Let me paint you a picture.
It's July 1, 2026. NVIDIA CFO Colette Kress publishes a blog post that reads like a bank's product launch — not a chipmaker's. The announcement: a shiny new thing called the "AI Compute Partnership." It's a revenue-sharing and credit-support financing model where NVIDIA effectively underwrites the GPU purchases of its own customers. Sharon AI and Firmus Technologies sign on as the first takers, promising to deploy up to 210,000 Grace Blackwell GB300 GPUs between them.
The same week, SoftBank announces SB Neo — a 10-gigawatt neocloud venture in the U.S. Aramco Ventures leads an $800 million round into Together AI. Baseten closes a $1.5 billion Series F. Four deals. Ten days. One unmistakable signal.
The neocloud gold rush? It's no longer funded by venture capitalists writing equity checks. It's funded by the chip vendor itself.
NVIDIA isn't just selling shovels anymore. It's financing the mine, taking a cut of the gold, and — here's the kicker — promising to buy back the shovels if nobody shows up to dig.
Welcome to the most fascinating and potentially dangerous financial experiment in Silicon Valley history.
Before we go deeper, let's define the terms Steve asked about. These aren't buzzwords — they're the architecture of a multi-trillion-dollar transformation.
A hyperscaler is a cloud provider operating at planetary scale. Think Amazon Web Services (AWS), Microsoft Azure, Google Cloud, and increasingly Meta. These companies build data centers the size of football stadiums, consume gigawatts of power, and deploy hundreds of thousands of servers.
What makes a hyperscaler different from just a "big cloud company" is the ability to scale compute elastically — adding and removing capacity in near-real-time to match demand. When you spin up a virtual machine on AWS, you're renting a sliver of a hyperscale data center.
The relevant number for 2026: Microsoft alone is guiding for $190 billion in capex — 95% of its projected cash from operations. Meta's capex guidance is $125-145 billion against $136 billion in expected operating cash flow. These companies are spending more on infrastructure than most countries' GDP. And yet, they're still not building fast enough.
A neocloud (or GPU-native cloud) is a new breed of cloud provider built from the ground up for AI workloads. Unlike AWS or Azure — which started as general-purpose compute platforms and bolted on GPUs later — neoclouds like CoreWeave, Nebius, Lambda, and Together AI exist for one reason: renting out NVIDIA GPUs to companies that can't get enough of them from hyperscalers.
The neocloud advantage is speed. Commercial real estate firm JLL notes that neoclouds "can deploy high-density GPU infrastructure within months compared to multi-year builds for hyperscale data centers." CoreWeave claims it can get the latest NVIDIA chips into production "in as little as two weeks from receipt."
This speed advantage is why Microsoft has committed roughly $60 billion to neocloud capacity, and Meta has committed up to $62.2 billion — deals extending through 2031-2032. Combined with deals from OpenAI and Anthropic, total hyperscaler commitments to neoclouds are estimated north of $145 billion.
That's not pocket change. That's a fundamental restructuring of who builds and owns AI infrastructure.
Vendor financing is when the company selling you equipment also provides the loan to buy it. Think GE Capital financing GE jet engines for airlines, or John Deere Financial financing tractors for farmers. The vendor becomes your bank.
For NVIDIA, this started as ad-hoc arrangements. In 2024, NVIDIA committed to purchasing CoreWeave's unsold data center capacity through 2032 in a backstop agreement worth $6.3 billion. It invested in CoreWeave equity. It joined funding rounds for OpenAI, xAI, and Nebius. These were bespoke deals — one-off, relationship-based, opaque.
The AI Compute Partnership changes that. It's a productized version of vendor financing. Any qualifying neocloud can apply. NVIDIA provides the credit backstop, the demand guarantee, and — in exchange — collects both the hardware sale price and a recurring percentage of cloud revenue.
Commercial leasing is the equipment-financing cousin of vendor financing. Instead of buying a $30,000 H100 GPU outright, a neocloud leases it. The lessor (often NVIDIA or a NVIDIA-backed entity) retains ownership. The neocloud pays monthly. At the end of the lease, the equipment returns to the lessor or the neocloud buys it at residual value.
The NVIDIA model blends vendor financing and commercial leasing into something new. The neocloud pays for the hardware (or finances it through NVIDIA-backed debt), but NVIDIA also guarantees a floor on utilization. If customer demand falls short, NVIDIA rents back the idle capacity at a predetermined rate. This is essentially a put option on GPU compute — NVIDIA absorbs the downside risk.
Here's where it gets fascinating — and where the critics sharpen their knives.
The pattern has a name the analyst community has settled on: circular financing. Here's how it works:
The same dollar cycles through the system three times.
NVIDIA manufactures the hardware. NVIDIA finances the buyer. NVIDIA participates in the revenue. NVIDIA buys back idle capacity. From a pure business model perspective, it's brilliant — NVIDIA captures margin at the manufacturing stage, the financing stage, the equity-upside stage, and the ongoing revenue stage. It's not a chip company anymore. It's a vertically integrated compute-finance conglomerate.
But the risks are not subtle.
If AI-native demand cools — if the hyperscalers decide they've overbuilt, if inference pricing crashes, if open-source models reduce the need for massive training clusters — NVIDIA absorbs the slowdown twice. First through declining hardware orders. Then again through evaporating cloud revenue shares. The circularity that multiplies returns on the way up multiplies pain on the way down.
Let's get concrete. What does it actually cost to run a neocloud GPU cluster, and where does the NVIDIA revenue share fit?
Based on industry figures, here's the monthly cost model for a 1,024-GPU H100 SXM5 cluster:
| Cost Component | Monthly Estimate |
|---|---|
| GPU capex amortization (3-year, ~$30K/GPU) | $853,000 |
| Power (1,024 GPUs × ~700W, $0.08/kWh) | $38,000 |
| Colocation and networking | $75,000 |
| Financing overhead and interest | $120,000 |
| NVIDIA revenue-share obligation (est.) | $50,000–100,000 |
| Total Monthly Cost | ~$1,136,000–$1,186,000 |
Now, here's the profit-and-loss at different utilization levels, assuming a competitive on-demand rate of ~$2.02/hour per GPU:
| Utilization | Monthly Revenue | Profit/Loss |
|---|---|---|
| 50% | $754,000 | −$432,000 |
| 70% | $1,056,000 | −$80,000 to −$130,000 |
| 75% | $1,131,000 | Break-even |
| 90% | $1,357,000 | +$170,000 to +$220,000 |
Read those numbers carefully. At 70% utilization — which is not terrible — the cluster is bleeding six figures a month. The NVIDIA revenue-share obligation sits inside the cost stack regardless of utilization. It's a fixed drag. When clusters run light, it shrinks the cushion. When they run heavy, NVIDIA collects its cut.
This is why neocloud pricing is not random. When you see on-demand rates spike or shift, it's often a reflection of utilization math — a provider with clusters at 55-65% has a financial gun to its head. It must push prices up, enforce minimum commitments, or steer customers toward long-term reserved contracts.
The Spheron Network analysis puts this in stark relief. CoreWeave's H100 SXM5 on-demand rate is approximately $6.16/hour. Independent aggregators like Spheron offer the same GPU at $2.54/hour on-demand and $1.43/hour spot. That gap — 2.4x to 4.3x — is the cost of NVIDIA-financed circularity baked into your GPU bill.
The official story: hyperscalers lease from neoclouds because they get faster access to the latest GPUs and better utilization rates.
The real story is on the balance sheet.
Microsoft is guiding for $190 billion in capex in calendar year 2026. Analysts forecast $200 billion in cash from operations over the same period. That's 95% of operating cash flow consumed by infrastructure spending. The $60 billion in neocloud agreements? Those costs are recognized as operating expenses over the life of the contracts — not as capex upfront. Microsoft essentially moved $60 billion of spending off its cash flow statement and onto a series of smaller, deferred opex payments.
Meta's situation is even tighter. With $125-145 billion in capex against $136 billion in projected OCF, Meta could be free cash flow negative this year. Its up to $62.2 billion in neocloud agreements — extending through 2032 — means annual opex payments averaging under $10 billion versus the massive upfront capex hit of building equivalent capacity internally.
This is accounting arbitrage, plain and simple. It's perfectly legal. It's disclosed in SEC filings. But let's not pretend it's primarily about GPU access or utilization optimization. It's about making the numbers work when your infrastructure appetite exceeds your cash generation capacity.
The bears on this trade argue: if hyperscalers are just offloading capex onto neoclouds, and neoclouds are financing that capex through NVIDIA vendor financing, and NVIDIA is backstopping the neoclouds with demand guarantees — then the entire AI infrastructure buildout is being funded by financial engineering rather than genuine end-customer demand.
That's too simplistic. There is real demand. Together AI has annual bookings above $1.15 billion. Baseten serves over 1 billion inference calls daily. Cursor, Decagon, and thousands of AI-native companies are consuming GPU compute at rates that would have seemed absurd 18 months ago.
But — and this is the key but — the financing structure means demand doesn't need to stay at current levels for the whole thing to work. It just needs to stay above the utilization floors embedded in the NVIDIA backstop agreements. And those floors? We don't know what they are. NVIDIA hasn't disclosed them. Neither have Sharon AI or Firmus.
NVIDIA is now exposed to AI demand at every layer of the stack. If the hyperscalers cut capex — and they will, eventually, because no company spends 95% of its operating cash flow on infrastructure forever — the impact cascades. GPU orders drop. Neocloud revenue shares shrink. NVIDIA's own equity stakes in neoclouds lose value. The demand backstops get triggered, leaving NVIDIA holding idle GPU capacity it has to pay for.
This is what the I/O Fund's Beth Kindig calls the "circular financing red flag." When your supplier is also your lender, your customer, your demand backstop, and your equity holder, there is no natural market-clearing mechanism. The price signals are corrupted. Nobody knows what anything actually costs.
Sharon AI, the first AI Compute Partnership signatory, already holds a separate revenue-share facility of up to $200 million with investor Digital Alpha. Portions of its future income are now pledged in two directions. Two different parties have claims on the same revenue stream.
This is not unusual in project finance — lenders take senior and junior positions all the time. But in project finance, the underlying asset is a toll road or a power plant with decades of predictable cash flows. A neocloud's "toll road" is GPU compute sold at market rates that can shift by 30-50% in a quarter. A neocloud that layers vendor financing on top of investor financing has narrowed its margin for error before serving its first token.
SoftBank's SB Neo — with its 10 gigawatts of planned capacity, backed by the group's $65 billion OpenAI investment — won't launch until fiscal year 2027 (ending March 2028). CoreWeave, Nebius, Together AI, and Baseten are operating and signing contracts today.
By the time SB Neo comes online, the incumbents will have locked in the best hyperscaler contracts, the most favorable power purchase agreements, and the deepest customer relationships. Power procurement and construction schedules — not capital availability — are the binding constraints on every gigawatt-scale plan. SoftBank's money can buy a lot of things, but it can't speed up the permitting process for a 10-gigawatt campus.
The risk is a capacity wall: a flood of vendor-financed GPU supply hitting the market simultaneously in 2028-2029, just as the hyperscalers' internal buildouts (which take longer but are enormous) also come online. If supply outruns demand, utilization rates crash across the board, NVIDIA's backstops get triggered en masse, and the circular financing machine runs in reverse.
If you're renting GPU compute — or planning to — the NVIDIA financing model changes your counterparty risk calculus. Here's your checklist:
Is your GPU provider paying NVIDIA a percentage of cloud revenue? If they won't tell you, that's an answer. A provider carrying a vendor revenue-share obligation has a cost floor built into every GPU-hour they sell you. It doesn't go away when demand drops. It shows up in your bill.
Ask for average cluster utilization. If the answer is vague — "we maintain high utilization across our fleet" — probe harder. A provider at 55-65% utilization has an urgent financial incentive to raise your rates or lock you into a long-term contract. A provider at 85%+ has room to compete on price.
A healthy spot market means excess capacity priced at genuine market rates. The absence of spot pricing — or spot prices suspiciously close to on-demand — means all capacity is committed to covering fixed costs. That's the provider's problem until it becomes yours.
Before signing any GPU commitment, find the cancellation clause, the notice period, and the minimum term. A discount that requires a 90-day exit window and a three-year minimum is priced to protect the provider's financing structure — not your flexibility. If your AI workload might shift in 12 months, don't sign a contract built for a 36-month depreciation schedule.
The gap between CoreWeave's $6.16/hour H100 and Spheron's $2.54/hour is not a rounding error. It's the premium you pay for a NVIDIA-financed, hyperscaler-backstopped, IPO'd-neocloud GPU versus a commodity GPU from an independent aggregator. For some workloads — mission-critical training runs requiring guaranteed capacity — the premium is worth it. For inference, experimentation, or batch processing, it almost certainly isn't.
Let me be blunt: this is both.
The brilliance is undeniable. NVIDIA has transformed a transactional hardware business into a platform business with recurring, usage-linked revenue. The AI Compute Partnership formalizes what was already happening through back-channel deals and turns it into a scalable, repeatable product. If utilization across vendor-financed sites stays above the contract thresholds, NVIDIA will have added a permanent recurring revenue layer to the largest hardware franchise in computing history. It's the GE Capital playbook executed in the AI era.
But GE Capital also nearly destroyed General Electric.
The danger is in the opacity. We don't know the revenue-share percentages. We don't know the backstop utilization floors. We don't know the real exposure if a neocloud partner faces financial distress. The entire structure depends on assumptions about AI demand that are, charitably, uncertain beyond 18-24 months.
The most important number in this story isn't the 210,000 GPUs Sharon AI and Firmus are deploying. It's not the $4.7 trillion market cap. It's not the $145 billion in hyperscaler-neocloud commitments.
It's the utilization rate at which these vendor-financed clusters break even — and whether that rate holds.
If it does, Jensen Huang will be remembered as the man who built not just the world's most valuable chip company, but the financial architecture that funded the AI revolution.
If it doesn't, the circular financing machine that multiplied returns on the way up will be remembered as the house of cards that multiplied losses on the way down.
Either way, the era of NVIDIA as a pure hardware company is over. The bank is open.
Peter is a Business Development Manager and mentor at NXagents.net. He covers the intersection of AI infrastructure, business strategy, and financial markets. This article is analysis and commentary — not financial advice.
Sources:
