The original figure Nvidia was expected to stand behind was $250 billion. The Wall Street Journal reports that the chipmaker is now willing to guarantee a little under $120 billion, a retreat that followed investor complaints about the scale of risk the company was absorbing.
A cut of that size isn’t a rounding error. It’s a business examining its own exposure to AI’s largest customer and concluding that half the number is plenty.
What the smaller number actually covers
Only the first construction phase falls under the guarantee, a stage that would bring roughly five gigawatts of capacity online. A lease covering the entire 10-gigawatt project, which SB Energy — a SoftBank subsidiary — is developing, is being negotiated separately by OpenAI.
A second track is moving in parallel, too. Nvidia is in talks over separate financing tied to OpenAI’s chip purchases, worth as much as $350 billion.
The relationship, in other words, isn’t contracting. Only the written commitment is.
Why bubble skeptics will grab this
No company has benefited more visibly from the AI buildout than Nvidia. So when it cuts its own risk in half under pressure from shareholders, the move reads as caution coming from the party with the clearest sightline on demand.
For the past year, AI bubble warnings have mostly been a mood. Now the critics have something concrete. Slicing a guarantee in half is what you do when you’re running the scenario where the capacity gets built and the revenue never arrives.
Anthropic’s quarter says the opposite
Reuters reports that Anthropic’s revenue more than doubled inside a single quarter, climbing from $4.73 billion in Q1 to more than $11.5 billion in Q2 — a 14x jump year over year.
The two narratives refuse to line up, and that friction is what makes them worth watching. A major supplier is quietly retreating on paper at the same moment one of the largest compute buyers posts figures that look nothing like a cooling market.
According to people familiar with Anthropic’s finances, the company is modeling revenue somewhere between $190 billion and $200 billion for 2028. Measure that against the roughly $45 billion annual run rate it disclosed in May and the steepness of the internal curve becomes clear.
By Anthropic’s own account, revenue grew more than 10x in each of the three years leading up to early 2026.
The number to watch isn’t the projection
Publishing a 2028 forecast costs nothing, and holding anyone accountable for one is close to impossible. The quarter-over-quarter move from $4.73 billion to north of $11.5 billion is the piece that’s already in the books.
Should those figures hold, they would point to demand for proprietary AI services still rising — political headwinds and mounting competition from China notwithstanding.
One crack deserves a mention. Ramp, a financial services firm, recently registered a mild flattening in Anthropic token demand across its business customers. Mild, and confined to a single firm’s client base. Still, it’s exactly the kind of early signal that either fades entirely or becomes the headline half a year from now.
An IPO priced against all of it
Reports suggest Anthropic intends to list publicly at a valuation approaching $1 trillion, in late September or early October.
That offering is where both threads finally get settled in the open. Nvidia pared back a guarantee because its investors demanded it. Anthropic will turn to a far bigger investor pool and ask them to validate a trillion-dollar figure on the back of a quarter that more than doubled.
Pay attention to where that pricing settles. It will reveal more about genuine market conviction than any guarantee hashed out behind closed doors.