Nvidia just lined up half a trillion dollars in financing for its own customers, and the list of banks writing the checks reads like a who’s who of Wall Street. Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR are all in. We walk through what the deal actually covers, why nobody seems totally sure whether this money is new or already spoken for, and what it means that the same bank helping arrange the financing is also the one publishing the CapEx forecasts everyone’s citing.
Nvidia announced six strategic partnerships this week—Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR—forming new AI-compute infrastructure financing platforms designed to raise more than $500 billion in third-party capital. The money funds infrastructure for Nvidia’s own customers, the data centers and compute build-outs that keep GPU demand growing.
Nvidia CEO Jensen Huang framed the logic simply: Compute generates revenue, and Nvidia’s compute holds value longer than most hardware because CUDA software keeps improving it after the sale. He pointed to real numbers backing that argument. One-year rental pricing for an H100 GPU climbed from about $1.70 per GPU-hour in October 2025 to roughly $2.35 per GPU-hour by March 2026, prices rising, not falling, on hardware that’s been shipping for years. Six-year-old A100 chips, an even older generation, still run in production today and still generate revenue for whomever owns them.
The announcement landed with less clarity than the dollar figure suggested. Bloomberg reported the deal days before Nvidia confirmed it, citing people familiar with the matter, and even then, nobody could say for certain whether this $500 billion connected to a separate $750 billion in Nvidia-arranged deals reported a month earlier, deals that had already revived concerns about vendor financing. Bloomberg’s own reporting admitted as much: It wasn’t clear which projects or companies the money would back, what form the funding would take, or whether $500 billion represented new commitments layered on top of existing ones.
That fuzziness matters more than it sounds like it should. Critics have already flagged how hyperscalers and AI companies blur the line between revenue and capital spending, cycling the same billions between each other in ways that make the underlying numbers hard to verify independently. Half a trillion dollars is large enough that whether it’s genuinely new money or already-committed capital wearing a fresh press release changes the real picture of how much capital is actually flowing into AI infrastructure right now, not just how much a headline announces. Six banks and asset managers signing onto one platform doesn’t, by itself, tell anyone whether that platform funds a new data center or simply repackages a commitment that already existed on someone’s balance sheet.
Goldman Sachs, one of the six firms backing the new financing platforms, published its own hyperscaler CapEx and free cash flow projections around the same time. The bank’s numbers show free cash flow running net negative across the hyperscaler group this year and next, with Oracle projected to remain the only member still deeply in the red once the group’s cash flow turns positive again in 2028. Capital spending growth outpaces cash flow from operations by a wide margin through that same stretch, which is the entire reason this financing conversation exists in the first place: Somebody has to cover the gap between what these companies spend building AI infrastructure and what their operations currently generate.
Goldman expects hyperscalers to issue roughly $400 billion in new debt next year, covering about a third of projected capital spending, with credit strategists specifically forecasting 35% of 2027 CapEx funded through debt. On top of that, Goldman projects another $300 billion in project finance issuance next year aimed specifically at data centers and chips. Worth stating plainly: Goldman sits close enough to this deal to know these numbers well, since Goldman also sits inside the same $500 billion consortium financing the spending its analysts are forecasting. That’s not evidence the numbers are wrong. It’s a reason to read them as informed, not neutral.
Debt markets are already signaling they’ll want concessions as this spending keeps escalating; those concessions haven’t turned punitive so far in the deals that are visible. Goldman itself expects equity raises to keep playing a role alongside debt, part of what it calls multi-year strategic plans to fund AI investment and protect balance sheet quality at the same time. Translation for anyone budgeting against this cycle: The hyperscalers aren’t betting on one funding source; they’re mixing debt, equity, and now these Nvidia-brokered financing platforms, spreading the exposure across enough instruments that no single lender or shareholder base carries the full weight of the build-out alone. That diversification is itself a signal worth reading: When five or six different capital sources all show up at once for the same build-out, it usually means none of them alone was willing to underwrite the whole bet. That’s not necessarily a red flag on its own, spreading risk across multiple lenders is standard practice for a project this size, and it does mean no single institution has staked its full credibility on this build-out succeeding, which matters if anyone’s trying to gauge how confident Wall Street actually is versus how loudly it’s announcing that confidence.
Meanwhile, Intel raised $20 billion of its own this week, upsizing a share offering from an original $15 billion target after investor demand ran past $100 billion, pricing shares at $95 to fund its foundry and chip manufacturing turnaround. Anthropic locked in a parallel bet on the compute side, a 20-year, $9.1 billion agreement with Riot Platforms for 191 MW of data center capacity out of Riot’s Rockdale, Texas, campus, running through June 2048. Different companies, different mechanisms, same underlying pressure—very major AI player is reaching for a different lever right now to raise capital at a scale that would have looked unreasonable two years ago.
None of this changes what silicon teams and CIOs actually need to plan around: GPU pricing that’s still climbing, not falling, and a financing structure complicated enough that tracking real capital flow now requires reading bank research alongside vendor press releases, not instead of them. The $500 billion headline is real. What it’s actually paying for stays genuinely unclear.
What do we think?
The $500 billion number matters less than who’s providing it and who’s explaining it. Goldman sits on both sides, backing the financing and publishing the CapEx forecasts that justify it, worth flagging every time anyone cites these numbers. For CIOs, the actionable signal isn’t the headline figure. It’s that GPU rental pricing kept climbing through this entire financing story.
Inflection point: Half a trillion dollars in vendor-arranged financing, with the arranging bank also publishing the CapEx math that justifies it, marks an inflection point in how AI infrastructure spending finds funding: The line between customer, lender, and analyst has disappeared. Watch GPU rental pricing as the signal instead of financing headlines. Huang’s numbers show H100 pricing rising even as newer chips ship; demand data that doesn’t depend on which bank is standing on which side of the transaction this week.
WHATYA THINK? LIKE IT? DON’T BE STINGY, SHARE IT WITH YOUR FRIENDS.