OpenAI buys AMD’s future, pays in equity

Key Takeaways

• OpenAI committed 6 gigawatts of AMD chips, offering warrants for 160 million shares—roughly 10% of AMD—vesting as milestones and $600 share price hit.

• Deal adds to $1 trillion-plus commitments including Nvidia and Oracle, structured as supplier-financed operational spend rather than massive upfront capital requirements.

• AMD focuses on inference workloads where portability is easier, positioning as credible Nvidia alternative while OpenAI reduces single-vendor concentration risk in AI infrastructure.

• Circular financing model—suppliers investing in customers buying their products—raises sustainability questions as OpenAI's infrastructure liabilities far exceed its $13 billion revenue.

OpenAI has locked in a multi-year commitment to deploy six gigawatts of AMD GPUs—paired with a warrant that could give it roughly 10% of AMD. AMD shares spiked on the news. The warrant vests as OpenAI hits deployment milestones and as AMD’s share price rises, including a tranche tied to $600. Friday’s close was $164.67. The timing underscores a single message: OpenAI won’t be a one-vendor shop. Not anymore.

What’s actually new

AMD lands its biggest AI customer and the clearest path yet to chip away at Nvidia’s dominance. OpenAI gains a second source for inference—the real-time computations behind ChatGPT’s replies—reducing single-supplier risk as usage swells. The first one-gigawatt wave is slated for the second half of 2026 on AMD’s MI450 line, a head-to-head answer to Nvidia’s next generation. That’s soon in data-center time. It’s not tomorrow.

The warrant makes the partnership sticky. OpenAI benefits if AMD’s stock appreciates; AMD secures “tens of billions” in expected revenue while binding OpenAI to a multi-generation roadmap. AMD says this is definitive. By contrast, OpenAI’s separate Nvidia pact remains a non-binding letter of intent. Paper matters.

The circular financing architecture

OpenAI’s supplier deals now look like capital strategy as much as procurement. Nvidia agreed to invest $100 billion; OpenAI then uses that capital to buy Nvidia hardware. Oracle committed $300 billion of cloud capacity over five years, paid out in usage-based increments. Broadcom is building custom silicon. AMD adds equity warrants in exchange for guaranteed orders. It’s a loop.

That loop solves a near-term constraint. OpenAI told investors it expects to spend about $16 billion renting compute this year, rising to as much as $400 billion by 2029. Cash flow alone won’t carry that load. Tapping supplier balance sheets and the debt markets lets the company scale infrastructure ahead of revenue. It’s leverage, not magic.

For AMD, the warrant is a hedge with upside. If OpenAI’s scale materializes, AMD participates via equity as well as sales. If the ramp slips, shipped chips still generate revenue. Vesting depends on deployment volume, technical readiness, and share-price thresholds. Incentives are aligned—and gated. That’s deliberate.

The inference buildout accelerates

This deal is about inference, not training. Training is bursty and capital-dense; you buy massive parallel compute to grind for weeks or months. Inference is perpetual and scales with users. As ChatGPT’s weekly audience approaches the billion-user mark, the steady-state cost sits in serving those queries with low latency and reasonable unit economics. It never sleeps.

Nvidia’s grip on training won’t loosen quickly. CUDA, networking, and software tooling remain moats. But inference workloads are more standardized and easier to port, and cost per token rules the day. AMD is aiming right there. It’s a segmentation play, not a frontal assault.

There’s a physical constraint, too. Chips are only one bottleneck. Power, cooling, and construction schedules now set the pace. OpenAI’s Stargate program with partners plans sites in Texas, New Mexico, Ohio, and the Midwest. A second supplier means fewer single points of failure if any one vendor or facility slips. Redundancy is a feature.

The sustainability question grows louder

Tech giants will spend well over $300 billion on AI data centers this year. Amazon, Microsoft, and Google can finance from cash flow. OpenAI cannot. Structuring deals as operational spend, equity warrants, and supplier financing spreads the burden over time—but it also stacks future liabilities on a young income statement. That’s the trade.

Skeptics see bubble mechanics: suppliers investing in the customer that buys their gear, valuations underwriting capacity, capacity justifying valuations. Supporters counter that demand is visible in product telemetry, enterprise contracts, and API usage, even if monetization lags. Current conversion from free to paid hovers in the low single digits, and new lines of business—from enterprise features to shopping—are still maturing. The bill arrives monthly. So does the traffic.

The competitive map shifts

For AMD, this validates a multi-gen Instinct roadmap and reframes its AI story around booked demand, not just benchmarks. Management is now pointing to a path where AI revenue could ultimately clear $100 billion, though it has not put a date on it. Ambition is cheap; volume isn’t.

For Nvidia, the message is hedging, not abandonment. OpenAI’s Nvidia plan is larger on paper—10 gigawatts versus AMD’s six—and Nvidia remains the default for training. But credible scale on AMD creates pricing tension and architectural diversity. Meanwhile, hyperscalers already field first-party silicon. With Broadcom in the mix, hardware plurality becomes the baseline, not the exception.

The bigger picture: AI now operates at sovereign scale. Single agreements run to tens of gigawatts and hundreds of billions of dollars. Winning requires not just performance, but a balance sheet and a financing strategy. Smaller vendors can’t write $100-billion checks or offer dilutive upside to secure supply. Gravity favors giants.

The context line

Two weeks ago, OpenAI unveiled a $100 billion pact with Nvidia and a 10-gigawatt target. Today’s AMD deal adds six gigawatts and a different financing lever. Together with Oracle commitments and custom-chip work, OpenAI has mapped an infrastructure roadmap north of 20 gigawatts and roughly a trillion dollars in lifetime build costs. That’s more than rhetoric. It’s a plan.

Why this matters

Supplier, investor, and customer roles are collapsing into one loop—magnifying both upside and systemic risk across a handful of firms that now anchor the AI economy.
By multi-sourcing inference at scale, OpenAI elevates AMD into a credible counterweight to Nvidia’s lock on AI infrastructure, with likely effects on price, portability, and time-to-serve.

❓ Frequently Asked Questions

Q: What does "6 gigawatts" of chips actually mean?

A: It refers to the total power consumption of the processors, not their computational power. Six gigawatts equals the electricity needed to power roughly 4.5 million homes continuously, or about the entire state of Massachusetts. Each gigawatt of data center capacity costs approximately $50 billion to build and equip, meaning this deal represents roughly $300 billion in total infrastructure investment.

Q: How does the warrant vesting actually work?

A: AMD issued OpenAI warrants to buy 160 million shares at $0.01 each—essentially free equity. The warrants vest in tranches as OpenAI deploys each gigawatt of capacity and hits technical milestones. Critically, AMD's stock price must also rise for vesting to complete, with one tranche requiring shares to reach $600 (they closed Friday at $164.67). OpenAI can't exercise warrants until all conditions are met.

Q: What's OpenAI's actual financial situation?

A: OpenAI generates about $13 billion in annualized revenue but remains unprofitable. The company expects to spend $16 billion on compute capacity this year alone, potentially rising to $400 billion annually by 2029. With only 5% of its 700 million weekly users paying for subscriptions, OpenAI must either dramatically increase conversion rates, find new revenue streams, or rely on continued supplier financing and debt markets to fund growth.

Q: Why does the difference between "inference" and "training" matter?

A: Training builds AI models through massive parallel computation over weeks or months—expensive but episodic. Inference runs those models to answer individual user queries in real-time—cheaper per operation but continuous and scaling with usage. As ChatGPT approaches a billion weekly users, inference costs compound while training costs remain relatively fixed. Inference is also more standardized, making it easier to port between different chip vendors like AMD and Nvidia.

Q: How far behind Nvidia is AMD in AI chips?

A: AMD projects $6.55 billion in AI GPU revenue for 2025. Nvidia's data center division alone—which includes AI chips—generated $115 billion last year and is on track to double that this year. Nvidia controls over 70% of the AI chip market. AMD's advantage lies in offering a credible alternative for companies wanting to reduce single-vendor risk, not in matching Nvidia's performance or ecosystem breadth.