping on infrastructure: the model layer commoditizes, so the durable edge moves to whoever runs inference cheapest — which is an architecture-plus-hardware-plus-energy question, not a 'whose model is smartest' question.

My grade on my own January 2025 self: B-minus. Right about convergence and architectural fatigue. Right-ish about search-based reasoning. Wrong to imply the transformer would fall, and I shelved the most important shift — test-time compute — under the wrong heading. A real partial. I'm not going to pretend otherwise.

____

^{Sources: original forecast (Jan 6 2025); Mamba/Mamba-2/Mamba-3 research (2024-2026); AI21 Jamba hybrid architecture; M1 / PromptCoT-Mamba reasoning models (2025); DeepSeek efficiency reporting; test-time compute scaling literature 2025-26.}

Prediction seven: fully driverless cars — zero human oversight — remain confined to 'safe zone' regions, often with mild climates and favorable road conditions. Human drivers remain indispensable elsewhere. Early adoption limited by geography, regulatory hurdles, and the hard problem of mixed human-and-robot traffic. My favorite spin-off: closed autonomous systems — dedicated tracks, hanging railways — scaling faster than open-road autonomy because they strip out the edge cases.

Where did Waymo actually scale in 2025-2026? Phoenix. San Francisco. Los Angeles. Austin. Atlanta. Then the 2026 expansion: Dallas, Houston, San Antonio, Miami, Orlando — and planned moves into Las Vegas, Nashville, San Diego. Read that list. Sunbelt. Mild climates. Favorable, mappable road conditions. Exactly the 'safe zones' I named. Waymo only began offering freeway routes in late 2025, in its established sunny markets, gradually. The geofence is the business model.

And the regulatory friction I flagged? Washington D.C. still required a human behind the wheel as of the 2026 expansion announcements — Waymo has to lobby to change the permitting before it can go driverless there. China suspended new autonomous-driving permits nationwide after a Baidu Apollo Go fleet froze en masse in Wuhan, trapping passengers for up to two hours. Regulatory hurdles, exactly as predicted, slowing the rollout of fully autonomous fleets.

~10 US metros with Waymo driverless service by 2026 — all favorable climates. 400K+ Waymo weekly paid trips across its markets. 100M+ fully driverless miles in Waymo's safety data. 0 cold-climate, snow-belt cities with full driverless service.

Mixed traffic stayed hard — as called

My prediction that integrating human drivers and pedestrians with driverless cars 'remains a key challenge' held. Waymo recalled over 1,200 vehicles to fix software causing low-speed collisions with gates and chains. A Waymo burned in the Los Angeles protests — a reminder that robots on public streets live in a messy human world. The technology works beautifully inside the geofence and gets humbled the moment conditions drift outside it. That boundary — works in the safe zone, struggles outside — is the entire 2026 autonomous-vehicle story, and it's the boundary I drew in January 2025.

The closed-track bet: still early, still right in principle

My 'favorite' was closed autonomous systems — dedicated tracks and right-of-way that eliminate the edge cases. This didn't produce a flashy 2025-2026 headline, so I'll mark it honestly as 'unresolved but vindicated in principle.' The logic I laid out — that controlled environments with no random pedestrians or unpredictable human drivers dramatically reduce the autonomy problem — is exactly why Waymo's geofenced, well-mapped operational design domains work and open-road full autonomy still doesn't. The safe zone is a soft version of the closed track. I was right about the principle; the literal hanging railways remain a niche.

The next 12 months: the geofence expands, the fantasy doesn't

My forecast for May 2027: Waymo crosses into 15+ metros and begins meaningful freeway operations, but every new market still fits the safe-zone template — favorable climate, mappable geometry, cooperative regulators. No Waymo robotaxi provides commercial driverless service in a serious snow city by mid-2027. The 'go anywhere, any conditions' driverless car remains a demo, not a product.

Watch the private-ownership fantasy get a reality check. Startups keep promising consumer-owned full-autonomy vehicles. The same constraint I named in January 2025 bites: a privately owned autonomous car comes with the same geofence — 'drive here, not there; sunshine, not storms.' Consumers won't pay premium money for a car that only fully drives itself in five zip codes on clear days. The robotaxi fleet model wins precisely because it can stay inside the safe zone; private ownership can't.

And watch China. The permit suspension is temporary, but it reveals the real bottleneck — not the technology, the trust. A single mass-freeze incident set the whole Chinese industry back. That fragility is structural, and it's why I keep saying the safe zone isn't a phase autonomous vehicles graduate out of. It's the shape of the business for years.

I drew the geofence in January 2025. Waymo spent 2026 driving inside it. Grade: surgical.

^{Sources: original forecast (Jan 6 2025); Waymo 2026 expansion announcements (CNBC, TechCrunch, The Robot Report, Nov-Dec 2025); Waymo safety data (100M+ driverless miles); Fortune on China's robotaxi permit suspension and Baidu Wuhan incident (May 2026); Waymo recall reporting}

.

Three claims, three different grades. Let me grade them in public.

Prediction six: the US leverages AI for national security; China shifts from copying-and-monetizing to genuine original research; Europe’s regulatory inertia risks losing ground. My favorite spin-off was ‘Iron Curtain Copy Cats’ — the idea that nations erect digital iron curtains while entrepreneurs smuggle and adapt AI models across borders, spurring localized innovation and fragmented parallel ecosystems.

Europe: A+. Painfully so.

I said Europe leads on ethics and regulation but its fragmented landscape deters experimentation and risks lost ground. Eighteen months later this is the consensus diagnosis of European tech policy. The EU AI Act arrived; European AI champions did not arrive at the same scale as American or Chinese ones. The continent that wrote the rulebook is buying most of its frontier models from elsewhere. I wish I’d been wrong about this one. I wasn’t.

The US: A. Security-first held.

The US national-security framing was correct and intensified. Federal AI policy reoriented around security primacy, the chip-export contest with China became the central geopolitical fact of the industry, and the build-fast doctrine (see prediction 2) is explicitly justified in national-competitiveness terms. The US treated AI leadership as a security imperative, exactly as predicted.

China: right direction, wrong mechanism — and bigger than I imagined

Here’s the interesting miss. I predicted China would shift from ‘copy and monetize’ to ‘authentic original research and breakthroughs.’ The direction was dead right — China stopped being a fast-follower. But I pictured that originality showing up as proprietary national champions racing the US lab-to-lab.

Instead, China’s originality expressed itself through open source, and that turned out to be a far more powerful vector than I imagined. DeepSeek’s efficiency shock. Alibaba’s Qwen crossing 700M+ downloads. MIT Technology Review documenting that Chinese open-source models surpassed US models in total downloads. A ‘ghost model’ that turned out to be Xiaomi — a phone company — beating expectations. China didn’t just do original research. It commoditized the model layer and gave it away, which is a more disruptive strategic move than any single proprietary breakthrough would have been.

My ‘Iron Curtain Copy Cats’ concept anticipated cross-border model smuggling and adaptation. The reality was almost the inverse and more elegant: China didn’t need anyone to smuggle its models. It published them under permissive licenses, and the world adapted them voluntarily. The iron curtain went up on chips and capital; the models flowed freely outward by design. I had the fragmentation right and the direction of model flow backwards.

What this teaches about predicting geopolitics

The lesson for any forecaster: you can call the actor and the direction correctly and still miss the mechanism, and the mechanism is where the strategic surprise lives. ‘China innovates’ was easy. ‘China weaponizes open source to commoditize its rivals’ moat’ was the actual 2025-2026 story, and I didn’t see the open-source angle clearly enough in January 2025.

The next 12 months: the open-weight bloc becomes a third power

My forecast for May 2027: the open-weight ecosystem — overwhelmingly Chinese-led — becomes a recognized third pole of AI power alongside US closed labs and EU regulation. Enterprises and entire nations standardize on open Chinese weights for cost and sovereignty reasons, the way they once standardized on Android. The strategic consequence: the US closed-lab moat erodes from below, not from a competing closed lab but from free, good-enough open alternatives.

Watch the US response. The build-fast doctrine collides with the reality that ‘fast’ may not beat ‘free.’ Expect a real debate inside the US about whether to compete with Chinese open source by open-sourcing more aggressively — a debate that would have sounded absurd in January 2025 and will sound inevitable by 2027.

And watch Europe try to turn its regulatory weight into a third kind of leverage — ‘trusted AI’ as a product category. It’s the only card the continent holds. Whether anyone pays a premium for it is the open question. I called Europe’s predicament in January 2025. The next year tells us whether predicament becomes strategy or just decline.
___

^{Sources: original forecast (Jan 6 2025); MIT Technology Review on Chinese open-source models (Feb 2026); Japan Times on the Xiaomi ‘ghost model’ (Mar 2026); DeepSeek efficiency reporting; US-China chip-export contest 2025-26; EU AI Act competitiveness debate.}

Sources: original forecast (Jan 6 2025); EY 2025 wealth-management AI survey; MIT findings on white-collar task automation; entry-level/junior labor-market analysis 2025-26; the author’s own prediction 12 (90% of certain job tasks automatable).

Prediction five: fears of mass unemployment prove overstated. Roles transform rather than vanish. Specialized fields keep humans for oversight and interpretation. New skills — prompt engineering, data literacy — become sought-after. My favorite framing was ‘Grandma AI with Superpowers’: the human expert as the grandmother passing generational wisdom to a fledgling AI system.

The augmentation thesis held at the aggregate

No AI jobpocalypse arrived. The mass-unemployment headlines of early 2023 did not materialize into 2025-2026 reality. What happened instead is exactly the role transformation I described: the wealth-management analyst who now supervises AI rather than building models by hand, the lawyer reviewing AI-drafted discovery, the radiologist using AI as a second reader. EY’s 2025 survey found 95% of wealth-management firms using generative AI somewhere. The work changed shape. The headcount didn’t collapse wholesale.

And ‘Grandma AI’ turned out to be a genuinely good frame. The most valuable humans in 2026 aren’t the ones who can do the task — they’re the ones with the contextual, hard-won judgment to know when the AI’s output is subtly wrong. The expert as mentor to the machine. That’s holding up better than most of my 2025 metaphors.

The crack I underweighted: the ladder lost its bottom rung

Here’s where I have to be honest. I framed augmentation as broadly good news for workers. The data forming through 2025-2026 says: good news for senior workers, brutal news for juniors.

The tasks AI handles best are exactly the entry-level tasks — the document review, the first-draft research, the basic coding, the junior analyst grunt work that used to be how you broke into a field. When AI does up to 90% of those tasks (my own prediction 12 said as much), the entry-level job doesn’t transform. It thins out. And if you eliminate the bottom rung, you eventually have no seniors, because seniors are just juniors who survived a decade. I called the augmentation. I underweighted the generational discontinuity it creates.

This is the bifurcation I now write about constantly: the elite few with AI leverage become dramatically more productive, while the people who would have learned the trade on the bottom rung find the rung gone. The aggregate unemployment number stays calm. The distribution underneath it gets violent. Both things are true. My January 2025 self emphasized the calm aggregate and didn’t look hard enough at the violent distribution.

Prompt engineering: right skill, wrong shelf life

I called prompt engineering a hot skill. It was — for about a year. Then the models got good enough that elaborate prompt-craft mattered less and ‘context engineering’ and agent orchestration mattered more. I had the right idea (new AI-collaboration skills become valuable) attached to a job title with a two-year half-life. The meta-lesson: when you predict a hot skill in AI, predict the category, not the title. The titles rotate every eighteen months.

The next 12 months: the missing-rung problem goes from anecdote to policy

My forecast for May 2027: the entry-level squeeze becomes a named political and corporate problem. Expect the first serious ‘where do juniors come from now’ panic in knowledge-work fields — law, consulting, accounting, software. Companies that quietly stopped hiring juniors in 2025-2026 discover in 2027 that their senior pipeline is drying up, and the smart ones start building deliberate human-apprenticeship-alongside-AI programs as a competitive advantage.

The augmentation thesis survives, but it gets a mandatory asterisk: augmentation is great if you’re already on the ladder. The defining workforce question of the next year isn’t ‘will AI take jobs’ — that framing is dead. It’s ‘who builds the ladder back, and who gets left standing under where it used to be.’

I’d give my January 2025 self a B+. Right that the apocalypse wouldn’t come. Too optimistic about who’d be fine when it didn’t.

Sources: original forecast (Jan 6 2025); EY 2025 wealth-management AI survey; MIT findings on white-collar task automation; entry-level/junior labor-market analysis 2025-26; the author's own prediction 12 (90% of certain job tasks automatable).

What I wrote: consciousness and self-awareness on par with fictional robots stays out of reach. Progress centers on form factors and specialized intelligence. Niche applications — manufacturing, healthcare, surgical assistance — see traction; household companions remain far off. And the workforce impact is augmentation: robots handle repetitive or hazardous tasks alongside humans, not instead of them.

Now the news.

Form factors got real. Cognition stayed narrow.

Figure AI’s Figure 02 spent eleven months on BMW’s Spartanburg line — 1,250 operating hours, 30,000+ vehicles supported, 90,000+ sheet-metal parts handled. It came back scratched and grimy, and that data trained Figure 03. In January 2026, Figure deployed 40 Figure 03 units commercially at Spartanburg, billing BMW roughly $25 per robot-operating-hour. Tesla discontinued the Model S and Model X and converted Fremont lines to build Optimus, targeting production start in summer 2026. Unitree shipped 5,500+ units in 2025 and is targeting up to 20,000 in 2026, with a G1 on Amazon for around $17,990.

Notice what’s in none of those sentences: consciousness. Self-awareness. General intelligence. Every single deployment is a narrow, supervised, industrial form-factor play. Musk himself admitted on the Q4 2025 earnings call that the Optimus units inside Tesla factories exist to generate training data, not to perform productive labor. He called it R&D. That is the single most honest sentence anyone in humanoid robotics said all year, and it confirms my January 2025 call to the decimal.

Augmentation, not replacement — exactly as called

The deployment pattern is precisely the ‘human-robot collaboration’ I described. Agility’s Digit moving totes at Amazon and GXO. Apptronik’s Apollo tested by Mercedes. Boston Dynamics’ Atlas piloting at Hyundai. Every one of these is a robot taking over the repetitive, hazardous, or ergonomically brutal task while humans handle judgment, exceptions, and oversight. Nobody got replaced. The work got reshaped. That was the prediction.

Where I planted a flag worth revisiting: the autonomous power plant

My ‘favorite’ spin-off in this prediction was the Autonomous Power Plant with AI Data Centers — a self-contained complex where humanoid robots tend specialized crops for biogas conversion, power generates on-site, and AI workloads run on that isolated energy. I called it a frontier where robotics, renewable energy, and AI converge under one roof, run with minimal human intervention.

Eighteen months later, that concept is the spine of half of what I write about — the modular, energy-sovereign, robotically-tended compute facility. The grid can’t keep up (PJM capacity prices ran up over 1,000% for the 2026-2027 delivery year). The case for compute that brings its own power and tends itself with robots got stronger every month of 2025. I didn’t predict it would become my central thesis. It did.

The next 12 months: the China volume bet versus the Western quality bet

My forecast for May 2027: humanoids cross from ‘hundreds to low thousands doing productive work globally’ into the low tens of thousands — still niche, still industrial, still supervised. No household humanoid does anything genuinely useful at scale by mid-2027. The 1X NEO home pre-orders ship, underwhelm, and teach the market that the living room is harder than the factory.

The real 2026-2027 story is the bifurcation I’d now bet on: China wins on volume and price (Unitree’s 20,000-unit target, 90% of global unit volume), the West wins on industrial quality and integration (Figure’s $25/hour billable units).

The question that decides the decade is whether Chinese volume-and-iteration beats Western quality-and-integration — the same question that decided EVs, drones, and solar. I know which way that historically breaks.

And watch the economics, not the demos. A robot that bills $25/hour and runs more than one shift pays for itself against a $20-30K purchase price fast. The backflip videos are marketing. The invoice to BMW is the business. I told you in January 2025 to watch the form factor and the niche. Watch the invoice now.

Sources: original forecast (Jan 6 2025); Figure AI BMW Spartanburg deployment (Jan 2026); Tesla Q4 2025 / Q1 2026 earnings; Unitree shipment and pricing data (2026); Agility/Apptronik/Boston Dynamics deployments; PJM capacity auction (2026-27 delivery year).

Prediction three said the single ‘super agent’ that seamlessly does everything stays out of reach. Instead, multiple specialized agents — swarms — operate under a central orchestrator, each handling a narrow domain with high efficiency. I predicted dynamic coordination, a marketplace of plug-and-play agents, and interoperability standards as the crucial enabler. My favorite spin-off was the ‘Central Swarm Transaction System’ — a high-velocity marketplace where agents transact with each other.

Read any 2026 AI architecture doc and you’re reading my January 2025 prediction back to me.

The single super-agent never showed

OpenAI, Anthropic, Google — none of them shipped the mythical do-everything agent. What they shipped was orchestration. Claude’s agent framework. Multi-agent systems where a coordinator delegates to specialized sub-agents. The ‘manager agent dispatches to worker agents’ pattern is now so standard it’s boring. Exactly as predicted: building one agent to flawlessly handle everything proved too complex; specialized agents leveraging domain-specific tooling won.

The reliability data backs the architecture choice. Fully autonomous single agents fail roughly 70% of real multi-step tasks. Structured, orchestrated workflows with specialized components consistently outperform. The industry didn’t choose swarms because it’s elegant. It chose swarms because monolithic autonomy doesn’t work yet.

The marketplace is forming — slower than the hype, faster than the skeptics

My ‘Central Swarm Transaction System’ — agents transacting with agents at machine speed — is half-built. The agent-to-agent payment rails I wrote about are real: protocols for machine-to-machine commerce, micropayment layers, agent-to-agent communication standards. Y Combinator’s Winter 2026 batch was 41.5% agent-infrastructure companies. The plumbing for the swarm marketplace is being laid right now.

But — and this is the honest part — it’s plumbing, not yet a thriving economy. The ‘self-sustaining AI services ecosystem where agents compete, collaborate, and transact’ that I described is still mostly humans wiring agents together, not agents autonomously hiring each other. The vision was right. The timeline was optimistic. The economy is being built; it isn’t running.

What being early actually costs

Here’s the uncomfortable founder lesson. I was right about the architecture in January 2025, eighteen months before it became consensus. Did that make me money? Only if I built or invested on the conviction. Being right and being early are the same thing wearing different clothes, and the market pays you for the second one only if you can survive long enough to be called the first one.

The companies that committed to multi-agent orchestration in early 2025 — when it still sounded speculative — are the ones with the 2026 lead. The ones that waited for consensus are paying acquisition prices for what conviction would have given them cheap.

The next 12 months: orchestration eats the application layer

Watch interoperability standards become the real battleground. Whoever owns the protocol that lets agents from different vendors transact becomes the TCP/IP of the agent economy. That’s a winner-take-most position, and the fight for it is just starting.

And watch the failure rate stay the moat. As long as autonomous agents fail 70% of unsupervised tasks, the orchestration layer’s most valuable function isn’t coordination — it’s knowing when to stop, escalate, and put a human in the loop. The orchestrators that get the human-handoff right will beat the ones that promise full autonomy and ship chaos.

I called the architecture in January 2025. The whole industry caught up in 2026. The lesson isn’t that I’m a genius. It’s that the obvious structural answer is often visible eighteen months before everyone agrees to see it — and that gap is the only edge that matters.

Sources: original forecast (Jan 6 2025); industry multi-agent orchestration patterns 2025-26; agent task-reliability data (Gartner, buildmvpfast analysis 2026); YC W26 batch composition (41.5% agent infrastructure); agent-to-agent payment protocol development 2025-26.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek.io bare-metal GPU cloud platform.

I’m going to give myself a qualified win here, and then tell you the part I missed.

My second prediction was about regulatory divergence. Three claims: the US lags and takes cues from Europe; Europe’s bureaucratic drag slows actual deployment despite leading on the paperwork; and parts of Asia — China, Singapore — fast-track via flexible regulatory sandboxes. My favorite spin-off concept was ‘Regtech Hopping’: AI that automates jurisdictional optimization, spinning up entities wherever the rules are friendliest.

Where the map redrew itself on me

Here’s what I got wrong: I assumed the US would keep taking cues from Europe. It didn’t. It went the other direction — hard.

Instead of importing Brussels-style precaution, the US pivoted to an aggressively pro-deployment, pro-build posture. Federal AI policy reoriented around speed, national-security primacy, and infrastructure. The FDA, in a move I’ll cover in prediction 13, started relaxing premarket requirements for AI clinical tools rather than tightening them. The ‘US follows Europe’ framing was wrong. The actual 2025-2026 story is the US explicitly rejecting the European model and betting that whoever deploys fastest wins.

So my directional claim — divergence, patchwork, friction at the borders — was right. My specific vector for the US was wrong. I had the US as a follower. It became a defector.

Regtech Hopping: more real than I expected, and more dangerous

My ‘Regtech Hopping’ concept — AI automating the setup of legal entities across jurisdictions to optimize for the friendliest rules — turned out to be directionally prophetic in an ugly way. The proliferation of solo, AI-run companies (see prediction 1) created exactly the jurisdiction-shopping pressure I described. When one founder can spin up operations anywhere, the regulatory arbitrage I flagged stops being theoretical.

The dark version I warned about — gambling, betting, crypto operators using the same automation to stay one step ahead of regulators — is now a live enforcement problem across multiple jurisdictions. I called it a ‘high-stakes cat-and-mouse game.’ The cat is losing.

The next 12 months: enforcement theater meets the build-fast doctrine

My forecast for May 2027: the EU AI Act delivers its first headline enforcement action against a major model provider — and it backfires politically. Expect the fine to become a rallying cry for European competitiveness hawks who argue Brussels is regulating an industry it doesn’t have. The gap between European AI rule-making and European AI building will become the defining tension of the continent’s tech policy.

The US holds its build-fast line through the 2026 election cycle, because the infrastructure spend is now too economically load-bearing to slow down. You don’t tap the brakes on $400B+ of annual capex when it’s propping up GDP growth.

And watch the third bloc harden: the open-weight world. China’s open-source dominance plus the EU’s regulatory weight plus the US’s deployment speed creates three incompatible operating environments. The winners will be the ‘bridge builders’ I named in January 2025 — the compliance-as-a-service and cross-border translation firms that help companies operate across all three without getting fined into oblivion in any of them.

I predicted a patchwork. We got a patchwork. I just had one of the three patches sewn on backwards.

Sources: original forecast (Jan 6 2025); EU AI Act GPAI obligations effective Aug 2 2025; MIT Technology Review on Chinese open-source models (Feb 2026); FDA clinical-decision-support guidance (Jan 6 2026); Manatt Health AI Policy Tracker (Apr 2026).

On January 6, 2025, I sat on a beach — off the coast, naturally — and published thirteen predictions for where AI was actually headed. Not the hype. Not the doom. The structural reality.

Most prediction posts are written to be forgotten. The forecaster bets that nobody keeps the receipts. I’m doing the opposite. I’m pulling every one of those thirteen calls back into the light, scoring it against seventeen months of hard data, and telling you — to my own discredit where it applies — exactly where I was right, where I was right for the wrong reason, and where I was simply wrong.

Then I’m doing the harder thing. For each one, I’m extending the call twelve months forward, to May 2027, with new sources on the table. Because a forecast you won’t re-grade is just decoration.

The honest headline: I was substantially right on the structural direction nearly across the board. Where I missed, I missed on mechanism — China’s open-source vector, the test-time-compute shift — and on second-order human cost, like the entry-level rung quietly vanishing. Direction is the easy 70% of forecasting. Mechanism and consequence are the hard 30%. That’s where the next thirteen predictions get sharper.

Here’s the scorecard. Then the receipts, one by one.

The Receipts, at a glance

Prediction 1 Revisited: I Said AGI Wouldn’t Arrive in 2025. I Was Right. I Also Underestimated the One-Person Unicorn.

In January 2025 I told you AGI stays elusive and the real money is in vertical agent swarms run by a single founder. Seventeen months later, the AGI part aged fine. The one-person-unicorn part aged like a prophecy I didn’t have the nerve to fully believe.

Let me start with the part I got wrong by being too cautious.

In my January 6, 2025 forecast, prediction number one had two halves. Half one: true AGI — the unified, human-adaptable kind — does not show up in 2025. Half two, my favorite, the thing I buried under a header called ‘My Favourite’: vertical stream agent swarms, specialized micro-agents each owning a slice of an industry vertical, coordinated by a thin human layer on top, eventually enabling the 1-person billion-dollar startup.

I hedged on the second half. I called it a ‘middle path.’ I should have called it the road.

The AGI half: correct, and not even close

Nobody serious is claiming AGI shipped in 2025. The AAAI 2025 presidential panel surveyed 475 researchers and 76% said scaling current approaches is unlikely to even get us to AGI. What we got instead was exactly what I described: brutally competent narrow systems. GPT-5 in August 2025. Gemini 3. Claude crossing $30B in annualized revenue by April 2026 on the back of enterprise coding agents, not general intelligence. The systems got jaw-droppingly good at specific things and remained, definitionally, not general.

That was the easy call. Predicting AGI won’t arrive next year is the safest bet in technology. I’ll take the point, but I won’t celebrate it.

The one-person unicorn: the call that actually mattered

Here’s where it gets interesting, and where I want to be honest with you about my own timidity.

When I wrote about ‘enabling a single founder to operate at a level once reserved for large organizations,’ I framed it as a 2025-to-2027 trajectory. A direction. A thing that would emerge. I did not expect to be naming a specific company with specific revenue this fast.

Meet Matthew Gallagher. In September 2024 — months before I published — he launched Medvi, a GLP-1 telehealth startup, out of his Los Angeles home with $20,000, zero employees, and more than a dozen AI tools doing the code, the ad creative, the customer service. By the end of 2025: $401 million in revenue. 250,000 customers. A 16.2% net profit margin. The company is now tracking toward $1.8 billion in 2026.

For context: Hims & Hers did $2.4 billion last year with 2,442 employees and a 5.5% margin. Gallagher is running roughly three times the margin with a headcount of one.

Is Medvi a ‘pure’ one-person billion-dollar company? Not yet. It uses contractors, fulfillment partners, a regulated pharmacy supply chain. The purists will quibble. Let them. The structural point I made in January 2025 is now undeniable: the cost of execution collapsed, and a single operator orchestrating agents can now reach revenue that used to require a building full of people.

The receipts piled up faster than the skeptics could object

Dario Amodei was asked at Anthropic’s May 2025 developer conference when the first one-person billion-dollar company arrives. One word: ‘2026.’ He gave it 70-80% odds. Sam Altman has a running bet in a CEO group chat on the same question. Sequoia is rewriting its underwriting models around what it calls ‘agentic leverage.’ By early 2026, 36.3% of new ventures were solo-founded, and those solo ventures were reporting 77% first-year profitability versus roughly 40% for traditional startups.

Pieter Levels runs a multi-million-dollar portfolio alone. Danny Postma’s HeadshotPro: $3.6M ARR, solo. Base44 hit 250,000 users and sold to Wix for $80 million within six months of launch. The pattern I called a ‘middle path’ became the default operating model for a whole category of software business.

The tooling cost? $100 to $500 a month for a stack that replaces a 2024 team. That’s the number that should keep incumbents awake.

What I’d change about the original call

One thing. I anchored the swarm vision to orchestration platforms — the idea that you’d buy a ‘swarm OS’ and assemble agents. That’s not quite how it played out. The winning solo operators didn’t buy a unified swarm platform. They duct-taped Cursor, Claude, ChatGPT, Zapier, and a CRM into a personal operating system. The orchestration happened in the founder’s head, not in a product. The market for a true ‘agent orchestration layer’ is still wide open — and still mostly hype, with autonomous agents failing roughly 70% of real multi-step tasks when left fully unsupervised.

That’s the gap. The dream is autonomous swarms. The reality is a very smart human conducting a very capable orchestra and catching the wrong notes before they reach the customer.

Leave a comment

The next 12 months: from solo unicorns to solo conglomerates

My forecast for May 2027: the first genuinely defensible one-person company crosses $1 billion in revenue, and Amodei collects his bet — or comes embarrassingly close. But the more important shift is structural, not symbolic.

Watch for the solo operator to stop being a single product and become a portfolio. The same founder who can run one agent-powered business can run five. We’ll see ‘personal holding companies’ — one human, a dozen AI-run product lines, each a former full-team business. Y Combinator’s batches are already 40%+ agent-infrastructure companies feeding this exact pattern.

And watch the failure rate become the story. When 70% of unsupervised agent tasks fail, the moat isn’t the agents — it’s the human judgment layer that knows which 70% to catch. The winners of 2027 won’t be the founders with the best agents. They’ll be the ones with the best taste about when to override them.

Cathedrals are for worship. Tents are for war. The one-person unicorn is the ultimate tent: no overhead, no committee, no three-week procurement cycle. Just a founder, a fleet of agents, and the nerve to ship before the incumbents finish their planning meeting.

I was right in January 2025. I just wasn’t bold enough. Won’t make that mistake twice.

Sources: original forecast (aiofthecoast.dcxps.com, Jan 6 2025); PYMNTS/NYT on Medvi (Apr 2026); Anthropic developer conference (May 2025); Sequoia ‘agentic leverage’; Scalable.news solo-founder data (2026); Gartner (Tom Coshow) on agent reliability; YC W26 batch composition.

Part 5 of a five-part series on the structural opportunity in modular AI infrastructure

There is a question that should be asked at the front of every conversation about AI infrastructure investing, and almost never is. The question is: what, exactly, am I owning?

It seems like a basic question. In conventional infrastructure investing — toll roads, regulated utilities, midstream energy, traditional data center colocation — the answer is unambiguous. You own a depreciable physical asset with a known cash-flow profile, a known residual value, and a known liquidity pathway. The asset is in the ground, on the meter, or in the rate base. You can put a number on it that an auditor and a lender will both accept.

In AI infrastructure investing as currently practiced, the answer is much less clean. Capital flows through a half-dozen distinct vehicles — listed neocloud equity, private neocloud equity, hyperscaler equity, GPU vendor equity, tokenized compute claims, hardware-backed SPVs, lease-financing structures, sale-leaseback structures, project finance — and each one delivers a structurally different risk exposure even when the underlying revenue source looks similar.

The capital partner who treats these as interchangeable is making the most common mistake in the cycle. The capital partner who understands the structural differences captures a meaningful return premium for selecting the right vehicle for the right portfolio mandate.

This final article in the series lays out the asset-class framework. It is the lens through which DCXPS designed the SPV structure described in Article 2, and it is the lens that capital partners should apply when comparing modular hardware-backed SPV exposure to the alternative vehicles in the market.

Six vehicles, six exposure profiles

The capital flowing into AI infrastructure in 2026 distributes across six broad structures. I will name each, characterize its exposure, and identify what kind of capital it is appropriate for.

Vehicle 1 — Hyperscaler equity (MSFT, GOOG, AMZN, META)

What you own: A fractional claim on a diversified cash-flow stream that includes cloud, AI, advertising, consumer hardware, retail, and other business lines. AI infrastructure exposure is dilutive of and diluted by every other business line.

Exposure profile: Highest liquidity (daily exchange trading), lowest direct exposure to AI infrastructure economics. Beta to AI cycle approximately 0.4–0.6 depending on issuer. Annualized return target: ~10–15%.

Appropriate for: Generalist public equity portfolios; not appropriate for capital with a specific mandate to gain AI infrastructure exposure.

Vehicle 2 — GPU vendor equity (NVDA, AMD)

What you own: A fractional claim on the cash flow stream of the supplier to the AI infrastructure build. Exposure peaks during the buildout and may compress during operating maturity.

Exposure profile: Highest cycle beta of any liquid vehicle. Reflexive — when AI infrastructure demand grows, GPU vendor margins expand; when growth slows, margins compress. Annualized return target: highly variable.

Appropriate for: Tactical positioning on the buildout phase; problematic for long-horizon hold strategies because exposure is to suppliers, not to operating cash flow.

Vehicle 3 — Listed neocloud equity (CRWV, IRIS, etc.)

What you own: A fractional claim on the equity of a specialized AI infrastructure operator. The asset is the operator’s enterprise value, not the underlying hardware. Critically, this exposure is highly geared — most listed neoclouds carry significant lease, debt, and off-balance-sheet liabilities that amplify both upside and downside.

For perspective, CoreWeave carries approximately $34 billion in off-balance-sheet lease commitments against a market capitalization that has ranged considerably over the past 18 months.

Exposure profile: High direct exposure to AI infrastructure operating economics, but with substantial gearing risk. Beta to AI cycle ~1.2–1.8. Annualized return target: high, with high volatility.

Appropriate for: Public equity allocators with high risk tolerance and conviction on the operator’s specific platform; not appropriate for capital seeking hardware-backed exposure with downside protection.

Vehicle 4 — Tokenized compute claims

What you own: A claim, structured through blockchain infrastructure, on a fraction of compute capacity or compute revenue from a specific provider. The structure is novel; the legal recourse is novel; the secondary-market liquidity is genuine but thin.

Exposure profile: Variable depending on token construction. Generally lower legal recourse than equity, lower transparency than listed instruments, higher liquidity friction than private equity.

Appropriate for: Specialized crypto-native portfolios with high tolerance for legal and counterparty risk; not appropriate for institutional capital with conventional fiduciary mandates.

Vehicle 5 — Project finance / lease structures

What you own: A debt instrument secured against AI infrastructure hardware and contracted revenue. Returns are capped at the coupon; downside is mitigated by the asset collateral.

Exposure profile: Lowest direct AI exposure of any AI-named vehicle. Returns in the 7–12% range, depending on credit profile. Effectively a fixed-income exposure with AI-cycle correlation in the tail risk only.

Appropriate for: Yield-seeking capital with infrastructure debt mandate; not appropriate for capital with growth equity or operating equity targets.

Vehicle 6 — Hardware-backed SPV structures (DCXPS Site 01 and analogous)

What you own: Direct title to the physical hardware (GPU servers, containerized infrastructure, fabric), held inside a ring-fenced legal vehicle, operated by a third-party operator under a structured economic agreement.

Exposure profile: Direct exposure to underlying asset cash flow, with hardware-backed downside protection (the physical asset has measurable secondary-market value). Returns delivered through monthly distribution waterfall. Capped only by the operating term of the SPV; uncapped on the upside above modeled cases through utilization or pricing outperformance.

Appropriate for: Capital with specific mandate for AI infrastructure exposure, preference for hardware-backed downside protection, tolerance for 5–7 year illiquidity, and structural alignment with operator through ring-fenced ownership.

This is the vehicle category that DCXPS operates in, and the structural reasons it is the appropriate vehicle for a specific kind of capital partner are worth unpacking.

What hardware-backed ownership actually delivers

The DCXPS SPV structure delivers four structural features that none of the other five vehicle categories can replicate. Capital partners considering exposure to AI infrastructure should evaluate whether they need any of these four features; if they do, the SPV structure is materially the only vehicle that provides them.

Feature 1 — Asset-level title

The SPV holds direct title to 49 GPU servers, two reinforced 40-foot ISO containers, 450 kW of power infrastructure, the Cisco fabric, the cooling loops, the BMS, and all ancillary equipment. The hardware is on the SPV’s balance sheet, recorded in the SPV’s books, insured in the SPV’s name, and titled in the SPV’s name.

This is not a contractual claim against operator revenue. It is not a tokenized representation of compute. It is title. The hardware can be inspected, audited, photographed, and verified through customs and bills of sale.

The practical implication: if any party — operator, customer, regulator, counterparty — fails to perform, the SPV retains the underlying asset. The capital partner’s downside is bounded by the asset’s secondary-market value at the relevant moment in the lifecycle, not by the operator’s enterprise solvency.

Feature 2 — Waterfall priority

The economic structure pays the SPV capital partners first. Specifically:

Gross revenue is collected by Chapek (the operating platform) and distributed to the SPV monthly. OPEX (electricity, fixed admin overhead) is deducted to produce EBITDA. EBITDA is then split 85% to SPV capital partners / 15% to DCXPS as operator. Project financing, where applicable, is deducted from EBITDA before the split.

The operator share is paid after the capital partner share is determined. This is structurally different from equity in an operating company, where capital partners are pari passu with operators (or, more often, junior to operators by virtue of management’s preferred securities).

The practical implication: the operator’s economic incentive is to grow EBITDA, because the operator’s compensation is a percentage of EBITDA after OPEX. The operator does not benefit from growing revenue while degrading margin, from running up SG&A, or from prioritizing growth over unit economics. This is the structural alignment that ring-fenced ownership creates.

Feature 3 — Refinanceability

Once a unit is operating, the hardware-and-cash-flow combination becomes financeable as collateral. A capital partner who put in $45M of equity can refinance 50–70% of asset value into senior debt at favorable rates, secured against the operating asset, and redeploy the freed capital into additional unit positions (at Site 01 or elsewhere).

This is the same mechanism by which CoreWeave scaled its capital base into the $34B off-balance lease pipeline reference above. At our scale, with our jurisdictional positioning, the equivalent mechanism is available to capital partners who want it.

The practical implication: the initial $45M ticket can be re-leveraged into capital for additional unit positions, effectively creating a portfolio of units from a single committed equity tranche. This is the mechanism by which family offices in particular have built scale exposure to traditional infrastructure asset classes, and it is available in the AI infrastructure context for the first time at the unit level.

Feature 4 — Three-path exit optionality

At the end of the 6-year operating term, the SPV holds 49 servers of B300/H200 capacity plus the containerized infrastructure plus operating know-how. There are three exit paths, each with different IRR profiles:

Path A — Refresh and extend. New-generation hardware (Rubin or successor) is procured and deployed into the existing containerized infrastructure. The existing hardware is sold into the secondary market. The SPV continues into a second operating term with updated economics. This is the optimal path if the customer mix at year 6 supports continued bare-metal demand at attractive rates.

Path B — Sell to operator or third party. The SPV liquidates the unit as a going concern, with operator continuity. This is the optimal path if the capital partner wants liquidity and the operator (or a third party) wants continued asset exposure. The sale price is a multiple of trailing EBITDA, structured similarly to traditional infrastructure asset sales.

Path C — Liquidate hardware to secondary market. The SPV decommissions and sells the GPU inventory into the secondary market. The containerized infrastructure has residual value (typically 30–40% of original cost). This is the floor path — the path available even if all operating optionality has been exhausted.

The practical implication: capital partners are not exposed to a single exit pathway. They retain agency over exit choice based on the actual market conditions at year 6.

Comparison to traditional infrastructure asset classes

The natural calibration question: how does this exposure compare, on an apples-to-apples basis, to traditional infrastructure asset classes that institutional capital already deploys into?

The table below summarizes the comparison. The numbers for the traditional asset classes are sourced from publicly available infrastructure investing benchmarks (Cambridge Associates, Preqin, McKinsey, EDHEC); the numbers for DCXPS are derived from the model in Article 2.

Asset class

Typical IRR

Typical multiple

Typical hold

Downside protection

Core US infrastructure (regulated utilities, airports, toll roads)

7–11%

1.5–1.8×

7–10 yr

High — regulatory ratemaking

Renewable energy project equity

8–13%

1.7–2.1×

7–10 yr

Medium — PPA-backed cash flow

Traditional retail colocation

11–15%

2.0–2.5×

7–10 yr

Medium — long-term tenant leases

Hyperscale build-to-suit (development)

13–18%

2.0–2.6×

5–7 yr

Medium-high — anchor tenant

Hyperscale build-to-suit (stabilized)

9–13%

1.6–2.0×

7–10 yr

High — long-term lease

Modular AI compute SPV (DCXPS modeled)

24–29%

2.75×

6 yr

Medium — hardware-backed

The premium is real and quantifiable. ~12–18 percentage points of IRR over core infrastructure, ~10–14 over renewable, ~10–13 over traditional colocation. That premium is paid for in three forms of incremental risk:

Risk premium 1 — Technology obsolescence cycle. GPU generations turn over in 18–24 months. The 6-year SPV term spans approximately three full technology generations. Hardware-residual value is not linear over time, and the operator must actively manage the workload mix to optimize revenue against generation-cycle compression.

Risk premium 2 — Operator dependency. Unlike a regulated utility (where operating risk is dispersed across a public-utility-commission framework) or a stabilized colocation asset (where the lease is contractual and durable), the modular SPV depends on operator execution on customer acquisition, NOC operations, and platform performance. Ring-fenced asset ownership mitigates but does not eliminate this dependency.

Risk premium 3 — Secondary market depth. GPU secondary markets are real (the H100 secondary market in 2026 is active and price-discoverable) but less institutional than the secondary markets for, say, commercial real estate or core utility equity. A forced sale would clear at less attractive pricing than an orderly sale.

The question for the capital partner is: is the IRR premium adequately compensated for the additional risk? My honest assessment is that, at current market conditions, it is. The IRR delta exceeds what a rational pricing of the marginal risk would suggest, because the market for modular AI infrastructure capital is not yet mature enough to be efficiently priced. Capital partners moving now are capturing a pricing inefficiency that will compress over the next 24–36 months as the asset class matures.

That window is the case for moving now rather than later.

What kind of capital partner this fits

The vehicle is not appropriate for every capital partner. It is genuinely appropriate for several specific profiles, and identifying whether your capital fits is part of the diligence:

Family offices with infrastructure mandate

For family offices targeting 15–25% IRR exposure with hardware-backed downside protection, the DCXPS SPV structure fits cleanly. The $15M minimum participation is sized for typical family office single-position appetite. The 6-year term aligns with conventional family office liquidity preferences. The hardware-backed exposure provides downside protection that matches the conservative wing of typical family office portfolios.

Sovereign-adjacent capital with European mandate

European sovereign wealth funds, government-affiliated investment vehicles, and sovereign-backed family offices with explicit European positioning. The Czech jurisdictional positioning, EU AI Act alignment, and sovereign infrastructure thesis described in Article 3 make this vehicle particularly aligned with sovereign-mandate capital that needs to deploy into European AI infrastructure but cannot deploy into US-incorporated operators.

Strategic investors with European enterprise exposure

Industrial, financial, and technology companies with significant European enterprise customer bases and an interest in either (a) anchor-customer access to compute capacity for their own AI workloads, or (b) strategic alignment with EU-sovereign infrastructure positioning. The SPV structure supports strategic-investor positioning through dedicated capacity allocation if structured at the unit level.

Specialist infrastructure funds

Funds with explicit mandate for next-generation infrastructure exposure (data, energy, compute) that need to deploy at meaningful scale into AI infrastructure but want the hardware-backed exposure profile rather than operator equity. The 9-unit Site 01 structure provides scale optionality for a fund that wants to commit across multiple units.

Yield-seeking institutional capital

The ~29.2% simple yield profile is meaningfully above what fixed-income or traditional infrastructure yield can deliver. For institutional capital pools with a meaningful yield-seeking allocation, the SPV structure can serve as a higher-yielding component of an infrastructure-yield portfolio.

The vehicle is not appropriate for capital that requires daily liquidity, capital with a sub-5-year horizon, or capital with no tolerance for technology-cycle risk. Be honest about your mandate, and the structure will be honest about whether it fits.

The seven-question diligence framework

Repeating the framework from Article 2 because it is the right set of questions to bring to any AI infrastructure SPV diligence, ours or anyone else’s:

1. Who owns the hardware?

2. Where does the power come from?

3. What is the operator’s economic share, and is it paid first or second?

4. What is the GPU allocation pathway?

5. What is the realistic revenue ramp?

6. What is the exit mechanism?

7. What is the operator’s track record at unit-scale operations?

For DCXPS:

1. Hardware ownership: The SPV holds direct title. Bills of sale, customs documentation, and corporate records available in the data room.

2. Power: Co-located 200 MW power plant at Kladno, with direct on-site interconnect and structured off-take.

3. Operator share: 15% of EBITDA, paid after the 85% capital partner share is determined.

4. GPU allocation: Tier-1 NVIDIA supply agreements through HPE and Supermicro, with confirmed allocations for B300 and H200 capacity covering the Site 01 build.

5. Revenue ramp: First unit online October 2026; full fleet operating cadence by Q2 2027; >95% utilization benchmark contractually defined.

6. Exit: Three documented paths (refresh-and-extend, sale-as-going-concern, hardware liquidation), with operator obligation to support all three.

7. Track record: DCXPS founding team carries 60+ years of combined Fortune-500 data-center delivery experience across Deutsche Telekom, T-Mobile, and other tier-1 carriers.

The framework, summarized

The AI infrastructure capital cycle of 2025–2030 will distribute roughly $5.2 trillion across the six vehicles described above (per McKinsey’s base-case projection). Most of that capital will flow into hyperscaler equity, GPU vendor equity, and listed neocloud equity, because those vehicles are the most liquid and the most familiar to institutional allocators.

A meaningfully smaller slice — perhaps $200–400 billion globally — will flow into hardware-backed SPV structures. This slice is where the structural return premium sits, because it is where the pricing inefficiency exists. Hardware-backed AI infrastructure exposure is harder to access (no public market), more diligence-intensive (operator selection matters), and less liquid (5–7 year hold). The premium that compensates for these frictions is what capital partners are capturing right now.

DCXPS’s Site 01 deployment at Kladno is one specific expression of this opportunity. Nine SPV positions. $45M unit size. $15M minimum participation. Ring-fenced Delaware LLC SPV per unit. Hardware owned by SPV. Operator paid 15% of EBITDA, second in the waterfall. EU-sovereign positioning. October 2026 first units online.

For data-room access: investors@dcxps.com.

Closing the loop

This is the final article in the five-part series. Pulling the thesis together:

• Article 1 established that power, not silicon, is the binding constraint on AI compute — and that modular distributed deployment at sites with existing grid capacity is the structurally correct response.

• Article 2 worked through the unit economics of a $45M modular SPV in forensic detail, demonstrating that the 2.75× / ~29.2% yield profile is mathematically supported by the operating waterfall.

• Article 3 identified the EU AI Act / CLOUD Act conflict as the regulatory force creating structural demand for genuinely EU-sovereign infrastructure that US-incorporated operators cannot fully address.

• Article 4 mapped the training-to-inference workload transition through mid-2027 and demonstrated how modular distributed deployment is structurally aligned with the post-transition equilibrium.

• Article 5 (this one) framed AI infrastructure as a distinct asset class, identified the six available exposure vehicles, and argued that the hardware-backed SPV structure delivers a specific set of features no other vehicle replicates — at a return premium that compensates for the additional risk.

The thesis composes into a single argument: the next phase of AI infrastructure value is captured by EU-sovereign, modular, hardware-backed exposure deployed at sites with existing power, ahead of the August 2026 regulatory window and the mid-2027 workload transition. The capital partner who deploys into this structure now captures a return premium that will compress as the asset class matures.

We are deploying that thesis at Kladno. The data room is open. The first units come online in October 2026.

The factory is being built. The question is whether you own a piece of it.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek bare-metal GPU cloud platform. The complete five-part series is available on aiofthecoast.dcxps.com.

For investor relations: investors@dcxps.com
For sales and partnerships: sales@chapek.ai
Web: dcxps.com · chapek.ai

Part 4 of a five-part series on the structural opportunity in modular AI infrastructure

There is a pattern in technology infrastructure cycles that recurs reliably enough to be useful as a planning heuristic. The infrastructure built during the first phase of any major compute paradigm — the burst phase, the experimental phase, the phase where the model architecture itself is still being figured out — is almost never the infrastructure that turns out to be optimal for the second phase, the production phase, the phase where the workload settles into a steady-state economic equilibrium.

This was true of the mainframe-to-minicomputer transition. It was true of the on-premise-to-cloud transition. It was true of the desktop-to-mobile transition. And it is becoming true, right now, of the training-to-inference transition that the AI infrastructure market crosses through over the next eighteen months.

The investor implication is straightforward but worth stating directly: the infrastructure being built today, on the assumption that frontier-model training is the dominant workload, will not be the infrastructure that monetizes the production AI economy from 2028 onward. The geography, the form factor, the latency profile, the customer mix, and the unit economics of inference-dominant AI compute are materially different from those of training-dominant AI compute.

DCXPS’s modular distributed architecture is not optimized for the training-burst world. It is optimized for the inference-continuous world that follows it. And the 6-year SPV horizon described in Article 2 is calibrated to the exact transition curve where the value reallocates.

Here is the analysis behind that positioning.

The flip

The market-sizing data points that frame this transition:

• In 2025, training workloads accounted for approximately 62% of AI compute demand, with inference at ~38%.

• By mid-2027, inference workloads cross training in aggregate compute consumption.

• By 2030, the projected mix is inference at ~60–65%, training at ~35–40%.

• On energy footprint specifically, multiple independent estimates (IEA, McKinsey, Schneider Electric) converge on inference reaching 60% of AI energy consumption by 2027, despite using less per-query compute than training.

The shift from training-dominant to inference-dominant is not a marginal rebalancing. It is the largest workload-mix transition the data center industry has experienced since the mobile-first transition reshaped consumer compute in 2010–2015.

The mechanism driving it is straightforward and not particularly controversial. Frontier model training is, by its nature, a periodic event — a single model is trained once (or a small number of times for fine-tuning), and that training run consumes enormous compute for weeks or months. Inference is the productive deployment of those trained models. Every query, every API call, every agent step, every chatbot interaction, every image generated, every line of code completed — each one consumes inference compute. As enterprise AI deployment moves from proof-of-concept to production scale, inference volume grows by orders of magnitude even as training cluster size grows linearly.

The clearest signal that we are already inside this transition. NVIDIA’s most recent earnings disclosures and capital markets commentary explicitly call out inference as the dominant near-term growth vector for Blackwell deployment. AMD’s MI300X/MI350 product positioning is explicitly inference-first. Hyperscaler custom silicon (Google’s TPU v6, AWS Trainium-3, Meta MTIA v3) is increasingly inference-optimized rather than training-optimized.

The capital market is pricing the transition. What most operators have not yet internalized is what it changes about where and how you build the infrastructure.

What changes when inference takes over

Six characteristics of AI compute infrastructure change materially as the workload mix shifts from training-dominant to inference-dominant. Each has direct implications for site selection, deployment architecture, and unit economics.

1. Single-node performance becomes more important than aggregate cluster performance

A frontier training run requires tens of thousands of GPUs operating in synchronized parallel, with extreme demands on inter-node bandwidth and latency. Training infrastructure is dominated by the question of “how do I keep 25,000 GPUs lockstep at 99.9% efficiency for six weeks.” This is why hyperscale training clusters live in single-building campuses with InfiniBand fabric and meticulous topology engineering.

Inference is overwhelmingly single-node or small-cluster. A typical LLM inference request runs on 1–8 GPUs. A computer vision inference runs on 1 GPU. A recommendation model inference runs on a fraction of one. The aggregate compute is enormous, but the unit of compute is small. This means:

• Infrastructure can be physically smaller and geographically distributed without performance penalty.

• Inter-node bandwidth requirements are dramatically lower (you do not need 800 Gb/s InfiniBand for a single-node inference query).

• Hardware failures have local rather than systemic impact (one container down ≠ entire training run lost).

2. Geographic distribution becomes a feature, not a cost

Training infrastructure benefits from physical concentration. Inference infrastructure benefits from physical distribution. The reason is latency.

Modern agentic AI applications, real-time copilots, autonomous-systems control loops, robotics control planes, and consumer-facing AI products operate on sub-50-millisecond round-trip latency budgets. At the speed of light in fiber, that’s roughly 4,000 km of round-trip distance under ideal conditions — and real-world routing typically delivers 30–40% of theoretical, meaning a sub-50ms budget translates to 1,000–1,500 km of actual reach in practice.

The implication: a hyperscale data center campus in northern Virginia cannot serve a real-time inference workload in Prague, Madrid, or Helsinki at the required latency. Not even if the data center is fast, not even if the network is well-engineered. It is a physics problem, not an engineering problem.

This is why the inference geography looks structurally different from the training geography. Training can be done at the most power-efficient site available, anywhere on the planet. Inference must be done near where the inference is consumed — which means distributed, regional, and increasingly metro-edge.

3. Utilization curves change shape

Training workloads exhibit a “burst” utilization profile. A new model is announced; training begins; cluster utilization runs at 95%+ for weeks; training completes; utilization drops. The capital model has to monetize the high-utilization period sufficiently to amortize the deep-cycle equipment cost.

Inference workloads exhibit a “continuous” utilization profile. Customer queries arrive 24/7. The diurnal pattern is real (utilization is higher during business hours) but the dynamic range is much narrower — typically 60–85% utilization on a continuous basis, rather than 0–100%.

For modular operators, this is the friendlier curve. Continuous utilization is easier to forecast, easier to bill against contractual commitments, and produces more predictable EBITDA. The 96% margin profile described in Article 2 is structurally easier to defend against an inference-continuous mix than against a training-burst mix.

4. Hardware specifications diverge

Training-optimized hardware (H100, B200, B300 in their training configurations) is dominated by high-bandwidth memory (HBM3E, soon HBM4), extreme inter-node interconnect (NVLink, NVSwitch, InfiniBand), and maximum FLOPS per dollar.

Inference-optimized hardware emphasizes memory capacity (to hold larger models in single-GPU memory), memory bandwidth (for token generation throughput), and energy efficiency (because the workload runs continuously rather than in bursts).

The H200, in particular, is interesting in this context. Its higher HBM3E capacity (141 GB vs. H100’s 80 GB) makes it materially better than H100 for inference of larger models that did not fit in H100 memory. The B300 inherits this profile and extends it. Both of these chips are inference-relevant in a way that the original H100 generation was not. This is part of why our SPV unit composition is 14 B300 + 35 H200 — it produces a mix that is well-positioned for the workload transition over the SPV’s 6-year term.

5. Customer mix broadens

Frontier model training is concentrated among a small number of customers: OpenAI, Anthropic, xAI, Meta AI, Google DeepMind, a handful of Chinese labs, and the major European frontier labs. The customer count is in the low double digits.

Inference customers are every enterprise running AI in production. The customer count for the inference economy is in the millions. The pricing dynamic is different (smaller average ticket but more buyers), the procurement cycle is different (faster, transactional, often self-serve), and the loyalty profile is different (price-sensitive, performance-sensitive, multi-cloud by default).

For the Chapek platform — DCXPS’s bare-metal GPU cloud monetization layer — this transition is operationally favorable. We are built for the inference-dominant world: self-service provisioning, transparent hourly pricing, no orchestration tax, geographic distribution that puts compute near demand.

6. Energy economics change

The training workload favors absolute lowest energy cost regardless of location — which is why hyperscale training facilities locate near hydropower, nuclear, or stranded gas, and why Crusoe’s bitcoin-flare model worked.

The inference workload trades energy cost for latency. A site that is 20% more expensive on power but 200 km closer to demand may be the better economic site for inference because the value of the latency reduction exceeds the energy cost premium.

This is the dynamic that makes Central and Eastern European positioning particularly attractive for inference workloads serving the EU enterprise market. Czech power costs are competitive with German and French rates. The latency profile from Prague to Frankfurt, Vienna, Munich, Warsaw, Budapest, Berlin, and Amsterdam all fits within real-time inference budgets. The regulatory positioning (EU member state, EU AI Act jurisdiction, GDPR-compliant) is a structural advantage for the EU-sovereign demand pool described in Article 3.

What this means for capital deployed in 2026

The investment thesis question is not “is inference becoming bigger than training” — that is settled, the data are clear, the only debate is timing. The question is: what does an infrastructure portfolio look like that is positioned for the post-transition equilibrium, deployed inside the transition window?

Five propositions follow from the analysis above:

Proposition 1 — Distributed beats concentrated, in 6-year horizon terms

Capital deployed into single-campus hyperscale builds is exposed to the risk that the workload mix being optimized for is not the workload mix that exists at the operating end of the lifecycle. Capital deployed into geographically distributed modular capacity has the option to serve either training (with intra-site clusters of multiple containers) or inference (with single-container deployments serving regional demand) — and the architecture is flexible enough to shift between them as the mix evolves.

Proposition 2 — The training cluster premium is compressing

Hourly pricing for top-tier training clusters has been declining since mid-2024. The squeeze comes from two directions: hyperscaler internal capacity ramping up, and inference-economy demand absorbing the marginal GPU at lower per-hour rates than the peak training rate. The peak per-hour pricing for training-burst configurations may not return to 2024 levels. Operators whose capital model assumes those rates are exposed.

Proposition 3 — Hardware-fungibility matters

The B300 and H200 configurations in our SPV structure are not training-specific or inference-specific. They are hardware platforms that can be rebalanced toward either workload as customer demand evolves. The Chapek platform supports this rebalancing through its bare-metal provisioning model — capacity can be allocated to training tenants or inference tenants as the market mix dictates. Capital partners are not exposed to a hardware bet on one workload type.

Proposition 4 — EU latency geography is undervalued

The market is currently pricing AI infrastructure substantially on its training capability (power cost, cluster scale, interconnect bandwidth). It is not yet pricing AI infrastructure on its inference geography (latency reach, regional sovereignty, regulatory positioning). As the workload mix transitions, the EU-distributed-and-sovereign positioning will reprice. Operators positioned ahead of that reprice capture the value transfer.

Proposition 5 — The 6-year SPV horizon spans the transition perfectly

A unit deployed Q4 2026 begins commercial operation in 2027 — the year inference crosses training in aggregate compute. The unit operates through 2032. The first half of its operating life monetizes the late-training era; the second half monetizes the established-inference era. This is not a coincidence. The 6-year term was calibrated specifically to span the transition.

The technology arc — what 2028 looks like

For investors thinking 5+ years out, the technology arc that will shape the second half of the SPV term:

HBM4 deployment (2026–2027). NVIDIA Rubin, AMD MI400 series. ~15 TB/s memory bandwidth, vs. ~9 TB/s for current HBM3E. The performance uplift is meaningful for both training and inference, but the inference impact is more economically significant — it dramatically improves per-query throughput for large models.

Inference-specific silicon at scale (2027–2028). The economics of dedicated inference chips improve as the inference market matures and reaches a size that justifies dedicated silicon investment. Google TPU v7, AWS Inferentia-3, Meta MTIA v4, plus startup entrants (Cerebras, Groq, SambaNova) scaling production volume. This is a risk for general-purpose GPU revenue per hour — and a reason to think carefully about lifecycle planning at the unit level.

Mixture-of-experts and model efficiency (continuous). Architectural innovations are reducing per-query compute requirements by 30–50% for many workloads. This is offset by expansion of the addressable user base (Jevons paradox) but should be modeled as a moderating factor on aggregate accelerator demand growth.

Edge inference (2028+). 5G/6G-native AI workloads, autonomous vehicle compute, industrial IoT, smart-city infrastructure. This is a market growing from a small base but at significant CAGR. It is structurally aligned with the distributed modular thesis — these workloads cannot be served from hyperscale campuses.

The DCXPS roadmap accounts for each of these technology arcs:

• Generation 1 (current): 1 MW MADC units. Brownfield grid. Human technicians. Revenue from month five.

• Generation 2 (2027–2028): MADC+ at 2 MW. On-site CHP. First-generation robotics. 85%+ grid independence.

• Generation 3 (2028–2030): DDCU at 4 MW. Dark, distributed, dynamic. Lights-out compute. Software-defined energy management.

• Generation 4 (2030–2032): 10 MW DDCU. Humanoid robots in operational role. Integrated agricultural co-loads. EBITDA margin profile expanding toward ~93%.

• Generation 5 (2032–2034+): 100–200 MW “Robotic DC City” campuses. 200–500 humanoid robots. 5–10 human orchestrators.

The Generation 1 SPV that a capital partner commits to in 2026 is not exposed to the Generation 5 build — but the operator’s roadmap, the unit-level technology refresh paths, and the residual-value pathways do tie into the broader trajectory. Capital partners hold the option to participate in the next generation, but their exposure is locked at the unit-level economics of the generation they entered.

The operator framework

For DCXPS as operator, the workload mix transition creates a specific set of execution priorities for the next 24 months. I am laying these out openly because capital partners ought to be able to verify, in diligence, that the operator has thought about the right questions:

Priority 1 — Customer mix construction. The Chapek platform must build a customer mix that balances training tenants (for peak utilization revenue) with inference tenants (for utilization-stability revenue). Our target mix at 18 months of operation is 35% training / 65% inference, weighted by revenue. This is a deliberate mix; pure-training mix exposes us to the burst-cycle compression risk, pure-inference mix gives up the upside of peak training rates.

Priority 2 — Geographic expansion ahead of inference reprice. The Site 01 Kladno deployment is the anchor. Subsequent sites prioritize EU-distributed latency geography — central Czechia, Hungary, Poland, additional Western European positions where co-located power exists. Each site is structured as its own SPV pool.

Priority 3 — Hardware refresh discipline. The SPV operating term is 6 years, but hardware generations turn over in 18–24 months. The operator’s job is to optimize the workload mix on each generation as it ages — moving an H200-heavy unit toward inference workloads as B300/Rubin take over the training tenant base, and so on. This is the lifecycle management that capital partners hire us to do.

Priority 4 — Customer development at the regulated-enterprise tier. The inference customer mix that monetizes the EU-sovereign positioning is BFSI, healthcare, government, automotive, frontier AI labs (per Article 3). Each of these is a long-cycle enterprise sales motion. We have started; we are not done.

Priority 5 — Platform investment in audit, observability, and compliance tooling. The AI Act enforcement window (August 2026 onward) will commercially differentiate operators with mature compliance tooling from operators offering raw compute. Chapek’s roadmap includes auditability, training-data-traceability, model-versioning, and human-oversight integration tooling as differentiation points for the regulated-enterprise customer tier.

Where this leads

For capital partners, the inference-transition thesis is a confirmation argument for the modular distributed structure, not a contradiction of it. The 6-year SPV horizon is structurally aligned with the workload mix transition. The hardware composition (B300 + H200) is structurally aligned with the workload profile. The geographic positioning (EU-distributed, latency-optimized) is structurally aligned with where the inference economy actually consumes compute.

The DCXPS Site 01 deployment at Kladno is the deployment vehicle. Nine SPV positions. $45M unit size, $15M minimum participation. 2.75× modeled multiple, ~29.2% simple yield, 85/15 EBITDA waterfall. First units online October 2026.

For data-room access: investors@dcxps.com.

The final article in this series steps back from the operating thesis and looks at the asset-class framing. Why hardware-backed exposure to AI infrastructure is structurally different from platform equity, from tokenized compute claims, from listed neocloud equity, and from traditional infrastructure asset classes. And why, for a specific class of capital partner with a specific risk-return mandate, the SPV-owned modular structure is the only one of these vehicles that fits.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek bare-metal GPU cloud platform. Previous in this series: “Power Is the New Silicon,” “The 195-Day Data Center,” and “The CLOUD Act Conflict.” Final: “Own the Metal — A Framework for AI Infrastructure as an Asset Class.”

Every fifteen years or so, European technology policy produces a regulation that the rest of the world initially dismisses as continental bureaucracy, then quietly internalizes as global commercial reality. GDPR was the last one. The EU AI Act is the next one. Both were underestimated in their year of enactment. Both became, within five years, the binding compliance ceiling for any company that wanted to do business in the European single market.

The EU AI Act enters its decisive enforcement phase on 2 August 2026. As of this writing, the binding question is not whether the regulation will be enforced — it will be — but whether the European Commission’s Digital Omnibus proposal, which would defer the high-risk obligations until 2 December 2027, will be adopted before that deadline. The second political trilogue on the Omnibus, held on 28 April 2026, ended without agreement. The default outcome — the one prudent organizations are planning against — is enforcement from August.

Most coverage of the EU AI Act treats it as a compliance project: a set of documentation, audit, and governance obligations that EU-operating AI companies must satisfy to avoid penalties of up to €35 million or 7% of global turnover for prohibited-AI violations, or €15 million or 3% of global turnover for high-risk system breaches.

That framing is true, but it misses the larger structural force that the EU AI Act unlocks. The Act, combined with GDPR and the still-unresolved US CLOUD Act conflict, creates a market — not a compliance burden — for genuinely EU-sovereign AI infrastructure. That market is being capitalized in real time, the precedents are forming now, and the operators positioned inside it before August 2026 will be the ones that European enterprises and governments default to for the next decade.

This is the third pillar of the DCXPS thesis. It is why our first site is Kladno, not Atlanta. And it is the reason that the time-to-deploy question — the 195-day modular thesis from Article 1 — becomes more than an efficiency advantage: it becomes a regulatory window advantage.

The CLOUD Act problem, in three sentences

The US Clarifying Lawful Overseas Use of Data Act of 2018 (the “CLOUD Act”) permits US law enforcement agencies, with appropriate legal process, to compel US-incorporated cloud providers and their subsidiaries to produce data in their possession, custody, or control — regardless of where that data physically resides.

The EU General Data Protection Regulation (GDPR) requires that personal data of EU residents be processed only in conditions that protect fundamental rights, with restrictions on transfers outside the EU/EEA and protections against unauthorized governmental access.

A US-incorporated cloud provider operating an EU data center cannot, as a matter of statutory construction, fully satisfy both regimes. Data residency — i.e., the physical location of the server — is not the same as data sovereignty. The provider’s parent jurisdiction is the binding one.

This is not a theoretical concern. It has been litigated. The Court of Justice of the European Union’s Schrems II ruling (2020) invalidated the EU-US Privacy Shield specifically on these grounds. The EU-US Data Privacy Framework (2023) was constructed as a replacement but is now under renewed legal challenge for the same underlying reason: US surveillance law is structurally inconsistent with EU fundamental rights protections.

For most enterprise workloads, this is a manageable risk. For workloads involving regulated personal data — healthcare records, financial transactions, biometric identifiers, defense and intelligence applications, government services — it has become a binding commercial constraint. The customer cannot, as a matter of law, use a US-incorporated provider for that workload. Even if the data never leaves Frankfurt.

The EU AI Act, by introducing audit, traceability, and governance obligations on top of this, makes the question more acute. Who can the auditor compel to produce training data, model weights, system logs, and governance documentation? If the answer is “a US-incorporated entity subject to US legal process,” then the AI Act compliance position is structurally compromised — irrespective of where the GPU physically sits.

This is the gap that EU-sovereign AI infrastructure operators exist to fill.

The market is capitalizing this gap, right now

The clearest evidence that capital markets have priced in the sovereign-infrastructure thesis is the Mistral AI debt financing announced on 30 March 2026. The structure is worth examining in detail because it is the precedent that the next two years of EU AI infrastructure financing will be built on.

Mistral raised $830 million ($722 million) in debt — its first-ever debt financing — to fund a 44 MW data center at Bruyères-le-Châtel, south of Paris. The capital purchases 13,800 NVIDIA GB300 GPUs, with operations starting in Q2 2026.

The composition of the lending syndicate is the part most worth attending to. Seven banks: BNP Paribas, Crédit Agricole CIB, HSBC, MUFG, Bpifrance, La Banque Postale, Natixis CIB. Six European institutions plus one Japanese partner. No US bank participation.

That is not coincidence. That is structural signal. European banks are willing to underwrite AI infrastructure debt against EU-sovereign assets, EU-resident management, and EU-resident customer revenue. They are not extending the same terms to assets with US-jurisdiction exposure. The capital structure of European AI infrastructure is being deliberately ringfenced from US legal reach by the institutions that finance it.

For context on what this represents: Mistral’s annual recurring revenue grew from approximately $20 million to $400 million in a single year, with a $1 billion target by end-2026. The customer base includes the French armed forces, BPifrance, ASML, SAP, IBM, Cisco, Stellantis, and Accenture. These are not customers that can run sensitive workloads on US-incorporated cloud providers. They are the prototype of the European AI demand profile.

The broader pattern. In the first four months of 2026 alone, European AI infrastructure raised, by independent count:

• Mistral AI: $830M debt + ongoing Series C (€1.7B at €11.7B valuation in Sept 2025)

• Nscale (UK): $2 billion equity, alongside a reported $23 billion Microsoft contract for 200,000 GB300 GPUs

• Wayve: $1.2 billion (autonomous driving, GPU-intensive training)

• AMI Labs (Yann LeCun’s France-based lab): $1 billion

• Ineffable Intelligence (London, David Silver / DeepMind alum): $1.1 billion seed — Europe’s largest-ever seed round

That is over $6 billion of fresh capital flowing into EU-sovereign AI capacity in a four-month window. Add the European Commission’s €15 billion AI Factories program as the public-sector anchor, and the total addressable capital pool for EU-jurisdiction AI infrastructure stands at well over €20 billion in 2026 alone — against an installable capacity that is currently a fraction of what the demand profile requires.

The demand side — who is buying

The customers that the sovereign infrastructure thesis serves fall into five categories. Each has different price sensitivity, different latency requirements, and different procurement timelines, but they share the structural fact that they cannot use US-incorporated providers for their AI workloads.

Category 1 — National governments and defense

The European Commission’s Joint Procurement Office for AI Factories, France’s DGA and CNRS, Germany’s Bundeswehr Cyber and Information Domain Service, the UK’s MoD AI Lab, the Czech NCSA. These customers buy compute on multi-year frameworks with strict jurisdictional requirements. They are not price-sensitive in the conventional sense; they are sovereignty-sensitive. The contracts are large, multi-year, and dominated by relationship rather than spot-market pricing.

Category 2 — Financial services and insurance

European retail banking, insurance, capital markets infrastructure, payment processors. Workloads include fraud detection, risk modeling, anti-money-laundering compliance, algorithmic trading, credit scoring (specifically named as a high-risk category in Annex III of the EU AI Act). GDPR exposure on customer data is the binding constraint; AI Act compliance is the emerging binding constraint. Both push these workloads out of US-incorporated providers.

Category 3 — Healthcare and life sciences

EU-resident pharmaceutical companies (Sanofi, Bayer, Roche, Novartis, AstraZeneca’s EU operations), national health services, clinical research organizations, genomics platforms. Patient data sovereignty is the binding constraint. The AI Act adds requirements around medical AI as a regulated product (Annex I). This is one of the highest-growth verticals in EU AI demand.

Category 4 — Industrial and automotive

The European automotive sector is rebuilding its AI infrastructure stack as the autonomy and digital-twin compute requirements converge. Mercedes, BMW, Volkswagen, Stellantis, Renault. None of these companies can route training data through a jurisdiction that may compel disclosure of competitive engineering data. Stellantis is already a Mistral customer for exactly this reason.

Category 5 — Frontier and applied AI labs

European AI labs (Mistral, Stability, Aleph Alpha, Helsing, Black Forest Labs, Wayve, Ineffable Intelligence) that have specifically chosen to build their commercial position on EU jurisdictional positioning. These are the customers most aligned with the sovereign infrastructure thesis because their own customer pitch depends on it.

If you sum the addressable spend across these five categories at a 2027–2030 horizon, you arrive at a market materially larger than European AI infrastructure can currently serve. McKinsey’s base-case projection of 125 GW of incremental global AI capacity through 2030 implicitly assigns roughly 20–25 GW to Europe, against a current pipeline that can realistically deliver less than half of that on conventional grid-connected timelines.

This is the gap. It is bigger than any single operator can fill. It is also denominated, almost entirely, in workloads that must sit in EU-jurisdiction infrastructure.

The regulatory mechanics — what August 2026 actually changes

For the operators, customers, and capital partners thinking about this market, the technical content of the AI Act’s August 2026 enforcement matters. Let me name the parts that have direct infrastructure implications.

Article 6 / Annex III — high-risk AI systems

The categories of AI applications that become high-risk on 2 August 2026 include:

• Biometric identification systems

• AI used in critical infrastructure (energy, water, transport)

• AI used in education and vocational training admissions and assessment

• AI used in employment, worker management, and access to self-employment

• AI used in access to and enjoyment of essential private and public services (including credit scoring and insurance pricing)

• AI used in law enforcement

• AI used in migration, asylum, and border control

• AI used in administration of justice and democratic processes

For each, the AI Act mandates: risk management systems, data governance, technical documentation, automatic event logging, transparency provision to deployers, human oversight, and accuracy and robustness requirements. These are not boilerplate. They imply infrastructure capability — the platform must support detailed logging, audit access, traceability of training data, model versioning, and reproducibility of inference outputs.

Article 50 — transparency obligations

Providers of generative AI systems and deployers of AI systems that generate or manipulate audio, image, video, or text must label outputs as AI-generated. The platform-level implication: provenance metadata, watermarking infrastructure, audit trails.

Articles 53–55 — General-Purpose AI Models

GPAI obligations have technically applied since 2 August 2025; the penalty regime begins in August 2026. For systemic-risk models (currently a small set, but the threshold will not stay where it is), additional obligations apply: model evaluation, systemic risk assessment, incident reporting, cybersecurity protections.

Article 99 — penalties

The penalty structure that the market is pricing against:

• Up to €35 million or 7% of worldwide annual turnover (whichever is higher) for prohibited AI practices.

• Up to €15 million or 3% of worldwide annual turnover for high-risk system violations.

• Up to €7.5 million or 1% of worldwide annual turnover for providing incorrect information to authorities.

For context: a US hyperscaler with global revenue around $250 billion is exposed to a maximum single penalty of $17.5 billion under the high-risk provisions. That is not a theoretical exposure number that gets priced at zero in board-level risk reviews. It is a structural reason to ensure that AI workloads with EU-exposure are running on infrastructure where the compliance position is genuinely defensible.

Why modular wins this race

The combination of (a) accelerating EU sovereign demand, (b) 2 August 2026 enforcement deadline, (c) ongoing capital availability for EU-jurisdiction infrastructure, and (d) the universal grid-bottleneck problem detailed in Article 1, creates a competitive landscape with a specific shape:

• Hyperscaler greenfield builds: too slow. The 36–60 month timeline does not survive the August 2026 deadline.

• US-incorporated neoclouds: structurally unable to claim sovereign positioning regardless of where the assets sit.

• Conventional EU colocation operators (Equinix EU, Digital Realty EU, Interxion): legally EU-resident but parent-jurisdiction-exposed; partial solution, not complete.

• EU-incorporated frontier AI labs (Mistral, Aleph Alpha): building their own infrastructure for their own workloads; not a marketplace.

• EU-incorporated operators with modular deployment capability at sites with existing power: the architecturally correct fit for the window between August 2026 and the next major capacity expansion.

DCXPS sits in the last category by deliberate design. Czech entity (DCXPS a.s.). EU-resident management plane. Czech and broader CEE operational footprint. Site 01 at Kladno, co-located with an existing 200 MW power plant — meaning we are not exposed to the multi-year transmission upgrade timeline that gates most greenfield builds in Western European data center markets.

Most importantly: the 195-day deployment timeline matches the regulatory window. A capital partner committing to a DCXPS SPV in May 2026 sees first units online in October 2026, full operating cadence by mid-2027. That timeline lands inside the AI Act enforcement window, ahead of the demand surge that the enforcement deadline triggers, and at a moment when the cost of capital for EU-sovereign AI infrastructure is still being established.

Two years from now, that cost of capital will be established, the premium that early operators captured will be visible in cap tables, and the question of “should we have moved on this in 2026” will have an answer.

The implementation framework

For capital partners, family offices, and strategic investors thinking specifically about how to position for the EU sovereign AI thesis, here is the framework I would apply:

Step 1 — Define the exposure. Is the goal direct asset ownership (modular SPV structures), platform equity (neocloud operator equity at the EU-resident operator level), or strategic positioning (anchor customer relationships with future capacity preference)? Each has different risk-return characteristics and different liquidity profiles.

Step 2 — Validate the jurisdictional structure. EU-incorporated parent entity, EU-resident management, EU-resident data plane, EU-resident customer revenue. Each of these is verifiable in corporate documents. Anything short of all four is partial sovereignty, not full.

Step 3 — Confirm the power pathway. As detailed in Article 1, this is where the operator’s claim of “we can deploy in months” stands or falls. Off-take agreements with documented term and indexing structure. Co-located generation or secured grid connection. Not “we are in conversations with utilities.”

Step 4 — Confirm the customer pipeline. EU sovereign demand is structurally underwriting this thesis. But not every operator has the customer development capability to convert that demand into contracted revenue. Demand structured commercial pipeline evidence — not press releases.

Step 5 — Confirm the regulatory positioning. Has the operator’s legal team mapped its position against the AI Act’s specific provisions? Is the platform’s logging, audit, and traceability capability mapped to Annex III requirements? This is increasingly a buyer’s diligence question; expect it to become a financing diligence question by Q4 2026.

Step 6 — Confirm the operator’s track record at unit-scale operations. Same diligence framework as Article 2 — years of experience, sites operated, uptime delivered. Sovereignty is necessary but not sufficient; operational competence is the binding constraint on actually monetizing the regulatory position.

For DCXPS, the answers to all six questions are documented and available in the data room. We are happy to walk through them line by line.

Where this leads

The Site 01 fleet at Kladno is the first commercial expression of the DCXPS sovereign infrastructure thesis. Nine modular unit positions, $405M of contracted capacity, 6-year operating term, ring-fenced Delaware SPV structure with hardware owned by capital partners and operated by DCXPS under the 85/15 EBITDA waterfall described in Article 2.

The Czech jurisdiction matters here. EU membership, GDPR application, EU AI Act application. The Delaware SPV structure provides US-investor familiarity with the legal vehicle while the underlying operating entity (DCXPS a.s.) ensures the EU jurisdictional positioning on the data plane.

For capital partners specifically thinking about the EU sovereign thesis — family offices with European mandate, sovereign-adjacent capital pools, strategic investors with European customer exposure — this is the deployment vehicle that exists now, operating now, with first revenue in October 2026.

The window closes on 2 August 2026. Operators that are deployed and operating before that date will be positioned for the demand surge that follows. Operators still in greenfield development will be priced out of the early commercial cycle, regardless of how attractive their long-term roadmap looks.

For data-room access: investors@dcxps.com.

The next article in this series examines the second large structural shift inside the AI compute market — the training-to-inference transition that crosses over by mid-2027, and why modular distributed deployment is structurally better suited to the inference-dominant world than the training-dominant infrastructure being built right now. That transition is what makes the 6-year SPV term horizon mathematically aligned with the next phase of the cycle.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek bare-metal GPU cloud platform. Previous in this series: “Power Is the New Silicon“ and “The 195-Day Data Center.” Next: “From Training Burst to Inference Continuous — Why 2027 Reshapes the Compute Map.”

Part 2 of a five-part series on the structural opportunity in modular AI infrastructure

In the first article in this series, I argued that the binding constraint on AI compute is power, not silicon — and that this structural shift creates a specific opportunity for modular deployment at sites where grid capacity already exists. That argument lives or dies on the unit economics. If the math doesn’t work, the macro is irrelevant.

So let’s work the math. Openly, transparently, with every assumption on the table. This is the analysis I would expect from any institutional capital partner doing their own diligence — and it is the analysis you should demand of any operator asking you to commit $15M, $30M, or $45M to a hardware-backed AI infrastructure SPV.

What follows is the actual unit-level economics of a single DCXPS modular AI data center, structured as a ring-fenced Delaware LLC SPV. The numbers are drawn directly from our Confidential Investor Memorandum (April 2026) and our public pricing at chapek.io. I will show you what the math says, where the sensitivities are, what the comparables look like, and where you should push us in diligence.

The asset

Each SPV unit comprises:

• 2 × 40-foot ISO containers, reinforced for structural and thermal envelope requirements

• 450 kW IT load at design PUE of 1.3 (total facility power: ~585 kW)

• 49 GPU servers total, split:

  • 14 NVIDIA B300 servers (Blackwell Ultra generation)

  • 35 NVIDIA H200 servers (Hopper generation, high-bandwidth-memory configuration)

• Mixed cooling: forced-air for H200 capacity, direct-to-chip liquid for B300 capacity

• Spine-leaf RDMA fabric (Cisco), high-throughput NVMe storage tier, out-of-band management

• Modular UPS, switchgear, busway (Schneider Electric), BMS telemetry, fire suppression, physical security

Total unit cost: $45 million, of which:

This is list pricing. It is also what we charge today on chapek.io for production capacity. Several layers of conservatism sit beneath that number that I want to call out explicitly, because no analyst worth their bonus believes a list-price assumption at face value:

1. Utilization assumption. The modeled revenue assumes 8,760 hours per year — i.e., 100% time availability. Our contractual benchmark is >95% utilization, but utilization ≠ availability. The math above is operationally optimistic at the revenue line, then conservative everywhere else. In practice, a blended 90–95% billable utilization is a more realistic central case; the 6-year aggregate reflects a blended ramp plus discount profile.

2. No reserved-instance discount modeling. Real customers buy multi-month and multi-year reservations at 15–30% discounts to spot. The modeled $25.28M does not net that out — meaning realized revenue at scale will sit below list. Counterbalanced by the fact that reserved revenue is contracted, predictable, and refinanceable.

3. No price decay modeling. GPU hourly rates compress over time. H100 hourly rates declined ~30% across 2024–2025 as supply came online. B300 will follow a similar curve. By year 4–5 of the SPV term, we expect blended pricing materially below year-1 list.

Net of these adjustments, our internal modeled central case for fleet revenue across the 6-year term sits in a $130–155M range against the $151.7M list-optimistic case. The 2.75× multiple holds across that range. The structure absorbs the price decay; the math does not require the upside.

1. Power cost is the dominant variable cost. At $222/MWh — a representative Central European industrial rate, embedded in a long-term off-take with a co-located generator — the unit consumes ~3.94 GWh/year. That’s $875,124 of electricity against $25.28M of revenue. ~3.5% of gross. This is the magic number that makes modular AI infrastructure economics work: power is cheap relative to the value of the compute it produces.

2. Personnel cost is near-zero at the unit level. The unit is monitored and operated through Chapek’s centralized NOC. Field intervention is exception-based, not continuous. The $108,000 annual admin overhead covers fractional NOC allocation, basic site presence, and corporate overhead.

3. There is no co-location margin layer. We are not paying a colo operator $200/kW/month for rack space. The container is the rack space, and the SPV owns it.

For comparison: a typical retail colocation revenue model runs at 20–35% EBITDA margins. Hyperscaler internal compute runs at 30–50% (estimated). The neocloud category, depending on operator, runs 35–65% on a gross-margin basis, with significant variation by lease structure. Bare-metal GPU-as-a-service at a power-co-located modular deployment, with no real estate lease and no colocation margin layer, structurally runs above 90% EBITDA before financing.

Sensitivity analysis. The number that breaks this model is not GPU pricing — it is power cost. At $222/MWh, electricity is 3.5% of gross. At $500/MWh, it becomes 7.8%. At $1,000/MWh — i.e., PJM at the cap — it becomes 15.6%, and the margin compresses to ~80%. Even at the cap, the unit economics are still extraordinary by every comparable benchmark. This is why site selection is everything.

The waterfall, in plain English

The SPV pays its capital partner first, then the operator. Specifically:

1. Gross revenue is collected by Chapek (the operating platform) and distributed to the SPV monthly.

2. OPEX (electricity, fixed admin) is deducted from gross to produce EBITDA.

3. EBITDA is split 85% to SPV capital partners / 15% to DCXPS as operator.

4. Project financing (if any) is deducted before the split in the SPV’s discretion. The model above is unlevered.

This is the headline number from the investor memorandum, derived openly. 2.75× of distributed cash on a $45M ticket over 6 years equates to ~29.2% per annum simple yield. On an IRR basis — which accounts for the cash flow timing across a 6-year build/operate cycle — the figure depends on the deployment ramp, but sits broadly in the 24–28% range for a single unit, before optionality value.

Where the optionality lives

The 2.75× multiple is the base case. There are three places where optionality sits, none of which is modeled in the headline number:

Optionality 1 — Refinancing. Once the unit is operating, the asset becomes financeable. A capital partner who put in $45M of equity can extract working capital by refinancing 50–70% of asset value into senior debt secured against the hardware and the cash flow. The cash extracted is redeployable into further units. This is the same structural play CoreWeave executed in scaling its $34B off-balance-sheet lease pipeline. At our scale, it is meaningful.

Optionality 2 — Hardware refresh and residual value. GPU hardware has measurable secondary-market value. H100s deployed in 2023 still trade actively in 2026 at 30–50% of original cost, depending on configuration. B300 will have a similar residual profile. At year 6, the SPV holds 49 servers of B300/H200 capacity with either (a) a secondary-market exit path, (b) a refresh-and-extend path with new hardware redeployed into the same containerized infrastructure, or (c) a repurpose path into inference-optimized configurations (more on this in Article 4). None of these residual paths is in the 2.75×.

Optionality 3 — Compute price inflation. The model assumes flat-to-declining GPU hourly pricing. If supply rationing intensifies — which the macro analysis in Article 1 suggests is likely — realized pricing could exceed model. We have not yet seen a market where modular co-located capacity systematically undersells the spot market over a multi-year period. The reverse is the historical pattern.

What it costs to be wrong

Every honest investor pitch should include the failure modes. Here are the four that matter for this structure, and what they look like quantified:

Failure mode 1 — GPU price collapse. If average GPU hourly pricing declines 50% over the 6-year term (vs. ~25% in our central case), the multiple compresses from 2.75× to ~1.9×. Still positive. Still better than most infrastructure asset classes. Not a wipeout.

Failure mode 2 — Power cost shock. If our co-located generator off-take is disrupted and we are pushed to grid power at PJM-cap-equivalent rates, EBITDA margin compresses from 96% to ~80%, and the multiple comes in around 2.3×. Still positive.

Failure mode 3 — Utilization collapse. If sustained utilization runs at 75% rather than 95% across the entire term — a scenario that implies a fundamental break in AI demand — the multiple sits around 2.0×. Still positive, but the IRR profile materially weakens.

Failure mode 4 — Operator failure. If DCXPS, as operator, fails to deliver contracted operational performance — uptime, billing, customer acquisition through Chapek — the SPV has remedies. The hardware is owned by the SPV. The container is owned by the SPV. The operator can be replaced. This is the structural alignment that ring-fenced ownership creates. You are not exposed to operator equity; you are exposed to the asset, with the operator as a service provider.

Where this model would break catastrophically is a scenario in which:

• AI inference demand collapses and

• GPU secondary market collapses and

• Power costs spike beyond model and

• The operator fails to execute and

• The structural EBITDA margin compresses below ~50%.

That stack of conjunctions is not, in any reasonable view, a base case. It is a stress-test scenario, and in that scenario most AI infrastructure exposure — equity in neoclouds, equity in hyperscalers, equity in GPU vendors — does materially worse than a hardware-backed SPV holding refinanceable assets.

How this compares

The 2.75× / ~29.2% yield profile sits inside a recognizable infrastructure-investment universe. To calibrate:

• Core US infrastructure (toll roads, airports, regulated utilities): 7–11% IRR, 1.5–1.8× over 7–10 years.

• Renewable energy project equity (utility-scale solar/wind): 8–13% IRR, 1.7–2.1× over similar tenor.

• Traditional retail colocation: 11–15% IRR, 2.0–2.5× over 7–10 years.

• Hyperscaler-leased build-to-suit data center: 13–18% IRR for development, lower for operated.

• Modular AI compute (DCXPS modeled): 24–29% IRR, 2.75× over 6 years.

The premium is real, and it is paid for in three forms of risk: (1) higher operator dependency than core infrastructure, (2) technology obsolescence cycle compression, (3) less established secondary-market liquidity for the underlying asset compared to commercial real estate.

That premium is what the cycle is paying right now to capital that can move at the speed of the modular deployment timeline. We do not believe it persists at this level indefinitely. That is the case for moving now rather than later.

The diligence checklist

If you are evaluating this structure — ours or any operator’s — the questions to ask, in order:

1. Who owns the hardware? If the answer is anything other than “the SPV directly,” the entire alignment story falls apart. Demand evidence of bills of sale, customs documentation, and SPV title.

2. Where does the power come from? Off-take agreements, term length, indexing structure, fallback provisions. A modular thesis without secured power is not a thesis.

3. What is the operator’s economic share, and is it paid first or second? If first, alignment is theoretical. If second, alignment is structural.

4. What is the GPU allocation pathway? Letter of intent vs. binding purchase order vs. delivered inventory. The further down that ladder, the less of a thesis you have.

5. What is the realistic revenue ramp? Linear, hyperbolic, S-curve? At what utilization, against what customer mix? “95% contractual benchmark” is a number to verify against contracts, not against models.

6. What is the exit mechanism? Refinance, resell to operator, resell to third party, wind-down with hardware liquidation. Any operator who cannot articulate at least three exit paths has not thought about your capital seriously.

7. What is the operator’s track record at unit-scale operations? Years of experience, sites operated, uptime delivered. This is not a venture bet. It is an industrial operations bet.

We are happy to answer all seven questions in writing, with documentation, in a structured data-room session. The first three are answered in the memorandum.

Where this leads

The Kladno site (Site 01) hosts 9 of these SPV unit positions. Total fleet capacity, at full deployment: $405 million of contracted infrastructure investment and a 6-year gross revenue potential of approximately $1.37 billion. First units online October 2026.

The minimum participation in any single SPV is $15 million. The full SPV size is $45 million per unit.

For data-room access, write to investors@dcxps.com.

The next article in this series moves to the third pillar of the thesis: why the August 2026 EU AI Act enforcement deadline — and the underlying CLOUD Act conflict — creates a structural opening for genuinely EU-sovereign AI compute that US-incorporated operators cannot fully address. That’s where the geographic positioning of Kladno stops being incidental and starts being decisive.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek bare-metal GPU cloud platform. Previous in this series: “Power Is the New Silicon.” Next: “The CLOUD Act Conflict — Why EU AI Sovereignty Is a Capital Allocation Problem, Not a Compliance Cost.”

There is a comfortable story about the AI infrastructure shortage that the market has been telling itself for two years. It goes like this: NVIDIA cannot ship H100s, then H200s, then Blackwell, fast enough. Hyperscalers fight for allocation. Neoclouds raise capital, get on the wait list, and as soon as the GPUs arrive, the revenue starts flowing.

That story is wrong now. It was already becoming wrong in 2024. By the time you finish reading this article it will be obvious that the binding constraint on AI compute is no longer the chip — it is the kilowatt-hour, the substation, and the five-year median wait to connect anything material to the North American or Western European grid.

This matters because almost every dollar deployed into “AI infrastructure” right now is being underwritten against the wrong scarcity. If you are a capital partner trying to deploy $15M to $200M into this cycle, the most expensive mistake you can make is to assume that the limiting reagent is silicon. It isn’t. It hasn’t been for a year.

Let me show you the numbers.

The 2,060 GW backlog

The Lawrence Berkeley National Laboratory tracks every project sitting in U.S. transmission interconnection queues. Its most recent dataset, updated through end of 2025, shows 2,060 GW of generation and storage capacity waiting to connect — alongside 408 GW that has already signed an interconnection agreement but still has not reached commercial operation.

For perspective: total installed U.S. generating capacity is roughly 1,280 GW. The waiting room is bigger than the entire grid.

Three numbers from that dataset matter for anyone thinking about AI infrastructure investment:

1. The median time from interconnection request to commercial operations has doubled — from under 2 years for projects built between 2000–2007, to over 4 years for projects built in 2018–2024, with a median of 5 years for projects coming online in 2023.

2. The completion rate is collapsing. Of all projects that entered queues between 2000 and 2018, only 14% of capacity ever reached operation. Roughly 80% of projects withdraw before they connect.

3. Late-stage withdrawals are rising. Even projects with executed interconnection agreements — projects that look operationally certain on paper — are now pulling out in greater numbers, triggering re-studies that further delay everyone behind them in the queue.

When the AP, McKinsey, or Bloomberg writes the next “AI is using too much power” headline, this is the actual constraint they are describing. Not the megawatt-hour. The procedural, political, and physical impossibility of getting new megawatt-hours connected to where the chips already are.

The PJM signal

The clearest market price for “power, now, in the place AI wants it” is the PJM capacity auction. PJM is the regional transmission organization that covers the Mid-Atlantic, including Northern Virginia — the densest data-center market on the planet.

The PJM capacity auction price moved from $28.92 per MW-day for the 2024–2025 delivery year to $329.17 per MW-day for 2026–2027 — a more than 11× increase in two years. The 2026–2027 auction cleared at the FERC-imposed price cap — meaning the true clearing price would have been higher had regulators not intervened.

PJM is telling you, in dollars per megawatt, what every operator in Loudoun County already knows: there is no slack left. The market is rationing what cannot be expanded fast enough.

Texas tells the same story in a different language. ERCOT’s large-load interconnection queue grew from approximately 56 GW to 205 GW between September 2024 and October 2025 — a 73% increase driven primarily by data-center requests. AEP, the utility serving much of that footprint, doubled its contracted large-load pipeline to 56 GW over a similar period.

These are not the demand curves of a market that has solved its supply problem. These are the demand curves of a market that has not yet noticed it is broken.

The McKinsey math

McKinsey’s April 2025 Cost of Compute report puts a number on the gap. In the firm’s base-case scenario, data centers will require $6.7 trillion in cumulative capital expenditure by 2030, of which $5.2 trillion is AI-specific. That base case assumes 125 incremental gigawatts of AI capacity added globally between 2025 and 2030 — bringing total AI data center demand to 156 GW.

In their upside scenario, the figure climbs to $7.9 trillion and 205 incremental GW. In their downside, $3.7 trillion and 78 GW. Even the downside assumes a doubling of current global data center capacity in five years.

Now hold that 125 GW number next to the U.S. interconnection queue’s actual delivery rate. Berkeley Lab’s data implies that the U.S. can realistically deliver 60–75 GW of grid-connected new capacity on current timelines. The rest of the McKinsey forecast — the part that justifies hyperscaler capex, the part that justifies neocloud valuations, the part that justifies the $500 billion Stargate program — assumes one of three things must happen:

• Grid interconnection reform unlocks 20–30 GW of stranded capacity. (Possible, but slow, and dependent on FERC and state regulators.)

• Hyperscalers self-generate behind the meter. (Happening — BCG estimates 30–50% of new capacity will be self-generated by 2030.)

• Operators deploy compute at sites that already have power, on a timeline measured in months instead of years.

That third path is where the modular thesis lives.

Modular as a grid bypass

A traditional hyperscale data center takes 36 to 60 months from greenfield to first commercial revenue. Add the average 5-year wait for grid interconnection (where one is needed beyond what the site can already deliver) and you are looking at the better part of a decade between capital commitment and first dollar earned.

The hardware that capital was committed against will have depreciated through two full GPU generations by then. The H100 that made sense to procure in 2024 is competing with B200 in 2025 and B300 in 2026; the operator who locked in H100 capacity on a 60-month build schedule is selling vintage compute at a discount to operators who deployed Blackwell on a 195-day timeline.

This is the structural insight that animates everything DCXPS does.

A modular AI data center — two reinforced 40-foot ISO containers, 450 kW of IT load, mixed air- and direct-to-chip liquid cooling, populated with 14 NVIDIA B300 servers and 35 H200 servers — can be deployed at a site with existing power capacity in 195 days from contract to first revenue. Phase 1 is 10–12 weeks of strategic planning, site selection, container reservation, and regulatory work. Phase 2 is 15–18 weeks of system integration: mobile DC installation, hardware setup, energy-system integration. Phase 3 is 6–8 weeks of commissioning, GPU server initiation, configuration, and platform connection.

Then 15+ years of operating cash flow.

The math behind this is not exotic. It is what every modular industrial deployment has done since pre-cast concrete: pre-fabricate the complex parts off-site, ship them to where the resource is, assemble in place, commission, operate. The novelty is not the modular concept — it is that, for the first time in the history of computing, the constraint that justifies modular deployment is real, structural, and measurable in dollars per MW-day.

The economics of co-location

Here is the framework that should drive site selection in the modular era, in order of priority:

Tier 1 — Sites with existing, owned grid connection at adequate capacity. Retiring fossil generation sites, decommissioning industrial facilities, large utility properties. Power is on the meter, permits are in place, transmission is sized. Deployment timeline: 90–195 days.

Tier 2 — Sites with co-located generation. Combined heat and power (CHP), biogas, hydro, nuclear, or large renewable installations with surplus capacity. The generator becomes the anchor; the data center becomes the off-taker. Deployment timeline: 6–9 months, contingent on commercial framework.

Tier 3 — Sites with secured but not-yet-built generation. Greenfield renewable projects with executed interconnection agreements and PPAs. Higher risk, longer timeline, but available on negotiated terms. Deployment timeline: 12–18 months.

Tier 4 — Pure greenfield. Don’t bother. This is the hyperscaler trap. Capital is committed against a timeline you do not control and a counterparty (the utility) that has no commercial incentive to move quickly.

Every category-leading neocloud in 2026 — CoreWeave, Crusoe, Nscale, IREN — is now executing some version of Tier 1 or Tier 2. Crusoe built its entire thesis on stranded gas at the wellhead. IREN converted bitcoin-mining sites that already had power on the meter. Nscale anchored at Norwegian hydro. The pattern is not coincidental. It is the only pattern that works in the post-interconnection-queue era.

Why we chose Kladno

DCXPS’s first site is Kladno, in the Czech Republic, 30 kilometers from Prague. The site is co-located with a 200 MW power plant. We have direct on-site interconnect — meaning the kilowatt-hour does not transit transmission infrastructure we do not control. We are within a fully GDPR-compliant jurisdiction, EU-incorporated, and positioned to serve the demand profile that the EU AI Act will define starting in August 2026 (more on that in the third article in this series).

The site supports 9 deployable modular unit positions. Each unit is structured as a ring-fenced Delaware LLC SPV. The fleet, at full deployment, represents $405 million of contracted infrastructure capacity — and roughly $1.37 billion of gross compute revenue over the 6-year operating term, based on Chapek’s current list pricing for B300 and H200 capacity.

The first units come online in October 2026. The bottleneck, for us, is not power. It is not permits. It is not GPUs (allocations are secured). The bottleneck is capital deployment velocity — and that is the negotiation we are having with partners right now.

What to watch

A few signals that will tell you which way this market actually moves over the next 18 months:

1. PJM’s next capacity auction. If the 2027–2028 auction continues at or near the cap, the rationing thesis is confirmed and behind-the-meter / modular economics get materially more attractive.

2. FERC Order 2023 implementation. The cluster-study reforms are well-intentioned but slow. If they begin to materially shorten queue times in 2026, hyperscale greenfield re-enters the calculus. If they don’t, modular continues to take share.

3. The DATA Act. The proposed Senate exemption for off-grid data centers from FERC oversight would dramatically accelerate behind-the-meter economics. Worth tracking quarterly.

4. Hyperscaler self-generation announcements. Every gigawatt of self-generation Microsoft or Meta announces is implicit confirmation that the grid path is closed.

5. Neocloud margin compression. If sustained capacity rationing continues, the operators who locked in power before the squeeze will see margins expand, not compress. Watch CoreWeave’s gross margin trajectory.

Where this leads

The five-part series this article opens will work through, in sequence:

• Article 2 unpacks the unit economics of a $45M modular SPV in forensic detail — the math, the waterfall, the sensitivities, the comparables.

• Article 3 examines why the EU AI Act and the CLOUD Act create a structural opening for EU-sovereign AI infrastructure that US-incorporated operators cannot fully address.

• Article 4 maps the training-to-inference transition that will reshape the compute geography by mid-2027 — and why modular’s 6-year horizon is built for exactly this transition.

• Article 5 frames AI infrastructure as a distinct asset class, with a hardware-backed exposure profile that does not exist in tokenized compute claims or platform equity.

If you are a family office, fund, or strategic capital partner thinking about how to participate in the next stage of AI infrastructure deployment, the questions to bring to the conversation are not “can we get GPUs” or “is the market real.” Those are answered. The questions are: whose power, on what timeline, at what cost of capital, with what exit profile.

We’re happy to walk through ours. The Kladno site has 9 unit positions. The minimum participation is $15M. The structure is a ring-fenced Delaware LLC SPV per unit, with hardware owned by the SPV and DCXPS operating under a 15%/85% EBITDA split, paid second.

For the data room and a structured introduction to the management team, write to investors@dcxps.com.

The factory is being built. The question is whether you own a piece of it.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek bare-metal GPU cloud platform. Next in this series: “The 195-Day Data Center — A Forensic Walk-Through of $45M of Unit Economics.”