Technology & SocietyAugust 11, 2026

NVIDIA Is Building a Market Beyond Its Biggest Customers

By Oguzhan Tekin•Back to Blog

On August 10, 2026, NVIDIA announced a plan with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize more than $500 billion for AI infrastructure. The fine print matters. These are memorandums of understanding, not funded commitments. The $500 billion is an objective, not money already available.

The ordinary explanation is real. AI data centers require enormous spending before they earn revenue. Buildings, power, cooling, networking, and GPUs must be paid for first; usage fees arrive later. Infrastructure funds know how to turn those future payments into financing today. NVIDIA is using their balance sheets and distribution networks instead of becoming a bank.

Then consider the scale. The Bank for International Settlements projects that the entire private-credit market for AI companies could reach $300 billion to $600 billion by 2030. NVIDIA's target is roughly the size of that whole market. This is not a chip company looking for one more source of loans. It is a company trying to shape how a market is built.

For NVIDIA, the immediate benefit is more buyers. Cheaper capital lets more projects order GPUs, networking, and software now rather than wait for their own cash flow. NVIDIA gains hardware sales and a larger CUDA installed base, while much of the construction, credit, and refinancing risk sits with project owners and investors. Finance turns customers who cannot afford the stack today into customers who can sign for it over time.

The hyperscalers are absent from this vehicle, not from NVIDIA's orbit. BlackRock already works with Microsoft in the AI Infrastructure Partnership, where NVIDIA served as technical adviser. NVIDIA also announced a multiyear Meta partnership covering millions of Blackwell and Rubin GPUs. Microsoft, AWS, Google, and Meta remain essential customers and partners. The interesting choice is to build a separate financing channel outside them.

Why build that separate channel? The four hyperscalers can finance their own data centers, negotiate from enormous scale, and design alternatives — Microsoft's Maia, AWS Trainium, Google's TPU, and Meta's MTIA. They are customers today and potential substitutes tomorrow. Enterprises, governments, AI labs, and independent clouds are different. Most cannot design chips. If financing makes their AI factories possible, they are more likely to buy the complete NVIDIA stack. So who would actually use this? The likely buyers are independent cloud companies like CoreWeave, national AI programs, and large firms building their own capacity — plus the small number of AI labs big enough to sign a long contract but not big enough to fund one alone. NVIDIA does not need many of them: at roughly $2 billion for a large site, $500 billion is about 250 projects.

A simple example. Say you run a small AI lab. You need a data center with 100 megawatts of power. The building, cooling, networking, and chips cost about $2 billion. You have $300 million.

Today you have three choices, and none are good. Sell a large share of your company. Rent from a big cloud that also competes with you. Or build something much smaller and fall behind.

Under this plan, a fund sets up a separate company. That company owns the building and the chips, not you. The fund puts in part of the money and lenders provide the rest. You sign a contract to use the capacity for five years and pay every month. You get the compute now, without raising $2 billion and without giving away a piece of your company.

The trade is real. Five years is a long time in AI. If a much faster chip arrives in year two, you are still paying for the old one. Most contracts make you pay even when the machines sit idle. And your price depends on how safe lenders think you are, so a young lab pays more than a large, steady business.

One thing is easy to miss. Money is often not the hard part. A site with power already connected is harder to find than cash or chips.

This resembles Meta's strategy. Meta gives away open models because models are not where Meta earns most of its money. Making that layer cheaper pressures companies that sell model access while strengthening Meta's advertising and distribution businesses. NVIDIA is making a similar move from the other direction. It is not giving away chips; it is making capital around the chips easier to obtain. Meta commoditizes models to protect its platform. NVIDIA commoditizes access to finance to expand the compute platform it controls.

The monopoly question needs precision. This is not proof of an illegal monopoly, and NVIDIA still faces AMD and custom silicon. But it is an attempt to extend market power. If lenders begin treating NVIDIA systems as the safest or most liquid compute collateral, rivals must compete on financing as well as performance. NVIDIA would influence not only which chips enter a data center, but which data centers are easiest to fund.

The telecom bubble gives the clearest warning. Lucent committed about $8.1 billion in vendor financing and Nortel about $3.1 billion, while Cisco's smaller program promised roughly $2.4 billion in customer loans. When cash-strapped telecom customers failed, Lucent and Nortel absorbed the more serious vendor-credit losses. Cisco's best-known damage was instead a separate inventory write-down. The comparison does not predict NVIDIA's outcome; NVIDIA's major customers are far stronger. It shows why financed orders are not the same as independent end demand, and why the final payer matters more than the announced order.

The circularity is measurable. NVIDIA is a founding investor in KKR's Helix platform. Its non-marketable equity securities — stakes in other companies that do not trade on public markets — rose from $3.4 billion to $22.3 billion in fiscal 2026. There is nothing inherently improper about a supplier investing in customers or infrastructure. The risk appears when ecosystem funding is counted as proof that unaffiliated customers are willing and able to pay.

So which explanation wins? The boring one explains why the financing is needed: AI infrastructure has a genuine capital gap. The strategic one explains why NVIDIA is leading it and why this vehicle sits outside its largest customers. For now, vendor finance — a seller helping buyers pay for its own product — remains the better description. Compute becomes a separate asset class only after repeat deals establish independent valuation, secondary liquidity, and credible loss histories.

The test is not the $500 billion headline. Watch who supplies the cash, who signs the contracts, how much capacity unaffiliated customers use, and who absorbs losses when a GPU generation becomes less competitive. NVIDIA is helping finance the buyers — but the larger goal is to build a market beyond its most powerful customers and carry its dominance from the server rack into the financing of the data center itself.

References