Nvidia now underwrites the concrete: $105B of residual value on capacity it will never own

Nvidia has agreed to guarantee the residual value of the buildings on the first 4.25 gigawatts of OpenAI's new Ohio campus, capped at $105 billion, in return for becoming the exclusive chip supplier across that half of the site and putting $1.5 billion into SB Energy ^[17][4]. It is not guaranteeing rent, which is the part that matters: Nvidia has not underwritten OpenAI's ability to pay, it has underwritten the resale market for empty server halls at the end of a six-year build that runs to 2032 ^[17][1].

The lease itself is 20 years, signed with SB Energy, SoftBank's power subsidiary, for the PORTS-Pike campus, roughly 8GW of compute capacity ^[2] and 10GW gross once cooling and supporting infrastructure are counted ^[6]. The site is a decommissioned uranium enrichment property in Pike County, and is the largest data centre project announced to date ^[5]. The guarantee sits at the end of a waterfall: if OpenAI walks, SB Energy has to find a replacement tenant first, then try to sell the facilities, and only then does Nvidia cover the value gap, and only on the first 4.25GW ^[3]. Divide the cap by the tranche and you get about $24.7 billion per gigawatt of guaranteed shell ^[14].

Compare instruments before comparing sizes. Google took the other side for Anthropic, guaranteeing the rents themselves on several data centre leases, an exposure Alphabet's quarterly report put at $43.8 billion as of 30 June ^[19]. Nvidia's cap is roughly 2.4 times that number ^[22], but it is insuring the market rather than the counterparty.

The $105 billion figure did not come out of an asset model. The structure originally contemplated a guarantee of about $250 billion across the whole project, and was cut after a 5% fall in Nvidia's stock and investor pressure, passing below the $120 billion discussed more recently ^[9]. That is a 58% reduction from the starting point ^[16], which makes the cap a measure of shareholder tolerance as much as of building value.

Set it against the sales side. Jensen Huang has talked about roughly 1.5 million GPUs per system generation, worth $150 billion to $200 billion in revenue ^[10], so the maximum guarantee is smaller than the expected revenue from a single hardware generation ^[13]. Huang puts OpenAI's total commitments at about 12GW of Nvidia compute by 2030, rising toward 16GW and some $600 billion if the option on Ohio's remaining 3.75GW is exercised ^[11]. He has also given the constraint a label, LPS, for land, power and shell, and argues those have replaced chips and networking gear as the binding limit on AI construction ^[7].

Note that the two accounts of this deal differ. The Neuron described Nvidia as backing roughly $105 billion in financing, vouching that OpenAI can repay, covering construction and lease costs but not the chips ^[18]. That is credit exposure to OpenAI. A residual value backstop is exposure to 2032 real estate. Operators pricing multi-year compute contracts should establish which one they are actually relying on.

Meanwhile the physical delivery is modest. Only 800MW is due in service in 2028 ^[1], about a tenth of the claimed compute capacity ^[15], against a total project cost that could exceed $500 billion including chips ^[20], on top of the $30 billion Nvidia has already invested directly in OpenAI ^[21]. One financial press analysis puts around $3 trillion of largely AI-related commitments off the balance sheets of nine large technology groups ^[12].

Watch whether the 3.75GW option is exercised, since that is where the $600 billion sits ^[11]; whether the cap moves again, given it has already moved twice under investor pressure ^[9]; and whether the 800MW actually energises in 2028 ^[1]. Jobs numbers, 35,000 construction roles through 2032 and 2,500 permanent ^[8], will be the visible proxy for whether the schedule is real.