The Tell: What “AI Agents” Are Quietly Admitting
A field investigation into the four assumptions holding up the AI trade, and why they are cracking at the same time.
The market is pricing AI as intelligence. The architecture the industry is actually shipping tells a different story. The bull narrative rests on four load-bearing assumptions, and all four are weakening at once. Not in sequence. At the same time.
Disclosure and vantage point: I consult for water and power utilities, I have worked in machine learning for over a decade, and I am currently building two AI systems. This is not a piece written from the cheap seats. It is written from inside the rooms where the power gets provisioned, the tokens get billed, and the models get shipped. Treat everything below as analysis, not investment advice.
The Narrative the Market Is Pricing
There is a version of the AI story that the market has been pricing for two years. It goes like this: scale the models, scale the data, scale the compute, and capability keeps compounding until the productivity gains pay for the build-out many times over. Buy the picks and shovels, buy the hyperscalers, buy anything with “AI” in the deck.
I want to take that story seriously, because most of the people telling it are not stupid and some of the demand underneath it is real. But after spending the last several months with one foot in utility infrastructure and one foot in actual model development, I think the bull narrative rests on four load-bearing assumptions, and all four are weakening at once.
The four assumptions are simple to state:
- The power and water are available.
- There is enough new data to keep training on.
- The unit economics eventually close.
- The output is reliable enough to deploy where the money is.
Crack One: The Physical Bill Is Coming Due
Start with the part I see firsthand. The demand for data center infrastructure in the United States right now is not aggressive. It is deranged.
The demand for data center infrastructure in the United States right now is not aggressive. It is deranged.
The numbers from the outside confirm what the inside feels like. A January 2026 Bloom Energy report projects U.S. data center power demand nearly doubling between 2025 and 2028, from roughly 80 gigawatts to roughly 150 gigawatts. That is the equivalent of bolting a mid-sized country’s worth of electricity demand onto the grid in three years. Consumer Reports counts 3,069 data centers already operating in the country with another 1,489 planned or under construction. The IEA and Brookings put global data center consumption around 415 terawatt-hours in 2024, roughly 1.5 percent of all electricity on earth, and project that figure climbing toward 945 to 1,050 terawatt-hours later this decade.
Here is the part the infrastructure decks gloss over. This load does not arrive evenly. It clusters. Roughly fifteen states, led by Virginia and Texas, accounted for about 80 percent of U.S. data center demand in the recent baseline. When that much constant, uninterruptible load lands in one region, somebody pays for the grid upgrades, and it is not the hyperscaler. Dominion in Virginia proposed its first base-rate increase since 1992. Residential electricity prices rose about 7 percent in 2025. There are documented cases of household bills tripling near major clusters, and analysts have clocked wholesale prices near some data center hubs up by triple digits versus five years ago. In 2026, lawmakers in more than thirty states introduced over three hundred bills aimed at data centers: moratoriums, tax fights, cost-allocation rules.
Then there is water. The cooling has to go somewhere, and a single large facility can consume as much water as tens of thousands of households. I have sat in the meetings where a utility realizes a proposed campus would draw more water than the town it is being built next to. These are not abstractions. They are permitting fights, rate cases, and ballot-box backlash that will throttle how fast capacity can actually come online, regardless of how much capital wants to build it.
The bull response is that efficiency and on-site generation solve this. Maybe, eventually. But the grid interconnection queue, the transformer backlog, and the political reality of voters watching their bills climb are physical constraints, not spreadsheet constraints.
You cannot scale your way out of a permitting timeline.
Crack Two: The Data Ran Out, and the Fix May Be Poison
The second assumption is that there is enough fresh, high-quality data to keep feeding the models. There is not.
Epoch AI’s work, which has held up reasonably well, estimates that the supply of usable human-generated text suitable for training runs out somewhere in the window of roughly 2026 to 2032. We are now inside the front end of that window. The labs know it. That is why you have seen the wave of licensing deals: Reddit to Google, News Corp to OpenAI. When companies start paying hundreds of millions for access to human writing, that is not a sign of abundance. That is the industry quietly conceding the well is running dry.
The proposed escape hatch is synthetic data, meaning data generated by models to train other models. And here is where my machine learning background makes me nervous, because the failure mode is well documented. The Shumailov work published in Nature in 2024 showed that models trained recursively on their own output suffer what researchers call model collapse. The tails of the distribution vanish first. The model forgets the rare and the unusual, drifts toward the average, and gets blander and less reliable with each generation. Other researchers nicknamed an analogous effect in image models “model autophagy disorder,” a deliberate nod to mad cow disease. The metaphor is apt. You are feeding the system its own rendered remains.
Synthetic data is not a bridge. In the domains where verification is hard, it is a feedback loop with a downward slope.
It gets worse, because the open web is no longer clean. By early-to-mid 2025, a large majority of newly created web pages already contained some AI-generated text. So even the labs that want to train on “real” human data are increasingly crawling a corpus contaminated by machine output. The training set is eating itself whether the labs intend it to or not.
The honest counterpoint: synthetic data works fine in narrow domains where answers are verifiable. Math, code, games. AlphaGo Zero trained on self-play and got superhuman. Multimodal data, meaning video and audio, could expand the supply for a while. Verifier models might extend the trick further. All true. But none of that solves the core problem for the open-ended, hard-to-verify reasoning that the trillion-dollar valuations are actually betting on.
Crack Three: The Unit Economics Do Not Close
The third assumption is that the economics eventually work, that today’s losses are a land grab and profitability is a switch you flip later. Look closely at the token math and that gets harder to believe.
The bull case points out, correctly, that the price per token has collapsed. By one widely cited figure, the average cost per million tokens across major providers fell from roughly ten dollars to roughly two and a half dollars in a single year. Nvidia’s next-generation Rubin platform targets another order-of-magnitude cut in inference cost. If you stop reading there, costs are falling and the problem solves itself.
But that is only half the equation, and it is the half that flatters the story. Consumption is exploding faster than price is falling. Sam Altman has said publicly that token budgeting flipped almost overnight from something nobody thought about to a genuine problem, with OpenAI’s heaviest single user now burning on the order of 100 billion tokens a month. The reason is the very thing being sold as the future: agentic workflows that fire off ten to twenty model calls per task and run continuously instead of once per prompt.
The enterprise receipts tell the story. Uber reportedly burned through its entire 2026 AI budget in four months as internal coding-tool adoption jumped from roughly a third to over four-fifths of its engineers, with monthly per-engineer costs running into four figures. That is not an efficiency story. That is a budget detonating.
This is not infinite intelligence. It is constrained computation, and someone is paying more than a dollar for every dollar it earns.
And the providers themselves are still underwater. By multiple accounts, OpenAI in 2025 generated around 3.7 billion dollars in revenue against roughly 5 billion in losses, spending on the order of 1.35 dollars for every dollar it earned, with 2026 losses projected far higher. Deutsche Bank has pegged the inference math at over two dollars spent per dollar of revenue. When every frontier lab is pricing inference below cost to grab share, you get an artificial floor under prices that has nothing to do with sustainable economics. The question is not whether costs fall. They do. The question is whether they fall faster than usage rises, and whether the below-cost pricing survives contact with public-market scrutiny once the IPO queue forces real disclosure. So far, the answer to both is unproven.
Crack Four: The Output Is a Legal Liability Now, Not Just a Quirk
The fourth assumption is the one I think the market is most badly underpricing, and it is the reason I started writing this in the first place.
For years, hallucination has been treated as a UX annoyance. A funny screenshot. Something you tell users to “verify.” That framing just took a serious legal hit. In June 2026, a German court issued a preliminary ruling holding Google liable for false statements its AI Overviews generated about two publishers, statements that effectively branded them as scams. As reported by Ars Technica via The Decoder, this appears to be the first time a court has held an AI company liable for content its own model fabricated.
The reasoning is what matters for investors, not the specific company. The court drew a line between traditional search, which has long been shielded because surfacing third-party links is considered unavoidable, and AI summaries, which it treated as an additional function the user does not actually need. The framing that gave the ruling its headline: nobody needs AI to search the internet. Because the AI layer is optional and the company is the only party that can fix the underlying model, the court found the liability shield does not extend to it. Google’s standard defense, that users know AI output can be wrong and should verify it, got turned against it. The court noted that the tool’s whole value proposition collapses if users have to independently check every line.
A system that confidently asserts false things, attributes them to real named sources, and is trusted by users who do not verify is not a product quirk. It is a liability surface.
Layer on the accuracy data and the exposure becomes obvious. An analysis cited in the same reporting found AI Overviews on a current frontier model were inaccurate roughly 9 percent of the time and included inaccurate source links the majority of the time, while surveys show most users never click through to check. You cannot deploy a probabilistic text generator into medicine, law, finance, or insurance underwriting at scale if every nonzero error rate is now a liability the vendor owns. And that constraint, more than any benchmark, is what reshapes where the value actually accrues.
The Backdrop: Why All Four Cracks Rhyme
Take a step back and the four cracks share a structure. The build-out is being financed in a loop.
By 2026, analysts have identified more than 800 billion dollars in what is now openly called circular financing. The pattern: a chipmaker or cloud provider invests billions into an AI lab, the lab spends that money buying the investor’s chips and capacity, and the resulting “revenue” gets reported as organic demand. Nvidia committed up to 100 billion dollars to OpenAI. AMD structured deals reportedly worth 200 billion with equity warrants attached. Oracle committed 300 billion in cloud. OpenAI’s disclosed infrastructure commitments run past a trillion dollars across roughly seven vendors through 2035. The cash circulates among a handful of interconnected firms.
We have seen this movie. Bloomberg and others have drawn the explicit parallel to late-1990s telecom, when equipment makers used vendor financing and “capacity swaps” to manufacture the appearance of demand. Names like Qwest and Global Crossing ended up in front of congressional investigators. The capacity sat dark for years.
I want to be fair to the other side, because an investigation that only steelmans the bear case is just a rant. UBS argues the circular deals, while large, are not overwhelming relative to the underlying business, estimating the OpenAI-Nvidia arrangement at around 13 percent of Nvidia’s projected 2026 revenue. Economists like Noah Smith argue this looks more like genuine vendor finance, which has bootstrapped real capital-intensive industries before, than like fraudulent round-tripping, partly because the revenue flows one direction and the companies face real disclosure. Jensen Huang has repeatedly called the demand real and the bookings deep. Anthropic is reportedly on track for its first profitable quarter. These are not nothing.
But the market itself is starting to twitch. In early June 2026, the Philadelphia Semiconductor Index dropped about 10 percent in a single session, erasing on the order of 1.4 trillion dollars in value, with Nvidia, Broadcom, and Marvell all taking sharp hits and the selloff spreading to Asian markets. It rebounded, and Huang called it a buying opportunity. Maybe it was. But a sector that can shed over a trillion dollars in a day on a hot jobs print and a geopolitical headline is a sector where the narrative is doing a lot of the structural work.
When the story is the load-bearing wall, you watch the story for cracks. That is the entire premise of how we read tape here.
The Tell: “AI Agents” Is a Confession
Now connect the fourth crack to the industry’s response, because this is the part almost nobody is saying out loud.
A pure large language model generates text probabilistically. It will be wrong some nonzero percentage of the time, by construction, forever. You cannot prompt that away. And as the German ruling makes clear, that error rate is migrating from “annoying” to “actionable.” So how is the industry actually shipping reliable products on top of an unreliable core?
By not trusting the core. The hottest phrase in enterprise AI right now is “AI agents,” and when you look at what an agent architecture actually is under the marketing, it is a deterministic scaffold that uses the language model in narrow, isolated, verifiable steps. Retrieval grounds the model in real documents instead of its own memory. Tool calls hand the actual computation to deterministic code. Structured-output constraints box the response into a validated schema. Validators and human review catch what slips through. The model is demoted from “the system” to “one component inside the system,” surrounded by guardrails whose entire job is to contain the part that makes things up.
“Agent” is, to a meaningful degree, a marketing word for “we wrapped the part that lies in deterministic plumbing.”
In other words, the industry’s flagship architecture for the next phase is an implicit admission that the end-to-end probabilistic approach does not work where reliability matters. This is the direction I am building in myself, and I am not alone. It is where serious deployment is going precisely because it is the only way to get a usable error rate in a regulated context.
That has a direct investment implication, and it is not the one priced into a basket of “AI” tickers. If the durable architecture uses the raw model in isolated cases inside a deterministic structure, then value does not accrue primarily to whoever has the biggest model. It accrues to:
- Orchestration and verification layers, the deterministic plumbing that makes probabilistic models safe to deploy.
- Proprietary, verifiable, domain-specific data, which is exactly what synthetic data cannot replace and exactly what the data wall makes scarcer and more valuable.
- Companies that own a deterministic workflow and use a model as a feature, rather than companies selling raw model access at below cost.
- The picks-and-shovels with real pricing power, as opposed to the ones whose revenue is partly the same dollars going around a circle.
The Mispricing: Narrative vs Reality
Bull Assumption
What the Cracks Reveal
Implication
Power and water scale on demand
Load clusters, grids strain, household bills rise, and states are legislating against it
Capacity is gated by permitting and politics, not capital
Endless fresh training data
High-quality human text is running out; synthetic data risks model collapse
Scaling-by-data is hitting a wall; verifiable proprietary data gains value
Unit economics close as token prices fall
Usage outruns price cuts; providers still spend more than a dollar per dollar earned
Below-cost pricing is unproven against public-market scrutiny
Output is reliable enough to deploy
Hallucination is now a court-tested legal liability, not a quirk
Probabilistic models get boxed into deterministic, supervised systems
The gap between these columns is where the AI premium lives. It is also where the risk sits.
Investor Takeaway
I am not calling a top. Timing a narrative-driven market on fundamentals is a good way to be right and broke. The demand for AI compute is partly real, the efficiency gains are real, and bubbles can inflate for a long time after the thesis curdles. The dot-com builders were directionally correct about the internet and still went to zero on the way to being right.
The trade is not about whether AI works. It is about understanding what it actually is, and pricing it accordingly. The market is assigning an AI premium based on the idea that these systems are autonomous, software-margin intelligence. The architecture the industry is shipping points to something else: layered, metered, supervised infrastructure where the model is one constrained component and everything around it is what makes it usable. The “AI agent” concedes that already.
The market is pricing intelligence. The product is infrastructure with a legal liability attached.
So watch the places where the four cracks show up as numbers rather than vibes, in roughly this order:
- Infrastructure-provider balance sheets. Rising debt, swelling lease commitments, and widening credit spreads are where circular financing strains first.
- The token gap. The distance between falling per-token prices and rising per-task token volume in the labs’ actual unit economics.
- The grid docket. The pace of state-level data center legislation and rate cases, because that is the real governor on how fast capacity comes online.
- The legal docket. The German ruling is one preliminary decision in one jurisdiction today, and a template the moment a US plaintiff’s lawyer reads it.
When that clarity spreads, the AI premium does not vanish. It moves. Away from companies valued purely on the promise of intelligence, and toward the ones that control the systems, the verifiable data, and the economic bottlenecks that make the intelligence usable.
Not investment advice. Analysis only. Do your own work before you size anything.
Narrative Intelligence Platform
Trade Like a Hedge Fund.
Understand Why Markets Move.Market Prism is a narrative forensic engine designed to track speculative cycles, narrative exhaustion, coordination risk, valuation disconnects, and market psychology in real time. Built for investors who want more than headlines.
Launch Market Prism →Read Research
120K+
Narratives scored and archived across market cycles.
170+
Tracked equities monitored daily through narrative physics models.
37K+
Daily market datapoints processed through deterministic systems.
Real-Time
Live narrative analysis built to identify crowd psychology shifts.