Compute Financing Is the New Battleground

There's an old bit of engineering wisdom that the bottleneck is never where you think it is. You optimize the hot loop for a week, and it turns out the whole thing was waiting on a single mutex the entire time. The AI infrastructure market spent three years convinced the bottleneck was silicon. The week of June 29 to July 6 quietly established that the bottleneck moved, and nobody sent a memo.

Three announcements landed close enough together to read as one story. NVIDIA started financing the companies that buy its chips. Meta hinted it might stop buying compute and start selling it. And Together AI raised $800 million to keep renting out the chips NVIDIA makes. The common thread isn't a chip. It's the question of who fronts the capital, and who collects the rent.

NVIDIA decides to own both ends of the trade

On July 1, in a blog post co-authored by CFO Colette Kress, NVIDIA formalized what it's been doing ad hoc for two years. The pitch, stripped of the "AI factories" varnish: an AI cloud buys NVIDIA hardware, and NVIDIA collects its usual margin on the sale plus an ongoing cut of the cloud revenue that hardware later generates. A revenue-sharing and credit-support model, in their words. Double-dipping on the same silicon, in less generous ones.

The first two partners are Sharon AI, an Australian neocloud deploying up to 40,000 Grace Blackwell GB300 GPUs, and Firmus, building a 360 MW campus in Batam, Indonesia, that's expected to house up to 170,000 GPUs. Combined, roughly 210,000 units. NVIDIA named Baseten, Fireworks AI, and Together AI as the demand side, the companies expected to actually consume that capacity.

Why do this at all, when you already run ~75% gross margins on the hardware? Because the constraint on your own growth is no longer your ability to make chips. It's your customers' ability to finance them. NVIDIA said the quiet part in its own release: even signed, long-term customer commitments haven't been enough to convince lenders to fund large-scale deployments. Smaller clouds can't borrow against demand they've already booked. So NVIDIA stepped in as the lender of first resort, absorbing the residual-value risk on GPUs whose depreciation curve nobody fully trusts yet, in exchange for a permanent seat at the revenue table.

The stoic read is that this is admirably within NVIDIA's circle of control. It cannot manufacture demand for a startup's tokens, but it can manufacture the financing that lets the startup exist long enough to generate them. It's de-risking its own order book by underwriting the people placing the orders. Elegant. Also, if you've read any history of vendor financing, faintly familiar in a way that should make you check the exits. The analysts reaching for the GE Capital and dot-com telecom comparisons aren't wrong to; the last time a dominant equipment vendor propped up its own customers' balance sheets at scale, the unwinding was memorable. The revenue-split percentages remain undisclosed, which is precisely the number that decides whether this is a genuine capital lift for the partners or a long, quiet lock-in.

Meta threatens to flip from buyer to seller

The same day, Bloomberg reported that Meta is building a cloud business, "Meta Compute," to sell its excess AI capacity, either as hosted model access or as raw GPU rental in the CoreWeave mold. The market did not wait for confirmation.

CoreWeave closed down 13.9% at $85.69. Nebius fell 17% to $229.18, a roughly $11.9 billion single-day evaporation, the worst one-day loss in the AI-infra complex. IREN dropped about 6.5%. Meta, credited with exactly the optionality subtracted from its suppliers, rose 8.8%.

The logic is brutal and simple, and D.A. Davidson's Gil Luria put it plainly: CoreWeave and Nebius lean on Meta for growth, and Meta may not need them anymore. Meta is a $21 billion CoreWeave customer and an up-to-$27 billion Nebius customer. When your single largest anchor tenant muses about becoming your landlord, every landlord on the street gets repriced.

Here's the part worth publishing rather than repeating: this is a stock move built entirely on a Bloomberg report. Meta has confirmed no product, no pricing, no timeline, no name. Zuckerberg said cloud was "definitely on the table" at May's shareholder meeting, which is a long way from a launch. And Rosenblatt, reiterating a Buy on CoreWeave, flagged the detail everyone selling ignored: Meta doesn't have the right to resell the capacity it leases from CoreWeave through 2032. So whatever Meta Compute becomes, it can't be built from the chips Meta is already renting. The near-term threat to existing neocloud contracts is smaller than $11.9 billion of panic implies. Treat any headline calling this a "Meta announcement" with suspicion. It's a plan someone described to a reporter.

Together AI raises $800M into the teeth of all of it

And then, on the same July 1, Together AI closed an $800 million Series C at an $8.3 billion valuation, led by Aramco Ventures, with NVIDIA among the participants. That's a 2.5x valuation step from its $3.3 billion Series B roughly sixteen months earlier. Annual bookings crossed $1.15 billion last quarter, and the company says open-weight model usage across the industry tripled in twelve months. Cursor, Cognition, and Decagon are named customers running production workloads on open models instead of closed APIs.

Two details matter more than the headline number. First, the lead investor is the venture arm of Saudi Aramco, and the round rhymes with a broader pattern of sovereign energy money treating compute capacity the way it treats oil reserves: a strategic reserve to be secured for decades. Second, and this is the one that ties the whole week into a bow: NVIDIA is an investor in Together AI. So NVIDIA is simultaneously financing neoclouds through its new revenue-share program and holding equity in the neoclouds that consume their capacity. It is on both sides of the trade, everywhere you look.

What it actually means

Line the three up and the picture resolves. NVIDIA is propping neoclouds up. Meta is threatening to flatten them. Capital is still flooding into them. The same company, NVIDIA, is quietly equity-aligned with the survivors on every side.

The FLOPS war is over in the sense that it stopped being the interesting variable. Everyone can get Blackwell now; the question is who can afford to keep it lit, at what utilization, financed by whom, and who collects the recurring cut when the tokens finally flow. NVIDIA has answered that question in the most NVIDIA way imaginable: by making sure that no matter which neocloud wins, and no matter whether Meta floods the market or doesn't, it holds a claim on the revenue either way.

For anyone building on this infrastructure, the takeaway is unglamorous and worth internalizing. When you evaluate a neocloud provider now, the first question isn't tokens-per-second. It's solvency: who backstops this provider if utilization disappoints, and for how long does that support run? Sharon AI, for one, has now pledged portions of its future revenue in two directions at once, to NVIDIA and to a separate $200 million facility with Digital Alpha. A provider stacking vendor financing on top of investor financing has narrowed its margin for error before serving its first token.

The chips were never the moat. The moat is who can write the check, and who gets paid back forever.

Sources: NVIDIA blog; Together AI / Business Wire; Tom's Hardware; CNBC; TechCrunch. The Meta Compute plan is Bloomberg-sourced and unconfirmed by Meta; stock moves are verifiable market data.