OpenAI scores another chip deal but dodges the $10 billion question

Broadcom’s stock jumped on partnership news—then its chip chief said OpenAI isn’t the mystery buyer

Broadcom and OpenAI unveiled a 10-gigawatt, multiyear custom-chip partnership that starts deploying in late 2026, with full rollout by 2029. Shares of Broadcom surged on the news. A wrinkle arrived minutes later: Broadcom’s chip president, Charlie Kawwas, said on TV that OpenAI is not the “mystery” customer behind a separate $10 billion purchase order flagged in September. The partnership is real; the $10 billion PO belongs to someone else.

The Breakdown

• OpenAI and Broadcom announce 10-gigawatt chip partnership deploying 2026-2029; OpenAI isn't the separate $10 billion mystery customer from September earnings

• OpenAI now has 26 gigawatts committed across three vendors—hundreds of billions in obligations against $13 billion annual revenue and $10 billion yearly burn

• Broadcom stock jumped 10% on announcement despite undisclosed payment terms, adding $150 billion in market cap on "multibillion-dollar" deal value

• Strategy treats partnerships as financing: announce deals, create vendor equity value, push payment timelines years out while raising more equity to cover bills

The market shrugged. Broadcom still closed up nearly 10% after the announcement, underscoring investor hunger for any credible, multi-year line of sight into AI infrastructure revenue. OpenAI, meanwhile, added another plank to an already aggressive compute buildout schedule that has it spreading bets across several chip vendors.

What’s actually new

Kawwas’s on-air clarification matters. It separates Broadcom’s OpenAI partnership from the September $10 billion order and implies two distinct demand streams: OpenAI’s co-designed hardware path and an undisclosed hyperscale buyer writing very large checks. Analysts have long pointed to Google, Meta, or ByteDance as Broadcom’s biggest web-scale customers; OpenAI is now a separate, confirmed lane.

Substantively, OpenAI is moving from buyer to co-designer. The company says it will bake “what we’ve learned from building frontier models” into silicon, while Broadcom builds racks and networking to match. It’s a shift from off-the-shelf accelerators to systems tuned for OpenAI’s inference workloads. The companies say first deployments land in the second half of 2026, with completion by the end of 2029. Timelines slip in this industry. But that’s the plan.

The claim vs. the math

OpenAI’s chip capacity commitments now total 26 gigawatts across Broadcom, Nvidia, and AMD. That dwarfs its current business: OpenAI expects about $13 billion in revenue this year and has indicated it won’t be cash-flow positive until late in the decade. The company’s internal north star is far larger—250 gigawatts by 2033, an ambition that third-party estimates peg at $10 trillion-plus at today’s build costs. Those are staggering gaps between vision and income statements.

That’s why the financing narrative looms as large as the technology. Each big-ticket commitment binds suppliers to OpenAI’s fate and, often, lifts their market caps on impact. The Financial Times’ tally suggests OpenAI’s spree could add $350–$500 billion on top of roughly $1 trillion in recent chip and data-center deals, spreading risk—and leverage—across multiple balance sheets. It’s vendor dependency as funding mechanism.

Equity-creation arbitrage, in plain English

Here’s the trade: OpenAI announces a multiyear build with a public supplier; the supplier’s stock jumps immediately; cash outlays and manufacturing risk sit years in the future. The announcement itself creates value on the vendor side today, while OpenAI buys time to raise more equity or secure cheaper credit tomorrow. As Matt Levine has argued, if you owe one company a little, it’s your problem; if you owe the entire supply chain a lot, it’s their problem—and they’ll work with you. It’s brazen, but coherent, so long as demand materializes.

From Broadcom’s perspective, the calculus is straightforward: validate the custom-AI roadmap, lock in a marquee co-design partner, and let the stock rerate while engineering ramps. From OpenAI’s side, the co-design push helps de-risk Nvidia reliance, pressures costs over time, and signals to investors that it’s building systems, not just renting GPUs. The risk is obvious. If deployment slips or AI demand underwhelms, the “announce now, pay later” loop tightens.

The three tells to watch

1) Schedules. First-of-kind chips miss dates. If “late 2026” slides, capex schedules and revenue assumptions slide with it.

2) Fundraising velocity. A fresh OpenAI round—size, valuation, and structure—will reveal whether private investors will still subsidize multi-hundred-billion compute plans in 2026–2027 conditions.

3) Supplier terms. Today’s releases omit payment mechanics. If vendors begin requiring bigger deposits or milestone cash, OpenAI’s vendor-as-lender model morphs into traditional financing, dulling the equity-creation edge.

Bottom line

OpenAI keeps turning supplier partnerships into implied financing and time. Broadcom gets validation, pipeline, and a rally. And the $10 billion “mystery customer”? Still out there—and definitely not OpenAI, per Kawwas. The arms race continues.

Why this matters

OpenAI is using partnerships as capital. Multiyear chip commitments create supplier dependence that substitutes for near-term cash, shifting risk into the ecosystem.
Investors are underwriting timelines, not just tech. Stocks jumped on announcements while delivery and financing stretch to 2026–2029, making execution risk the new macro.

❓ Frequently Asked Questions

Q: Who is Broadcom's actual $10 billion mystery customer if it's not OpenAI?

A: Broadcom doesn't disclose web-scale customers, but analysts point to Google, Meta, or ByteDance—its three largest hyperscale clients who buy billions in custom AI chips annually. These companies write actual purchase orders with payment terms, unlike OpenAI's announcement-first approach. The September $10 billion order likely came from one of these established buyers.

Q: How much computing power is 26 gigawatts actually?

A: Twenty-six gigawatts would meet New York City's summer electricity needs more than twice over. For context, the entire U.S. data center industry uses roughly 17 gigawatts today. OpenAI's commitments across three vendors equal 150% of current national data center power consumption—a staggering bet on exponential AI demand growth.

Q: Why build custom chips instead of just buying more Nvidia GPUs?

A: Custom chips let OpenAI embed learnings from running ChatGPT and Sora directly into hardware, optimizing for inference workloads rather than general-purpose computing. It also reduces Nvidia dependency—critical when you're competing with Google and Meta, who already run custom AI chips. Custom silicon takes longer but costs less at scale.

Q: How does OpenAI's "vendor dependency as financing" strategy actually work?

A: OpenAI announces massive deals, vendor stocks rally (Broadcom added $150 billion in market cap Monday), and payment obligations sit years out. Vendors become invested in OpenAI's success through stock gains and future revenue expectations. When you owe hundreds of billions across multiple suppliers, they'll work with you on terms—it's their problem too.

Q: What happens if OpenAI can't pay for all these chip commitments?

A: OpenAI would need to raise more equity (diluting current shareholders like Microsoft and SoftBank) or restructure payment terms with vendors. The company's $500 billion valuation provides cushion, but with $10 billion yearly burn and cash-flow positive targeted for "late decade," missing revenue projections could force renegotiations or payment delays starting 2027-2028.