There’s a comfortable assumption running the entire AI trade: whoever owns the biggest pile of GPUs wins. That single belief is inflating a three-quarters-of-a-trillion-dollar-a-year bubble — and two Chinese labs just gave the whole thing away for the price of a coffee.
Start with the belief, because it’s the load-bearing wall. Every mega-cluster, every gas peaker fired up to feed it, every stretched depreciation schedule rests on the premise that intelligence is scarce — that you can hoard the compute, dig a moat around the model, and charge rent on the frontier forever. Get there first with the most silicon and you own the future.
That is exactly the mentality driving the bubble. And a bubble isn’t a lie — it’s a true story with the wrong price on it. The models work, the demand is real; what’s mispriced is the scarcity. Because in the first half of 2026, China quietly proved the whole premise false. They didn’t out-spend anyone. They just shipped weights.
Here’s the trouble with reading it clearly: almost everyone with a microphone in this debate has a position to talk. The chipmaker has silicon to move. The hyperscaler has a capex bet to defend. The lab has a model to pump and a valuation to protect. The permanence story is load-bearing for all of them, so of course they keep telling it. Set that aside for a second — no book to talk, nothing to sell you — and just read the tape.
The commodity nobody wanted to call a commodity
On June 13, Z.ai — the lab formerly known as Zhipu — dropped GLM-5.2, a 744-billion-parameter open-weight model under a plain MIT license. Within days it topped the open-weight intelligence rankings and landed fourth overall, roughly five points under Anthropic’s Fable 5 on the industry’s headline index. Not “impressive for open source.” Impressive, full stop. You can download it. You can run it. Nobody can revoke it.
DeepSeek had already knocked out the price floor two months earlier — V4 Flash lists at fourteen cents per million tokens in, twenty-eight out. One shop swapped its production agents off a closed frontier model onto GLM and watched its inference bill fall seventy-two percent for the same outputs. That’s not a discount. That’s a repricing of the entire category.
- 744B — GLM-5.2 params · MIT license · free to run
- $0.14 — DeepSeek V4 Flash · per 1M input tokens
- −72% — inference cost, closed frontier → GLM swap
The part the capex crowd keeps skipping: the frontier stopped being a club. When the best downloadable model on Earth is standing inside the frontier instead of chasing it, “own the model” stops being a business. The model is the giveaway. The moat is supposed to be somewhere else — and increasingly, nobody can find it.
Outdated in six months, financed for six years
Flip to the balance sheet, because that’s where the real story lives. The biggest hyperscalers are on track to deploy somewhere between $765 billion (Goldman’s count) and $800 billion (BofA’s) in capital this year, most of it AI-specific; Goldman’s model runs the cumulative build past $7.6 trillion by 2031. Against that, the AI revenue anyone will actually disclose is an order of magnitude smaller — Microsoft, the most forthcoming of the bunch, claims a roughly $37 billion annual run-rate, and it’s the only one bragging. Free cash flow at the top five spenders just went negative for the first time in about thirty-five years. The Fed added AI to its list of top systemic risks. Those aren’t my adjectives; they’re the footnotes.
A GPU loses 64% of its rental value in eighteen months. You’re financing it over five to six years.
That mismatch is the whole game. Hyperscalers book these chips over five to six years. Michael Burry — early and right in 2008 — argues the real economic life is closer to two or three, and that the industry is understating depreciation by roughly $176 billion across 2026–2028. Stretch the schedule and today’s earnings look clean; mark it to reality and the payback math gets ugly fast. Microsoft’s own CEO waved the flag without meaning to, telling investors the newest models run fine on the older fleet. Read that slowly: management is telling you the shiny new silicon isn’t the edge they sold.
And the exit ramp is already paved. GLM-5.2 squeezes onto a single high-end consumer card at aggressive quantization; the smaller open models — Qwen, Gemma — run on one GPU, or a phone. The frontier is walking out of the datacenter and onto a desk. I’m running near-frontier weights on a workstation in a spare room. That’s not a flex; it’s the thesis. When the model is free, the weights fit on hardware you already own, and the price of a token is rounding toward zero, what exactly is the trillion-dollar moat protecting?
The ugly kicker isn’t financial — it’s physical. We’re strip-mining the grid for this: water for cooling, gas peakers firing to hold load, transmission queues jammed, whole counties handing over power that could’ve gone to housing. All of it justified by permanence. But you don’t build a cathedral on a two-year clock. The cleanest read of the tape is that a large share of this buildout gets stranded — not as fiber that quietly becomes the next decade’s backbone, but as depreciating racks of obsolete silicon in gas-guzzling sheds, made redundant by a fourteen-cent API and a machine on somebody’s desk. We’re burning rivers to build the Blockbuster of compute.
Even the house is pointing at the exit
Here’s the tell that ties a bow on it. On June 29, Palantir and NVIDIA — not two open-source idealists, but the defense-data giant and the company selling every shovel in the gold rush — launched an engine to run NVIDIA’s Nemotron open-weight models inside sovereign, on-premise, air-gapped environments. Agencies and critical-infrastructure operators get to train on their own data, keep full ownership of the weights, and never let a token leave the building. Karp’s pitch, stripped of the flag-waving: don’t bake your secrets into somebody else’s closed model. Huang’s: open, controllable AI is a national-security requirement. Translation from both: the future is a model you own and run behind your own walls — not one you rent by the token from a hyperscaler.
Now read it like a skeptic, because that’s the job. NVIDIA doesn’t care whether you rent the frontier or own it — it sells the Blackwell either way. Preaching “local” and “sovereign” and “open” isn’t a change of heart; it’s demand engineering. Capex growth at the top decelerates hard from here — roughly 51% this year, ~13% next, low single digits after — and Jensen needs a fresh buyer. “Every federal agency and utility runs its own on-prem cluster” is one hell of a backstop. That’s the single crack in my own thesis: the shovel-seller is hedged. Fade the players renting out centralized closed-model capacity. Do not blind-short the company that wins whether the compute lives in a hyperscaler shed or a SCIF.
Beijing gives the model away. Washington’s own vendors sell you a box to run it in. Nobody’s renting you the frontier anymore.
That’s the whole signal in one line. It isn’t two stories — cheap Chinese weights over here, sovereign American deployment over there. It’s one story arriving from both directions at once: the “centralized, scarce, trust-us-it’s-ours” intelligence trade doesn’t have a friend left in the room. Not in Hangzhou. Not in D.C. Not on my desk.
The margin illusion
Wall Street keeps trying not to say the quiet part: AI doesn’t have software economics. Real software costs a fortune to build and almost nothing to serve — write it once, ship the ten-millionth copy for free. That’s where fat margins live. AI is the opposite. Every query burns compute. Cost of goods scales with revenue, not against it — serve more, spend more, roughly in a straight line. Cheaper per-token pricing hasn’t saved anyone either; usage just explodes to swallow the savings. Enterprises that budgeted at 2024 token rates are getting agentic-workflow bills that are multiples of the spreadsheet. Unit economics that looked fine in the pilot stop working at adoption.
And the hardware can’t rescue the margin, because the hardware is the margin problem. Each generation costs more than the one it replaces — Hopper to Blackwell to Rubin, every step more silicon, more power, more dollars per rack, priced accordingly. So the cost floor climbs with every launch while the sellable price of a token falls toward zero, dragged there by a free open-weight model that does 90% of the work. Rising cost per generation, collapsing price per token: no vendor on that curve can ever widen its margin — not Rubin, not the generation after it. You cannot out-silicon a problem where the newest silicon is the most expensive input and the output is a commodity. The token-expenditure indices already show it: aggregate spend keeps rising, but it’s migrating hard toward the cheap models. The money flows to whoever serves intelligence for the least, not whoever builds the most. That is the opposite of a moat.
Which is why the marquee names bleed. OpenAI is projected to lose around $14 billion in 2026 on roughly $25 billion of revenue, with cumulative burn near $115 billion through 2029 and no cash-flow profit before the end of the decade. Nine hundred million weekly free users torching GPU time isn’t a business; it’s a customer-acquisition bonfire funded by private capital. Every frontier lab has priced inference below cost to buy share — a Turing-winning researcher put inference cost forward as the single thing blocking profitability.
But the lazy version of this take is wrong, and the correction is sharper than the take: it is not true that no closed-frontier model can make money. Anthropic is the counterexample. SemiAnalysis estimates it crossing $1 billion in operating profit by Q3 2026, gross margin swinging from about −94% in 2024 to roughly 60% now — driven by inference efficiency, not price hikes, on a book that’s ~75–85% enterprise API. The lesson isn’t “AI can’t profit.” It’s that the consumer-subsidy model — rent us your attention, we’ll eat the compute — has never made a dime and structurally may never, while the enterprise-API model just flipped positive. One of those is a business. The other is a valuation waiting for an IPO to test it.
Then stack the financing on top, because this is where it goes circular — literally. NVIDIA doesn’t just sell shovels; it funds the diggers. It has put money into OpenAI, CoreWeave and Nebius — outfits that turn around and spend it on NVIDIA GPUs. CoreWeave’s own IPO filing lists NVIDIA as both a top customer and a shareholder. A Mizuho analyst called it what it looks like: pre-funding the purchase of your own chips. By 2026 analysts tag more than $800 billion of interlocking commitments across the chain — chip maker to AI lab to cloud and back, the same dollars booked as revenue on multiple sides. Some of NVIDIA’s demand is NVIDIA’s own money making the round trip.
Fair is fair: Jensen calls the concern “ridiculous,” and he’s not entirely wrong — NVIDIA’s ~$2 billion into CoreWeave is only about 6% of CoreWeave’s single-year capex, so it’s not pure self-dealing. But “not pure” isn’t “clean.” And the reason you can’t size it precisely is the quiet scandal underneath all of it: nobody is disclosing the real AI numbers. OpenAI didn’t even hand over Q2 projections. Hyperscalers won’t cleanly split AI revenue from cloud. And the one figure that would settle the whole bubble debate — utilization, how full these clusters actually run — is disclosed by no one. When the house won’t show you the cards, assume the hand is worse than the betting.
China distills the check you’re writing
Look at the intelligence leaderboard and two things are true at once. America still reigns at the very top — the three highest-scoring models are all U.S. closed systems. But drop one rung to everything you can actually download and the board goes red: GLM-5.2 at #1 among open weights, then MiniMax, DeepSeek, Kimi, Qwen — a wall of Chinese labs, with NVIDIA’s Nemotron the lone American open entry hanging on. Hugging Face now says Chinese models have passed U.S. models in downloads outright, roughly 41% of the platform’s pulls over the past year. The measured capability gap at the top is down to a few percent — while the Chinese models run at roughly a quarter of the price.
And here’s the part that should sting. The U.S. is outspending China by orders of magnitude — three-quarters of a trillion dollars of capex against a DeepSeek V3 training run the lab claims cost around $5.6 million — and a chunk of that American spend is effectively subsidizing the competition. The mechanism is distillation: use a strong “teacher” model’s outputs to train a cheaper “student” that mimics it, skipping most of the cost of building the teacher. DeepSeek did it in the open with its own R1. The accusation from the U.S. labs is that the Chinese labs have been doing it to us.
I’ll flag my own bias, since MarketCrystal runs on Claude and Anthropic is a party to this fight: treat what follows as allegation, not settled fact. In February, Anthropic accused DeepSeek, Moonshot and MiniMax of mining its model at industrial scale — some 16 million Claude exchanges through roughly 24,000 fake accounts — to accelerate their own training. In April the White House science office went further, naming it outright theft of American AI IP; OpenAI made the same charge against DeepSeek. The counter-nuance is real: distillation through an API doesn’t transfer a frontier model’s deepest reasoning — you get most of the behavior, not all of the soul. But directionally the free-ride argument holds. You paid to draw the map. They photographed it.
Which sets up the question underneath the whole capex bill: most people don’t need the frontier at all. The vast majority of real work — the tool calls, the retrieval, the boilerplate, the summarization — is served fine by a good-enough open model routed intelligently. Keep the closed frontier for the genuinely hard problems; let the cheap student handle the other 90%. If that’s how the market actually consumes intelligence — and the routing data says it is — then the trillion-dollar bet on ever-scarcer, ever-pricier frontier capacity is aimed at a sliver of demand, not the mass of it.
Sidebar · Meanwhile, on the factory floor
While the capex machine argues about token prices, the ground truth on whether AI even delivers is getting messy.
Ask Ford. It thinned its ranks and leaned on AI-driven quality inspection — then quietly rehired and promoted around 350 “graybeard” engineers after the automated systems missed defects at an unacceptable rate. The returning humans reprogrammed the very systems brought in to replace them, and Ford went on to top J.D. Power’s initial-quality study for the first time in sixteen years. CEO Jim Farley — who’d just declared AI would “replace literally half of all white-collar workers” — credited the reversal with hundreds of millions in lower warranty and recall spend. Not an isolated mea culpa: Klarna replaced 700 support agents with an AI assistant, watched quality fall, and started hiring humans back; IBM moved to triple entry-level hiring in roles it had forecast as AI-replaceable.
And the productivity gains everyone assumes? The most rigorous test we have — a randomized controlled trial from METR — found experienced developers were about 19% slower using AI tools while believing they were 20% faster: a thirty-nine-point gap between what they felt and what the clock said. (Caveat: early-2025 models, small sample, and METR itself later softened the causal claim.) But the durable finding survives every caveat — we are unreliable narrators of our own AI productivity. Which lines up with what every honest heavy user eventually admits: half the time, babysitting the model through a task burns more clock and more tokens than just doing the thing yourself. Feels fast. Isn’t.
Trend Readout · The AI Capex Trade
────────────────────────────────────────────────
TREND TOPPING · score −58
MOMENTUM Weakening — margin story cracking
COST CURVE Collapsing — inference −98% since '22
UNIT ECON Costs scale w/ revenue — no software margin
MARGIN Can't widen — each gen (Rubin) costs more, token $→0
DEMAND Shifting to cheap models — spend chases lowest cost
PROFIT Consumer-subsidy: never · Enterprise-API: just flipped
DEPRECIATION Understated — book 5–6yr / real 2–3yr
FINANCING Circular — >$800B loops, NVDA on both sides
DISCLOSURE Opaque — no one publishes utilization
CATALYST Open weights (GLM, DeepSeek) at commodity price
THE PIN A hyperscaler cutting AI capex guidance (not yet)
POSITIONING Incumbents pivoting to owned/sovereign · PLTR×NVDA
HEDGE RISK NVDA wins rent OR own — not a clean short
VOLATILITY Expanding — single-name concentration
────────────────────────────────────────────────Mark’s Take. China didn’t predict the AI bubble. It’s deflating it — in real time, in public, for free. The demand is real and the technology is real, which is exactly why this is dangerous: it lets everyone pretend the price is real too. It isn’t. The value is migrating off the infrastructure layer, where the moat was supposed to be, and onto the model layer, where it’s being given away — while the labs that bet on renting scarce intelligence to nine hundred million free users keep proving they can’t make the math close.
I don’t forecast. I read what’s in front of me: a cost curve in freefall, a depreciation schedule stretched to hide it, a financing loop where the biggest winner quietly funds its own demand, the best models on the planet shipping under a license that says help yourself — and now Palantir and NVIDIA racing to build you a private box to run them in. When Beijing and Washington’s own vendors agree the future is owned, not rented, the rental trade is finished arguing.
But watch what calls the pop, because it hasn’t printed yet: the day a hyperscaler cuts its AI capex guidance. The whole edifice is held up by the promise of ever-more spend — $765–800 billion this year, more next. The moment one of the big five looks at its negative free cash flow, its stretched depreciation and its utilization numbers and says we’re pulling back, the permanence story is over — because the permanence story was the spend. Growth is already decelerating on schedule (51% this year, ~13% next, low single digits after); deceleration is the setup, an outright cut is the pin. That’s the number to watch — not the token price, and not a Fed meeting.
SIGNAL · FADE THE PERMANENCE TRADE
Own the workflow, the data loop, the thing that gets smarter as it runs. Own compute only if you rent it back at a spread. Do not own the commoditizing brick. The first pin was already a file on a server in Hangzhou that anybody could download. The second is a line in an earnings call. Neither one is priced in.
And now the part we buried up top. Everyone loud in this debate is selling you something — a chip, a cloud contract, a model subscription, a story that keeps their valuation intact. That’s the whole reason a read like this one is worth the paper: MarketCrystal isn’t in the hype trade. We don’t own a GPU cluster to defend, we don’t take the vendors’ ad money, and we don’t get paid whether the permanence story holds or folds. We price the thing in the real unit and tell you what the tape says, full stop. When the people with everything to sell all agree the number is real, the one voice worth hearing is the one with nothing to sell you. That’s the whole business — and it’s the only kind of business that can afford to be honest with you about a bubble.
Not financial advice. MarketCrystal provides trend analysis for informational purposes only. Equities and crypto are volatile and you can lose money. Figures are drawn from public 2026 reporting and vendor disclosures and were current at publication; distillation and IP claims are allegations by the named parties, not adjudicated facts; markets and model rankings move fast, so verify before you act. Always do your own research. Past trends do not guarantee future results. Mark reads the tape — he does not predict it.