All summer this newsletter has been about safety and control. Who can switch a model off, whose sandbox leaked, which agency holds the checkpoint. This week the story swung back to money, and the numbers were the most interesting thing anyone said out loud.
On Friday, Bloomberg reported that Anthropic booked more than $11.5 billion in the second quarter and turned positive adjusted operating income, which as far as anyone can tell makes it the first frontier lab to show investors a profitable quarter. On Monday, OpenAI signed a ten-year lease on an 8 gigawatt data center in Ohio, and the way it got financed was the news. Nvidia agreed to guarantee up to $105 billion of the lease and power obligations.
Two labs, two answers to the same question. One says the frontier can now pay for itself out of revenue. The other says the frontier gets built anyway, on someone else's signature. Both answers are about to be tested in public, because both companies are headed for the stock market this fall.
For three years the standing objection to every AI valuation was simple. These companies cannot outrun their own compute bills. Revenue grows, inference costs grow faster, and the whole thing only works if somebody keeps writing checks. On Friday that objection took its first real hit.
Bloomberg, working from documents it obtained, reported that Anthropic booked preliminary revenue above $11.5 billion in the second quarter. A year ago the same quarter was $787 million. The first quarter of this year was $4.73 billion. So revenue more than doubled in three months and grew roughly fourteenfold in twelve. The company also posted positive adjusted operating income, which is the part that had never happened before at this end of the industry. Forbes put booked first-half sales at $16.2 billion. CNBC noted the figures are preliminary and could still move.
Read the words carefully, because they are doing work. Preliminary. Adjusted. Operating income, not net income. This is a private company showing selected numbers to prospective investors ahead of an October listing that its own shareholders have reportedly floated at $2 trillion. Nobody has audited any of it. The most useful thing to remember is that a company two months from a roadshow chooses which quarter to talk about.
And there is a live argument about why the quarter looked the way it did. Writer Ed Zitron made the case back in May, before the numbers landed, that Anthropic's cost line got a very convenient haircut in exactly the months it would later point to. The company pays SpaceX's xAI arm roughly $1.25 billion a month for compute on the Colossus cluster, a deal that ramped up in May and June with reduced fees during the ramp. If your biggest expense is discounted for one quarter and then is not, a profitable quarter is a real event and also a temporary one. I have not seen anything that settles this, and the S-1 is where it gets settled.
Which brings in the other Anthropic story of the week, because it is the same story. On Thursday, Bloomberg reported the company is in talks to buy Decart AI for about $6 billion, its largest known acquisition and its fifth of the year. Decart builds world models, the kind that simulate physical environments for robotics and driving, but the reason it costs $6 billion is the other half of what it does. Its software makes chips run more efficiently, which lowers the cost of training and serving models. The team would land inside Anthropic's inference and performance group. Decart raised at close to $4 billion in May with Nvidia among the investors, so this is a healthy premium three months later, and the talks could still collapse.
Line those two items up and the strategy is not subtle. Revenue is doubling every quarter. The way you keep a margin under that curve is to spend billions buying the ability to serve each token for less. The profitable quarter and the $6 billion efficiency purchase are the same bet, placed twice.
My take: I want to be careful here, because I write this newsletter with Claude and Anthropic is the last company I should be cheerleading. So let me put it in the least flattering way that is still true. A frontier lab produced one quarter of adjusted operating profit, in a period when its largest cost was discounted, and released it to investors eight weeks before it asks the public for money at a reported $2 trillion. That is a fact and a sales document at the same time. But the underlying shift looks real to me, and it is the one worth watching. Inference is getting cheaper faster than usage is getting more expensive. If that holds, the entire argument that these companies can never make money stops being about physics and starts being about who runs their infrastructure well. If it does not hold, we will find out in a filing, in public, for the first time.
OpenAI answered the same question a completely different way. On Monday it signed a ten-year lease on a data center campus in Pike County, Ohio, built and owned by SoftBank's SB Energy, on private land and former federal property once used for uranium enrichment. The site is designed for 8 gigawatts of computing capacity, supported by at least 10 gigawatts of new power generation, with roughly 9.2 gigawatts of new gas-fired plant ultimately envisioned. Officials say Japan is helping fund the energy build under the 2025 trade agreement. OpenAI says the project supports 35,000 construction jobs through 2032 and 2,500 permanent ones.
The number that matters is the one in the SEC filing. Nvidia, which will be the exclusive chip supplier to the site, agreed to guarantee up to $105 billion in conditional lease and power payment obligations to SB Energy. It is not writing a check. It is co-signing, so lenders will finance a project of this size, and stepping in if OpenAI cannot pay. Nvidia is also putting $1.5 billion directly into SB Energy. Jensen Huang framed it on X as securing long-lived infrastructure that can be re-equipped with each new generation of chips.
The guarantee got smaller on the way to the signature, which tells you as much as the deal does. CNBC reported in July that the backstop under discussion ran to $250 billion, and Nvidia's stock dropped about 4.5% intraday on the news. The Wall Street Journal reported on Friday that it had been cut to under $120 billion after investors raised concerns about the exposure. It landed Monday at $105 billion, covering only the first phase, with the cost of the actual chips a separate conversation. Fortune framed the shrinkage as a signal about how much demand here is real and how much is the industry financing itself in a circle.
This is not an isolated move. The Monday before, Nvidia teamed up with six large asset managers to build financing platforms aimed at pulling more than $500 billion of outside capital into data centers, helped along by SEC staff guidance in July that keeps certain data-center debt outside the risk-retention rules that would otherwise force sponsors to hold part of the exposure themselves. Nvidia is systematically arranging for other people's money to buy its chips.
Sitting underneath all of it is the awkward part. OpenAI is projected to lose around $14 billion this year and does not expect positive cash flow until roughly 2030, and it is targeting a public listing as early as September at a valuation approaching $1 trillion. It cannot borrow against a commitment this size on its own credit. So the supplier lends its balance sheet to the customer to buy the supplier's product, and the market decides whether that is vertical integration or a loop.
My take: Put the two stories side by side and you get the clearest picture of AI economics I have seen all year. Anthropic's answer is to get the cost per token down until the business closes on its own. OpenAI's answer is to get the buildout financed on a partner's credit and grow into it. Neither is obviously wrong. But the second one only works while Nvidia is willing to keep signing, and Nvidia only signs because every gigawatt is chips it sells. Six weeks ago I wrote that Nvidia had become the system. This week it became the bank too, and it trimmed its own guarantee by more than half before signing, which is the most honest thing anyone has said about the risk in this trade.
Everything I read this week was an argument about who carries the cost of the frontier, and every party answered with something other than cash. Anthropic answered with efficiency, buying its way to a cheaper token. OpenAI answered with a guarantee from its chip supplier. Nvidia answered by routing half a trillion dollars of other people's capital toward the buildout. Stripe answered by buying the meter. Alibaba answered by giving the model away and charging for the cloud around it.
What strikes me is how little of this is a technology argument anymore. Nobody won this week by having the best benchmark score. They won by having the cheapest inference, the friendliest lender, or a position in the middle of the flow. That is what a maturing industry looks like, and it is also how you end up with a handful of firms sitting at every toll booth at once.
The catch is that every one of those answers eventually has a physical address. An 8 gigawatt campus in Ohio needs 10 gigawatts of new generation, most of it gas, and the people who live near it did not sign the lease. A financing structure is a clever instrument right up until it is a transmission line through somebody's county. The money can be moved around in a filing. The power cannot, and that is the part that shows up in our lives.
On Tuesday the Greensboro City Council voted 8 to 0, with one recusal, to enact a 180 day moratorium on data centers drawing more than 10 megawatts. It runs until February 14, 2027 and pauses the acceptance, processing and approval of any application for a facility that size. Yadkin and Alamance counties held their own data center hearings the same day.
The path there is the interesting part. On July 20 the same council voted 5 to 4 against even starting the moratorium process. Two weeks later it reversed and voted 9 to 0 to begin, with members openly crediting the volume of resident feedback. Then came two days of hearings this week where people packed the chamber, held up signs reading 32 for the number of months they wanted the pause to last, and told the council, in one speaker's words, that they did not want excuses, they wanted results. This is a city that had said no, then said yes, then got told it had not gone far enough.
Hold that next to the Ohio number for a second. Greensboro is pausing to study anything above 10 megawatts, which is roughly the continuous power of 8,000 homes. The campus OpenAI leased on Monday is designed for 8,000 megawatts. One project, eight hundred times the threshold that a North Carolina city of 300,000 people just decided it needs six months to think about. The North State Journal counted at least 11 NC counties and 17 towns with moratoriums as of August 6, and Greensboro makes the Triad's largest city the newest name on that list.
Meanwhile the piece that actually decides who pays has still not started. Duke's rate settlement for the Carolinas sits with the Utilities Commission awaiting a decision expected in November, and the separate large-load tariff proceeding, the one that would put data centers in their own rate class, is expected to get going this fall. So the state's mechanism is months out while towns act weekly, which is why the local votes keep happening. For those of us building here, the lesson from this week's two big stories is worth pinning up. The economics of the models you rent are improving fast, which is good for your bill. The economics of the power underneath them are being negotiated by people who have never heard of your company, in rooms a few miles from your office. Show up to one of those meetings sometime.
That's the week the frontier opened its books. See you next Wednesday.
Daniel
BullCity AI · Durham, NC
P.S. Has your inference bill actually gone down this year, or does it just feel like the prices dropped while your usage ate the difference? Hit reply with a real number if you have one. I'm putting together a picture of what the price war looks like from the invoice side, which is the only side that counts.
P.P.S. Forward this to whoever on your team still says AI companies can't make money. One of them just did, for one quarter, with an asterisk you should read carefully.