Based on NVIDIA’s official essay, “Securing the Infrastructure of Intelligence,” by Jensen Huang.
In his official essay “Securing the Infrastructure of Intelligence,” Jensen Huang addresses the circular-financing charge directly — and his rebuttal is stronger than critics allow, weaker than NVIDIA would prefer, and rests on a single load-bearing assumption that nobody can yet verify.
What Happened
In an official essay published by NVIDIA, Jensen Huang lays out what he calls the LPS model — land, power, and shell — as the next strategic resource layer in AI infrastructure, and he formalizes NVIDIA’s arrangement at the PORTS-Pike Technology Campus in Portsmouth and Pike County, Ohio. The essay is NVIDIA’s self-interested argument, not a neutral document, and every number in it should be read as such. NVIDIA is partnering with SoftBank-backed SB Energy to secure the physical substrate of the site; OpenAI is named as the tenant building on NVIDIA’s full-stack DSX platform. The broader arrangement is reported as still being finalized and could change.
Huang does not dance around the financing question. He writes it plainly: “Is this circular financing? No. OpenAI will pay the lease.” He then describes the structure of NVIDIA’s guarantee: it covers defined portions of lease and power payments, plus a residual-value commitment, over a twenty-year site life. The exposure phases in only as data centers come online between 2028 and 2030, and it declines as OpenAI makes payments. By NVIDIA’s own account, this is not a guarantee of the full cost of the site, nor of all of the tenant’s obligations.
The scale Huang attaches to this structure — ~$150–200B in potential NVIDIA revenue per generation of systems, ~$600B of NVIDIA compute through 2030 across OpenAI’s projected commitments — carries a caveat that matters: these are NVIDIA’s opportunity projections and sales-case figures. They represent what the buildout could mean for NVIDIA if every assumption holds. They are not booked revenue, not committed orders on record, and not a forecast any independent party has verified.
The key insight: Huang’s “no” to circular financing is stronger than critics allow — the guarantee is limited, declining, phased, and tied to a real long-lived asset — but his rebuttal concedes the substance of vendor financing while disputing the label, and it rests entirely on the assumption that NVIDIA’s compute stays fungible enough to re-tenant if OpenAI falters. That assumption is untested in a downturn.
The Structural Read
Start with the steelman, because it deserves one. The naive version of the circular-financing story imagines NVIDIA handing OpenAI cash that returns immediately as GPU purchases — manufactured demand with no real underlying asset. That is not what Huang describes. The guarantee is capped to defined slices of lease and power, not the full build cost. It declines as the tenant pays. It phases in over years, not all at once. OpenAI is contractually on the hook for the lease. And the site has a twenty-year life, meaning it can host multiple GPU generations and multiple tenants — it is not a one-product warehouse that becomes worthless after a single cycle. On these points, Huang’s “no” is a serious answer, not a dodge.
But steelmanning the rebuttal does not make it the whole story. Naming the structure honestly cuts the other way. Huang’s own essay explains why NVIDIA is providing this backstop: frontier labs are, in his words, “growing faster than their balance sheets and long-term credit profiles can support.” That is the textbook condition under which a supplier extends its own balance sheet to enable a customer to buy its product. Whatever label you prefer — vendor financing is the accurate one — the structure exists because the customer cannot finance the infrastructure independently. The semantic question Huang answers (“Is it circular?”) is easier than the structural one his rebuttal quietly rests on.
Jensen Huang — “Securing the Infrastructure of Intelligence” (NVIDIA, 2026)
“Frontier labs are growing faster than their balance sheets and long-term credit profiles can support… NVIDIA will back defined portions of lease and power payments, as well as provide a residual-value commitment over the twenty-year term.”
The load-bearing assumption underneath everything Huang argues is fungibility. His deepest point is that because CUDA constitutes a standardized, broadly adopted compute platform, the capacity at PORTS-Pike is re-tenantable: if OpenAI exits or defaults, NVIDIA’s ecosystem provides a pool of alternative qualified tenants, which makes the asset financeable and the residual-value backstop defensible. That is a genuinely reasonable argument today, when demand for AI compute is broad, every gigawatt finds a buyer quickly, and the resale market for infrastructure is liquid. The problem is structural: fungibility and residual-value assumptions are exactly the ones that fail first in a downturn, when demand thins, competitive models proliferate, and specialized infrastructure finds fewer buyers at the worst possible moment.
FDE Framework — Enabler Extending Downstack
The LPS Move Is a Stack Extension, Not Just a Financing Decision
NVIDIA began as a chip Enabler. CUDA made it a platform. Systems and NVL-scale racks extended it to infrastructure. The LPS framing — land, power, shell — pushes the stack all the way to the ground. By underwriting the physical substrate, NVIDIA is no longer just selling into a data center; it is defining what a data center is for its ecosystem. That is a competitive position, not merely a balance-sheet decision. The risk is that the further down the stack an Enabler extends, the more its returns depend on the operating assumptions of the customers it is enabling.
The LPS framing also does something subtle strategically. By coining a new term for the physical layer — land, power, shell — Huang signals that NVIDIA intends to treat this layer as a standardized resource within its ecosystem, the way it treated CUDA as a standardized compute layer. If that framing takes hold, NVIDIA becomes the party that defines, secures, and guarantees the ground beneath AI factories, not just the silicon inside them. That is a significant platform extension, and it is also the mechanism by which NVIDIA’s exposure to any single tenant becomes, in principle, diversifiable. In principle.
Three Implications
IMPLICATION 1 — THE GUARANTEE STRUCTURE MATTERS MORE THAN THE LABEL
Whether you call it “circular” or “vendor financing,” the operative question for analysts and investors is the same: what is NVIDIA’s actual exposure, under what conditions does it grow, and what does the residual-value backstop assume about resale demand? Huang has provided more structural detail than most vendor-financing arrangements disclose publicly. That detail supports a more precise risk assessment than the headline charge allowed — and a more precise one than NVIDIA’s own “not circular” framing invites.
IMPLICATION 2 — NVIDIA’S PLATFORM BET IS NOW LEGIBLE IN PHYSICAL INFRASTRUCTURE
The chips-to-systems-to-CUDA-to-factories-to-LPS progression shows a coherent platform logic: each layer down the stack increases switching costs and makes NVIDIA’s ecosystem stickier. The risk of this extension is symmetric — if CUDA’s installed base and the ecosystem’s breadth are the mechanism by which compute stays fungible, then any development that fragments the AI compute market (new architectures, sovereign alternatives, competing open standards) narrows the very resale market that makes the guarantee defensible.
IMPLICATION 3 — THE PROJECTIONS ARE A SALES CASE, AND SHOULD BE READ AS ONE
NVIDIA’s figures — ~$150–200B per generation, ~$600B through 2030 — are the party with the most to gain telling you what it hopes to earn. That does not make them wrong; it makes them a claim. The appropriate response is neither to adopt them as a forecast nor to dismiss them as fiction, but to track which assumptions (sustained AI demand, OpenAI’s financial trajectory, compute fungibility) are required for even a substantial fraction of them to materialize. Those assumptions are the story going forward, not the numbers themselves.
