Nvidia Cuts OpenAI Backstop as Anthropic Posts First Profit

In the same 48 hours, Nvidia told the market it's less willing to underwrite OpenAI's infrastructure buildout than it said a month ago — and Anthropic told the market it turned a profit for the first time.

Neither headline is really about the other company. But read together, they're the clearest signal yet of how capital markets are starting to price the two dominant strategies in frontier AI: OpenAI's bet that scale now justifies losses later, and Anthropic's bet that discipline compounds faster than hype. This week, the market rewarded discipline and quietly punished scale.

What Nvidia actually walked back

The project is a planned 10-gigawatt data center in southern Ohio, developed alongside SB Energy, the SoftBank-owned infrastructure arm, as part of OpenAI's broader Stargate-style buildout. The full project is expected to cost north of $500 billion once chips are included. Nvidia's role was never to fund the chips — it was to backstop the lease and debt financing that make the rest of the project bankable, essentially co-signing the loan so lenders would extend credit against a data center that doesn't generate revenue until it's built and filled with GPUs.

When a $250 billion version of that guarantee first leaked in late July, Nvidia's own stock dropped roughly 5% in a single session. That's shareholders doing the math themselves: a quarter-trillion-dollar contingent liability tied to one customer's ability to keep raising money is a very different risk profile than Nvidia's core business of selling chips for cash upfront. The pushback worked. Nvidia and OpenAI are now finalizing a structure where Nvidia backstops only the project's first phase, cutting the guarantee to under $120 billion — less than half the original number.

Anthropic's profit is real, but it comes with an asterisk

Anthropic's Q2 2026 revenue landed around $10.9 billion, more than double the roughly $4.8 billion it posted in Q1 — pushing the company to an estimated $559 million operating profit for the quarter, its first ever. That's a genuinely fast climb: Anthropic's annualized revenue run-rate has now grown more than tenfold in about eighteen months, driven by enterprise API demand and Claude's growing share of coding and agent workloads.

The asterisk matters, though. Part of the Q2 jump is attributed to a large compute resale arrangement with SpaceX, worth a reported $1.25 billion a month, that inflated the quarter on a ramp-up discount unlikely to repeat at the same margin. Anthropic itself has reportedly told investors not to expect sustained profitability in the quarters immediately following — it's still spending aggressively on its own infrastructure. The honest read is that Anthropic proved profitability is achievable at its current scale and pricing, not that it has permanently flipped from cash-burning lab to cash-generating business.

Two AI labs, two completely different capital strategies

OpenAI's model has been to lock in as much compute, as many data centers, and as much guaranteed capacity as possible now, funded by a stack of financing partners including Nvidia, Oracle, Microsoft, and SoftBank, betting that revenue scales into the commitments before the bill comes due. It's reportedly on pace to lose about $14 billion in 2026 even as revenue keeps climbing — a gap that only widens as more of these infrastructure guarantees convert into real, drawn-down obligations. That's the strategy an over $1 trillion IPO valuation, reportedly targeted for as soon as September, is being built on.

Anthropic's model has been narrower: fewer consumer-facing bets, heavier weighting toward enterprise and API revenue, and infrastructure commitments sized closer to demonstrated demand. It's a less aggressive land grab, and it shows in market share versus OpenAI's ChatGPT — but it's also why Anthropic can post a real operating profit off $10.9 billion in quarterly revenue while OpenAI, reportedly generating comparable or greater top-line revenue, still loses billions. Neither approach is obviously right. But this week is the first time the market visibly repriced the risk of one of them.

Why this matters beyond two companies

Every AI infrastructure deal signed in the past year has depended on someone with a strong balance sheet effectively co-signing risk that a startup, even a $500 billion one, can't carry alone. Nvidia has been the most important co-signer in the industry — its willingness to backstop customer financing is arguably as important to the AI buildout as its chips. When Nvidia's own shareholders force it to cut that backstop in half within weeks of it being reported, that's a real constraint appearing in a market that, until now, looked like it had none.

That constraint will show up first in the deals just below this one — mid-size inference and neocloud players who've been raising debt against multi-year compute contracts on the assumption that a hyperscaler or chipmaker would always be there to backstop the downside. If the top of the market is getting more conservative about concentration risk, the guarantees smaller companies can extract get smaller too, and the bar for proving revenue coverage before signing goes up.