The $500 Billion Nvidia Financing Push Is the Real AI War

Nvidia is building a financing machine that can buy the computers first and let the AI company pay later. On Aug 10, 2026, Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to set up independent compute-financing platforms designed to mobilize more than $500 billion of third-party capital for AI infrastructure over time. That figure is an aggregate mobilization target, not cash sitting in one Nvidia fund and not a signed final commitment. The partnerships remain subject to final agreements. The point is the structure. Jensen wants Wall Street to finance Nvidia systems the way it finances power plants, factories, planes, and ships.

A letter went with that move. On July 24, 2026, Microsoft hosted a coalition letter titled "Open Weights and American AI Leadership." It asked Washington to protect open-weight AI and to avoid premature restrictions that freeze the market around a few providers. Meta, Microsoft, Nvidia, and Palantir were among the launch signers. Anthropic, the company behind Claude, did not sign. That is the tell. OpenAI, Google, and SpaceX were absent at launch. They appear on Microsoft's later expanded list, which had grown past 270 names as of Aug 3. Elon publicly backed the letter. SpaceX, not Musk personally, is the later corporate signer. Farzad puts the late-August stack of the largest public companies on that letter at well over $20 trillion. That is his estimate, not a filing. The four launch-day public names sit closer to $11 trillion. Tesla and private OpenAI do not belong in that count. The financing platforms and the letter are the same bet: models should get plentiful and cheap, and the factories that print tokens should stay busy no matter which lab is winning this month.

What the $500 billion actually buys

A company that wants a giant AI system still needs land, a data center, power, cooling, networking, and racks of chips. That bill can hit billions, sometimes tens of billions, before the first dollar of product revenue. The largest tech companies can write that check. A startup, a smaller cloud, a drug company, or a country with real demand often cannot.

Under the new structure, the platform looks at the customer, the use case, the cash the work is supposed to throw off, and what the equipment might still be worth years later. If the numbers work, outside investors pay to build it. The customer pays for compute over time. Those payments go back to the investors. That is asset-backed finance. Lend against a useful asset and a stream of cash. Airlines do it with planes. Shipping companies do it with ships. Jensen wants to do it with intelligence factories.

In some cases Nvidia may provide residual-value support covering up to 25% of an individual opportunity, assessed project by project. In plain English, Nvidia may eat part of the loss if the hardware is worth less later than the model assumed. That support is limited and optional. It is not a hard guarantee on the full $500 billion target. That is not charity. If you sell cars, financing sells more cars. If you sell the systems inside an AI data center, hundreds of billions of new financing creates customers who could never have bought those systems with cash.

Lenders do not need one model to win

The part most people miss is what the lender is betting on. The bet is not that OpenAI wins. It is not that Anthropic wins. It is that the chips stay busy.

Training a model is expensive. Inference is the meter that runs every time a user asks a question, writes code, studies an image, or completes a task. That work is tokens. Electricity goes in. Tokens come out. Jensen's compression: compute equals revenue. A half-empty data center is a bad loan. A factory that sells useful intelligence all day is a cash machine.

Would you rather finance a factory that only one model company can use, or a factory that can run thousands of closed and open models for startups, banks, governments, hospitals, robot companies, and clouds? The second factory has more customers. If one provider loses, another model takes the load. If one customer fails, the compute moves. Nvidia makes money if OpenAI wins, if Anthropic wins, if SpaceXAI wins, and if a Chinese open model gets good enough for thousands of companies to download. The winner at the model layer can rotate. The chips still run.

Open-weight models are how that demand gets created. A company can download the settings inside the brain, run it on hardware it controls, keep private data inside its walls, and stop depending on a closed lab's servers. Farzad named open-weight systems such as Moonshot AI's Kimi K3 in that bucket. Every new open model that is good enough is another reason to buy or rent the factory underneath it. When Jensen backs open models, it is not a vibe. It is demand creation for the hardware.

Four CEOs who win if intelligence gets cheap

In a July 23, 2024 essay, Zuckerberg wrote that selling access to AI models is not Meta's business model. The money is on the products above the model. That still describes ads and apps. It is not a 2026 statement. Meta has since begun selling API access to Muse Spark. If strong AI gets cheap, Meta can put more of it inside every product. It can give developers an open model, let the outside world improve it, and capture the value in the apps and hardware above it.

Alex Karp is looking at the same stack from the application layer. Palantir connects models to the private data and daily work of governments and large companies. A customer does not only need a smart brain. It needs software that knows which factory is short on parts, which unit needs help, which bed is free, and who is allowed to decide. On July 1, 2026, Karp said customers want control of their computers, models, data, and proprietary knowledge.

Satya Nadella has built Microsoft around a multi-model world. Azure offers thousands of models, more than 11,000 as of Microsoft's FY26 Q4 update on July 29, 2026. On the April 29, 2026 earnings call, Nadella said more than 10,000 customers had used more than one model on Foundry and 5,000 had used open-source models. Those are Microsoft's counts, not an independent census.

Elon's version is the whole stack. Grok at the model layer. Giant AI computers. X for distribution. He also has a coding product: Cursor. SpaceX closed its acquisition of Cursor's parent, Anysphere, on Aug 14, 2026. Cursor said it is joining the SpaceXAI team. That is a SpaceX and SpaceXAI deal, not a separate xAI merger and not a rebrand of Cursor into Grok. SpaceX, Starlink, Tesla, robots, vehicles, energy, and a path to put compute almost anywhere. If Grok is the smartest model, he wins. If a different model is better and it runs through a product or machine he owns, he still wins.

None of those players need one model to be the best forever.

Anthropic needs the opposite

Anthropic needs intelligence to stay expensive enough to fund a giant closed lab. Farzad's read is blunt: a company that lives on the price of intelligence being as high as possible takes a hit when the rest of the industry finances a world of cheap, plentiful models. Investor Gavin Baker said in mid-August 2026 that people he trusts told him Dario Amodei had said Anthropic might someday be the only private company left. An Anthropic employee publicly called that account false. It is a secondhand anecdote, not a company filing. The trend in the numbers is pointing the other way.

A few weeks before this video, Grok was still treated as second tier. On Aug 12, 2026, Artificial Analysis scored SpaceXAI's Grok 4.6 at 61 on its Intelligence Index, in line with OpenAI's GPT-5.6 Sol at max, behind Anthropic's Claude Opus 5 at 63 on max and Claude Fable 5 at 62 on max with fallback.

You can see the crowding without picking a winner. Farzad said that in early June only Anthropic and OpenAI had models scoring 51 or higher on Artificial Analysis's Intelligence Index, which mixes reasoning, coding, knowledge, and real work. Six weeks later, by the July 17 snapshot, six labs were above 50. During one eight-day stretch, July 8 through 16, 2026, the dollars per Intelligence Index task for near-frontier intelligence fell by two to three times. In mid-June, Z.ai's open-weight GLM-5.2 crossed 50, scoring 51 at max. On July 16, Moonshot AI's Kimi K3 hit 57. The race is moving toward a rotating set of frontier players and a much larger set of models that are good enough for most work, and that second group will soak up most of the tokens.

Most companies do not spend all day solving new physics. They summarize meetings, search files, sort documents, draft email, answer common questions, check invoices, write software, and point workers at the next action. If a smaller open model can do that job, keep the data private, and cost a fifth as much, the customer uses the smaller model. Software can make that choice automatically. The user may never know which brain finished which step.

When suppliers get easy to switch, their margins usually fall. A pure model company can spend billions to train the smartest system on Earth. Once a rival closes the gap, or an open model becomes good enough, more work leaves. Price per token falls toward the cost of running the chip, which is mostly electricity. The best models can still charge a premium for the hardest work. Training a model and renting it by the token is becoming a brutal standalone business. Jensen's financing system makes that problem worse, not better, for anyone whose valuation assumes scarcity.

The political fight is the economic fight

Frontier labs have a real safety argument. Once a powerful open model is out, the company cannot pull it back. Someone can strip the safety limits. A future system can make cyber attacks easier or help people build dangerous biological tools. Those risks deserve tests and rules.

Farzad said Anthropic's Dario Amodei, in a July 27, 2026 position, does not support a blanket ban on open models. He treats models without dangerous abilities as a public good. He wants mandatory tests for sufficiently capable systems, stronger limits on advanced chips going to authoritarian countries, and tighter action against industrial-scale distillation, which is when one model learns from another model's answers. Labs can see that as copying years of expensive work. A smaller lab can see it as a way to catch up.

Then add the money. A closed model company needs to recover research and training costs. Scarcity helps it charge more. A rule that makes it expensive to release or train a competing model may improve safety. It also protects the leader's margins. An infrastructure company has the opposite incentive. Nvidia wants more models and more users. Clouds want more workloads. Application companies want more choice. The firms financing the token factories want the machines running 24 hours a day.

That is why Nvidia, Microsoft, Meta, and Palantir signed the July 24 letter asking Washington to protect open-weight AI. OpenAI, Google, and SpaceX joined Microsoft's later expanded list. Anthropic stayed off it. The war over open models will decide who controls intelligence and whether the factories being financed today have a broad market of models and customers tomorrow. It is also about who owns the machines, who owns the work, and who gets the wealth when intelligence gets cheap to run.

The loop Jensen is trying to lock in is simple. More capital creates more chips. More chips create more models. More models make intelligence cheaper. Cheaper intelligence creates more demand. More demand makes the chips easier to finance. Easier finance creates more capital. He is trying to turn the falling price of intelligence into a rising volume of compute.

This Exclusive is from the long-form at https://www.youtube.com/watch?v=H0NNRZd_vK4.