Nvidia becomes the AI central bank—will this deepen the AI bubble?
The Economist’s recent piece, “Nvidia is the central bank of AI,” really did spell out Nvidia’s AI finance for the UK market:
Nvidia took 30 years to reach a market cap of $1 trillion
Then it hit $2 trillion in just nine months
Now at $5.4 trillion, it’s one of the most valuable companies in the world, and analysts have even predicted Nvidia could generate annual revenue of $1 trillion by 2029
But The Economist’s discussion of growth here isn’t just about manufacturing process improvements and the CUDA moat. The real driver is Nvidia’s financial engineering in recent years—something it also touched on previously in “How Google turned chips and data centers into financing projects to sell TPU.”
Over the past three years, Nvidia has basically done two things:
Committed more than $70 billion in investment to startups
Provided financial support on the order of roughly $300 billion to customers, including loan guarantees, buybacks of idle capacity, revenue floors, residual value support, and more
This looks very similar to what Google did to sell TPU. And now Nvidia has gained a nickname in the market: “the central bank of AI.”
The essence of an AI central bank is that Nvidia can decide who gets cheaper loans, whether assets can be used as collateral, and whether there is enough liquidity. This is what Nvidia is doing to “financialize” computing power.
Nvidia’s current playbook is quite direct. First, it takes equity stakes in customers, turning buyers into a shared-interest community. Big customers like Amazon, Google, Meta, and Microsoft contribute half of Nvidia’s revenue, but they are also developing their own chips. Their costs may be only one-fifth to one-third of Nvidia’s, according to Bloomberg Intelligence, and by the end of the decade, custom chips could make up half of the AI processor market.
So Nvidia’s countermeasure is to make large equity investments. Last year it completed 90 deals; this year, so far, it has already done 60. Nvidia is also betting on open-weight model ecosystems—making major investments in Poolside and Hugging Face, for example—so more workloads can continue to run on its stack.
Second, Nvidia provides credit enhancement for “new cloud” providers (neocloud). Companies like CoreWeave face high borrowing costs because they aren’t investment-grade—a point we discussed earlier when covering Google’s TPU strategy. Nvidia does this via a chain of moves:
Direct investment: takes an 11% stake in CoreWeave and puts in more than $200 million
Promises to buy back idle capacity within a certain time window—up to $6.3 billion through 2032
Sets a “floor price” for data center revenue and shares in upside gains
With this kind of chain, financing costs for downstream players can fall, and they can afford to buy GPUs. The result is that Nvidia ships faster. As SemiAnalysis folks also said earlier, this is essentially “buying time” for lenders.
And then there are residual value/lease guarantees for mega projects. The most extreme recent case involves a massive data center in Ohio:
SoftBank’s SB Energy is responsible for building it; OpenAI will lease it for 20 years. Nvidia provides guarantees of up to $105 billion, and in return the project uses a large quantity of Nvidia chips—on the order of about 1.5 million processors.
Moreover, Nvidia isn’t paying OpenAI’s rent directly. Nvidia provides residual value guarantees on built assets such as the land, electricity, and shell. If the tenant leaves, can’t sublease, or can’t sell, Nvidia makes up the shortfall. Nvidia also invested $1.5 billion in SB Energy to lock in exclusive supply.
Finally, Nvidia has turned GPUs into an “investable asset class.” It partners with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, mobilizing more than $500 billion in third-party capital into AI infrastructure. Nvidia offers up to 25% residual value support for some investments, while also providing technical endorsement.
As mentioned earlier, Nvidia’s biggest customers are all becoming rivals. Cloud providers build their own chips, and OpenAI also has its own “Mexican chili” chips—so Nvidia has to strengthen its moat from elsewhere. That’s why Nvidia is using its own profits and cash flow to supply the broader ecosystem with something like quasi-public goods—enabling small and mid-sized businesses to access credit at lower cost for financing.
It’s similar to how automakers’ finance companies work: if they don’t lend, cars can’t be sold.
But the problem highlighted in the reporting is this: is it satisfying demand, or manufacturing demand? Nvidia has now moved along the thin line between “enabling demand” and “creating demand”—and it’s already close to crossing the line. If chip demand growth slows and supply comes online, causing prices to fall, Nvidia’s profits—and its ability to support customers—could be harmed alongside the customers.
That said, Nvidia can still weather this for now: it has roughly $99 billion in cash and securities, and about $200 billion in operating cash flow. The guarantees are spread over many years, and it can also find replacement tenants. But if this financial model keeps running—especially with the loops of cross-holdings and circular demand—could it eventually trigger a financial crisis? That is a very serious question.
Nvidia isn’t the only one doing something like this. AMD and Google, for example, have similar structures, but none entangles this deeply the way Nvidia does.
The Economist’s recent piece, “Nvidia is the central bank of AI,” really did spell out Nvidia’s AI finance for the UK market:
Nvidia took 30 years to reach a market cap of $1 trillion
Then it hit $2 trillion in just nine months
Now at $5.4 trillion, it’s one of the most valuable companies in the world, and analysts have even predicted Nvidia could generate annual revenue of $1 trillion by 2029
But The Economist’s discussion of growth here isn’t just about manufacturing process improvements and the CUDA moat. The real driver is Nvidia’s financial engineering in recent years—something it also touched on previously in “How Google turned chips and data centers into financing projects to sell TPU.”
Over the past three years, Nvidia has basically done two things:
Committed more than $70 billion in investment to startups
Provided financial support on the order of roughly $300 billion to customers, including loan guarantees, buybacks of idle capacity, revenue floors, residual value support, and more
This looks very similar to what Google did to sell TPU. And now Nvidia has gained a nickname in the market: “the central bank of AI.”
The essence of an AI central bank is that Nvidia can decide who gets cheaper loans, whether assets can be used as collateral, and whether there is enough liquidity. This is what Nvidia is doing to “financialize” computing power.
Nvidia’s current playbook is quite direct. First, it takes equity stakes in customers, turning buyers into a shared-interest community. Big customers like Amazon, Google, Meta, and Microsoft contribute half of Nvidia’s revenue, but they are also developing their own chips. Their costs may be only one-fifth to one-third of Nvidia’s, according to Bloomberg Intelligence, and by the end of the decade, custom chips could make up half of the AI processor market.
So Nvidia’s countermeasure is to make large equity investments. Last year it completed 90 deals; this year, so far, it has already done 60. Nvidia is also betting on open-weight model ecosystems—making major investments in Poolside and Hugging Face, for example—so more workloads can continue to run on its stack.
Second, Nvidia provides credit enhancement for “new cloud” providers (neocloud). Companies like CoreWeave face high borrowing costs because they aren’t investment-grade—a point we discussed earlier when covering Google’s TPU strategy. Nvidia does this via a chain of moves:
Direct investment: takes an 11% stake in CoreWeave and puts in more than $200 million
Promises to buy back idle capacity within a certain time window—up to $6.3 billion through 2032
Sets a “floor price” for data center revenue and shares in upside gains
With this kind of chain, financing costs for downstream players can fall, and they can afford to buy GPUs. The result is that Nvidia ships faster. As SemiAnalysis folks also said earlier, this is essentially “buying time” for lenders.
And then there are residual value/lease guarantees for mega projects. The most extreme recent case involves a massive data center in Ohio:
SoftBank’s SB Energy is responsible for building it; OpenAI will lease it for 20 years. Nvidia provides guarantees of up to $105 billion, and in return the project uses a large quantity of Nvidia chips—on the order of about 1.5 million processors.
Moreover, Nvidia isn’t paying OpenAI’s rent directly. Nvidia provides residual value guarantees on built assets such as the land, electricity, and shell. If the tenant leaves, can’t sublease, or can’t sell, Nvidia makes up the shortfall. Nvidia also invested $1.5 billion in SB Energy to lock in exclusive supply.
Finally, Nvidia has turned GPUs into an “investable asset class.” It partners with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, mobilizing more than $500 billion in third-party capital into AI infrastructure. Nvidia offers up to 25% residual value support for some investments, while also providing technical endorsement.
As mentioned earlier, Nvidia’s biggest customers are all becoming rivals. Cloud providers build their own chips, and OpenAI also has its own “Mexican chili” chips—so Nvidia has to strengthen its moat from elsewhere. That’s why Nvidia is using its own profits and cash flow to supply the broader ecosystem with something like quasi-public goods—enabling small and mid-sized businesses to access credit at lower cost for financing.
It’s similar to how automakers’ finance companies work: if they don’t lend, cars can’t be sold.
But the problem highlighted in the reporting is this: is it satisfying demand, or manufacturing demand? Nvidia has now moved along the thin line between “enabling demand” and “creating demand”—and it’s already close to crossing the line. If chip demand growth slows and supply comes online, causing prices to fall, Nvidia’s profits—and its ability to support customers—could be harmed alongside the customers.
That said, Nvidia can still weather this for now: it has roughly $99 billion in cash and securities, and about $200 billion in operating cash flow. The guarantees are spread over many years, and it can also find replacement tenants. But if this financial model keeps running—especially with the loops of cross-holdings and circular demand—could it eventually trigger a financial crisis? That is a very serious question.
Nvidia isn’t the only one doing something like this. AMD and Google, for example, have similar structures, but none entangles this deeply the way Nvidia does.