When you put in 1 and out comes 100, is that really a circular model?
Jensen Huang addressed the topic last week at a Goldman Sachs conference. His remarks echoed those made at the end of August during NVIDIA’s quarterly results announcement. In broad terms: small investments generate large value.
A “durable” infrastructure? “We can always rent Volta and Ampere”
Beyond the fact that NVIDIA “knows how to run all the models,” this value is all the greater when it is “durable,” he explains. Today, one can still rent Volta and Ampere, he insists, in contrast to estimates that peg the current economic life of a GPU at a maximum of three years.
Jensen Huang reprises a figure released at the end of August: $60 billion as the indicative cost to install 1 GW of compute power. Assuming a six-year project, he says that each $10 billion tranche of investment, at the current pace, yields $50 billion. The takeaway: compute is highly profitable. So the race to acquire it is on. If one believes him, global spending could reach $3,000 billion to $4,000 billion by 2030.
To illustrate the value generated, the NVIDIA chief cited another reference from late August: about twenty AI companies were now surpassing $1 billion in trailing annual revenue. He mentioned Cursor, Figma, and Together AI.
The big piece: financial guarantees
To sustain demand, NVIDIA acts mainly at three levels. Among them are equity stakes in its customers. It would consider, for example, investing up to $10 billion in Anthropic’s IPO. There are also supplier credits; in other words, extended payment terms.
NVIDIA earns such margins on GPUs that these two mechanisms have so far consumed only a small portion of cash. The bulk of the risk is carried by a far larger mechanism: financial guarantees. The largest officially disclosed to date stands at $105 billion. It concerns the Portsmouth campus (Ohio), the planned site for OpenAI. NVIDIA is the exclusive compute supplier there.
Revenue sharing with the neo-clouds
To turn compute power into a long-term financeable asset, NVIDIA has struck deals with major asset managers. Specifically Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The result should be vehicles that raise “more than $500 billion” to finance infrastructure construction*.
NVIDIA says it is all the more confident because it has visibility into these projects. By the end of August, it cited Portsmouth campus as an example. In partnership with SoftBank Energy, it helped secure land, energy, and real estate for an initial deployment expected to be around 4 GW.
This tracking of client activity makes NVIDIA a “well-informed investor,” to borrow Jensen Huang’s words. It also places its bets in the neo-clouds (“small” providers specialized in AI clusters), with a revenue-sharing model. It offers a firm take-or-pay commitment on part of the data-center capacity, guaranteeing a minimum revenue. This allows neo-clouds to secure their bank financing without locking up all of their capacity with a single customer. In exchange for this guarantee, NVIDIA captures a share of revenues above the fixed minimum. So, in the end, it’s a double stream of income, without directly lending out money.
A quarter with $89 billion for NVIDIA’s data-center business
In NVIDIA’s accounting, the datacenter segment accounts for roughly 93% of revenue (about $89 billion). The share from hyperscalers rose in the last quarter (+13%, to $49 billion), but grew more slowly than the share from the rest of the customer base (+25%, to $40 billion). Dell is among them, having just raised its annual earnings guidance in light of the size of its AI server order book.
In this context, NVIDIA is aiming for about a 70% revenue increase for its next fiscal year (roughly 2027). In the current year, its customer base remains highly concentrated: in the first half, three players accounted for nearly half of revenue. The mix could diversify as AI labs and clouders develop their own chips. For instance, AWS has just announced a multi-year deal worth $4 billion with Qualcomm.
* The Wall Street Journal recently counted about $3,000 billion in off-balance-sheet AI commitments.
