- 20-year term (LPS support) → NVIDIA blog — NVIDIA supports PORTS-Pike LPS for about 4 gigawatts over 20 years
- 2028 to 2030 → NVIDIA blog — guarantee becomes effective in phases as data centers are placed in service
- through 2030 → NVIDIA blog — OpenAI’s committed NVIDIA deployments
- October 20 to 22, 2026 and Wednesday, October 21 → NVIDIA GTC — GTC Berlin conference dates and Jensen Huang keynote
A frontier AI lab can have soaring demand and a fast-growing product, yet still hit a wall no model architecture can solve: it cannot sign the 20-year land, power and shell contract an AI factory needs. NVIDIA’s newest commitment moves that risk onto its own balance sheet — and in doing so hands outsiders a rare, fully disclosed template for judging such deals.
NVIDIA has detailed a partnership with SB Energy to secure land, power and shell (LPS) at the PORTS-Pike campus in Portsmouth, Ohio, where OpenAI will be the tenant. Read the fine print and you get a working method for evaluating any AI-factory infrastructure guarantee — what is actually covered, what is excluded, and where the real financial exposure sits.
Why frontier labs stall on the site, not the silicon
NVIDIA draws a sharp line between two kinds of AI builder, the company says. The largest cloud providers and investment-grade enterprises have the balance sheets and decades-long contracts to secure LPS themselves. Frontier labs are different: extraordinary compute demand, but balance sheets growing slower than their revenue.
The practical upshot is that infrastructure financing, not model quality, decides who can scale. A lab that cannot show a 20-year power contract cannot light up the next GPU generation no matter how strong its research is. That is the bottleneck worth measuring before any capacity announcement.
That gap is why a lab’s growth can be capped not by algorithms or customers but by the physical site. NVIDIA frames compute as the constraint on intelligence, and the site as the constraint on compute. The first evaluation step is to locate exactly which bottleneck you are actually facing before reading a single term sheet.
Separate the land-power-shell promise from the compute
The announcement covers LPS — the real estate, the electricity, and the building shell — not the GPUs inside. NVIDIA is explicitly not guaranteeing the full cost of the site or all of the tenant’s obligations. Its support is limited to defined portions of lease and power payments plus a specified residual-value commitment.
So when you read “NVIDIA is securing the AI factory,” translate it precisely. The chipmaker is securing the container, not the contents. A tenant such as OpenAI still pays the lease; NVIDIA uses its scale and supply-chain visibility to lock the site so its own compute can be deployed there. The discipline mirrors how it already manages semiconductors.
For an investor reading the release, the distinction changes what to underwrite. A guarantee on the building is a real but bounded promise; a guarantee on the chips would be a far larger balance-sheet commitment. NVIDIA is explicit that it stops at the site, which keeps the exposure measurable.
Read the 20-year term as phased, not lump-sum
The LPS support runs for roughly 20 years and covers about 4 gigawatts, but it does not switch on all at once. The guarantee becomes effective in phases as data centers are placed in service between 2028 and 2030, and as the tenant pays, NVIDIA’s remaining exposure declines.
That phasing matters for risk. Treat the obligation as a declining curve tied to placement dates, not a single headline number. The dates are the mechanism, not decoration: a buyer or analyst should model exposure falling as capacity comes online, rather than assuming a fixed multi-decade liability from day one.
Phasing also protects the guarantor. NVIDIA’s obligation only matures as capacity goes live and the tenant pays, so the company is never exposed to an empty building. The risk curve slopes down, which is the opposite of an upfront lump-sum commitment.
Find the ceiling NVIDIA put on its own exposure
NVIDIA answers the obvious question — “is this circular financing?” — with a flat no. The tenant pays the lease; NVIDIA simply secures the site using demand visibility. The more useful diligence point is the residual-value commitment: a defined backstop, bounded by the specific payments it covers.
When you evaluate a similar guarantee elsewhere, look for that ceiling. Unlimited or vaguely worded backing is the red flag; a commitment scoped to lease, power, and residual value is the disciplined version. The absence of a stated cap should lower your confidence, not raise it.
The discipline mirrors how NVIDIA already manages its chip supply chain. It secures critical inputs where it has visibility into durable demand, rather than spreading thin. Applying the same logic to real estate is what makes the PORTS-Pike commitment look like procurement, not philanthropy.
Test whether the capacity can be re-let
The strongest protection on a 20-year site is not the guarantor — it is fungibility. NVIDIA argues its compute is versatile and broadly adopted, so if one tenant leaves, the capacity can be resold across its ecosystem of clouds, enterprises, and startups. CUDA is the lever that makes the hardware rentable and financeable.
For any site you assess, ask the redeployability question directly. A bespoke, single-customer build is a 20-year hostage to one tenant’s fortunes; a standardized, developer-dense platform is an asset with a secondary market. Versatility is what turns a liability into collateral.
The redeployability test also explains why NVIDIA keeps its platform standardized. A site wired for one lab’s custom stack would lose that option; a site running the same CUDA base as every other NVIDIA customer keeps a deep buyer pool. The secondary market is the quiet insurance policy on a 20-year term.
Do the gigawatt and GPU math yourself
The PORTS-Pike numbers are large enough to deserve their own check. Initial deployment is expected to provide 4.25 gigawatts of AI factory capacity NVIDIA blog. Each generation of NVIDIA systems there could mean roughly 1.5 million GPUs, or about $150 billion to $200 billion in NVIDIA revenue.
Zoom out and the picture grows. OpenAI’s existing and planned commitments represent about 12 gigawatts of NVIDIA compute, with room to expand to roughly 16 gigawatts if the PORTS-Pike arrangement extends beyond its initial 4.25 gigawatts NVIDIA blog. At those levels the opportunity is on the order of $600 billion of NVIDIA compute through 2030.
The 20-year clock is not really about one set of machines. Each generation placed at PORTS-Pike can be swapped for the next, so the site’s value compounds across upgrade cycles rather than expiring with the first GPUs. That is why NVIDIA frames the land as the durable asset and the compute as the renewable one.
Put PORTS-Pike next to the wider build-out
One site is easier to judge inside a portfolio. NVIDIA says most customers will keep securing their own LPS, and the vast majority of NVIDIA-powered sites will stay customer-financed. PORTS-Pike is selective: exceptional land where visible, durable demand can support multiple generations.
NVIDIA is also extending the model abroad. It just opened Indonesia’s first university AI center with Indosat and Universitas Gadjah Mada NVIDIA blog. “The Indonesia AI Center of Excellence reflects our long-term vision to position Indonesia as a nation that not only adopts AI but also develops and contributes AI innovations to the world,” said Meutya Hafid, Indonesia’s Minister of Communication and Digital Affairs.
Source: NVIDIA — GTC Berlin 2026 speaker roster
For hands-on exposure to the stack behind these sites, NVIDIA GTC Berlin runs October 20 to 22, 2026, with Jensen Huang’s keynote on Wednesday, October 21 NVIDIA GTC. The conference sessions span the five-layer AI stack, from energy and chips to physical AI.
Source: NVIDIA — GTC Berlin 2026 open-registration kit, from the PORTS-Pike announcement page
The selectivity is the tell. A guarantee attached to a one-off, single-generation site is weaker than one tied to a campus built for upgrade cycles. NVIDIA may extend PORTS-Pike to its remaining 3.75 gigawatts, so treat the headline 4.25 gigawatts as a floor, not a ceiling, when you model supply. The same logic applies to the broader push to cut AI power waste through higher-voltage standards NVIDIA backs 800V DC and the wider AI-factory financing wave NVIDIA and partners are unlocking NVIDIA and partners unlock AI-factory financing.
When the next capacity pledge lands, run this checklist before quoting the gigawatt headline. Confirm what is guaranteed, find the exposure cap, test resale, and check the generation plan. A deal that fails any one of those four questions deserves a discount, not a cheer.
The template outlives this one deal. Every frontier lab now racing to lock capacity will be judged by the same four questions — what is covered, where the exposure ceiling sits, whether the site can be re-let, and how the gigawatts convert to usable compute. The labs that answer them in public, as NVIDIA just did, are the ones worth underwriting.
