Two men described the same market this week, and only one of them was telling on it.
Ali Ghodsi called demand "crazy" as Databricks closed a $5 billion round at $190 billion. Then, almost in the same breath, he explained why: "This token maxing has freaked out the CFOs." Those two statements do not sit comfortably together. Crazy demand is supposed to be a story about hunger. Freaked-out CFOs is a story about pain. Ghodsi is selling both, and most readers only heard the first.
Compress the week and the euphoria looks total. Databricks at $190 billion, up from $134 billion six months earlier, on a $7 billion revenue run rate growing 80% year-over-year. CoreWeave up 24.6% on a $104 billion backlog and 112% revenue growth. Lovable at $13.3 billion for vibe-coding, more than doubling its $6.6 billion mark. Every one of these numbers is a bet on a demand curve.
Here is the figure that appears in none of the pitch decks: the customer running those workloads is paying roughly 2% of what the compute actually costs.
That number comes from a real user, not a model. Jay Schulman pulled his usage and found 9.6 billion tokens a month across 40-plus agents. At published API rates, that workload costs $8,434 a month. He pays $200. A 42x gap.
At published API rates that workload costs $8,434 a month. He pays $200. A 42x gap between price and cost.
And he is not an edge case. He is, in his words, "exactly the customer OpenAI and Anthropic are designing for" — the professional who moved from trying AI to running operations on it.
So when Ghodsi says demand is crazy, ask what that demand is denominated in. It is denominated in tokens priced at a fraction of their cost. Some meaningful slice of the "crazy" is not appetite. It is arbitrage — buying dollars for two cents and building your month-end close on top of them.
This is the part the euphoria skips over, and there is now academic language for why it happens. A game-theoretic paper calls it the Trainer's Dilemma: unlike classic SaaS, where marginal cost rounds to zero, inference in generative AI can consume 50% to 70% of revenue. That single fact turns monetization into a first-order operational problem rather than an afterthought.
The paper's third result is the load-bearing one. Competition induces defensive over-subsidization — platforms push free usage past the point any monopolist would choose, racing to lock in users before rivals poach them. Cheap tokens are not a promotion. They are a hostage situation between competitors, and the customer is the hostage everyone is bribing.
Which reframes the smartest tell of the week. Demand for Databricks' AI Gateway — its tool for controlling model use and costs — rises precisely because token maxing freaked out the CFOs. Ghodsi said so directly: skyrocketing AI costs are boosting demand for the Gateway and for open-source and even Chinese models that the same CFOs dismissed a year ago.
Sit with the mechanism. The Gateway sells because the underlying pricing is unsustainable. Its total addressable market is the anxiety created by the subsidy. Databricks is quietly short the pricing model that its neighbors on the valuation leaderboard are long.
The pure-play inference names are long that model whether they like it or not. CoreWeave's $104 billion backlog is a promise to deliver compute, and that compute costs what it costs. The company carried $35 billion in debt at quarter end, took a $2.6 billion term loan, and posted a $626 million net loss on $2.58 billion of revenue. Its "real" gross margin, after infrastructure expense, was 7.3%.
That debt matters here only as evidence, not as a financing puzzle. It tells you inference has a hard floor of cost that someone must eventually pay in cash. CoreWeave is at least honest about this — Mike Intrator says pricing and margins on Blackwell SKUs are "setting new highs," and finance chief Nitin Agrawal confirmed they are passing component price hikes straight to customers.
That is the tell inside the tell. CoreWeave, selling to sophisticated buyers like Meta, Anthropic, and Jane Street, raises prices and passes through costs. The frontier apps selling to Jay Schulman charge $200 for $8,434 of compute. Same input, opposite pricing discipline. One layer is priced to its cost; the other is priced to acquire dependency.
We have watched this movie before. Uber subsidized rides to kill taxis. Amazon priced Prime below cost to own default behavior. Salesforce gave away seats to become the system of record, then repriced once migration was unthinkable. Price low, acquire dependency, reprice when switching costs exceed pain tolerance.
The difference with AI, as Schulman notes, is that you knew the Uber ride would end. Nobody built a warehouse assuming Prime stayed $119 forever. But finance teams are now automating month-end close and audit firms are building review pipelines on subscriptions priced at 2% of underlying cost — operational dependencies masquerading as SaaS.
And because it is the Trainer's Dilemma, not a choice, the repricing does not arrive as a negotiation. It arrives as a blog post on a Tuesday: "exciting updates to our pricing structure to better align with value delivered." Thirty days, maybe sixty. Your automated workflows will not care.
When Schulman asked three finance leaders what budget variance triggers a workflow review, all said 15% to 20%. A move from $200 to cost-based pricing is a 4,200% increase. That is not a variance. That is a business-model event.
Now map that event onto the week's valuations and sort them by exposure.
The most exposed are the ones built purely on usage that only exists at 2% of cost. If a chunk of a vibe-coding platform's engagement is people generating tokens they would never generate at full freight, the demand curve underneath a $13.3 billion mark is partly a pricing artifact. It reprices when the tokens do.
CoreWeave is exposed to demand risk, but at least its revenue is priced to its cost and passed through — it is not selling the subsidy, it is selling the picks. Its problem is leverage and a short cash runway, not a fictional unit economic.
And then there is the layer that gets paid when the fear arrives. Constellation Gate is building exactly what Schulman's CFOs are about to demand: usage by team and by model, cost per workflow run instead of cost per seat, cache-aware prompt compression that trims token use, and a record you can retrieve without asking the person who ran the prompt. Flexera found only 31% of organizations have accurate visibility into AI spend, and 59% said wasted AI spend rose year over year.
That is the whole thesis in one statistic. The panic is documented, growing, and unmonetized — and the tooling that meters, controls, and routes around premium models is selling into it.
The winners of the repricing are not the platforms carrying inference cost on their balance sheet. They are the pick-and-shovel layer that sells CFOs a way to survive the very unit economics that break the pure-plays. Databricks' Gateway, cost governance, portability tooling — these are bets that the subsidy ends, packaged as bets that AI adoption continues. Both can be true at once.
So return to Ghodsi's two sentences. "Demand is crazy" is what you put in the press release at $190 billion. "Token maxing has freaked out the CFOs" is what you actually built a product to sell.
He told you which one he is betting on. The market only wrote down the first.