OpenAI has lined up more than a trillion dollars of computing deals, according to the Financial Times, cementing Sam Altman’s bid to lock down the AI supply chain from chips to cloud to power. Partners including Nvidia, AMD and Oracle are on the ticket, underscoring how capital commitments, not just algorithms, are now the competitive moat in artificial intelligence.
The headline figure signals a shift from opportunistic GPU scrounging to long-dated, industrial-scale procurement. The company is tying up capacity across silicon, networking, data centers and cloud services to train and run ever-larger models. That playbook started years ago with Microsoft’s $1 billion partnership in 2019, giving OpenAI privileged access to Azure’s infrastructure. It’s now expanding into a web of prepayments, take-or-pay agreements and build-to-suit arrangements designed to guarantee throughput when the next model reaches training time. The direction is clear: compute is the currency, and OpenAI intends to control its exchange rate.
Nvidia (NVDA) and AMD (AMD) are positioned as primary beneficiaries. Nvidia’s accelerated compute stack remains the default for state-of-the-art training, while AMD’s MI300-class chips are breaking into deployments that prize price-performance and supply diversification. Oracle (ORCL), a core cloud partner for GPU-dense workloads, has leaned into specialized clusters, high-bandwidth networking and data gravity, making it a practical complement to Microsoft (MSFT) Azure in a multi-cloud strategy. Together, these vendors sell more than chips or rack space; they sell time-to-train and reliability at a scale measured in gigawatts. For OpenAI, overcommitting beats being short one procurement cycle. For suppliers, it means backlog, better pricing power and tighter integration.
The spending arc matches the industry’s capital curve. Financial analyses peg training costs for leading foundation models at roughly $646 billion and inferencing at $487 billion through 2032, a combined load that easily supports trillion-dollar commitments across the stack. The spend doesn’t stop at GPUs. It flows into optical interconnects, networking fabric, HBM memory, liquid cooling and the real estate and power contracts that keep clusters fed. Every new model class compounds the demand: training is episodic but intense, while inference is a metronome that never stops. That dual pressure explains the size and structure of OpenAI’s deals and why vendors prefer long-term, capacity-reserving agreements with meaningful prepayment.
The trillion-dollar tally rhymes with the Stargate supercomputer plan announced in January 2025, a $100 billion project that could scale to $500 billion by 2029. Stargate is less a single machine than a blueprint for a distributed AI utility—purpose-built campuses, hardened supply chains and energy contracts that resemble those of hyperscalers and chip foundries. OpenAI’s infrastructure moves in March 2025—expanding capacity by acquiring cloud provider CoreWeave in a five-year, $11.9 billion deal—fit the same pattern: secure compute, abstract complexity, then scale. CoreWeave’s IPO in the same month raised $1.5 billion, underscoring how public markets are willing to bankroll the physical backbone of AI when visibility on utilization is high.
OpenAI has also bolstered its balance sheet via secondary liquidity. Allowing employees to sell about $10.3 billion in stock at a $500 billion valuation signals investor confidence and gives the company flexibility to navigate a capex-heavy cycle. That kind of paper value, paired with committed spend and vendor financing, becomes a negotiating tool. Suppliers want anchor tenants with guaranteed burn, and OpenAI wants priority in constrained supply nodes—HBM availability, next-gen GPU allocation, and accelerator module lead times. Deals that blend capacity reservations, price locks and co-development workstreams align incentives and reduce the risk of training delays when the next model checkpoint arrives.
For Oracle, sustained AI demand validates a differentiated cloud strategy. High-density GPU regions, low-latency networking and customer willingness to colocate data and models translate into a stickier platform and a swelling backlog. Microsoft’s early stake gives it a durable edge as OpenAI’s preferred hyperscaler, channeling AI workloads into Azure and seeding product integrations across the Microsoft stack. The coexistence with Oracle is strategic: multi-cloud reduces single-vendor risk for OpenAI and creates price tension in negotiations. For investors, the tell will be forward capex guides, disclosures on AI-related backlog and the cadence of capacity coming online in 2025–2027.
The risks are execution and physics. Chip supply remains tight, HBM packaging is a bottleneck, and yields on bleeding-edge accelerators can wobble. Power is the harder constraint: siting new campuses near available generation and grid interconnects is increasingly challenging, and AI’s thermal profile is forcing rapid adoption of liquid cooling with its own cost and operational complexity. Competitive risk is real as well. Google, Amazon and Meta are all scaling custom silicon and tuned inference stacks. If model scaling hits diminishing returns or monetization lags, long-dated take-or-pay obligations could turn into a drag. Regulatory scrutiny over supplier concentration and energy usage could reshape the pace of deployment.
The near-term payoff is predictability. OpenAI reduces the risk of hitting a compute wall mid-training, suppliers lock in multi-year demand at premium pricing, and the ecosystem builds around known delivery schedules. The medium-term prize is faster model cadence and feature velocity, which should improve unit economics for AI products as inference efficiency rises. If training costs per incremental capability step fall slower than expected, the benefits shift to those who control the full stack—silicon, systems, data and distribution. That is the logic behind Altman’s bet: secure the inputs at scale, drive the learning curve down and make competitors pay up to catch up.
One number does not reprice a sector, but it refocuses it. A trillion dollars in compute deals crystallizes AI as an infrastructure cycle, not a software fad. For chipmakers, watch allocation priority and the ramp to next-gen accelerators. For Oracle and Microsoft, track AI backlog conversion and new region announcements. For energy and data center operators, expect more AI-specific projects with firm offtake and financing anchored by model training schedules. OpenAI’s commitments raise the bar for anyone claiming leadership in frontier models. The premium now goes to whoever can prove they will have the compute, when they need it, at the scale their roadmaps demand.
