The most important sentence in OpenAI’s DevDay announcements was not about capability. It was four words in the enterprise rollout notes: “off by default.”

Read that again. The most aggressively marketed product category of 2026 (always-on agents that learn your work and carry it overnight) ships with the switch down, waiting for a workspace admin somewhere to accept responsibility for what happens next.

That switch is the whole story.

In the first week of October, Meta’s personal agent Muse hit #1 on the U.S. App Store. OpenAI shipped “dots,” always-on agents that plug into 4,000+ apps and sit inside Slack and Teams. Microsoft’s agent works while you sleep and handles your follow-ups before you’ve had coffee. Instinct, an invite-only Muse competitor, raised $1 billion at a $10 billion valuation — four times what it was worth a month earlier, per Axios. The agents shipped. All of them, everywhere, at once.

And then there’s the Oppenheimer survey from the same week: 58% of Americans won’t share a password with any AI agent. Eight percent would trust Meta with theirs.

Eight.

The gap between those two numbers (the valuation curve and the trust curve) is where the next decade of company-building lives. I’ve spent thirty years and 110+ companies watching technology adoption stall in exactly this gap, and I can tell you what fills it. It is never a better model. It is an employer.

Capability was never the constraint

Here’s what the agent vendors got right: the work is real. An agent that monitors channels overnight, drafts the follow-ups, reconciles the invoices, and prices the capacity is not a demo anymore. We run a fleet of them. I watch their output land in approval queues every morning, and the volume they carry would have required a back office of a dozen people three years ago.

Here’s what almost every buyer is discovering: capability without accountability is not labor. It’s liability with a chat interface.

Think about what an employer actually does. Not the payroll part — the boring, load-bearing part. An employer defines the job before the worker exists. Sets the probation period. Decides what the worker may touch, signs for what the worker does, measures output against a standard, and fires what underperforms. Strip those functions away and you don’t have a workforce. You have a crowd.

Now look at the agent economy as shipped in October 2026. Meta will sell you a crowd. OpenAI will sell you a crowd with enterprise SSO. What nobody sells (because it cannot be bought, only built) is the org chart that makes the crowd employable: who approves which action class, what the agent may never touch, what evidence retires it, whose name is on the damage.

Ben Thompson’s read on Stratechery is that agent lock-in will be brutal: once you’ve invested in one, switching costs compound, so most people and firms will keep exactly one. He’s right, and it cuts deeper than app-store dynamics. The investment that locks you in isn’t the chat history. It’s the management system you build around the agent — the permissions, the review rituals, the institutional knowledge of where it fails. That system is the asset. The agent is a plug-in.

“Off by default” is a confession

OpenAI’s dots launched in beta for Enterprise, Edu, and Healthcare — admin-enabled, off by default. Amazon and OpenAI built Bedrock Managed Agents so enterprises can run the whole thing inside their own AWS perimeter, because apparently the way to sell an autonomous workforce in 2026 is to promise it never leaves the building.

These are not product decisions. These are confessions. The vendors know the mandate hasn’t arrived — that somewhere between the model card and the quarterly board meeting, there is no human whose job description says “accountable for what the agents did on Tuesday.”

Amazon, meanwhile, blocked Muse outright. Thompson’s explanation is clean: an agent shopping on your behalf never sees an ad. A quarter-century of aggregator economics, dead on contact with a buyer that doesn’t scroll. I’d push it one step further: Amazon understands, faster than most enterprises, that whoever employs the agent owns the relationship. They are not blocking a chatbot. They are blocking someone else’s employee from walking their aisles.

Full disclosure of my own position, since I’m arguing the book I built: DCXPS runs as an AI-native operator — agents carry the operational volume of an infrastructure company, four humans hold the signatures. We sell compute, and we sell the operating model that consumes it. Discount accordingly. But the scars are real, and they’re the useful part.

What employing an agent actually took us

When we committed to running the company this way, the engineering was the easy half. The hard half was writing, in advance, the rules a human workforce accumulates over decades of labor law and management practice — because none of it exists for agents yet.

Every agent starts as a written specification: problem, scope, tools, failure modes, an evaluation set — before any agent exists. Then it shadows: it does the work, a human still does the work, outputs get compared against a golden set. Then it’s supervised, every single action queued for individual human approval, for weeks. Only after sustained clean performance does it graduate to batch review. Autonomy comes last, behind interlocks that never move: machines never move money, never sign, never send anything outside the company.

A commissioning ladder, in other words. We stole the idea from our own cooling loops — FAT then SAT, no system goes live ungated. The corporate version turned out to be identical. Specified, shadowed, supervised, then trusted. Never the other way around.

Two more pieces, because they’re the ones everyone skips. First, the kill bar: an agent that can’t beat 20% of loaded human cost on its deliverable within ninety days is retired. Not coached. Retired. Sentiment is for humans. Second, the failure memo: before we hire a person, someone writes down which agent was tried, at what level of the ladder, and why it failed. The memo requirement does more for organizational honesty than any OKR system I’ve deployed in three decades.

Total agents that survived their first quarter on our floor: fewer than the number we commissioned. That ratio is the tuition. Nobody shipping a $500/month Pro plan mentions it.

The mandate is coming from the worst possible direction

Enterprises had two years to build this architecture voluntarily. Most didn’t. The Harvard Business School number floating around this fall (roughly one-third of corporate AI experiments succeed) is usually read as a technology failure. It isn’t. Models don’t fail at 67% rates. Managements do. An experiment with no owner, no golden set, no kill bar, and no approval queue isn’t an experiment. It’s a hobby.

So the mandate will arrive the other way: from regulators, courts, and insurers, priced in liability instead of productivity. More on that Saturday — the California subpoena deserves its own piece. The preview is one sentence long: the question “who employed this agent?” has started appearing in legal filings, and “nobody, it’s a tool” is about to stop being an acceptable answer.

When that happens, the companies that treated agent governance as bureaucratic overhead will discover it was the product all along. The oil majors learned this about process safety. The banks learned it about trading desks: the money was never in hiring traders, it was in the risk architecture that let you run ten thousand positions without one of them killing you. Electricity didn’t reorganize the economy when generators got cheap; it reorganized the economy when the systems around it (metering, insurance, standards, operators) made it boring enough to depend on.

Agents in October 2026 are generators without meters. Spectacular, cheap, and uninsurable.

Where this goes

Three predictions, falsifiable, on the record.

First: within eighteen months, “agent governance” stops being a slide and becomes a budget line, and the spend flows to whoever can show an audit trail — every action attributed, timestamped, reviewable. The vendors will retrofit this. The AI-native operators already run on it.

Second: the first Fortune 500 “Chief Agent Officer” title appears, is mocked on tech Twitter for a week, and is standard within two years — because someone has to own the approval queue, and legal will refuse to let that remain ambiguous past the first subpoena.

Third: valuations bifurcate. Agent capability commoditizes on the model curve; every lab ships comparable dots within quarters of each other. Agent employment compounds like seniority — the trust architecture, the commissioning record, the audited history of what this fleet did under supervision. Instinct quadrupling to $10 billion in a month is a bet on the first category. The second category doesn’t have a ticker yet.

The dark company thesis was never that machines replace people. It’s that the org chart is the product: a handful of humans holding signatures, money, relationships, and witness, commanding fleets of agents commissioned like industrial equipment. In 2026 that reads as eccentric. The week Meta’s agent topped the App Store while 58% of its potential users refused to hand over a password, it started reading like the only bridge across the gap.

Everyone shipped an agent.

The next trillion dollars goes to whoever ships the employer.

I run DCXPS, an AI-native infrastructure operator — we build modular intelligence factories and run the company the way we run the racks. The agent-fleet numbers above are our own build telemetry; our commercial interest in this thesis is total.

Three pieces built the case: the org chart is the real moat, most operating models have no line for machine-handled work, and the ratio that actually matters, agents per human, never makes it onto a diligence checklist. This piece is where we stop describing the gap and show what’s already running inside it.

We commission agents through five graded levels, specified, shadow, supervised, batch, autonomous, each one earned against a real gate: a golden-set match rate, weeks of clean rejection rates, a security review, before the next level even unlocks. Four humans hold what never moves to a machine, money movement, signatures, client relationships, physical witness. That ladder isn’t a slide. It’s the operating structure running our own facility build right now, the same one behind the 8-month commissioning timeline from earlier in this newsletter. We’re building toward 15 agents for every human by the end of 2027, a target, not a claim already banked, but the one the ladder above is specifically built to earn.

Most of this category is still writing headcount plans for a business model that doesn’t require the headcount anymore. We built the ladder instead, one gate at a time, with the permanent interlocks that make it safe to climb: nothing autonomous ever touches money, a signature, or anything leaving the company, no matter how high up the ladder the fleet gets.

That’s the difference between a target on a roadmap and a system already running. We’re not asking you to believe the org chart works. We’re pointing at the ladder and the interlocks that make it real.

Get the full operating model, the commissioning ladder, and the build-log detail behind this series in our market report.

Subscribe now

Share

Every diligence checklist in this category asks about headcount. Almost none of them ask about the ratio that actually determines whether headcount scales with growth or stays flat while revenue climbs: agents per human.

Ask a prospective operator what that ratio is today, and what it’s targeting eighteen months out. Most won’t have a number ready. The ones that do are telling you something real about whether their cost base scales with facility count or decouples from it.

A traditional operator adding its fifth facility adds staff roughly proportional to the first four. An operator running toward a real agents-per-human ratio, if the commissioning ladder behind it is real, graded, and gated the way it’s supposed to be, adds facilities without adding headcount at the same rate, because the marginal unit runs on the same fleet that proved itself on the first one.

That ratio doesn’t show up in a capex table or a GPU spec sheet. It shows up years later, in whether an operator’s overhead grew with its footprint or stayed flat, and by then the diligence window has closed. Ask for it now, while it’s still a forward-looking number instead of a lagging one.

Subscribe for the closing piece in this series, or download the market report for the full operating framework.

Subscribe now

Share

The interesting thing about a permanent rule is what it rules out. “Machines never sign” sounds like a compliance footnote. It’s actually the entire architecture of how this business scales without the trust problem that kills most attempts at running infrastructure with a skeleton crew.

Here’s why that rule can’t be a setting, something a future version toggles off once the agents prove themselves. The moment a machine can sign, move money, or send something outside the company without a human in the loop, you’ve converted a productivity tool into a liability nobody signed up for, and you’ve done it exactly at the moment the fleet has earned the most trust and the least scrutiny. That’s the failure mode this rule exists to prevent, permanently, not provisionally.

Four humans hold the functions that stay human no matter how far the agent fleet advances: money movement, signatures, client relationships, physical witness. Everything else, the reconciliation, the provisioning, the pricing math, the paperwork, is built to be handled by agents that have earned their way up a graded ladder, one gate at a time, with rejection rates and security reviews standing between one level and the next.

That’s not a limitation on how far automation can go in this business. It’s the reason automation can go as far as it has without the whole thing collapsing the first time an agent makes a confident, well-formatted, completely wrong decision with real money attached.

Subscribe for the next piece, or download the market report for the full commissioning-ladder detail.

Subscribe now

Share

Pull up a standard data center operating model and look for the line where software agents handle reconciliation, provisioning, and pricing without a human touching every action. It’s not there. Most operating models in this category were written for a world where headcount scaled with facility count, because that’s the only model anyone had ever run.

That gap matters more than it looks like on paper. If your operating model doesn’t have a line item for what a machine does versus what a human does, you’re not modeling the actual cost structure of the business you’re evaluating, you’re modeling the cost structure of the last generation of data center operator, applied to a company that doesn’t work that way.

The operators doing this seriously aren’t “adopting AI tools” the way that phrase usually gets used, bolting a chatbot onto an existing process. They’re commissioning agents the way you’d commission a cooling loop, through five graded levels: specified, shadow, supervised, batch, and only then autonomous, each one gated behind a real test, a golden-set match rate, weeks of clean rejection rates, a security review, before the next level is even available. None of it skips a step, and the permanent interlocks never move: machines never move money, never sign, never send anything outside the company, never touch restricted data. Every one of those actions stays a human act, always.

That ladder is the actual line item missing from most diligence models in this category. If it’s not in yours, you’re pricing a headcount structure that doesn’t match the business in front of you.

Subscribe for the next piece, or download the market report for the full commissioning framework.

Subscribe now

Share

Every competitor in this category is racing to lock down the same GPU allocations from the same handful of suppliers. That’s not a moat. That’s a queue everyone’s standing in together, waiting on the same NVIDIA shipment schedule as everyone else with a checkbook.

The actual moat is underneath the hardware entirely: how the company that operates the facility is organized once the concrete’s poured and the racks are live. A handful of operators in this category are experimenting with what you could call a dark company: a small number of humans holding the functions that can never be automated away, money movement, signatures, client relationships, physical witness, with a fleet of software agents handling everything else: reconciliation, provisioning, pricing, paperwork.

That’s not a headcount efficiency story. It’s a different category of business entirely. A traditional data center operator scales headcount roughly linearly with facility count, more racks, more staff, more overhead compounding against margin at every step. An operator running a real dark-company structure doesn’t, because the marginal cost of operating unit two, three, and ten isn’t another team, it’s another agent instance running against a commissioning process the first unit already proved out.

Anyone can eventually buy the GPUs. Almost nobody has built the org chart that makes each additional facility cheaper to run than the last one, instead of more expensive. That’s the part of this business that doesn’t show up in a spec sheet, and it’s the part that actually compounds.

Subscribe for the next piece, or download the market report for the full operating model.

Subscribe now

Share

Every piece in this series has made the same argument from a different angle: the sticker price, the capex headline, the GPU generation, none of it matters if the build doesn’t clear on time, at the efficiency the model assumed. This piece is the receipt.

Our Gen1 modular unit: 392 GPUs, 112 B300 and 280 H200 NVL, roughly 450 kilowatts of IT load inside a 1-megawatt envelope, PUE at or below 1.12 on closed-glycol cooling. Order to commissioning, 8 months, documented in a 5-phase engineering program, not a target on a roadmap slide.

We didn’t get there by finding a cheaper GPU. We got there by treating the build timeline and the cooling architecture as the actual product, the thing that determines whether the hardware turns on current or turns on obsolete, whether the power bill eats 12% of spend or 30%. That’s the structural cost advantage this whole series has been pointing at: not a lower number on an invoice, a better number on everything the invoice doesn’t show.

Four pieces built the case that the market’s pricing the wrong things. This one is what it looks like when the right things get priced correctly from the start.

Get the full technical breakdown, including the complete unit specs and build timeline, in our market report.

Subscribe now

Share

The top four hyperscalers spent roughly $350 billion on capex in 2025. That number is heading toward roughly $400 billion in 2026. Every allocator watching this space has that figure memorized, because it’s the headline number, the one that gets quoted to justify why AI infrastructure is the trade of the decade.

Here’s the problem. That capex figure tells you how much money is being spent. It tells you nothing about how much of that spend converts into compute that’s actually online, actually powered, actually generating revenue on the timeline everyone’s modeling. You’re underwriting a market off the size of the check being written, not off the fraction of that check that clears the constraints actually gating deployment, land, permits, and above all, interconnection.

$400 billion chasing capacity that’s stuck 60-plus months deep in a grid queue isn’t $400 billion of realized infrastructure. It’s $400 billion of demand competing for a supply of clearable projects that’s a fraction of the headline number, and the gap between those two figures is exactly where every piece in this newsletter has been pointing.

The right number to underwrite isn’t total capex committed. It’s clearable capacity, adjusted for actual queue-clearance rates, adjusted for actual build timelines. That number is smaller, less exciting to quote in a pitch deck, and considerably more useful for figuring out where the real returns in this sector are going to land.

Subscribe for the closing piece in this series, or download the market report for the full capex-versus-clearance breakdown.

Subscribe now

Share

Somewhere in this market right now, someone is choosing a GPU-hour rate purely because it’s the lowest number on the page. That’s the wrong variable to optimize, and it’s costing them more than the discount is worth.

Here’s the counterintuitive part. List rates in this market for H200 NVL run around $50.44 per server-hour, B300 nearer $80.00. Both numbers sit above some of the stripped-down capacity you can find elsewhere. And both are still the better deal, because the price on the page isn’t the price you actually pay, it’s the price you pay per hour the hardware is actually available and running at the utilization you were promised.

A cheaper rate on capacity that’s oversubscribed, throttled, or sitting behind a queue you didn’t know existed isn’t a cheaper rate. It’s a deferred cost with a worse SLA attached. The real comparison isn’t dollars per GPU-hour listed. It’s dollars per GPU-hour actually delivered, at the reliability you were sold, and that number rarely appears anywhere in the sales conversation.

This is the same mistake as pricing the chip instead of the month it takes to turn on, one section over. The sticker price is the easy number to compare. The delivered number is the one that determines whether the deal made sense a year later.

Structural cost advantage was never about being the cheapest line on an invoice. It’s about what a dollar actually buys once you account for everything the invoice doesn’t show.

Subscribe for the next piece, or download the market report for the full pricing comparison.

Subscribe now

Share

PUE, power usage effectiveness, is the ratio of total facility power to the power actually reaching the compute. A PUE of 1.3 means 30% of everything you’re paying for goes to cooling, distribution, and overhead before a single GPU cycle happens. A PUE of 1.12 means that number drops to 12%.

That 18-point gap sounds like an engineering footnote. It isn’t. On a facility running hundreds of GPUs continuously, that gap compounds across every hour of every year the facility operates, and it shows up nowhere on a GPU spec sheet, nowhere in a financing summary, and almost never in the two-page teaser that gets passed around before a deal reaches diligence.

Most legacy data center infrastructure was designed for PUE assumptions closer to 1.3, sometimes worse, because it was built for a world where power was cheap and compute density was low. Neither of those assumptions holds anymore. Compute density has gone up an order of magnitude, and power is now the binding constraint on this entire category, which means every point of PUE inefficiency isn’t a rounding error, it’s a direct tax on the one resource everyone in this market is fighting over.

Closed-loop liquid cooling can push PUE to 1.12 or better with equipment available today, the technology isn’t the bottleneck. What’s missing is a diligence habit that treats cooling architecture as a line item instead of an afterthought, because power efficiency stopped being a nice-to-have the moment power became the scarcest input in the entire value chain.

If your diligence process has a line for GPU generation and no line for PUE, you’re measuring the part of the deal that’s easy to see and skipping the part that actually determines the margin.

Subscribe for the next piece, or download the market report for the full unit-economics comparison.

Subscribe now

Share

Two AI infrastructure deals can have the exact same GPU count, the exact same chip generation, and produce wildly different returns, for a reason that never shows up on the spec sheet: how long it takes to go from signed order to a running rack.

GPU performance doubles every 18 to 24 months. Traditional data center construction runs 48 to 72 months. Do that math once and it stops being abstract: a facility built on the old timeline is sized for hardware that’s already two generations behind by the time it’s live. You didn’t just wait four years. You bought four-year-old compute at today’s price.

The industry has spent a decade optimizing GPU procurement, allocation strategy, generation timing, and almost no attention on the build timeline itself, as if it were a fixed cost of doing business rather than the actual lever that determines whether the hardware you bought is still competitive when it turns on.

The industry doesn’t lack the components to build faster. It lacks the incentive to prioritize it, because build speed doesn’t show up on the same spec sheet as GPU count, and nobody’s diligence process rewards a builder for shaving eighteen months off a timeline nobody asked them to shorten. That’s not a claim about speed for its own sake. It’s the difference between deploying hardware that’s current and deploying hardware that’s already aging out, before it’s even switched on.

Price the month it takes to turn on. The chip count was never the scarce resource. The clock was.

Subscribe for the next piece, or download the market report for the full build-economics breakdown.

Subscribe now

Share

86% of queued capacity in U.S. interconnection queues never generates a watt. That’s not a forecast. That’s the historical base rate, and the five pieces before this one walked through why the market is only now starting to price it correctly.

We went from order to commissioning in 8 months.

Not a roadmap slide. Not a target on a slide deck. A completed, 5-phase build program, documented in the engineering brief that governs our Gen1 modular unit: 392 GPUs, roughly 450 kilowatts of IT load, PUE at or below 1.12 on closed-glycol cooling. The receipt exists. We’re not asking anyone to take our word for the timeline. We’re pointing at the spec.

This isn’t a thesis about what should happen to queue risk once the market fully prices it. It’s proof that a version of this exact problem is already solved, at a specific unit economics, on a specific timeline, by a team that’s already done it once. Everyone else in this series has been describing a mispriced risk from the outside. This is what it looks like from the other side of it.

If interconnection is the real diligence line item, and the last five pieces made the case that it is, the follow-up question is the only one left: who’s already cleared it, and are you talking to them.

Get the full breakdown in our market report, the grid data, the auction pricing, and the build specs behind this series, all in one place.

Subscribe now

Share

Capacity prices already moved once. PJM: $28.92 per megawatt-day to $329.17, one auction cycle. That happened because the market caught up, at least partially, to how much queued capacity never converts to actual generation.

It’s going to move again. And every repricing event rewards whoever solved the underlying constraint before the market noticed it was the constraint, not after.

That’s the part most allocators in this space keep missing. The upside doesn’t accrue evenly across “AI infrastructure” as a category, the way it might if this were a simple beta bet on a growing sector. It accrues specifically to operators who had already cracked permitting and interconnection before capacity got repriced, because they’re the only ones positioned to deliver against demand at the new price. Everyone else is still explaining, politely, why their queue position slipped another year.

I watch this exact split play out every time I sit in a diligence conversation now. There’s a specific kind of allocator who gets to say, two years from now, “we saw this coming in 2026.” There’s another kind explaining to their investment committee why they didn’t.

Being early to solve interconnection isn’t a nice-to-have feature of a deal. It’s the entire trade. Everything else, financing terms, GPU generation, logo slides, is negotiable. Queue position solved ahead of the repricing curve isn’t negotiable, because you can’t retroactively clear a queue you haven’t cleared yet.

So: who already cleared it, and who’s still telling you they will?

Subscribe for the closing piece in this series, or download the market report for the complete pricing timeline.

Subscribe now

Share

The old checklist for an AI infrastructure deal ran: land, power contracts, hardware, financing. Roughly in that order.

The new checklist has to start earlier. Before land. Before hardware. Queue position, verified, not asserted.

Here’s why the old order breaks down. Land, power contracts, and hardware are all things a good lawyer and a good term sheet can paper over, restructure, or insure against. Queue position can’t be papered over. It’s a physical, adjudicated fact about grid capacity, decided by a utility or ISO on a timeline that nobody at the closing table controls, no matter how good the lawyers are.

I’ve sat in diligence calls where queue position came up as a footnote, if it came up at all, buried after twenty minutes on financing structure. That ordering isn’t an accident. It’s what happens when a checklist gets built for a world where interconnection wasn’t the binding constraint, and nobody’s gone back to rewrite it for the world we’re actually in now.

And it’s the variable everyone quietly assumes away. Sponsors assume their project clears because the last one did, or because a developer said so, or because a deck says “Q3 2026 energization” with no footnote explaining what that’s actually conditioned on. Assumed-away risk is still risk. It’s just risk nobody’s put a number on.

2,300-plus gigawatts is stuck in queues today. If your diligence process doesn’t have a specific line item asking exactly where a project sits in that queue, and exactly what’s required to clear it, you don’t have a diligence process for this asset class. You have a diligence process for a different one, one where physical grid capacity wasn’t the scarce resource, and you’re applying it to a market where it is.

Subscribe for piece five, or download the market report for the full diligence framework.

Subscribe now

Share

14% of queued capacity in U.S. interconnection queues historically reaches commercial operation. That means 86% doesn’t. Not “gets delayed.” Doesn’t happen. Ever.

And that failure rate isn’t sitting quietly on the sidelines anymore, it’s getting priced directly into the survivors. PJM’s capacity auction moved from $28.92 per megawatt-day to $329.17 in a single cycle, and that jump is what 86% of queued capacity quietly withdrawing looks like once a market starts pricing scarcity correctly.

Here’s the part that should change how you read your own portfolio. The projects that survive interconnection aren’t just lucky. They’re capturing value that the 86% who didn’t survive would have competed away if they’d made it to the finish line. Every dead project in the queue is, in a strange way, a subsidy to the ones that clear.

If you’re allocating capital into this sector right now and you can’t tell me what percentage of your portfolio’s underlying capacity has actually cleared interconnection, as opposed to being “on track” or “in process,” you don’t know what you own. You know what a sponsor told you that you own, which is a different thing entirely, and the gap between those two things is exactly 86 points wide.

Your LPs are going to ask about AI infrastructure exposure this quarter, this year, or next. “It’s in the fund” stops being an answer the moment the LP has read a piece like this one. “It cleared interconnection” is the only answer left standing.

Subscribe for piece four, or download the full market report for the complete auction-pricing data.

Subscribe now

Share

Open any diligence deck in this sector and count the pages. GPU generation, financing terms, maybe a logo slide near the back. Now count how many pages are dedicated to the one variable that predicts whether the project exists at all.

Usually zero.

Interconnection.

Grid interconnection in the U.S. averages 60-plus months. In parts of Europe, industrial-scale connections run three to seven years. Meanwhile GPU performance doubles every 18 to 24 months. Do the arithmetic: by the time a traditional facility clears the queue and energizes, the compute it was sized for is two full generations obsolete. That’s the best case. The worse case is it never clears at all.

You can have the tightest financing terms in the market, a term sheet your lawyers are proud of, and still be underwriting a project that never generates a watt. The financing structure doesn’t know that. The GPU spec sheet doesn’t know that. Neither one carries any information about whether the project clears the one gate that actually matters.

Queue position does. It’s the one line item most diligence checklists have never had a row for, because it’s not a number a sponsor controls, and sponsors don’t love putting uncontrollable variables in the middle of their own pitch.

That’s exactly why you have to ask for it before you ask for anything else. Not because it’s comfortable. Because it’s the only number in the deck that isn’t self-reported.

Subscribe for piece three, or download the market report for the full risk breakdown.

Subscribe now

Share

Everyone in this market is underwriting GPUs. Nobody’s underwriting the queue.

Here’s the number that should be on the wall of every allocator looking at AI infrastructure right now: 2,300 to 2,600 gigawatts of generation capacity is sitting in U.S. interconnection queues today. Historically, 14% of that ever reaches commercial operation.

Sit with that. Not “14% is delayed.” 14% is the ceiling. The other 86% doesn’t slip a quarter and recover. It dies, usually years after the capital behind it already committed.

So when a deck lands on your desk with chip allocations, power-purchase letters, a rendered site plan, you’re looking at the wrong asset. The GPUs are a commodity. Anyone with capital and a relationship with NVIDIA can buy silicon. What nobody can buy on the open market is a cleared queue position, because that’s adjudicated by a utility or an ISO on a timeline nobody writing the check controls.

Compute capacity that hasn’t cleared interconnection isn’t an asset. It’s an option on an asset. And most options in this market expire worthless.

Here’s where it gets interesting for pricing. PJM’s capacity auction cleared at $28.92 per megawatt-day for the 2024-25 delivery year. One auction cycle later, for 2026-27, it cleared at $329.17. Eleven times higher. That’s not inflation. That’s a market waking up to how much of its queued supply was never going to show up, and repricing the supply that did.

The framing that’s about to change: “AI infrastructure exposure” as a category. It’s not one thing. It’s a spectrum from “chips on paper” to “watts on the grid,” and almost all the actual value sits at the far end of that spectrum, in a place most diligence decks don’t even ask about.

If you’re allocating into this space and your process doesn’t have a specific line asking where a project sits in the queue, right now, you’re not underwriting AI infrastructure. You’re underwriting a hope.

The queue is the product. The chips are the packaging.

Subscribe for the next piece in this series, or download the full market report for the complete grid-queue breakdown.

Subscribe now

Share

If it existed.

Last week somebody forwarded me a write-up of ARIA — “Artificial Recursive Intelligence Architecture,” a 238-million-edge reasoning core that allegedly runs on commodity CPUs at roughly $2,000 a month, hallucinates at under 0.0001%, and routes attention across 4,751 cognitive agents mapped to the wiring diagram of a fruit fly’s brain. The piece called it “the superintelligence Silicon Valley is desperately trying to ignore.” It claimed 1,200+ production deployments across banking, aerospace, and Fortune 500 environments.

I wanted it to be real. I’ve spent the better part of a year arguing that the industry is overpaying for intelligence, and here was a document arguing the same thing, louder. So I did the thing I do before anything goes out under my name: I went looking for the primary source.

Here is what I found. And here is the crack the ARIA story is pointing at, which is real, even though ARIA almost certainly is not.

Disclosure up front, because it matters for what follows: I build and operate sovereign AI compute infrastructure. I have a direct commercial interest in where the “cheaper intelligence” argument lands. Read me with that in mind, and check my links.

What the source actually is

The entire factual basis for ARIA is a single self-deposited preprint on Zenodo — record 20590584, credited to a Nicola Vettorato of “Counsellor Service Holding LTD.” Zenodo is a legitimate, CERN-operated open repository. It is also a place where anyone can upload anything. There is no peer review. There is no gatekeeper. The “citable source” the article congratulates itself on having is a document that cites only itself.

So I searched for independent corroboration. Anyone replicating the benchmark. A repository. A named client. A single hostile engineer poking at the <0.0001% hallucination claim. There are at least six unrelated projects called “ARIA” in AI research right now — a data-analysis framework out of Tsinghua, a math auto-formalization agent, an augmented-reality museum tool. The $2,000-superintelligence ARIA appears in exactly two places on the entire internet: its own Zenodo file, and the SEO article written to promote it. That article, by the way, ships with a “Meta Title” and “Meta Description” baked into the top of the document. It was built to rank, not to inform.

Then there’s the architecture itself. The claim is that you can map an “Economic Attention Network” onto the Drosophila melanogaster connectome and get verifiable reasoning out the other end. The fly connectome is real, beautiful science — FlyWire mapped roughly 140,000 neurons and published it in Nature in 2024. Bolting a commercial reasoning engine onto that name is the oldest move in the book: borrow the credibility of a verified result to launder an unverifiable one. The “No comment” when a reporter asked Vettorato to name one of the 1,200 deployments isn’t intrigue. When you ask a company for one verifiable customer and the answer is silence, the silence is the finding.

A hallucination rate of “structurally impossible.” A trillion-dollar problem solved by one man and a fruit fly. No code, no benchmark, no client. I’ve seen this pattern in maybe thirty pitch decks over twenty years. It has a name, and the name is not “superintelligence.”

So: ARIA goes in the bin. But don’t throw out the question it’s exploiting, because the question is the most important one in this industry.

The crack is real, and it isn’t in the model layer

ARIA is selling a true anxiety. The cost of intelligence at the model layer is collapsing, and everyone who built a business on the opposite assumption is quietly terrified.

The numbers are not in dispute, and they don’t come from a Zenodo PDF. Stanford’s 2025 AI Index found that the inference cost of a GPT-3.5-level system fell more than 280-fold between November 2022 and October 2024. Open-weight models closed the gap to closed models from 8% to 1.7% on some benchmarks in a single year. NVIDIA’s own analysis cites the same Index: hardware cost per unit down about 30% a year, energy efficiency up about 40% a year. Independent trackers put comparable-capability inference at roughly a 10x annual decline — from about $0.06 per thousand tokens in early 2025 to about $0.006 by mid-2026.

The model layer is commoditizing from below. That part of ARIA’s pitch is correct.

What ARIA gets wrong — what the entire “clever architecture beats the GPU monopoly” genre gets wrong — is where the escape hatch is. They think the moat you break is the model. It isn’t. The model is the thing dissolving on its own. A 13-billion-parameter model now hits 95% of GPT-3’s MMLU score. DeepSeek undercut incumbent pricing by roughly 90% and the sky did not fall on NVIDIA; Blackwell allocation is still constrained. Cheaper serving doesn’t shrink compute demand. It expands the set of products worth building, which pulls more compute, not less. That’s the part the CPU-miracle crowd never models.

The thing that stays scarce when intelligence gets cheap is not a better algorithm. It’s a megawatt you can actually plug into, and an operator who can turn it into revenue before the hardware goes obsolete.

Where the real bottleneck lives

Here is the number that should be on the wall of every AI infrastructure investor, and it has nothing to do with a fruit fly.

As of the end of 2024, roughly 2,290 gigawatts of generation and storage capacity sat stuck in U.S. interconnection queues — close to twice the entire installed U.S. generating fleet — per Lawrence Berkeley National Laboratory’s Queued Up report. The median project reaching commercial operation in 2024 spent about 55 months in the queue. Historically, only around 13% of the capacity that entered the queue from 2000 to 2019 had been built by the end of 2024. Most of it withdraws.

Now hold that against the clock the technology runs on. GPU performance roughly doubles every 18 to 24 months. A traditional data center takes 48 to 72 months to build. Grid interconnection averages five years. By the time the centralized megacampus is energized, the silicon it was poured for is two generations stale. PJM’s capacity auction told the same story in price: clearing rates jumped from $28.92/MW-day in 2024–25 to $329.17/MW-day for 2026–27. That is not a market signal. That is a market screaming.

This is the crack. Not the model. The interconnect. The mismatch between how fast compute demand compounds and how slowly grid-dependent infrastructure gets built is structural, and it does not get solved by building the old thing faster, or by a CPU graph engine nobody can reproduce.

It gets solved by moving the compute to where the power already is.

The answer has scars, not a preprint

This is where I disclose my hand fully, because the rest is what I actually do for a living.

Moving a megawatt-hour roughly a thousand miles costs about $41.50. Moving the data costs essentially nothing. So you stop dragging energy to compute and you drag compute to energy: retired power-plant campuses, industrial brownfields, stranded generation, substations with spare headroom. You manufacture the intelligence factory in a shop, truck it to the power, and run a live workload in roughly eight months instead of five years.

Our first unit is one ~1 MW modular intelligence factory: 392 GPUs, around 450 kW of IT load, closed-loop liquid cooling at a design PUE at or below 1.12, co-located at a power-plant site in Europe. Under contract, in construction, go live targeted for Q4/2026. One ~$45M asset. Four people on the operating side.

That last number is the part the ARIA story and the hyperscaler story both miss from opposite ends. ARIA fantasizes that you can make the compute almost free. The real lever is making the operator almost free — software agents carrying reconciliation, reporting, capacity pricing, network operations, commissioning paperwork, while humans hold the four things machines must never hold: money movement, signatures, client relationships, physical witness. Every machine action lands in an approval queue. Nothing gets sent, signed, or wired without a person. The agent that can’t beat 20% of a loaded human’s cost within ninety days gets retired. That’s not a slogan; it’s a kill rule we run against ourselves.

I’ll own the obvious objection: this is a hard thing to do, and most of who tries it will get the human-interlock part wrong and either move too slow or hand a machine a checkbook. The discipline is the product. The infrastructure is buildable. The operating model is where companies will actually die.

The tell that separates the two stories

ARIA published a hallucination rate and hid the engine. We publish the operating metrics and hide nothing that matters: fleet revenue per orchestrator, agents per human, human-touch minutes per closed deal, cost per deliverable against the human alternative. When a number and a derivation disagree, the derivation wins. When we miss, the miss gets more words than the win.

That’s the difference between a claim and a build. One asks you to trust a fruit fly. The other shows you the queue data, the auction prints, the contract date, and says: check it yourself.

The cheaper intelligence gets, the more the advantage migrates away from whoever owns the cleverest model and toward whoever owns the megawatt and can operate it without a thousand people. ARIA is right that the old assumption is dying. It’s just selling you the wrong funeral.

The scarce thing in 2027 won’t be intelligence. It’ll be a trustworthy operator standing next to a power plant.
_______________________________________________________________________

^{Sources: LBNL Queued Up (2025 edition, data through 2024); Stanford HAI 2025 AI Index; NVIDIA inference economics; FlyWire connectome; the ARIA Zenodo record — judge it for yourself.}

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

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

Healthcare: the loud, clear win

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

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

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

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

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

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

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

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

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

Closing the scorecard

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

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

____________________________________________

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

First, the chart everyone is misreading.

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

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

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

The bubble question is the wrong question

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

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

The bubble was real.

The value was realer.

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

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

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

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

The founders went quiet and started typing

Now the part that should actually worry the bubble callers.

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

They started coding again.

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

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

Why?

Because the abstraction penalty flipped.

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

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

Follow the money into the ground

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

Concrete.

Transformers.

Cooling loops.

Twenty-year assets.

Speculators buy what they can sell by Friday.

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

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

The wealth landscape, 2031

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

It will close the same way.

Quietly.

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

Share

The uncomfortable close

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

Same test, new era.

Yes, parts of the AI trade will burn.

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

The skeptics are writing think-pieces.

The believers are writing code.

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