I want to describe something that doesn’t fully exist yet but will exist because the physics and economics are already pointing directly at it. Not a data center. Something that has outgrown the category

The Five Generations in Summary: What We’re Progressing Toward

Generation 1 (2026–2027): The MADC. 1 MW modular units at brownfield sites with existing grid. Human technicians on-site. Revenue from month five. The embryo.

Generation 2 (2027–2028): The MADC+. On-site CHP power generation. First-generation robotics. 85%+ grid independence. The cluster begins to breathe on its own.

Generation 3 (2028–2030): The DDCU Gen 1. Zero grid dependency. Zero permanent human presence. Full robotic maintenance. The term “Dark” becomes real: dark to the grid, dark to the operator, dark in the AI sense — self-optimizing, emergent. The agricultural loop begins: waste heat from immersion cooling powers adjacent greenhouse structures.

Generation 4 (2030–2032): The 10 MW DDCU. Scale leap from 1 MW to 10 MW per unit. Humanoid robots performing both compute maintenance and agricultural labor. Integrated farms producing food for 2,000–5,000 people per cluster annually. EBITDA margins expanding toward 93%.

Generation 5 (2032–2034+): The Robotic DC City. Multiple DDCU clusters converging into 100–200 MW campuses managed by fleets of 200–500 humanoid robots. Human population: 5–10 orchestrators. Dark Factories manufacturing next-generation DDCUs on-site. AI designing the hardware that the next AI will run on.

The data center is a cocoon. What emerges from it is something else entirely.

Why Big Data Centers Will Fail — The Structural Argument

Read more

Read more

The DDCU architecture doesn’t have a ceiling problem. It has a flywheel.

Here’s the difference: infrastructure companies wait for customers. The flywheel builds them.

The Four Turns That Compound Into an Economy

Every dollar of infrastructure profit, 20% flows into DCXPS Ventures. This is not a corporate philanthropy program. It is the first turn of a compounding flywheel that transforms infrastructure profits into captive customers into more infrastructure profits.

Turn 1: Profit to Capital

The arithmetic is straightforward. Year 2 (2027): $11 million to Ventures from $56 million infrastructure profit. Year 5 (2030): $540 million from $2.7 billion profit. Year 9 (2034): $7.94 billion from $39.7 billion profit. Year 15 (2040): $27 billion in a single year from $135 billion profit.

This is not optimistic projection. It is the compound consequence of a 75–89% margin infrastructure business at scale. The compute business generates money faster than it can deploy it internally. The Ventures channel exists to productively compound the surplus.

Turn 2: Capital to Companies

Read more

I want to be precise about why. Not dismissive — the engineering behind both concepts is genuinely interesting. But interesting engineering and viable infrastructure architecture are different things, and the AI era cannot afford to confuse them.

Project Natick: What Microsoft Actually Learned

In 2018, Microsoft deployed a 12-rack, 864-server data center pod in 117 feet of water off the Orkney Islands in Scotland. Project Natick Phase 2. They retrieved it in 2020. The results were published, celebrated, and then quietly filed away.

Read more

The grid didn’t care about North Dakota. North Dakota didn’t care about the grid.

This was not a tactical win. This was the announcement of a new architectural religion.

The Inversion That Changes Everything

For sixty years, data centers followed a simple logic: build near the grid. Build near cities. Build near the workers who maintain the hardware and the customers who consume the compute. Power comes from centralized generation, travels over transmission lines, arrives at substations, serves buildings.

This logic works beautifully when electricity is abundant and cheap and universally distributed. It fails catastrophically when the grid is overloaded, permits take years, and transmission costs $41.50 per MWh per 1,000 miles of distance.

The data point that reframes everything: moving electricity costs $41.50/MWh per 1,000 miles. Moving data costs essentially nothing.

Compute should go where power is abundant. The workloads follow over fiber. This is not a clever optimization. It is a structural inversion of sixty years of infrastructure assumption.

Every stranded wind farm becomes a potential compute node. Every underutilized biogas plant, every industrial site with spare generating capacity, every agricultural region with methane from waste that would otherwise decompose into the atmosphere — all of them become viable locations for frontier AI compute. Not because anyone decided to put them there. Because the physics of moving power versus moving data makes them the correct answer.

What “Stranded” Actually Means — And How Much of It Exists

Stranded energy is generation capacity that exists but cannot reach buyers. The transmission network — built over decades for a different demand profile, in a different economic environment, for different customers — has not kept pace with renewable buildout.

In the United States, curtailment of wind and solar power reached a record in 2024 — meaning generation that was available but literally switched off because the grid couldn’t absorb it. In Texas, curtailment events occur hundreds of times annually. In California, solar panels are routinely disconnected from the grid during peak generation periods because the grid cannot accommodate the supply.

The Lawrence Berkeley National Laboratory estimates that 40–60% of proposed renewable energy projects in the U.S. are delayed or cancelled due to interconnection challenges. This is power that wants to exist. Power that has been built, that has investors, that generates electricity when the sun shines or the wind blows, and that has no viable path to customers.

This is not a temporary condition. It is structural. The transmission infrastructure required to move renewable power from where it’s generated (Great Plains wind, Southwest solar, offshore mid-Atlantic wind) to where it’s consumed (coastal population centers, dense industrial corridors) would cost trillions of dollars and require decades of permitting.

Or you could bring the compute to the power. In 120 days. On a flatbed truck.

The Modular Unit: What It Actually Is

Read more

This is not a success story. This is a structural trap being mistaken for a strategy.

The AI Arms Race Built on Quicksand

The hyperscale buildout began in earnest in 2022 when ChatGPT demonstrated that AI services had consumer demand.

What followed was the largest coordinated capital deployment in the history of technology infrastructure — faster than the telegraph network, faster than electrification, faster than the fiber optic buildout of the 1990s.

The scale is genuinely staggering. As of mid-2025, more than 1,000 data centers were under construction or permitted in the U.S. alone, with combined capacity of 75 GW — comparable to the peak daily demand of New York City. The global AI infrastructure market requires an estimated $6.7 trillion in investment through 2030, and roughly $24.6 trillion by 2034.

Every major technology company, every sovereign wealth fund, every infrastructure REIT, and a remarkable number of people who had never thought about cooling towers before 2022 poured capital into the same bet: that large, centralized, grid-connected data centers were the correct substrate for AI.

They were betting on a technology that had already started making their bet obsolete.

The Five Structural Failures of Hyperscale

Let me be specific, because this matters. Hyperscale AI data centers are not slightly suboptimal. They are structurally wrong for the workload they’re designed to serve. Here’s why.

1. The Grid Has Already Failed Them

Read more

The Mainframe Church, 1964–1985

The IBM System/360 launched in 1964. It cost $5 billion to develop — roughly $46 billion in today’s money — and it was, without question, the most audacious bet in the history of computing. Thomas Watson Jr. called it “the biggest risk that IBM has ever taken.”

He was right. He was also building a cathedral.

Mainframes required raised floors. Specialized cooling. Dedicated power substations. Rooms measured in thousands of square feet for machines that, by 2026 standards, have less compute than the chip inside your car’s keyless entry fob. IBM’s first commercial data centers consumed anywhere from 50 kW to 5 MW of power — for systems that, at peak, could process roughly 10 MIPS (million instructions per second).

Your phone processes 15,000 MIPS. Your phone is not climate-controlled in a purpose-built concrete fortress.

The mainframe era established a paradigm that would haunt us for sixty years: compute must be centralized, expensive, controlled, and surrounded by bureaucracy proportional to its power.

This wasn’t irrational. In 1964, compute was genuinely scarce. Centralization was efficiency. Sharing a mainframe across an entire university or corporation made mathematical sense when each transistor represented a miracle of precision manufacturing.

But the paradigm outlived its rationale by decades.

The PC Revolution Didn’t Kill Centralization — It Deepened the Contradiction

When the IBM PC arrived in 1981 and the Mac in 1984, the conventional wisdom was that we’d democratized compute. Decentralization. Power to the people. Every desk its own universe.

Wrong.

What actually happened: the PC created demand for networked services that no desktop could serve alone. Email servers. Database servers. File servers. Application servers. By the mid-1990s, every company that had eliminated its mainframe was building a server room. By 2000, those server rooms had become data centers. Smaller than IBM’s cathedrals, yes. But still centralized. Still dependent on utility power. Still requiring raised floors and precision cooling and specialized staff.

The distributed revolution created a distributed hardware problem, and the industry’s answer was to centralize it again. Just in smaller boxes. Called differently.

In 1998, the average corporate data center consumed 1–2 MW of power for what we would now recognize as trivially modest workloads. Heat was managed with perimeter air handlers and hopes. Power was managed with UPS systems and prayers.

Efficiency, as measured by Power Usage Effectiveness (PUE), hovered around 2.0 — meaning for every watt of useful compute, another watt was wasted on cooling, lighting, and the existential dread of the facilities manager.

The Internet Boom Built the Wrong Infrastructure at Maximum Speed

1996 to 2001. The industry collectively lost its mind. And built a lot of data centers.

Between 1998 and 2001, U.S. investment in data center infrastructure exceeded $45 billion — the equivalent of roughly $80 billion today. Buildings were raised in suburban office parks. Carrier hotels emerged in city centers. Colocation exploded. Every company needed “Internet-scale infrastructure,” a phrase that meant approximately nothing but justified approximately everything.

The assumption baked into every build: these facilities would run at capacity within 18 months. The assumption was wrong approximately 90% of the time. When the dot-com collapse arrived in 2001, something like 40% of all newly built data center space in the U.S. sat dark — unused, half-leased, consuming power for cooling systems that had nothing to cool.

The dot-com boom proved something important: infrastructure built for anticipated demand, at industrial scale, on speculative timelines, is infrastructure that fails expensively and publicly.

The physical plants were fine. The fiber was in the ground. The buildings stood. What was wrong was the model: build massive centralized infrastructure, pray demand fills it, lock capital in concrete for 20-year depreciation cycles.

Nobody learned the lesson. Because the next wave was already coming.

Virtualization and the Cloud: Efficiency Theater

VMware shipped its first x86 virtualization product in 1999. Amazon Web Services launched in 2006. These were genuine revolutions in how compute was consumed — and genuine disasters for how compute was built.

Virtualization allowed one physical server to run dozens of virtual machines. This was extraordinary. It meant data centers could do far more with existing hardware. Utilization rates climbed from the abysmal 5–15% typical of dedicated servers toward 60–70% for well-managed virtualized environments.

AWS meant a startup could launch without owning a single server. The capex-to-opex transformation was real, substantial, and genuinely valuable for customers.

For the underlying infrastructure? Different story.

Amazon’s first data center in Ashburn, Virginia — the so-called “Data Center Alley” that would become the most concentrated compute geography on Earth — consumed roughly 30 MW.

By 2010, Amazon, Google, Microsoft, and Facebook were collectively consuming more than 1 GW globally.

By 2015, that number crossed 10 GW.

Subscribe now

By 2024, hyperscaler data centers alone consumed an estimated 45–50 GW worldwide.

The architecture remained identical to 1964’s cathedral. Bigger.

More efficient (PUE had improved to 1.3–1.5 by 2015).

Fundamentally the same structure. Centralized. Grid-dependent.

Permitting-constrained.

Built to last 20 years in landscapes that were changing every 18 months.

The Structural Problems Nobody Was Counting

Here’s what the cloud era obscured: the physical infrastructure was becoming increasingly divorced from the economic and environmental reality around it.

Grid interconnection queues in the United States had grown from manageable to catastrophic. By 2024, the Lawrence Berkeley National Laboratory reported 2,300 GW of projects waiting in U.S. interconnection queues — with only 14% historically reaching commercial operation. Permitting timelines for utility-scale power connections averaged 60+ months.

Data center construction itself required 48–72 months from groundbreaking to operations.

That math is brutal: by the time a traditional hyperscale data center completes construction, the compute it was designed to host has gone through two full GPU generations.

NVIDIA’s H100 — state of the art in 2023 — was already being superseded by the Blackwell architecture before many 2023-vintage facilities came online.

Meanwhile, PJM capacity clearing prices — the wholesale price of grid power capacity in the mid-Atlantic region — surged from $28.92 per MW-day in the 2024–25 planning year to $329.17 per MW-day in 2026–27.

A 10x increase in one year.

Texas enacted SB-6, requiring legislative approval for data centers exceeding 100 MW. Europe implemented permitting cycles of 3–7 years for industrial power connections.

We built an industry on the assumption that grid power was infinite, cheap, and immediate. By 2026, every one of those assumptions had been catastrophically disproven.

The Pre-AI Legacy: What We Inherited

To understand where we need to go, you have to sit with what we inherited.

The pre-AI data center industry — everything from IBM mainframes through the hyperscaler cloud — built infrastructure on four axioms that were true in 1964 and increasingly false by 2024:

• One: Compute should be centralized near population centers and corporate headquarters.

• Two: Power comes from the grid, and the grid will scale to meet demand.

• Three: Buildings should be permanent, purpose-built, and depreciated over 20+ years.

• Four: Human staff are required for operations, maintenance, and security.

Each of these axioms made sense in its original context.

Each of them became a strategic liability as AI workloads arrived with their exponential, geography-agnostic, perpetually-accelerating compute demands.

Share

The AI era didn’t just need more data centers. It needed a different architecture entirely.

And the industry, locked into 60 years of infrastructure assumptions, was spectacularly unprepared.

The cathedrals were magnificent.

They just couldn’t survive the reformation.

Subscribe now

As we learned in Europe almost 600 years ago.

JF.

Whether AutoGPT-style workflows eliminate ServiceNow.

Wrong companies.

Wrong threat vector.

The real carnage from AI agents isn’t happening at the top of the SaaS market. It’s happening in the $50-100B mid-market SaaS layer — the category you’ve never thought about because it’s not sexy enough to make conference keynotes.

Vertical SaaS.

Workflow automation.

Industry-specific platforms that charge $15,000/year per seat for software that does one thing: move structured data between systems and generate reports.

Those products are dead.

They just don’t know it yet.

What Agents Actually Replace

Let me be specific about what an AI agent does that software workflows can’t. It doesn’t just move data. It interprets, adapts, handles exceptions, and takes action on ambiguous instructions without a human in the loop for every edge case.

The $45/month Zapier workflow breaks when the input format changes.

The agent handles it.

The $40,000/year legal contract management platform tracks deadlines, flags renewals, and generates compliance reports.

An agent connected to your document store, your calendar, and your email does all of that — and adapts when contract formats change, when legal requirements shift, when your team’s workflow evolves.

The platform charges $40K.

The agent costs $200/month in API fees if you’re running it efficiently.

That’s not a feature gap.

That’s a business model collapse.

The Numbers That Should Scare Mid-Market SaaS

GitHub Copilot has 1.8M paid subscribers. Developer hiring at AI-native companies is up 12% year over year. This is important: the productivity gains from AI aren’t eliminating the knowledge workers. They’re eliminating the software the knowledge workers used to need.

The McKinsey Global Institute’s 2025 analysis puts the addressable automation opportunity for AI agents in “task orchestration and workflow management” at approximately $2.4 trillion of current software spend globally. That’s not the total TAM for software. That’s the portion of current software revenue that agents can credibly replace within 5 years.

$2.4 trillion.

The companies in that crosshairs aren’t Salesforce (which is already building agent infrastructure) or ServiceNow (same). They’re the 3,000+ vertical SaaS companies that raised $50-500M during the 2021 SaaS bubble, built a category-leading product in restaurant management, legal ops, supply chain visibility, compliance automation, or healthcare workflow — and are now facing a world where their entire value proposition can be replicated by a well-prompted agent with MCP access to their customer’s data.

Subscribe now

The Salesforce Pivot Nobody’s Crediting Properly

Here’s the contrarian take inside the contrarian take: Salesforce might actually be fine. Not because their CRM is irreplaceable — it isn’t — but because they’re aggressively building the agent runtime that mid-market SaaS companies will pay for.

Agentforce, announced in late 2024, is Salesforce’s bet that enterprise companies want their AI agents running inside a trusted, audited, enterprise-grade orchestration layer.

That’s a different business than CRM.

Smaller addressable market in some ways, but higher defensibility.

Enterprise AI governance is a real need.

Companies deploying agents on customer data at scale need audit trails, compliance controls, and someone to call when something goes wrong.

Leave a comment

Salesforce can be that layer.

Your $50M ARR vertical SaaS startup in legal ops cannot.

Share

The Mid-Market Graveyard Is Coming

Here’s what I expect to happen in the next 24 months. The first wave of agent-powered “SaaS killers” will be built by AI-native startups targeting exactly these mid-market vertical categories.

They’ll charge 80% less than the incumbent, deliver 90% of the functionality, and add native agent capabilities the incumbent can’t retrofit.

The incumbents will respond by building “AI features.” Copilot buttons.

Summarization.

Smart search.

None of that addresses the actual threat, which is that the entire workflow the software was designed to automate can now be replicated by an agent that doesn’t require their specific platform to function.

The acqui-hire wave will follow.

The PE rollup wave will follow that, buying distressed mid-market SaaS assets for 2-3x revenue when they were trading at 12-15x two years ago.

I’ve seen this before.

Different technology, same pattern.

Subscribe now

The companies celebrating the AI agent wave should check whether their product is the thing agents are replacing.

The answers should terrify anyone building AI infrastructure right now.

AI data centers are on track to consume 8% of US electricity by 2030, up from roughly 2% in 2022. That’s not a forecast from some think tank hedging its bets. That’s from the Department of Energy’s own modeling, backed by utility capacity planning documents that are already being executed.

Utilities are signing 20-year power agreements.

AI companies are not.

That gap — between the infrastructure commitments that power AI and the contract structures that pay for it — is where the next crisis in this industry is being built in slow motion.

The Physics of the Problem

Training a single large language model frontier run costs between $50-100 million in compute. That’s the headline. Nobody talks about the fact that 40-60% of that cost is electricity. At scale, AI infrastructure is an energy business that happens to produce tokens.

NVIDIA’s H200 GPUs — the current gold standard — draw about 700 watts each under full load. A standard 8-GPU server: 5.6 kilowatts. A rack of eight servers: ~45 kilowatts. A modest 100-rack cluster: 4.5 megawatts.

The hyperscaler data centers being announced now?

200, 500, 1,000 megawatts.

Microsoft’s announced AI data center commitments for 2025-2026 alone: $80B capex.

Where does that electricity come from?

In the near term: natural gas peakers. Coal in some regions. Nuclear where available. Renewables where the grid can handle it. The clean energy story most AI companies tell their investors doesn’t match the power purchase agreements they’re actually signing.

The Utility Lock-In Nobody’s Talking About

Here’s the structural trap.

Electric utilities are regulated monopolies.

When a data center operator comes to them needing 500 megawatts of new capacity, the utility has to build generation, transmission, and distribution infrastructure to deliver it. That infrastructure has a 20-40 year lifespan.

Utilities are not building that infrastructure speculatively. They are requiring power purchase agreements — long-term contracts, typically 10-20 years, with take-or-pay provisions. You commit to paying for the power whether you use it or not.

Subscribe now

The data center operators building for the AI boom are signing these agreements. Some of them are AI companies themselves. Many of them are independent operators who signed LOIs and supply agreements with AI company customers on much shorter terms — 1-3 year agreements, with renewal options.

Share

Do you see the mismatch?

The utility contract: 20 years.

The AI customer contract: 1-3 years.

If the AI customer doesn’t renew — because the model changed, the competitor got cheaper, the startup ran out of money — the data center operator eats the difference. Locked into paying for power that nobody needs.

The European Version of This Problem

In the EU, the problem has an additional layer. Energy prices are structurally higher — averaging 2-3x US industrial rates in Western Europe, and volatile in ways the US market isn’t. The combination of high base energy cost, carbon pricing, and intermittent renewables creates margin compression that makes the US AI infrastructure economics look generous.

The EU data sovereignty requirements are creating demand for EU-jurisdiction AI infrastructure. That’s genuinely good for operators who can deliver it. But the energy economics mean that EU-based GPU compute is approximately 40-60% more expensive per token than equivalent US infrastructure — a structural disadvantage that doesn’t go away with better software.

The operators who figure out how to amortize that energy premium through platform services — fine-tuning, inference serving, developer tooling — are the ones who can make the economics work.

Bare-metal renting alone doesn’t.

Who Holds the Bag

Three groups are exposed here in different ways.

First: hyperscalers who signed the 20-year power agreements. If AI capex rationalizes faster than expected — and it will, because it always does — they’re stuck with stranded power capacity. Microsoft, Google, Amazon all have the balance sheets to absorb it. It hurts. It doesn’t kill them.

Second: independent data center operators who levered up to build AI-specific infrastructure on the assumption of multi-year customer contracts. These are the ones keeping me up at night. The private equity-backed builds, the regional operators, the “AI data center REIT” structures being pitched to institutional investors right now. When the AI customer churn hits, the power contracts don’t churn with them.

Third: the AI companies themselves who signed multi-year compute commitments and then face a world where open-source models halved the compute requirement. They’ve prepaid for capacity they no longer need. That’s a balance sheet problem.

The Opportunity in the Rubble

None of this means don’t build.

It means build differently.

The operators who survive the inevitable rationalization will be the ones who didn’t build for one customer, didn’t sign 20-year power agreements on the assumption of 3-year customer relationships, and built platform layers that create genuine switching costs.

Someone is very wrong. I want to talk about who.

I’ve been through enough boom-bust cycles — 110+ startups, three continents, two crypto collapses — to know what a valuation story looks like when the story is doing more work than the fundamentals. The OpenAI $300B story requires you to believe four things simultaneously: that they maintain model superiority, that enterprise customers don’t switch when cheaper alternatives arrive, that the Microsoft divorce doesn’t crater their distribution, and that the regulatory environment stays benign.

Miss even one of those.

One.

The whole thing reprices.

The Microsoft Divorce Is More Serious Than Anyone’s Saying

The $250 billion “betrayal” — that’s what the coverage is calling it. Microsoft built Azure AI services on OpenAI’s models. OpenAI is now competing directly with Microsoft’s enterprise customers. The exclusivity arrangement is unwinding. Microsoft is investing in Mistral, in its own internal model development, in open-source alternatives.

This is not a lover’s spat.

Subscribe now

Share

This is a distribution channel that generates the majority of OpenAI’s enterprise revenue actively building alternatives to the product it’s supposed to be selling.

When I ran distribution-dependent businesses, the moment your primary channel starts building a competing product is the moment your revenue forecast becomes fiction. Not bad fiction. Optimistic fiction, which is worse because people believe it longer.

The Microsoft-OpenAI relationship generated roughly $3-4B of OpenAI’s 2025 revenue through Azure commitments and reselling arrangements. That revenue doesn’t evaporate overnight.

But it reprices.

The moment Microsoft can offer enterprise customers a comparable model at Azure pricing without the OpenAI premium, enterprise procurement teams — who are already under budget pressure — will make the math.

Leave a comment

The Commoditization Clock Is Running

Here’s what’s happening in the actual market. DeepSeek R1, released January 2025, matched GPT-4 level performance at a fraction of the training cost. Llama 4 is closing the gap. Mistral is winning European enterprise deals on price and data residency. Google’s Gemini 2.0 is embedded in every Android device on earth.

The model itself is becoming infrastructure, not product.

When oil becomes a commodity, Exxon still makes money — but not 21x revenue multiples. When bandwidth became a commodity, nobody was paying 20x ARR for fiber. The commodity transition doesn’t kill the market. It kills the premium.

OpenAI’s bet is that they escape the commodity trap through the application layer — ChatGPT consumer, operator API, enterprise contracts with stickiness built in. That bet might be right. It requires them to successfully become a consumer software company, an enterprise SaaS company, and a frontier research lab simultaneously.

That’s three different businesses. Three different cultures. Three different competitive moats.

Simultaneously.

I’ve never seen a company pull that off cleanly. Google almost has. Microsoft almost has. Both of them took twenty years and some very expensive failures along the way.

What the $300B Actually Requires

Let me build this simply. For a $300B valuation to make sense at any reasonable multiple by 2030, OpenAI needs to be doing somewhere between $30-50B in annual revenue with healthy margins.

Today: $14B ARR, reportedly burning $5B+ per year in compute and operations.

The path to $30-50B ARR in four years requires either:

(a) holding or expanding per-token pricing as commoditization accelerates — unlikely;

(b) massive volume growth that outpaces price compression — possible but capital-intensive; or

(c) a genuine application-layer breakthrough that captures revenue from the productivity gains AI creates — unproven at this scale.

Option (c) is the interesting one.

Also the most uncertain.

And “uncertain” is doing a lot of work in a $300B valuation.

The Uncomfortable Conclusion

I’m not saying OpenAI fails. I’m saying that at $300B, you’re buying the best-case scenario at full price with no discount for execution risk, competitive risk, regulatory risk, or the very real possibility that the Microsoft divorce gets messier before it gets cleaner.

The railroads changed America. Railroad investors lost everything. Doesn’t mean you don’t build railroads.

Means you don’t pay 21x ARR for a railroad in 1870 when the land grants haven’t been decided and three other railroads are laying track in the same direction.

Here’s what I’m betting DCXPS on:

By 2028, more than 40% of new enterprise AI compute capacity will deploy on modular, energy-sovereign infrastructure.

• Not hyperscale data centers waiting for grid connections.

• Modular containers with integrated renewable power that deploy in weeks rather than years.

• This prediction is specific enough to be wrong.

• That’s intentional.

The Structural Forces

Five irreversible trends support this prediction. Each trend has observable markers and falsification conditions.

Force One: Grid Infrastructure Exhaustion

PJM’s interconnection queue contains 2,300 GW of requests.

ERCOT’s large load requests exploded from 56 GW to 205 GW in thirteen months.

Grid infrastructure investment required to serve projected AI demand exceeds $30 billion in Texas alone.

• The queue isn’t clearing.

• Grid investment isn’t accelerating fast enough.

• Traditional infrastructure cannot deploy at the rate AI demand is growing.

Falsification condition: If grid interconnection queues drop below 1,000 GW aggregate by 2027, the supply-side constraint loosens faster than projected.

Force Two: Power Density Transformation

AI racks are moving from 7-10 kW to 50-150 kW today, with projections reaching 1 MW per rack by 2028. Traditional data center infrastructure cannot support these densities without fundamental redesign.

The building stock that houses current data centers cannot be upgraded to support next-generation AI hardware. New purpose-built facilities are required.

Falsification condition: If average AI rack density stabilizes below 200 kW by 2027, existing infrastructure adaptation extends viability longer than projected.

Force Three: Deployment Velocity Requirements

AI capabilities are improving 10x annually.

Infrastructure that takes 5 years to deploy serves technology that’s been obsolete for 4 years by the time it’s operational.

The timing mismatch between AI development velocity and traditional infrastructure deployment creates structural demand for faster alternatives.

Falsification condition: If AI capability improvement slows to 2x annually by 2027, longer deployment timelines become acceptable.

Force Four: Community Opposition Scale

$64 billion in data center projects blocked or delayed by local opposition.

188 grassroots organizations actively fighting AI infrastructure across 28 states.

Bipartisan political alignment against hyperscale development.

The locations that can build traditional data centers are shrinking. The political economy increasingly favors infrastructure that communities want.

Falsification condition: If blocked/delayed projects drop below $20 billion annually by 2027, community opposition moderates faster than projected.

Force Five: Regulatory Sovereignty Momentum

EU AI Act requirements favor auditable infrastructure. GDPR enforcement favors local data processing. Singapore, Dubai, and emerging AI hubs are implementing their own frameworks.

Centralized offshore compute becomes harder as regulatory fragmentation increases.

Falsification condition: If regulatory harmonization creates unified global AI governance by 2027, sovereignty premium disappears.

The Prediction Mechanics

For 40% of new enterprise AI compute capacity to deploy on modular energy-sovereign infrastructure by 2028, several things must happen:

• Modular infrastructure must prove operational reliability at scale. We’re deploying 2 MW by June 2026, with 20 MW contracted for follow-on in Q1/2027. Plans and capacity secured stands right now at 2 GW for 2028.

  • The operational track record must validate the architecture.

• Energy integration must demonstrate cost competitiveness. Our unit economics show 42% cost advantage over CoreWeave, 70% advantage over hyperscalers.

  • These numbers must prove out in production.

• Enterprise procurement must accept modular deployment models. The sales cycle for our compute is 30 days versus 12-24 months for traditional infrastructure.

  • Procurement processes must adapt.

• Community opposition must continue making traditional development difficult.

  • The trends support this, but political dynamics can shift.

The Bet

I’m not predicting modular infrastructure wins because I want it to. I’m building modular infrastructure because the structural forces make traditional alternatives increasingly non-viable.

The hyperscalers will adapt. They’ll offer modular products, energy-sovereign options, faster deployment models. Their resources are immense. Their engineering capabilities are excellent.

But their architecture assumes centralized infrastructure at scale. Distributed modular deployment requires different operational models, different unit economics, different market approaches. Those adaptations take time.

The Personal Stakes

I’ve founded 110+ startups. I’ve seen predictions fail. I’ve bet on trends that reversed.

This prediction carries my reputation. If I’m wrong, it will be visible. The timestamps exist. The falsification conditions are specific.

That visibility is intentional. Predictions without accountability are marketing. Predictions with accountability are strategy.

By 2028, we’ll know whether the structural forces I’ve identified produce the outcomes I’m predicting.

If they do, DCXPS will have built significant market position during the adaptation window. If they don’t, the falsification conditions will reveal which assumptions failed.

• Either way, the prediction is useful.

• Correct predictions guide resource allocation.

• Incorrect predictions with clear falsification conditions reveal strategic blind spots.

What This Means For You

If you’re evaluating AI infrastructure options, consider the structural forces regardless of what you think about my prediction.

• Grid constraints are real.

• Community opposition is real.

• Regulatory fragmentation is real.

• Power density requirements are real.

• Deployment velocity mismatches are real.

Subscribe now

How you respond to those forces depends on your use case, your risk tolerance, and your strategic priorities.

The infrastructure decisions you make in 2026 will determine what AI capabilities you can deploy in 2028. Traditional paths are closing. Alternative paths are opening.

JF is a C-level executive and serial entrepreneur who has founded 110+ startups. He runs the AI Executive Transformation Program in Prague and writes about uncomfortable truths in AI implementation at AI Off the Coast (

AI of the Coast: The 5-Year Roadmap to General AI

Your strategic advantage in the race to General AI. Weekly intelligence on infrastructure, investments, and breakthroughs before your competitors see them. For leaders who won't be caught off guard by the AI revolution. Join 15,000 worldwide AI leaders.

By Jiri "Skzites" Fiala

).

The AI Paris Summit emphasized “environmentally sustainable AI” as policy priority. The EU AI Act requires transparency that centralized offshore compute makes harder to demonstrate.

Sovereignty isn’t a buzzword anymore. It’s a deployment constraint.

The Regulatory Fragmentation

Every major AI market is developing its own regulatory framework. The requirements don’t align. The compliance burdens compound.

EU AI Act categorizes AI systems by risk level and imposes requirements from transparency to human oversight. Systems categorized as high-risk face compliance obligations that require detailed documentation of training data, development processes, and deployment characteristics.

Demonstrating compliance is harder when compute runs in jurisdictions with different legal frameworks. When your model trains on infrastructure you don’t control in locations you can’t audit, the compliance evidence chains get complicated.

This isn’t theoretical future risk. It’s current procurement reality. Enterprises in regulated industries are asking where compute runs, who controls it, and what jurisdiction governs it.

The Middle East Emergence

While Western markets debate data center permitting, the Middle East is building at scale.

UAE is planning a 5 GW AI campus. Saudi Arabia committed $100 billion to its Transcendence AI Initiative. Gulf states that previously imported most technology are becoming AI infrastructure exporters.

The strategic implications matter.

Grid constraints in Western markets risk pushing AI infrastructure toward jurisdictions with different values and different governance. If you can’t build in Virginia because of community opposition and can’t build in Texas because of grid reliability, the alternative locations may not align with your organizational values or your customers’ expectations.

The sovereignty premium includes avoiding dependency on infrastructure in jurisdictions you wouldn’t otherwise choose.

Subscribe now

The Hyperscaler Jurisdiction Problem

When you deploy on AWS, your data processes in Amazon’s infrastructure. When you deploy on Azure, your data processes in Microsoft’s infrastructure. When you deploy on GCP, your data processes in Google’s infrastructure.

You can select regions. You can request contractual commitments. But the operational reality is that your AI compute runs on infrastructure controlled by organizations whose incentives may not perfectly align with yours.

For most use cases, this is fine. Hyperscaler security is excellent. Contractual protections are robust. The jurisdiction complexity is manageable.

For regulated industries, for government applications, for use cases where data exposure creates existential risk, the hyperscaler model creates dependencies that dedicated infrastructure eliminates.

Sovereignty means control. Control means ownership. Ownership means infrastructure that runs where you choose, under governance frameworks you accept.

Share

The Location Stack

Traditional infrastructure thinking treats location as a cost variable. Build where electricity is cheap. Deploy where land is available. Optimize for unit economics.

Sovereignty thinking treats location as a capability variable. Build where regulatory frameworks enable your use cases. Deploy where data governance meets your compliance requirements. Optimize for operational freedom.

The location stack includes: physical infrastructure, grid connectivity, regulatory environment, community relationships, and geopolitical stability. Each layer affects what you can do with your compute.

A modular data center in agricultural Poland offers different capability than the same hardware in suburban Virginia. Not because the compute is different, but because the location stack is different.

Leave a comment

The Renewable Sovereignty Advantage

Energy-sovereign infrastructure changes the location calculus.

When your compute requires grid connectivity, location choices are constrained to places with available grid capacity. Those locations are increasingly contested, regulated, and expensive.

When your compute integrates renewable generation, location choices expand to anywhere renewables are available. Agricultural regions with biogas potential. Rural areas with wind resources. Sites with solar capacity that grid infrastructure never planned to serve.

This isn’t just about energy cost. It’s about location optionality.

The biogas facility in Slovakia that would never get grid allocation for traditional data centers becomes viable infrastructure when the data center brings its own power. The rural community that would fight hyperscale development welcomes infrastructure that solves their energy production problem while solving your compute problem.

The Sovereignty Pricing

Dedicated sovereign infrastructure costs more per unit of compute than hyperscaler deployment. That’s the pricing reality.

The question is what you’re buying with the premium.

You’re buying regulatory clarity: your data processes under governance frameworks you select. You’re buying operational control: your infrastructure runs according to your policies. You’re buying location flexibility: your compute deploys where your requirements permit.

For workloads where those capabilities matter, the premium pays for itself in avoided compliance cost, reduced regulatory risk, and operational freedom that hyperscaler constraints would prevent.

The Strategic Implication

As AI regulation matures, sovereignty will increasingly differentiate infrastructure options.

The hyperscalers are responding with local region offerings and on-premises deployments. They see the demand shifting. But their architecture fundamentally assumes centralized infrastructure that customers rent rather than own.

Modular sovereign infrastructure offers an alternative that some use cases require. The market segment that needs it is growing as regulatory frameworks tighten and geopolitical uncertainty increases.

Location isn’t just where your compute runs. It’s what your compute can do.

JF is a C-level executive and serial entrepreneur who has founded 110+ startups. He runs the AI Executive Transformation Program in Prague and writes about uncomfortable truths in AI implementation at AI Off the Coast….

AI of the Coast: The 5-Year Roadmap to General AI

Your strategic advantage in the race to General AI. Weekly intelligence on infrastructure, investments, and breakthroughs before your competitors see them. For leaders who won't be caught off guard by the AI revolution. Join 15,000 worldwide AI leaders.

By Jiri "Skzites" Fiala

).

This isn’t a minor methodological concern. It’s systematic fraud dressed up as scientific measurement.

The Contamination Mechanics

Language models train on internet-scale data. That data includes the benchmark datasets published by researchers to evaluate language model performance.

The contamination is almost impossible to prevent. Benchmarks get published. They propagate across GitHub repositories, academic papers, and discussion forums. Models trained on web crawls ingest them.

When those same models get evaluated on those same benchmarks, they’re not demonstrating general capability. They’re demonstrating recall of training data.

SWE-Bench Pro addressed this by creating new evaluation tasks that didn’t exist in any training corpus. The results were illuminating: models that appeared capable on contaminated benchmarks performed dramatically worse on clean evaluation.

The capability gaps that benchmarks were supposed to reveal had been hidden by contamination that benchmarks were supposed to prevent.

The Investment Implications

Billions in capital flowed based on benchmark performance that may have been systematically overstated.

Investors who lack technical sophistication relied on leaderboard rankings as objective capability measures. Higher scores justified higher valuations. Competitive positioning depended on benchmark improvements that may have reflected contamination more than capability.

I’m not suggesting intentional fraud. Most contamination happens passively through training data collection processes no one fully controls. But the effect is identical: decision-makers received misleading information that shaped capital allocation.

How many AI investments would have occurred differently if benchmark results had reflected actual capability rather than contaminated measurement?

The Enterprise Trust Problem

Enterprise buyers evaluated AI vendors based partly on published benchmarks. Procurement processes included capability assessments derived from public evaluation results.

Those assessments may have been wrong.

The model that scored highest on contaminated benchmarks might not be the model that performs best on your actual use case. The capability gaps that mattered for your deployment were hidden by measurement artifacts that didn’t apply to your data.

This creates a delayed discovery problem. Enterprises select vendors based on benchmark performance. They discover capability limitations months later during implementation. By then, contracts are signed, integrations are built, and switching costs are substantial.

The benchmark contamination wasn’t just an academic measurement issue. It was a market information failure that shaped billions in purchasing decisions.

The Methodological Response

Clean benchmark development is harder than contaminated benchmark development.

Creating evaluation tasks that don’t exist in any training corpus requires novel task generation. Maintaining evaluation set security requires operational discipline that academic institutions often lack. Preventing future contamination requires ongoing task rotation that multiplies benchmark maintenance costs.

These are solvable problems. They’re also expensive problems that the industry has under-invested in because contaminated benchmarks served short-term marketing needs.

The models that won on contaminated benchmarks got the press coverage, the investment rounds, and the enterprise contracts. The models that would have won on clean benchmarks got less attention because the measurement system was broken.

The Capability Uncertainty

Here’s what benchmark contamination means for practical planning: we don’t actually know how capable current models are.

The published numbers overstate capability by unknown amounts. The true performance on novel tasks is systematically lower than benchmark performance suggests. The gap between measured and actual capability varies by model, by task type, and by contamination exposure in ways that resist simple correction.

Enterprise AI implementations should assume that model capability is lower than benchmarks indicate. Build margin for capability gaps that contaminated measurements didn’t reveal. Plan iteration cycles for the performance disappointments that clean evaluation would have predicted.

Subscribe now

The Accountability Void

No one has been held accountable for benchmark contamination.

The model developers benefit from inflated scores. The benchmark creators lack enforcement mechanisms. The investors who relied on misleading data have no recourse. The enterprises that made procurement decisions based on contaminated measurements absorb the consequences.

This accountability void means the incentives that created contamination remain in place. Future benchmarks will face the same pressures. Future measurements will be subject to the same artifacts.

Until the industry develops evaluation infrastructure that resists contamination and accountability mechanisms that punish misleading measurement, benchmark results will remain unreliable guides to actual capability.

The Practical Response

If you’re selecting AI vendors, benchmark performance should be one input among many, not the primary selection criterion.

Pilot deployments on your data, with your use cases, produce evaluation signal that benchmarks cannot. The model that scores highest on public leaderboards may not be the model that performs best in your environment.

Share

Reference customer conversations reveal capability in production, not capability in controlled evaluation. The vendors with the best benchmark scores may not be the vendors with the best deployment track records.

Technical evaluation of architecture, training methodology, and contamination exposure provides signal beyond headline benchmark numbers. The models built with rigorous contamination prevention may underperform on public benchmarks while overperforming in production.

The benchmark contamination problem isn’t getting fixed soon. Plan accordingly.

JF is a C-level executive and serial entrepreneur who has founded 110+ startups. He runs the AI Executive Transformation Program in Prague and writes about uncomfortable truths in AI implementation at AI Off the Coast…

AI of the Coast: The 5-Year Roadmap to General AI

Your strategic advantage in the race to General AI. Weekly intelligence on infrastructure, investments, and breakthroughs before your competitors see them. For leaders who won't be caught off guard by the AI revolution. Join 15,000 worldwide AI leaders.

By Jiri "Skzites" Fiala

).

Integration infrastructure consumed 60% of engineering effort. Algorithm development consumed 40%. The ratio surprised everyone except the engineers who’d seen this pattern before.

• We eliminated 750,000 phantom work orders.

• We achieved 5-month ROI on a €1.2M implementation.

• We delivered €279.5M in annual savings.

And the AI, the part everyone wanted to discuss, was the easier part of the project.

The Integration Reality

Enterprise AI doesn’t deploy into greenfields. It deploys into decades of accumulated technical decisions, data architecture choices, and integration debt that no one fully documents.

The facility management system we transformed had seventeen data sources. Building management systems from four vendors spanning three decades. Work order systems that had been upgraded, migrated, and patched so many times that no one knew which fields were authoritative. Sensor networks installed by contractors who’d left no documentation.

Before any algorithm could process this data, someone had to make it coherent. Field mapping. Data quality validation. Temporal alignment. Entity resolution across systems that used different identifiers for the same physical equipment.

This isn’t glamorous work. It doesn’t make conference presentations. It doesn’t generate academic citations. But without it, the most sophisticated AI produces garbage outputs from garbage inputs.

The Consultant Misdirection

AI vendors love algorithm discussions. They can demonstrate impressive capabilities in controlled environments with clean data and standardized interfaces.

The demo environment has nothing in common with your production environment.

When vendors scope AI implementations, they often underweight integration complexity. Their estimates assume data will be available in expected formats, that APIs will behave as documented, that edge cases will be edge cases rather than 30% of production volume.

Then implementation starts, and the integration work explodes. Timelines slip. Budgets expand. The algorithm that worked perfectly in demo struggles with data quality it wasn’t designed to handle.

The 95% AI project failure rate correlates directly with integration estimation failures. Projects don’t usually fail because the AI doesn’t work. They fail because the integration to make AI work costs more than anyone budgeted.

The Architecture Lesson

Our facility management success wasn’t algorithmic innovation. It was architectural discipline.

We built the integration layer first. Data quality pipelines that validated inputs before any model saw them. Entity resolution systems that created consistent identifiers across seventeen source systems. Temporal alignment mechanisms that made time-series data from different sources comparable.

Then we deployed relatively conventional AI. LLMs for knowledge management and work order interpretation. Gradient boosting for predictive maintenance. LSTM networks for time-series forecasting. Nothing exotic. Nothing that would impress an ML conference reviewer.

The conventional AI worked because the integration architecture worked. The same algorithms that fail in chaotic data environments succeed when the data architecture is sound.

The Hybrid Architecture Imperative

The facility management implementation taught us something else: edge computing matters more than cloud sophistication.

Thirty-seven percent of facilities had intermittent connectivity. Work order creation had to function when network access failed. Predictive alerts needed local processing to avoid latency that made them useless.

We deployed edge nodes that ran inference locally when connectivity dropped. The architecture was more complex than pure cloud deployment. It was also the only architecture that actually worked in production.

This pattern generalizes. Enterprise AI deploys into environments with constraints that cloud-native architectures ignore. Intermittent connectivity. Air-gapped security requirements. Latency constraints that centralized processing can’t meet.

Subscribe now

The integration work includes infrastructure integration, not just data integration. Understanding where compute has to run, how it connects to data sources, what happens when connections fail.

The Budget Reallocation

If you’re planning an AI implementation, consider this budget distribution:

Integration architecture and data quality: 35%

Infrastructure and deployment: 25%

Algorithm development and model training: 25%

Testing, validation, and iteration: 15%

Share

This distribution shocks organizations that assume AI projects are mostly about AI. But it reflects what successful implementations actually require.

The organizations that budget 70% for algorithms and 30% for everything else are the organizations that experience 300% cost overruns when integration complexity becomes visible.

The Strategic Implication

The integration iceberg creates a moat for organizations that master it.

Competitors can license the same algorithms. They can hire the same data scientists. They can deploy the same cloud infrastructure. What they can’t replicate is your accumulated knowledge of your data landscape, your integration architecture decisions, your hard-won understanding of where the data quality problems hide.

The 60% of AI success that comes from integration isn’t a cost to minimize. It’s competitive advantage to invest in. Organizations that view integration as unglamorous preliminary work miss the strategic value it creates.

Subscribe now

The AI is table stakes.

The integration is the moat.

JF is a C-level executive and serial entrepreneur who has founded 110+ startups. He runs the AI Executive Transformation Program in Prague and writes about uncomfortable truths in AI implementation at AI Off the Coast…

AI of the Coast: The 5-Year Roadmap to General AI

Your strategic advantage in the race to General AI. Weekly intelligence on infrastructure, investments, and breakthroughs before your competitors see them. For leaders who won't be caught off guard by the AI revolution. Join 15,000 worldwide AI leaders.

By Jiri "Skzites" Fiala

).

It was whether AI decisions went through a committee.

The Committee Disease

I’ve watched this pattern across forty AI implementations in the past two years. The organizations that succeed treat AI deployment like engineering decisions: owned by leadership, evaluated on merit, executed with autonomy.

The organizations that fail treat AI deployment like strategic initiatives: owned by steering committees, evaluated by consensus, executed after everyone agrees on everything.

Committees optimize for risk mitigation. They surface objections, document concerns, create space for dissent. These are valuable functions for decisions with irreversible consequences.

AI deployment isn’t irreversible. It’s iterative. You deploy. You learn. You adjust. You redeploy. The cost of a failed experiment is measured in weeks, not years.

But committees don’t distinguish between reversible and irreversible decisions.

They apply the same approval processes to a $500,000 pilot that they apply to a $50,000,000 acquisition.

The overhead becomes the bottleneck.

The Velocity Trap

AI implementation requires velocity. Not because speed is inherently valuable, but because AI capabilities are improving faster than organizational learning cycles.

The model you evaluate in Q1 is obsolete by Q3. The integration architecture you design in January faces new API capabilities by June. The training data requirements you document in spring don’t reflect the fine-tuning approaches available by fall.

Organizations that committee their way through AI decisions are always implementing last year’s technology with last year’s assumptions. By the time they deploy, the competitive landscape has shifted beneath them.

The 78% organizations deploy quickly, learn from production, and iterate toward capability. The 23% organizations study carefully, approve methodically, and deploy into markets their competitors already captured.

The Competence Displacement Problem

Committees create another pathology: they shift decision authority from people who understand AI to people who have organizational power.

The executive who controls budget allocation often lacks technical depth to evaluate AI proposals on merit. The legal team that must approve data usage doesn’t understand model architecture trade-offs. The procurement function that negotiates vendor contracts can’t distinguish meaningful capability differences from marketing claims.

These stakeholders have legitimate concerns. Budget discipline matters. Legal compliance matters. Procurement efficiency matters.

But when their concerns become approval gates rather than input to technical decisions, competence gets displaced by hierarchy. The people who understand the technology lose control to the people who control the process.

I’ve seen brilliant AI teams produce mediocre results because their proposals were filtered through stakeholders who couldn’t evaluate them. The work that survived committee review wasn’t the best work. It was the most politically palatable work.

The Alternative: Technical Ownership with Business Alignment

The 78% organizations don’t ignore business concerns. They address them differently.

Technical leaders own deployment decisions within defined boundaries. Those boundaries address budget, compliance, and strategic alignment. But within boundaries, technical judgment prevails.

Business stakeholders provide input rather than approval. They surface concerns that technical teams address. They don’t have veto power over technical decisions they can’t evaluate.

Executive sponsorship creates air cover for velocity. When deployment creates friction with traditional processes, executive authority resolves it. The technical team doesn’t have to committee their way through bureaucratic obstacles.

This isn’t anarchy. It’s appropriately scoped autonomy. The boundaries exist. The accountability exists. The competence to make decisions also exists.

Subscribe now

The 95% Failure Rate Context

Industry research suggests 95% of AI projects fail to deliver expected business value. That number is real, and it’s frequently cited to justify committee oversight. If most AI projects fail, shouldn’t we have more review, not less?

The causation runs the other direction.

Most AI projects fail because they’re designed by committees optimizing for internal stakeholder alignment rather than market reality. They’re scoped to satisfy procurement requirements rather than user needs. They’re architected to meet compliance checklists rather than performance objectives.

The committee process that’s supposed to reduce failure risk actually increases it by optimizing for the wrong objectives.

The 5% that succeed usually share a common characteristic: technical leadership with enough autonomy to optimize for outcomes rather than approval.

Share

Practical Implications

If you’re running AI initiatives through committees, you’re choosing the 23% path. That’s not necessarily wrong. Some organizations can’t tolerate the governance risk of technical autonomy. Some cultures can’t support the conflict that autonomous technical leadership creates.

But recognize what you’re choosing. You’re choosing slower deployment, older technology, filtered innovation, and optimized mediocrity.

If you want the 78% path, you need different organizational design. Technical leaders with budget authority. Executive sponsors who clear obstacles. Boundaries that enable rather than constrain.

Leave a comment

The organizations winning at AI aren’t the ones with the best committees.

They’re the ones that figured out how to deploy without committees at all.

JF is a C-level executive and serial entrepreneur who has founded 110+ startups. He runs the AI Executive Transformation Program in Prague and writes about uncomfortable truths in AI implementation at AI Off the Coast (

AI of the Coast: The 5-Year Roadmap to General AI

Your strategic advantage in the race to General AI. Weekly intelligence on infrastructure, investments, and breakthroughs before your competitors see them. For leaders who won't be caught off guard by the AI revolution. Join 15,000 worldwide AI leaders.

By Jiri "Skzites" Fiala

).

But the virtualization tax isn’t just about performance. It’s about control, security, and the illusion of isolation that multi-tenant environments create.

The Security Theatre Problem

Multi-tenant cloud promises isolation. Your model training runs on hardware that’s logically separated from competitors. Your data never touches their data. Your processes can’t see their processes.

Logically separated isn’t physically separated.

Side-channel attacks exist. Memory leaks happen. Noisy neighbors degrade your performance during peak utilization. The mathematical isolation that cloud providers guarantee operates on physical hardware that doesn’t respect those boundaries the way software does.

I’m not suggesting cloud providers are negligent. They’ve invested billions in isolation technology. Their security teams are among the best in the industry.

But the fundamental architecture requires trade-offs. The same resource sharing that makes cloud economics work creates attack surfaces that dedicated infrastructure eliminates entirely.

For commodity workloads, this trade-off makes sense. Email servers don’t need physical isolation. Web hosting doesn’t require hardware-level security boundaries. The efficiency gains justify the theoretical exposure.

AI model training is different.

Your training data represents competitive advantage. Your model architecture embodies intellectual property worth millions in development cost. Your hyperparameters reveal strategic priorities that competitors would love to understand.

Running that workload on shared infrastructure means trusting that isolation technology will continue to outpace adversarial research. That’s a bet. Some organizations are comfortable making it. Others shouldn’t be.

The Control Illusion

“Enterprise-grade” cloud instances promise dedicated resources. Premium pricing buys priority scheduling and capacity reservation.

What it doesn’t buy is actual control.

You can’t choose which physical servers run your workload. You can’t verify the firmware versions on hardware you’re paying for. You can’t audit the supply chain that delivered the components processing your data.

For regulated industries, these limitations create compliance complications that cloud providers address through certifications and audit reports. The paperwork exists. The actual verification doesn’t.

Dedicated AI infrastructure changes this equation. When you own the hardware, you control the stack from silicon to software. When you operate the facility, you audit the supply chain yourself. When you manage the network, you verify isolation isn’t just promised but implemented.

This matters more as AI regulation evolves. The EU AI Act requires explainability that’s harder to demonstrate when you can’t fully characterize the compute environment. SOC 2 compliance becomes more complex when “compute environment” spans infrastructure you’ve never seen and can’t inspect.

The Real Cost Comparison

Cloud pricing looks straightforward. Per-GPU-hour rates. Volume discounts. Reserved instance savings.

But the comparison isn’t cloud pricing versus owned infrastructure pricing. It’s cloud pricing plus virtualization overhead plus security risk plus control limitations plus compliance complexity versus owned infrastructure with its different cost structure.

When you add the 5% average performance loss, the premium you’re paying for “enterprise” isolation that doesn’t actually isolate, the consulting costs for compliance verification you can’t fully achieve, the numbers shift.

Not for every workload. Small-scale experimentation still makes sense in cloud. Variable workloads with unpredictable demand benefit from elastic scaling. Teams without infrastructure expertise shouldn’t build their own.

But at scale, with consistent utilization, for workloads where security and control actually matter, the cloud premium becomes a cloud penalty.

Subscribe now

Share

Leave a comment

The Hybrid Architecture Reality

I’m not arguing against cloud. I’m arguing against the assumption that cloud is the default correct answer for AI compute.

The organizations achieving best results deploy hybrid architectures. Development and experimentation in cloud, where flexibility matters more than efficiency. Production training and inference on dedicated infrastructure, where consistent performance and security justify operational complexity.

This isn’t a radical position. It’s how serious AI organizations have operated for years. What’s changing is the availability of dedicated infrastructure that doesn’t require building and staffing your own data center.

Modular AI infrastructure makes hybrid architecture accessible to organizations that couldn’t previously justify dedicated deployment. A 400kW container isn’t a data center. It’s a dedicated compute node that can scale your hybrid strategy without massive capital commitment.

The Questions Technical Buyers Should Ask

If you’re evaluating AI compute options, start with these:

What’s your actual utilization rate? Below 60%, cloud economics usually win. Above 70%, dedicated infrastructure deserves serious analysis.

How sensitive is your training data? If competitive exposure would cause material harm, physical isolation might be worth the premium.

What’s your compliance trajectory? Regulations tightening? Control requirements increasing? The compliance complexity gap between cloud and dedicated is widening.

Can you staff dedicated infrastructure? Operational expertise requirements are real. If you can’t hire or develop infrastructure skills, cloud’s operational overhead reduction matters more than its performance overhead.

How stable are your workloads? Variable demand favors cloud elasticity. Consistent demand favors dedicated efficiency.

The Strategic Shift

Two years ago, the question was whether your AI team should use cloud or build data centers. That was a false binary that assumed infrastructure required massive scale to justify dedicated deployment.

Modular architecture eliminates that assumption. Dedicated AI compute now scales from single containers to multi-megawatt installations. The entry point isn’t “build a data center.” It’s “deploy a compute node.”

This changes the calculus for organizations that previously had no alternative to cloud. The virtualization tax becomes optional. Physical security becomes achievable. Control becomes real rather than contractual.

The cloud providers see this shift coming. They’re responding with dedicated instance offerings, on-premises deployments, hybrid connectivity options. The market is forcing them to acknowledge what technical buyers have understood for years.

Multi-tenant compute is appropriate for some workloads. For AI at scale, it’s increasingly not.

JF is a C-level executive and serial entrepreneur who has founded 110+ startups. He runs the AI Executive Transformation Program in Prague and writes about uncomfortable truths in AI implementation at AI Off the Coast.

AI of the Coast: The 5-Year Roadmap to General AI

Your strategic advantage in the race to General AI. Weekly intelligence on infrastructure, investments, and breakthroughs before your competitors see them. For leaders who won't be caught off guard by the AI revolution. Join 15,000 worldwide AI leaders.

By Jiri "Skzites" Fiala

).

Cheap electricity. Business-friendly regulation. No state income tax. Land for days. The perfect formula for compute expansion while the rest of America choked on permitting delays.

• Then ERCOT published its October 2025 forecast.

• 205 gigawatts of large load interconnection requests.

• Up from 56 gigawatts just thirteen months earlier.

• A 227% increase, with 73% attributed to data centers alone.

The grid that was supposed to power America’s AI revolution now needs “thousands of miles of new transmission” at costs exceeding $30 billion.

I spent three hours last week on a call with partners evaluating Texas market entry. The numbers that looked compelling six months ago now look like traps.

The Illusion of Deregulation

Texas electricity isn’t deregulated.

It’s differently regulated.

ERCOT operates as an island, isolated from national interconnections that let other regions share load during emergencies. This independence was a selling point: no federal oversight, no cross-border complexity, pure market economics.

The 2021 winter storm exposed the vulnerability. Four million homes lost power. Hundreds died. The political response created oversight requirements that now complicate large-load interconnections.

Data center developers who chose Texas specifically to avoid regulation discovered they’d traded federal oversight for state oversight that can change based on which way the political wind blows after the next freeze event.

Subscribe now

SB-6 requires large electricity consumers to demonstrate reliability contributions before receiving firm power commitments. Sounds reasonable until you realize “reliability contributions” means either on-site generation capacity or contractual arrangements that effectively create the same grid dependencies Texas was supposed to avoid.

The WhiteGlove Problem

Last week, we discussed Texas market entry with partners who know the terrain. Their pitch: Economic Development Corporations across the state are hungry for data center investment. Land deals are available. Local officials want the tax revenue.

But here’s what the pitch didn’t address: Texas’s independent grid means every deployment bet carries binary weather risk. The same isolation that eliminated federal compliance created a single point of failure that no amount of local political goodwill can solve.

Share

When ERCOT calls for conservation events, your compute goes down. When transmission constraints limit power delivery to your region, your SLAs become suggestions. When the next freeze event triggers another political crisis, the regulatory response will affect your operational economics in ways you cannot model today.

• The EDC relationships are real.

• The land availability is real.

• The political friendliness is real.

The grid reliability problem is also real, and it doesn’t care about political friendliness.

The Urkott Factor

Texas power infrastructure isn’t owned by ERCOT. It’s managed by utilities operating under ERCOT coordination. The largest transmission providers in targeted development regions have their own capital constraints, permitting timelines, and strategic priorities.

Convincing a county to approve your project is step one.

Convincing the transmission provider to prioritize your interconnection is step two.

Those are different political economies with different stakeholders and different timelines.

The partners we spoke with acknowledged this directly: data center developers in Texas are discovering that local approval doesn’t guarantee grid access. The bottleneck has shifted from planning commission to transmission infrastructure investment decisions that happen in utility boardrooms, not county courthouses.

The European Alternative

While Texas debates grid expansion financing, European markets with established utility partnerships offer something Texas cannot: locked pricing without weather roulette.

Six-year fixed power agreements. Grid distribution cost elimination through local power plant connections. Dedicated infrastructure without queue competition.

The unit economics favor Texas on paper. Electricity costs look lower. Land costs are minimal. Tax incentives appear generous.

But paper economics assume grid availability. When that assumption fails, Texas’s cost advantage becomes a liability advantage. Your cheaper power becomes no power at all, and your contracts don’t care about force majeure clauses you thought would protect you.

The Real Opportunity in ERCOT’s Chaos

Texas will remain attractive for certain use cases. Training runs that can tolerate interruption. Workloads with genuine flexibility. Companies willing to accept reliability risk in exchange for cost reduction.

But enterprise SLAs require enterprise reliability. AI inference serving production traffic cannot tolerate conservation events. Financial services computing cannot wait for grid recovery after ice storms.

The gap between what Texas promises and what Texas delivers creates demand for alternatives that seemed unnecessary two years ago. Energy-sovereign infrastructure that doesn’t depend on ERCOT coordination. Modular deployment that can relocate when regulatory environments shift.

Texas taught the AI industry that cheap electricity isn’t cheap if it’s unreliable. That lesson is reshaping infrastructure strategy across the sector.

The state that wanted to be AI’s capital is becoming AI’s cautionary tale.

Strategic Implications

If you’re evaluating Texas deployment, ask questions your broker won’t volunteer.

The answers might still support Texas investment.

But they won’t be the answers you’d have gotten twelve months ago, and they definitely won’t be the answers you’ll get twelve months from now.

Texas is still building AI infrastructure.

Whether it can reliably power AI infrastructure remains an open question that $30 billion in transmission investment is supposed to answer.

Subscribe now

I wouldn’t bet my compute on “supposed to.”

JF is a C-level executive and serial entrepreneur who has founded 110+ startups. He runs the AI Executive Transformation Program in Prague and writes about uncomfortable truths in AI implementation at “AI Off the Coast".

AI of the Coast: The 5-Year Roadmap to General AI

Your strategic advantage in the race to General AI. Weekly intelligence on infrastructure, investments, and breakthroughs before your competitors see them. For leaders who won't be caught off guard by the AI revolution. Join 15,000 worldwide AI leaders.

By Jiri "Skzites" Fiala

).

$64 billion in data center projects sit frozen across America. Not because the technology failed. Not because financing collapsed.

Because Susan in Prince William County called her city councilman.

The most sophisticated companies in human history, wielding $315 billion in combined annual infrastructure capex, are being stopped by planning commission meetings in buildings that still use overhead projectors.

I’ve watched this pattern before.

In 2006, I watched a €40M manufacturing expansion die because a local politician wanted payback for an unrelated slight.

In 2014, a logistics hub I advised imploded when three retired schoolteachers organized a Facebook group. The technology was irreplaceable. The economics were undeniable. The project died anyway.

The Mathematics of Local Opposition

Data Center Watch counted 188 grassroots organizations actively opposing AI infrastructure across 28 states.

That number grew 125% in Q2 2025 alone.

Not organized by political affiliation.

Not funded by competitors.

Just people who don’t want humming boxes near their homes.

The arguments vary by partisan lens. Republicans cite tax abatements they fund through property assessments. Democrats raise environmental concerns. But the underlying psychology is identical: strangers are industrializing where I live.

Virginia’s Prince William County has become ground zero. The “Digital Gateway” proposal, worth $24.7 billion, faces three active lawsuits. Data centers already consume 25% of statewide electricity. The issue has triggered recalls, resignations, and primary defeats of elected officials who supported expansion.

In Tucson, Arizona, residents booed executives out of council chambers when Project Blue died unanimously. City councilman Rocque Perez admitted he learned more about the data center from angry constituents than from the developers themselves.

This isn’t NIMBY.

This is organized political infrastructure opposition at scale.

The Hyperscaler Blindspot

Amazon, Google, Microsoft, and Meta deploy armies of lobbyists in Washington. They have sophisticated government relations teams managing regulatory relationships across continents. They’ve mastered federal permitting, state incentive negotiation, international trade compliance.

Subscribe now

They have almost no capacity for Mrs. Henderson’s concerns about her water bill.

The skills that work at federal scale actively hurt at local scale. Washington influence tactics read as condescension in Mount Pleasant. Glossy presentations designed for institutional investors alienate volunteer fire department chiefs worried about traffic during emergencies.

Fortune magazine quoted local activists describing Big Tech as “robber barons at the turn of the century.” The comparison isn’t about wealth accumulation. It’s about perceived contempt for communities whose land, water, and electricity these projects will consume.

The hyperscalers fundamentally misunderstand the battleground. They’re optimizing for technical permitting when the real fight is emotional legitimacy.

Why Grid Constraints Are Now A Secondary Problem

For two years, the AI infrastructure narrative focused on grid interconnection queues.

PJM’s 2,300 GW backlog.

ERCOT’s 205 GW request explosion.

Seven-year wait times for transformer delivery.

These remain real constraints.

Share

But community opposition moves faster than grid upgrades.

A project can survive a 24-month utility interconnection timeline. It cannot survive a recall election that replaces three county commissioners with candidates who campaigned explicitly on blocking your proposal.

The grid constraint is predictable. You can model transformer lead times. You can project permitting timelines. Community opposition is chaotic, emotional, and path-dependent in ways that defy financial modeling.

Northern Virginia’s data centers consume more electricity than many European countries. The grid delivered. The neighbors revolted anyway.

The Emerging Arbitrage

Here’s what the market hasn’t priced: not all locations face equal opposition risk.

Rural communities with declining populations and agricultural employment often welcome industrial development their parents would have opposed. The same proposal that dies in suburban Virginia might sail through a depopulating county in eastern Montana.

But those locations lack grid infrastructure. Traditional data centers require grid connections that rural locations cannot provide within relevant timelines.

Modular, energy-sovereign infrastructure changes this calculus entirely.

A 1 MW container deploying to a biogas facility in agricultural Slovakia faces zero community opposition because it creates zero community burden.

No grid strain.

No water consumption.

No diesel backup generators humming through the night.

The locations that avoid NIMBY opposition are precisely the locations that traditional hyperscale architecture cannot serve. Energy sovereignty isn’t just about cost efficiency or speed of deployment. It’s about accessing locations where the political economy permits construction.

The Strategic Implication

Big Tech’s $64 billion problem creates a structural opportunity for infrastructure that communities actually want.

When your deployment creates local jobs without straining local resources, you’re not fighting the community. You’re partnering with it. When your electricity comes from agricultural waste that would otherwise decompose, you’re solving a local problem while solving your compute problem.

The hyperscalers will eventually learn this. They’ll hire community relations specialists. They’ll fund local benefit programs. They’ll adapt.

But adaptation at hyperscale takes years.

In the meantime, the companies that solve the neighbor problem today will have infrastructure deployed while their competitors are still fighting planning commission battles.

Subscribe now

The AI infrastructure race isn’t just about who can build fastest. It’s about who can build where building is welcome.

JF is a C-level executive and serial entrepreneur who has co-founded 110+ startups. He runs the AI Executive Transformation Program in Prague and writes about uncomfortable truths in AI implementation at AI Off the Coast

AI of the Coast: The 5-Year Roadmap to General AI

Your strategic advantage in the race to General AI. Weekly intelligence on infrastructure, investments, and breakthroughs before your competitors see them. For leaders who won't be caught off guard by the AI revolution. Join 15,000 worldwide AI leaders.

By Jiri "Skzites" Fiala

).

The future isn’t a datacenter.

It’s a dark economy.

Somewhere in Changping, China, a factory produces one smartphone per second. No humans. No lights. No breaks. One phone every heartbeat of every hour of every day.

Machines building products in perfect blackness, and nobody at Anthropic seems to have noticed that Dario Amodei’s “country of geniuses in a datacenter” already has a working prototype. It just doesn’t look the way he imagines.

Amodei published “The Adolescence of Technology” in January 2026.

Thirty thousand words mapping five categories of civilizational risk from powerful AI.

• Biosecurity nightmares.

• Autonomous weapons.

• Labor displacement that could shatter the social contract.

• Power concentration that would make Rockefeller blush.

• Unknown unknowns lurking in the acceleration.

I read it with the strange vertigo of recognizing someone describing your house from the outside. Getting the number of windows right while completely missing the foundation.

After twenty years across Fortune 500 boardrooms, dozens of startups, and watching 95% of enterprise AI projects die not from technical failure but from organizational incomprehension, I can tell you: Amodei’s risk taxonomy is intellectually honest, technically rigorous, and built on an architectural premise that physics has already vetoed.

• His essay assumes centralization.

• The future is dark.

• And distributed.

• And already here.

Subscribe now

The Dark Economy Doesn’t Wait for Permission

Amodei envisions fifty million superintelligent instances in a datacenter. A singular “country of geniuses” running at 10 to 100x human speed, coordinated centrally, requiring societal governance frameworks to prevent catastrophe.

I envision something stranger.

Something already forming in the phosphorescent glow of server racks scattered across European farmlands, and Texas industrial parks and abandoned European manufacturing sites.

A dark startup is a company where one founder orchestrates dozens of AI agents. AI Product managers synthesizing market signals into specifications while the founder sleeps. AI Architects proposing system designs for morning review. AI Developers shipping features before breakfast. AI QA catching regressions before dawn. AI Marketing adjusting campaigns based on overnight performance data.

One human.

Dozens of AI machines.

Complete operational continuity.

This is not Amodei’s centralized singleton.

This is a thousand founders, each commanding their own miniature country of geniuses, deploying from modular containers that bypass grid dependency entirely. The “country” doesn’t materialize in one place. It metastasizes everywhere, simultaneously, in facilities that operate without light.

Gartner estimates 60% of manufacturers will adopt lights out operations by 2026. The dark factory market hit $119 billion in 2024. China deployed over one million industrial robots by 2025, more than the rest of the world combined.

Amodei worries about a hypothetical future.

The dark economy is a measurable present in making...

Where Amodei Is Absolutely Right

Let me be clear about something uncomfortable: most of his risks are real. I partially agree with his thesis. In some places, violently.

His biosecurity framework is the most responsible thing any AI CEO has published. The classifiers Anthropic deploys to block bioweapon related outputs cost 5% of inference compute. That’s genuine commercial sacrifice, not safety theater. His warning that AI could walk someone through the entire bioweapon production process interactively over weeks or months is not hypothetical terror. It’s engineering probability.

His labor displacement analysis tracks precisely with what I see. AI models went from struggling with basic arithmetic to writing production code in three years. His prediction that 50% of entry level white collar jobs face disruption within one to five years aligns with MIT’s 2025 research showing top performers doubling productivity while the bottom third gained almost nothing. AI amplifies whatever capabilities you already possess. If you lack deep judgment, AI will not supply it. It will expose the gap.

His power concentration warning deserves more attention than it’s getting. The IMF’s Working Paper 2025/068 confirms what Piketty theorized: capital income and wealth inequality always increase with AI adoption. The wealth Gini coefficient rises 2 to 7 percentage points depending on adoption scenario. Not over decades. Over years.

These are not hypothetical risks from a hypothetical future.

These are current trajectories with measurable acceleration.

Where I diverge is not on whether these dangers exist.

It’s on the architecture through which they manifest.

And architecture determines everything about how you defend against them.

The Centralization Fallacy

Amodei’s five risk categories all assume a single architectural premise: concentrated power in concentrated infrastructure. His “country of geniuses” sits in one place, controlled by one entity, requiring one governance framework to contain.

Physics has already rejected this application.

There are 2,300 gigawatts of AI infrastructure sitting in US grid interconnection queues. Median wait time: five years. The entire US grid operates at roughly 1,200 gigawatts. To deploy what’s already under contract would require nearly doubling national grid capacity. Community opposition has blocked $64 billion in Big Tech infrastructure projects. These are not regulatory failures. They are the collision between exponential technology demands and linear infrastructure capacity.

I build modular AI data centers. The physical kind: 1MW compute containers with NATO grade security, biogas and solar microgrids, deploying in 120 days while traditional facilities wait five years for an interconnection agreement.

We achieve 42.3% cost advantages over centralized approaches like CoreWeave.

Our five container deployment model delivers 187.2% ROI over four years.

This matters for Amodei’s thesis because it changes the geometry of risk.

A distributed architecture of thousands of several megawatt’s facilities replacing a dozens of gigawatt hyperscale centers doesn’t just reduce cost.

It restructures power.

Multiple actors control multiple facilities across multiple jurisdictions.

No single entity accumulates the concentrated compute that makes Amodei’s autonomous weapons swarm or global surveillance apparatus feasible at the scale he fears.

Amodei wants surgical regulation to prevent concentration.

I’m watching physics enforce distribution whether regulation exists or not.

The Risk He Never Named

Here’s where it gets strange. And where I depart from comfortable agreement into territory that should keep every executive reading this awake tonight.

Amodei’s five risk categories miss a sixth: the dark economy itself as a civilizational force.

When Piketty’s r > g formula meets the dark startup model, you get something unprecedented. A single founder commanding AI agents can produce the output of fifty knowledge workers. The return on capital doesn’t just exceed economic growth. It eclipses it. And the infrastructure to enable this doesn’t require Amodei’s hypothetical datacenter singleton. It requires a modular container, a power source, and a fiber optic connection.

Three resources will define who accumulates wealth at this new velocity: land for data centers, access to AI compute, and electricity. Whoever controls these inputs controls the means of production in the dark economy. This is not metaphor. This is industrial economics applied to the age of artificial intelligence.

Amodei worries about a world where AI companies capturing $3 trillion in annual revenue creates dangerous concentration. He’s right to worry. But his framing assumes concentration happens through centralized AI providers. The darker scenario is fragmentation: ten thousand dark startups, each running their own compute infrastructure, each operating 24/7 without employees, collectively displacing entire industries while no single entity is large enough to regulate.

• This is the risk Amodei’s centralized framing cannot see.

• Not a country of geniuses in a datacenter.

• A diaspora of geniuses in darkness.

Share

The Four Phases Meet the Five Risks

Let me map what I see coming against what Amodei fears.

Augmentation (2025 to 2027). We’re here now. Team sizes compress 5x to 10x. McKinsey reports 20% to 45% productivity gains. Amodei’s classifiers and constitutional AI work at this phase. The infrastructure bottleneck limits proliferation. His risks are real but constrained.

Autonomy (2027 to 2030). Dark startups become dominant. 60% of US venture capital in 2025 already flowed to rounds of $100 million or more because capital concentrates where velocity concentrates. Amodei’s labor displacement manifests here with brutal precision. But the threat vector isn’t “AI replaces your job.” It’s “a founder with forty seven agents replaces your company.”

Infrastructure Wars (2030 to 2033). The battle shifts from software to physics. Land, energy, compute become contested resources. Whoever pioneers energy sovereign, modular deployment at scale gains asymmetric advantage that makes algorithmic capability secondary.

The New Feudalism (2033 to 2036). Wealth stratification completes. Those who own AI infrastructure occupy positions analogous to landowners in agrarian economies. But the path runs through distributed dark infrastructure, not centralized datacenter kingdoms. The feudalism is architectural, not political.

What Amodei Should Build

I respect Anthropic’s work on Constitutional AI. Training Claude at the level of identity and values rather than instruction lists is architecturally sound. Their interpretability research represents genuine science. The public disclosure of system cards reflects commitment most competitors lack.

But Amodei treats infrastructure as someone else’s problem.

His risk framework assumes the datacenter exists. His policies assume centralized control points. His economics assume AI revenue concentrates in identifiable companies. All fail in a dark economy where compute distributes across energy sovereign nodes that bypass traditional infrastructure.

Anthropic should be thinking about how Constitutional AI applies when Claude runs in a modular container on Czech farmland, operated by a single founder with no employees, serving clients across three continents while the founder sleeps. That’s a deployment architecture we’re building right now at DCXPS.

The question isn’t whether AI models will have good values. The question is whether governance frameworks designed for centralized providers survive contact with distributed, dark, energy sovereign compute.

The Real Rite of Passage

Amodei frames humanity’s test as whether our institutions can mature fast enough to govern superintelligent AI. It’s a beautiful formulation. Worthy of Carl Sagan’s Contact scene he opens with.

But the test is already different from what he describes.

The real rite of passage is whether we can build distributed infrastructure fast enough that no single actor accumulates the concentrated power Amodei rightly fears. Whether the dark economy emerges as a federation of sovereign compute nodes rather than a monarchy of hyperscale datacenters. Whether the ten thousand founders with their forty seven agents each create a new form of distributed capitalism or a new feudalism where infrastructure owners replace landowners as the permanent aristocracy.

Amodei wrote an essay about controlling a giant. The future is a swarm. You don’t govern swarms with constitutions written for giants.

I’ve founded 110+ startups. The pattern never changes. The technology everyone debates is never the technology that transforms. Centralized mainframes gave way to distributed PCs. Monolithic applications gave way to microservices. Hyperscale datacenters will give way to modular, energy sovereign compute nodes scattered across the physical world like digital mycelium.

Amodei’s risks are real. His architecture is wrong. And the dark economy forming in the spaces between his assumptions will reshape civilization before his governance frameworks finish their first draft.

Somewhere in Changping, another phone slides off the line.

• No lights.

• No workers.

• No ceremony.

• The future doesn’t announce itself.

• It just ships.

JF is a serial entrepreneur, founder of DCXPS modular AI data centers, and author of the AI Executive Transformation Program in Prague.

He writes at AI Off the Coast about the uncomfortable truths of AI implementation and our AI future.

AI of the Coast: The 5-Year Roadmap to General AI

Your strategic advantage in the race to General AI. Weekly intelligence on infrastructure, investments, and breakthroughs before your competitors see them. For leaders who won't be caught off guard by the AI revolution. Join 15,000 worldwide AI leaders.

By Jiri "Skzites" Fiala

€15 million potential investment.

Healthcare SaaS company.

Financial model review, market analysis, competitive assessment, technical architecture evaluation, IP verification, team background checks, regulatory compliance, customer reference calls.

Standard work.

Six-week timeline.

The Analyst spent four days building the financial model.

Reconstructing three years of historical performance.

Projecting five-year scenarios.

Sensitivity analysis on key assumptions.

Revenue recognition analysis.

Burn rate modeling.

The kind of Excel work that junior analysts do: meticulous, time-intensive, critical.

The AI completed the same analysis in forty-seven minutes. Found three errors in the company’s historical financials that she would have missed. Identified accounting treatment inconsistency that masked deteriorating unit economics. Generated twelve scenario models versus her three. Flagged regulatory risk that wasn’t in her checklist.

Then the AI found the material IP dispute. Patent infringement lawsuit filed three months prior. Not disclosed in data room. Buried in European patent office records. The company was being sued by larger competitor for €8.3 million. Case had merit. Would materially impact valuation.

She’d have missed it. The AI found it in nineteen seconds by cross-referencing 150 million patent records.

The partner pulled the investment. €15 million saved. The analyst kept her job. For now. But she understood. Four days of financial modeling versus forty-seven minutes. Critical IP issue she missed that AI found in seconds. Her €68,000 salary versus €35,000 monthly platform subscription serving the entire fund.

The mathematics were brutal.

She was obsolete.

She just didn’t know when the partner would realize it.

This is Investment Due Diligence AI.

Where algorithmic analysis proved human investment judgment is pattern matching that machines do better, faster, cheaper.

Welcome to automated due diligence.

Where three founders commanding AI agent teams replace investment teams of fifteen.

Where six-week analysis becomes four-hour reports.

Where 100% of data gets reviewed instead of 15%.

Where the VCs and PE firms deploying this technology are financing their own obsolescence.

The irony is exquisite.

The trajectory is inevitable.

The analysts are already redundant.

The €4.8 Billion Due Diligence Waste Problem

European venture capital and private equity conduct approximately 8,400 deals annually. Average due diligence cost per deal: €180,000 in analyst labor, third-party reports, expert consultations, legal reviews. Total: €1.51 billion annually spent on pre-investment analysis.

The outcome is catastrophically inefficient. Traditional due diligence analyzes only 15% to 20% of available data. Humans can’t process everything in six to eight weeks. They sample. They prioritize. They miss things. One-third of critical issues go undetected because analysts didn’t look in the right places or didn’t have time to review everything.

Investment analysts spend 60% to 70% of time on routine tasks. Financial model building. Market sizing research. Competitor analysis. Customer reference synthesis. Patent searches. Regulatory compliance checking. The work is repetitive. Pattern-based. Exactly the cognitive labor that AI excels at replacing.

The due diligence timeline is human-limited. Six to eight weeks from term sheet to closing. Not because the work requires six weeks. Because humans work sequentially. Financial analysis, then market analysis, then technical review, then legal verification.

With meetings.

With context switching.

With sleep.

Subscribe now

The error rate is structural. Analysts miss issues because they’re processing massive information volumes under time pressure. The healthcare SaaS company hid the IP lawsuit in plain sight. The analyst didn’t find it because she wasn’t cross-referencing European patent records. She was focused on financial models and market analysis. The work got compartmentalized. The critical risk fell through the cracks.

Scale this across European investment markets. If 30% of due diligence work is wasted through incomplete data coverage, missed issues, and inefficient analysis, that’s €453 million annually in flawed pre-investment work. Plus the opportunity cost of bad investments that due diligence should have prevented.

More devastating: calculate the value of investments that shouldn’t have happened. Deals where proper due diligence would have surfaced fatal flaws. Conservative estimate: 8% to 12% of deals proceed with material undisclosed issues. At €78 billion annual European VC/PE deployment, that’s €6.2 to €9.4 billion invested in companies with hidden problems.

Share

The industry knows this.

Read more