OpenAI has committed to buy more than 10 million high-end processors across Nvidia, AMD and Broadcom, an order book tied to roughly 26 gigawatts of compute capacity that would drink power on the scale of about 20 nuclear reactors. The spending dwarfs its revenue. The company expects billions in losses this year on about 13 billion dollars in sales and does not see profitability until 2029. That gap is forcing novel financing. Nvidia is talking investment and leasing. AMD is dangling equity-like warrants. One analyst called it circular financing. Another warned CEO Sam Altman has the power to crash the global economy or deliver the promised land. Markets now have to underwrite the difference.
The headline number is the power draw: 26 gigawatts of AI infrastructure in under a month. That implies more than 10 million units from Nvidia, AMD and Broadcom and data center buildouts that rival hyperscaler expansions. For context, most cloud regions today measure capacity in single-digit gigawatts. OpenAI’s purchases alone would pressure component supply, power grids and data center construction timelines. It also tests the capacity of foundries and advanced packaging. Even for Nvidia, the world’s hottest chip vendor, meeting those orders requires sustained output at TSMC and partners. For AMD and Broadcom, it is a once-in-a-decade chance to pry share from Nvidia’s dominant stack. The business question is simple: who pays, on what terms, and how long until the compute earns its keep.
OpenAI is forecast to lose billions this year and has told investors it does not expect to break even until late in the decade. That forces creative structures. Analysts and investors have latched onto a circular loop: suppliers invest in OpenAI, which then buys the suppliers’ products, and suppliers recoup their outlays via equity appreciation and long-term purchase agreements. Nvidia has floated up to 100 billion dollars of investment over several years tied to 10 gigawatts of capacity for OpenAI, according to people familiar with the plans. In parallel, the two sides have discussed a chip-leasing model, reported by The Information, that would let OpenAI access GPUs without paying cash upfront. The appeal is obvious in a high-rate world and a hot hardware cycle. The risk is equally clear: if AI demand or yields disappoint, the loop tightens at the wrong time.
AMD’s approach is even more unusual. OpenAI has lined up a multiyear plan to deploy roughly six gigawatts of AMD GPUs, and in return AMD granted warrants that could equate to as much as 160 million shares, or about a 10 percent equity stake, as deployment and adoption milestones are hit. The final tranche is tied to AMD’s stock price reaching 600 dollars a share, implying roughly a 1 trillion dollar market cap. A UBS analyst estimated that if OpenAI holds through final vesting and the price target is met, the stake alone could be worth about 100 billion dollars. That would effectively let AMD’s own stock appreciation subsidize OpenAI’s hardware bill. It is clever optionality for AMD, a shot at narrative parity with Nvidia, and a financing valve for OpenAI. It is also a bet on execution by both companies across silicon, software and customer uptake.
For Nvidia, investing in OpenAI and leasing chips could keep unit velocity high without blowing up customer budgets. Leasing allows faster deployment and smoother cash flows, and it helps Nvidia manage allocation among top buyers without forcing immediate capex checks. But leasing transforms Nvidia, in part, from an equipment vendor to a capital provider with duration risk. Margins, accounting and regulatory scrutiny shift when a chip sale becomes a long-dated lease backed by residual values and counterparties. If OpenAI later slows purchases, delays deployments or pivots to custom silicon, Nvidia would own more balance-sheet risk than in a traditional sale. That is manageable if the broader AI demand curve holds and secondary markets absorb used accelerators. It is painful if the upgrade cadence slows or software stacks leave older hardware behind.
Even at a 500 billion dollar private valuation, selling equity will not fund this build entirely. Gil Luria at D.A. Davidson estimates OpenAI will need hundreds of billions of dollars to honor current obligations. That pushes debt. One avenue floated by investors is using chips as collateral. In a world where H100s and their successors clear at premium prices and wait lists persist, lenders might accept accelerators as security. The terms would hinge on depreciation curves, firmware locks and secondary demand. Another lever is prepayment from anchor customers. If corporate buyers commit to multiyear AI capacity, OpenAI could securitize those contracts. Each route carries execution risk. The company also lacks the non-AI ad cash machines that Google and Meta use to self-fund. Without that internal flywheel, OpenAI must convince capital markets that its usage growth converts to durable cash flow.
Chipmakers are leaning in because the upside is enormous. For Nvidia, OpenAI is not just a customer; it is a showcase that pulls the entire software ecosystem toward CUDA, networking and systems it controls. For AMD, the warrant package is a shortcut to relevance, a way to bind a marquee tenant and validate its MI300 roadmap. For Broadcom, custom accelerators and networking silicon attached to OpenAI projects could lock in multi-year revenue and justify outsized capex at packaging and substrate partners. All three gain pricing power if OpenAI’s appetite sets market-clearing levels. Yet all three also take on concentration risk. If OpenAI changes course or the AI workload mix shifts, backlog can morph from asset to drag. The suppliers’ calculus is that diversification across hyperscalers, enterprises and sovereign AI programs will offset any single-buyer wobble.
Comparisons to the dot-com era are inevitable when spending outruns revenue. The differences are material. There is real, daily AI usage, with ChatGPT counted in the hundreds of millions of users and enterprises piloting and deploying AI copilots and agents. Harvard’s Josh Lerner has argued that demand this time is concrete in a way 1999 was not, even if sizing the payoff remains guesswork. Still, the mechanics rhyme with past cycles: vendors financing customers, customers stretching to meet perceived land grabs, and public markets extrapolating first-wave winners. Bernstein’s Stacy Rasgon captured the polarity in a recent note: Sam Altman could either crash the global economy for a decade or take us to the promised land. That duality is precisely why credit terms, not just unit shipments, matter now.
There is also a hard cap the 1990s never hit: power and land. Building 26 gigawatts of compute is not just a purchase order; it is a utility problem. Even with aggressive efficiency gains, AI clusters demand dense power, advanced cooling and proximity to fiber. Utilities are already signaling delays and grid upgrades in key regions. Data center developers are bidding up parcels near substations and chasing modular, on-site generation. If power delays push project timelines, financing costs rise as commitments sit idle and depreciation clocks start ticking. That is an underappreciated risk for both OpenAI and its suppliers, who are forecasting revenue against delivery schedules that depend on municipal approvals and transformer lead times, not just fab capacity.
Near-term, watch three things. First, the structure and scale of any committed debt against hardware or capacity contracts. A secured term loan backed by accelerators would formalize chips-as-collateral and set a pricing benchmark. Second, the details of Nvidia’s leasing program. The accounting and residual assumptions will tell you how much risk the vendor is absorbing to keep shipments flowing. Third, AMD’s warrant vesting cadence. If early tranches trigger on time, it signals OpenAI’s deployments are hitting milestones and customers are taking capacity. If not, the equity subsidy shrinks. Also watch whether OpenAI keeps diversifying suppliers to reduce dependence on a single stack. Competitors with ad cash cows can fund their own arms race from operating income. OpenAI has to make the numbers work in capital markets. That is the difference between a hype cycle and an investable buildout.
