GPU Cloud IPOs Inherit Nvidia's Merchant Risk
With the chipmaker's revenue-sharing backstop on hold, Yotta and ESDS are testing whether public markets will underwrite an order book as if it were a hyperscaler contract.
Two Indian GPU cloud operators are about to ask a question Nvidia has just stopped answering: whether a merchant cloud's order book can do the work of a hyperscaler contract. ESDS has a fully subscribed Rs 720 crore IPO that will fund data center capacity for exactly that bet, while Yotta's public listing puts a $7.5 billion Nvidia order book in front of investors with no way to verify what sits behind it. The timing is not accidental: Nvidia's July revenue-sharing backstop for merchant GPU clouds is on hold, and its $105 billion guarantee for OpenAI's Ohio campus shows where the balance sheet will still go—behind a hyperscaler, not a merchant cloud.
The pause redraws the capital stack for the AI buildout, because for the past two years a merchant GPU operator could point to Nvidia's willingness to share revenue risk as a proxy for creditworthiness. Now that backstop is gone and the financing burden has moved to public markets in its place, ESDS's subscription in full suggests demand for AI infrastructure exposure remains hot—but the structure of that exposure has changed: a hyperscaler contract carries a name, a site, and a contractual revenue stream, while an order book is a list of intentions.
Yotta puts a number to the leap. Its IPO will test whether a $7.5 billion Nvidia order book can stand in for a hyperscaler-backed revenue contract, and PWD's tracking shows GPU orders outpace disclosed demand. That gap between orders and actual, contracted offtake is precisely where the risk concentrates—and public investors are now underwriting it.
The contrast with hyperscaler financing could not be sharper. Amazon took 199 megawatts of Swedish wind in one deal and another 200 megawatts in a separate contract, using power-purchase agreements to anchor long-term infrastructure revenue; hyperscalers still sign contracts that make projects financeable, with the wind farms getting a reliable buyer and the capital stack working. Merchant clouds have no such anchor: their revenue projection is a hope that GPU demand materializes at the right price, with no commitment from the buyer side.
A backstop on hold
Nvidia's decision to pause the revenue-sharing program is quiet but decisive. The program was a financial engineering innovation: the chipmaker took a slice of a cloud's future revenue in exchange for lower upfront GPU costs, effectively sharing the merchant risk. With that mechanism suspended, the merchant cloud's balance sheet becomes the first line of defense and the IPO becomes the second—which makes ESDS's buildout a test of whether public investors will fund capacity Nvidia itself has declined to guarantee.
The numbers in the prospectus do not tell the whole story. X2M, an Australian IoT company, announced its first GPU facility contract this week with no named customer, no confirmed site, and no capacity figure—the omissions are not incidental but the disclosure equivalent of a blank space where the revenue model should be. If a public company can announce a GPU facility without saying who will buy the compute or how much power it will draw, the market is being asked to price a narrative, not an asset.
The SLA penalty math adds a sharper edge. A 45-minute outage at a $144 million facility can cut cash flow by 41.7%, meaning a single event can erase nearly half a year's earnings, and traditional insurance does not cover it; that is a tail risk with no hedge. For a merchant cloud, an outage functions less as a service interruption than as a direct hit to the only collateral the IPO offers, and public investors are being asked to underwrite that tail without the data to price it.
The week also brought a counterpoint from the hyperscaler side. Alibaba planted two data centers in Brazil, betting on local AI services and a $53 billion global buildout, while Volato's $500 million AI merger priced Ohio power at $3.25 million a megawatt, a reminder that powered land has value independent of any revenue contract. Those transactions are the exception; most merchant GPU capacity is being financed without a named tenant, a confirmed site, or a capacity figure, and the public market is swallowing it.
Continue this analysis
Get the complete Private Infrastructure Daily analysis and every detail that follows.
Enter a valid work email to continue reading.
Already a reader? Sign inPrivate Infrastructure Daily's daily briefing. Unsubscribe anytime.