One hundred and four point two billion dollars. That is CoreWeave's revenue backlog, up 246 percent year over year, and it does not include the roughly $25 billion in new commitments the company added in the first weeks of this quarter. On the same earnings call, chief executive Michael Intrator said near-term capacity is effectively sold out. On the same balance sheet, the company reports a $626 million quarterly loss driven by $640 million of net interest expense, and $46.7 billion of property and equipment that is mostly Nvidia GPUs. That is today's issue in one filing: demand that is provably real, financed by debt, secured against hardware with an unknown shelf life, and rippling outward into land, turbines, vacuum pumps, and university campuses.
The backlog is the bull case and the risk, on the same page
- CoreWeave's quarter was genuinely strong and genuinely leveraged. Revenue rose 112 percent to $2.58 billion, just past the $2.56 billion analysts expected, and the stock jumped more than 14 percent after hours. Intrator said prices for Nvidia's Blackwell and Vera Rubin chips are "setting new highs" while older inventory rents at levels last seen years ago. The bear case sits in the same release: the loss, the interest expense, and the depreciation question on $46.7 billion of GPUs. Management's answer is inference. Running models rather than training them keeps older silicon earning, and Intrator argues CoreWeave has "an embedded advantage because of our control over the silicon."
- The financing arrived the day before the earnings did. Nvidia, which owns close to 13 percent of CoreWeave, announced a plan to mobilize $500 billion for AI infrastructure with Apollo, Blackstone, BlackRock, and Brookfield. When the chip supplier organizes the capital that buys the chips, the demand signal and the funding signal stop being independent measurements. Counterpoint's Neil Shah is on Bloomberg today discussing exactly that circularity.
- The private marks are moving in the same direction. Cognition is in new funding talks at a $40 billion valuation, and Lovable, the vibe-coding startup, has hit $13 billion. Both are coding companies. The capital is concentrating in the one application category where the productivity claim is easiest to check.
- The supply chain is confirming it in earnings, not press releases. Hon Hai, Nvidia's manufacturing partner, beat on profit thanks to sustained AI spending. Tencent's revenue growth is holding well enough to bankroll its own costly AI investment. These are the least promotional numbers in the issue, and they point the same way.
The scarcity keeps moving further down the stack
- GE Vernova is the clearest case of a company repriced without changing its product. Equipment the firm makes generates about 25 percent of the world's electricity across gas, wind, steam, hydro, and nuclear. The stock is up more than 600 percent since the April 2024 spin-off. What Scott Strazik actually did was convert customers into R&D partners: "The technical hurdle with how they want to run AI factories is much more complex than our standard-fare offering, and that's great," he told Fortune, because that technology "will ultimately apply to the broader grid over time. But they will have funded a lot of the learning curve." Nvidia and nuclear startup Blue Energy are paying for grid research no utility would have financed.
- Below the turbines, the list gets unglamorous fast. Bloomberg's new AI winners today are vacuum pump makers and industrial gas producers. Zetwerk, a manufacturing services company, is planning a $400 million India IPO. This is what a genuine industrial boom looks like from underneath: the returns show up in components nobody wrote a thesis about.
- Sovereign money is now measurably exposed. Norway's sovereign wealth fund got a lift from tech investments, and South Korea is laying out national AI investment plans. When a country's pension returns and a country's industrial policy are both indexed to the same trade, a correction stops being a sector event.
The land grab has reached the campus
- George Washington University sold its 120-acre Ashburn campus to Amazon Data Services for $427 million and initially declined to name the buyer or the price, citing confidentiality, until the student newspaper reported both by the end of the day. The University of Michigan is pursuing a $1.25 billion computing center in Ypsilanti Township. "Universities are large landholders, and there is a land grab going on with the data centers these days," former GW professor Ellen Scully-Russ told Fortune. Her account of the Ashburn campus before the sale is the sadder detail: five or six years of attrition, faculty not replaced, teaching assistants eliminated. The campus was hollowed out first and sold second. Full-time faculty got $2,500.
- Local resistance is now a named risk on an earnings call. Eighteen states have restricted or are considering significant restrictions on data center construction. Intrator's position is that this changes where centers get built, not whether demand gets met, and that companies should offer sweeteners like paying for grid upgrades so costs do not land on residents. That is a reasonable answer and also an admission that the siting fight is real enough to be priced.
- Meanwhile the schools that are not selling land are running venture funds. Crystal Springs, a 569-student day school between San Francisco and Silicon Valley, is one of several Bay Area schools with a miniature VC fund, capitalized by community donations and steered by parent-investors from Sequoia, Lightspeed, Notable Capital, and Battery. One set of institutions monetizes its acreage, another monetizes its parent network. Both are converting an educational asset into exposure to the same trade.
What people actually want from the machine
- ChatTJB is the best-argued critique of AI published this week, and it contains no argument at all. Tucker Bryant, a 32-year-old former Google project manager, built a chat box with no model behind it and a $6,000 San Francisco billboard advertising a "leading chat interface powered by AI," asterisked to reveal that AI stands for "average individual." Over 100,000 prompts have come in, peaking at 5,000 an hour, and more than 10,000 people have applied to answer them, with the waitlist growing by about a thousand a day. Bryant has mostly stopped answering himself because one person could not keep up.
- The finding is buried in the first night. Someone on their honeymoon wrote in to say they did not feel totally relaxed and asked whether that was okay. Bryant says that was the moment the project stopped being satire. A meaningful share of what gets typed into a chat box is not a retrieval problem. It is a request to be answered by someone. The 10,000 volunteers are not a joke about AI. They are demand for the thing AI was substituting for.
- Set against that, the datamaxxers. The Journal profiles people feeding every health metric they generate into AI systems, which is the exact opposite posture: maximum instrumentation, minimum human interpretation. Both groups are responding to the same conditions and reaching contradictory conclusions about what the technology is for.
- And the surveillance layer keeps compounding quietly. The Times reports on Flock cameras, which can effectively track every car in America. Police departments love them; the citizens being tracked do not. It is the only story in today's issue where AI infrastructure is already deployed at national scale, fully operational, and nobody is arguing about its valuation.
The Throughline
The connection running through today's issue is that AI has become a claim on physical assets, and the people writing those claims are moving faster than the institutions holding the assets. CoreWeave's $104.2 billion backlog is a promise to deliver compute that does not exist yet, in buildings that in many cases are not built, drawing power that is not generated. Every layer beneath it is now signing correspondingly long commitments. GE Vernova is building generation on the assumption that AI factories keep ordering. Amazon paid $427 million for a campus. Zetwerk is going public on manufacturing demand. The backlog is real; what it triggers underneath is a chain of decade-scale bets, each justified by the layer above it.
What makes this different from an ordinary capex cycle is how much of it is internally financed. Nvidia owns 13 percent of CoreWeave and, the day before CoreWeave's earnings, announced it was mobilizing $500 billion with four asset managers to fund AI infrastructure. Intrator's price commentary, that Blackwell and Vera Rubin rentals are "setting new highs," is a statement about the pricing power of a supplier who is also an investor in the buyer. That does not make the demand fake. It does mean the usual test, that an independent customer paid a market price, is not doing the work people assume it is doing.
The GW story is the one worth sitting with, because it shows how these decisions actually get made. A university did not weigh its academic mission against Amazon's offer and lose. It let a satellite campus erode for five or six years, through attrition and cut teaching assistants, until the property's highest use really was a data center. The sale was a formality after a slow institutional decision nobody had to defend at the time. Scully-Russ's "drip, drip, drip" is a better description of how AI reshapes an institution than anything in the earnings transcripts. Nobody chose to trade a campus for compute. They chose a series of small unfunded years, and then the market chose for them.
And then there is Bryant, who spent $6,000 on a billboard and got a hundred thousand prompts. Trillions of dollars have gone into building systems to answer questions, and the thing that surfaced when the model was removed was that people will wait for a human and 10,000 people will volunteer to be one. That is not a refutation of the technology. It is a fairly precise measurement of what the technology was actually asked to do, and how much of it was never a compute problem in the first place.
The Bigger Picture
The AI industry is completing its transition from a software business to a capital-intensive industrial one, and the tell is on CoreWeave's income statement: a $640 million net interest expense in a single quarter. Software companies did not have interest expenses of that shape. They had gross margins, and they could revise a roadmap in a quarter. A company carrying $46.7 billion of depreciating hardware, financed with debt, against a backlog stretching years out, is a utility with a technology multiple. It will be judged by industrial failures rather than product ones: stranded assets, cost overruns, contracts that outlive the demand that justified them.
That shift moves the real decisions somewhere less visible. The consequential AI policy this year is not being written in an AI bill. It is 18 states deciding whether to restrict data center construction, a county deciding what an interconnection queue looks like, a university board deciding whether to sell 120 acres in Loudoun County. Those decisions have fifteen-year consequences and near-zero public deliberation, which is precisely why Intrator can say confidently that moratoriums change siting rather than demand. There is no lever anywhere for "should this land be a campus or a data center." There is only a price.
The distributional question follows directly. GW's faculty got $2,500 apiece from a $427 million sale. Norway's sovereign fund and South Korea's national investment plan capture the upside at the country level, and Crystal Springs parents capture it through a school VC fund. The gains from this build-out are being allocated through asset ownership, and the institutions that historically converted knowledge into opportunity are participating as landlords rather than as beneficiaries. A university that sells its research campus to fund an endowment has not made a bad financial decision. It has made a financial decision about a thing that was not supposed to be primarily financial.
Which is what makes ChatTJB the right story to end on. The most oversubscribed service in today's issue is not a model. It is 10,000 people volunteering to answer strangers' questions by hand, at no pay, faster than one man can onboard them. Every number elsewhere in this issue describes an attempt to manufacture that supply industrially. The honeymoon message suggests a share of the demand was never for an answer.
What to Watch
- Whether CoreWeave's backlog converts on schedule, or gets renegotiated. $104.2 billion booked plus $25 billion added in early Q3 is the industry's strongest evidence that demand is durable. The test is delivery. Watch for backlog that grows while revenue recognition slips, which is what a contracted promise looks like when the power, the land, or the chips arrive late.
- Whether the 18 states restricting data center construction hold. Intrator's position is that siting restrictions move projects rather than kill them, and he is probably right in the near term. The thing to watch is whether "pay for the grid upgrades" becomes a standard condition of approval. If it does, the cost of a gigawatt goes up permanently and quietly, and the current backlog was priced before that.
- Whether more universities sell. GW got $427 million and Michigan is putting up $1.25 billion of its own. If a third and fourth transaction land in the next two quarters, this stops being opportunism and becomes a category, and higher education's balance sheet gets restructured around AI infrastructure without anyone having debated it.
