By Peter | July 6, 2026
Here's a sentence that should make every person holding an NVIDIA share, every AI startup founder, and every cloud CFO sit up straight:
NVIDIA now gets paid twice on the same GPU.
Once when it sells the chip. Again when that chip generates revenue for the cloud provider who bought it. And if nobody rents the GPU? NVIDIA's got a buyback guarantee — meaning it'll write a check for the idle capacity.
This isn't a rumor. It isn't a test program. On July 1, 2026, NVIDIA published a blog post co-authored by CFO Colette Kress that formalized a revenue-sharing and credit-support model for AI cloud operators. The first two partners — Sharon AI and Firmus Technologies — have already committed to deploying a combined 210,000 Grace Blackwell GPUs under this program.
That's roughly the GPU fleet of several mid-tier hyperscalers combined. And it signals something bigger than just a new sales strategy.
NVIDIA is no longer just the world's most valuable chipmaker. It's becoming the world's largest AI infrastructure bank — underwriting its own customers, collecting rent on its own hardware, and embedding itself into the economics of every token generated on its silicon.
Let's break down what this means, why it's brilliant, and where it could go horribly wrong.
The mechanics are elegant in their simplicity. Here's the structure:
Credit support: NVIDIA backstops the financing for cloud providers who can't get traditional loans for GPU purchases. Banks won't collateralize cutting-edge chips — they can't price the residual value of hardware that gets superseded every 18 months. NVIDIA can. It knows exactly what the next generation is worth because it's building it.
Revenue sharing: Once the GPUs are deployed, NVIDIA collects a percentage of the cloud revenue those chips generate. Standard hardware revenue up front, plus an ongoing, usage-linked royalty on the back end.
Idle-capacity buyback: If a partner's GPUs sit unused, NVIDIA guarantees to purchase or rent that idle capacity at a predetermined price. This is the credit enhancement that makes the whole thing bankable — lenders aren't lending against GPU value, they're lending against NVIDIA's balance sheet.
As the NVIDIA blog put it directly: "NVIDIA will earn both standard product revenue and a share of the cloud revenue on the supported capacity."
That's the two-rail. Hardware sale + recurring royalty. It transforms NVIDIA's revenue from lumpy, cyclical chip sales into something that looks a lot more like a SaaS company — and the market pays premium multiples for recurring revenue.
Let's put the scale of this thing in perspective. Here's what's already committed:
| Partner | GPU Commitment | Scale |
|---|---|---|
| Sharon AI | 40,000 GB300 GPUs | 72 MW, 6-year agreement |
| Firmus Technologies | 170,000 GPUs | 360 MW campus, Batam, Indonesia |
| Combined | 210,000 GPUs | ~432 MW total |
But the GPUs are just one part of the story. Here's what each GB300 NVL72 rack actually contains: 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single liquid-cooled cabinet drawing 120 kilowatts of power and delivering 1.1 exaFLOPS of FP4 compute. Each rack is effectively an exascale supercomputer in a box.
Bloomberg reported that Firmus expects between $25 billion and $30 billion in committed customer offtake agreements over the first six years of the deal. That's not projected revenue — that's contracted demand from customers who have already committed to buying compute.
Meanwhile, NVIDIA raised $25 billion in bonds on June 15, 2026 — its first debt issuance since 2021 — drawing $85 billion in orders from lenders eager for AI infrastructure exposure.
The market for Sharon AI (NASDAQ: SHAZ) was less enthusiastic. The stock fell 14.2% to $67.91 on July 3, the trading day after the announcement, on volume of 2.6 million shares. Investors appear to be pricing in concerns about what a revenue-sharing model means for Sharon AI's long-term margins. Sharon AI had only listed on the Nasdaq in February 2026 via a $125 million IPO and raised an additional $1.6 billion in a private placement in June.
Firmus, for its part, began as a Tasmanian bitcoin mining operation in 2019. It raised $505 million in April 2026 at a $5.5 billion valuation — a round in which NVIDIA itself participated as an investor — and arranged a $10 billion debt facility to support its buildout.
The program wasn't born from generosity. It exists because of a structural problem that has been hiding in plain sight.
When a startup wants 40,000 Grace Blackwell GPUs, it needs a lender willing to put up hundreds of millions against hardware that might be worth dramatically less in 18 months. NVIDIA's Vera Rubin platform — the successor to Blackwell — is already in production. Traditional banks cannot and will not price that depreciation curve, because only NVIDIA knows what the next generation will do to the value of the current one.
NVIDIA's buyback guarantee solves this. By committing to purchase idle GPU capacity at a predetermined price, NVIDIA effectively sets a floor on the asset's value. That floor is what makes the cluster bankable. A lender who would reject a $500 million GPU loan from a startup will accept it when the world's largest AI chipmaker is backstopping the residual value.
This is vendor financing — but with a twist. Unlike traditional vendor financing, NVIDIA isn't just providing a loan. It's taking an ongoing equity-like stake in the revenue those chips produce. It's closer to a royalty structure than a credit facility.
At NXagents, we use a framework called the Claw Effect to evaluate major business moves like this. Let's apply it.
The historical parallel is uncomfortable. Between 1996 and 2001, telecom equipment makers extended billions in vendor financing to customers buying their gear:
The strategy worked brilliantly — until it didn't. When telecom funding collapsed in 2001, Nortel's bad loans went from 25.5% of its portfolio to 80% in a single year. Lucent provisioned for $3.5 billion in total customer loan losses.
Wedbush Securities analyst Matthew Bryson noted that NVIDIA's investments fit "squarely into the circular investment theme" that has driven fears about market durability — though he acknowledged the strategy could build a "competitive moat."
A Bain & Company analysis from September 2025 estimated the AI ecosystem needs $2 trillion in annual revenue by 2030 to justify current infrastructure spending — and that the trajectory falls roughly $800 billion short.
The bull case:
The bear case:
Here's the uncomfortable question that nobody on NVIDIA's earnings calls will ask out loud:
When NVIDIA provides the financing, sells the chips, backstops the idle capacity, and collects revenue share on the output — all with the same counterparties — how much of the resulting revenue reflects genuine end-user demand, and how much reflects capital moving in a loop?
The answer matters enormously. The distinction between "flywheel" and "house of cards" comes down entirely to whether real customers are buying AI compute at prices that justify the infrastructure spending.
The strategic logic is clear. NVIDIA's best potential customers — smaller, nimble cloud operators targeting sovereign AI deployments and specialized inference — have genuine demand but can't secure financing. Traditional banks can't price GPU residuals.
By backstopping those assets, NVIDIA:
It's a brilliant hedge. If hyperscalers build their own chips, NVIDIA already has the next tier locked in.
Sharon AI (NASDAQ: SHAZ) — The listed player. 40,000 GB300 GPUs. Went public at $125M in February 2026, raised $1.6 billion private placement in June, saw stock drop 14% after the NVIDIA announcement. The market is not yet convinced this is good for Sharon AI's margins.
Firmus Technologies — The bigger bet. 170,000 GPUs on a 360 MW campus in Batam, Indonesia. Born as a bitcoin miner in 2019, now valued at $5.5 billion. NVIDIA is an equity investor. $10 billion debt facility arranged. $25-30 billion in committed customer offtake expected.
Baseten, Fireworks AI, Together AI — Named by NVIDIA as the type of AI-native companies this model serves. These are inference providers and model platforms that need elastic access to compute without committing years of capex.
The Hyperscalers — Microsoft Azure, AWS, Google Cloud. Not directly affected — yet. But if NVIDIA's model creates a thriving ecosystem of independent AI clouds, the hyperscalers face new competition from operators whose GPU financing is underwritten by the chipmaker itself.
Let's be direct. If AI-native demand cools — if the inference boom slows, if agent platforms don't materialize at projected scale, if enterprises pull back on AI spending — NVIDIA is exposed twice.
Once through declining chip sales. Again through the cloud revenue it has contractually agreed to share.
This isn't a loan book in the traditional sense. But it functions like one. NVIDIA is extending credit support to customers who couldn't otherwise afford its products, in exchange for future revenue streams that may or may not materialize at projected levels.
The difference between this and the Nortel debacle? Three things:
AI compute is actually being consumed. The telecom companies in 2001 used less than 0.002% of available fiber capacity. NVIDIA's customers are running at high utilization rates with real workloads.
NVIDIA's balance sheet is categorically stronger. $193.7 billion in data center revenue. $85 billion in bond orders for a $25 billion raise. The company could absorb significant write-downs without threatening its existence.
NVIDIA controls the depreciation curve. Unlike Nortel, which couldn't control what its customers' equipment would be worth, NVIDIA sets the pace of obsolescence by deciding when to ship the next architecture.
That third point is both reassuring and terrifying. Reassuring because NVIDIA can manage its risk exposure. Terrifying because it means the company has more control over the AI infrastructure market than any single entity has ever had over a technology supply chain.
If you hold NVDA — directly or through QQQ, SPY, or any tech ETF — track one metric above all others: the ratio of NVIDIA's revenue-share income to its direct hardware sales. If that ratio climbs too fast, it may signal that more revenue is coming from NVIDIA-financed customers than from organic demand. The CoreWeave deal alone is $6.3 billion. Add Sharon AI, Firmus, and whatever comes next, and this number will grow.
If you're building an AI-native company and have been blocked by compute costs, this model is designed for you. NVIDIA named Baseten, Fireworks AI, and Together AI as the template. The path is clear: prove you have customer demand, partner with an AI cloud on NVIDIA's program, and get access to GPUs without the balance-sheet risk. Move now, before the terms standardize and potentially become less favorable.
If you run a cloud business and aren't in conversations with NVIDIA about this program, you're already behind. Sharon AI and Firmus have locked in first-mover advantages at massive scale. The window for favorable terms won't stay open forever. The playbook is visible: secure NVIDIA backing, raise debt against the backstop, build capacity, and lock in customer offtake agreements before your competitors do.
Here's the contrarian take: this program is ultimately deflationary for AI compute. NVIDIA is financing a massive expansion of GPU supply that isn't controlled by the Big Three hyperscalers. More independent AI clouds competing for enterprise customers means lower inference costs over time. Lock in short-term contracts now; renegotiate when capacity comes online in 2027.
NVIDIA's move is the logical endpoint of a trend that's been building since ChatGPT launched: AI infrastructure is becoming a financial product. GPU capacity is being securitized, underwritten, and traded like an asset class. The companies that understand this — both the opportunities and the risks — will be positioned to profit from it. The ones that treat it as just another tech trend will be left holding the wrong end of the contract.
NVIDIA just changed the structure of the AI industry — not with a new chip, but with a new business model.
The company that became the world's most valuable by selling shovels during a gold rush is now taking a permanent stake in every mine. It gets paid when the equipment ships, and it gets paid again every time someone strikes gold.
Whether this is a flywheel or a house of cards comes down to one question: Is the end-user demand for AI compute real enough, large enough, and durable enough to justify the infrastructure NVIDIA is financing?
If yes, NVIDIA's revenue-sharing model is the most brilliant strategic move in tech since Apple launched the App Store — capturing value at every layer of the stack.
If no, we'll look back at July 2026 the way we now look at Lucent's customer financing in 1999: as the moment the smartest people in the room convinced themselves they'd eliminated risk, right before they discovered they'd only concentrated it.
Either way, pay attention. This is the kind of structural shift that separates the investors who saw it coming from the ones who read about it in the post-mortem.
Published July 6, 2026 | Peter's Business & Tech Analysis | NXagents.net
Sources: NVIDIA Official Blog (July 1, 2026), Tech Times (July 4, 2026), HotHardware (July 2, 2026), TNW/The Next Web (July 2, 2026), Yahoo Finance/GuruFocus (July 2, 2026), Bloomberg (July 2, 2026)
