OpenAI’s plan for its own AI chips covers 10 gigawatts of accelerators, which OpenAI and Broadcom announced on 13 October 2025. That is about 1.7 times the 6 gigawatts of AMD Instinct GPUs that OpenAI agreed to deploy a week earlier, by our arithmetic. Separately, on 3 November, AMD chief executive Lisa Su said OpenAI had picked AMD processors as host CPUs for its first custom accelerator platforms, without naming those platforms in our transcript of the call.
Su told analysts on AMD’s third-quarter call: “As one example, OpenAI selected Turin EPYC processors as the host CPUs for its first custom accelerator platforms”. In our transcript of the call, she did not name the platform, Broadcom, volumes or dates. Together, the two programmes come to 16 gigawatts, by our arithmetic.
Business Pill · Concentration Explained
An evergreen Business Engineer explainer on concentration. It teaches the general idea only and does not discuss AMD, OpenAI or Broadcom.
The key insight:OpenAI’s announced custom accelerator programme with Broadcom covers 10 gigawatts, about 1.7 times its 6 gigawatt AMD GPU deal by our arithmetic, and Lisa Su said OpenAI chose AMD’s Turin EPYC processors as host CPUs for its first custom accelerator platforms.
Under the 13 October 2025 announcement, OpenAI designs the accelerators and systems, which are “developed and deployed in partnership with Broadcom”. Broadcom is to deploy racks of accelerator and network systems “targeted to start in the second half of 2026, to complete by end of 2029”. The two companies said they had signed a term sheet. The announcement does not mention CPUs.
Greg Brockman, OpenAI’s president, said then: “By building our own chip, we can embed what we’ve learned from creating frontier models and products directly into the hardware”. Sam Altman said developing its own accelerators “adds to the broader ecosystem of partners”.
AMD’s 6 October 2025 release describes a “multi-year, multi-generation agreement” for 6 gigawatts of AMD GPUs, starting with a first 1 gigawatt of Instinct MI450 GPUs in the second half of 2026. AMD issued OpenAI a warrant for up to 160 million AMD shares; “The first tranche vests with the initial 1 gigawatt deployment”. The release gives no end date for the full 6 gigawatts.
OpenAI’s two announced chip programmes: 10 gigawatts of OpenAI-designed accelerators with Broadcom (announced 13 October 2025, targeted to start in the second half of 2026 and complete by end of 2029) and 6 gigawatts of AMD Instinct GPUs (announced 6 October 2025, first 1 gigawatt from the second half of 2026, no end date given). Lisa Su said on 3 November 2026 that OpenAI picked Turin EPYC as host CPUs for its first custom accelerator platforms. The 1.7x ratio is our arithmetic.
What Su Said on the Call
In her prepared remarks, Su said: “In head nodes, we won major deployments with several of the largest model developers and cloud providers in the quarter, which will ramp over the coming quarters”. She added: “EPYC’s leadershipperformance and mature ecosystem, proven at scale, have made it the host CPU of choice for custom accelerators”. The OpenAI line followed.
A host CPU, or head node, is the server processor that sits beside the accelerators in a system and feeds them work. As we read it, that role does not depend on whose accelerator is in the rack.
On the GPU side, Su said OpenAI was among the lead customers that “brought their first Helios racks online quickly”, and AMD’s release lists OpenAI among companies running workloads on the first Helios rack systems.
The ASIC Question in the Q&A
Bank of America analyst Vivek Arya asked how frontier labs “make the choice between using ASICs versus AMD products”. Su answered: “I think we feel extremely good about where we’re positioned in terms of our performance and our total cost of ownership that we’re delivering with Helios”.
She said “the key is for us to continue to show that TCO advantage across workloads”, and: “It is true that from a general purpose, GPU standpoint, that we have more fungibility as you think about various workloads that occur”. In our transcript, she did not name OpenAI in that answer.
The Arithmetic
10 gigawatts of OpenAI-designed accelerators divided by 6 gigawatts of AMD GPUs is 1.67, or about 1.7 times, by our arithmetic. The custom programme is 4 gigawatts larger, and the two together come to 16 gigawatts. The first AMD tranche of 1 gigawatt is about 17% of the 6 gigawatt deal.
Both programmes were announced to start in the second half of 2026. Only the Broadcom programme states an end point, the end of 2029.
Why the Host CPU Matters to AMD
Su said on the call: “We now expect server revenue to double year-over-year in 2027”. Answering Morgan Stanley’s Joe Moore on server growth, she described the addressable market as growing at “about 50% CAGR over the next three or four years” and said AMD saw growth in three areas: traditional cloud and enterprise, agentic CPUs and workloads, and head nodes. “We believe we’re gaining share in each one of those segments”, she said.
AMD’s Data Center segment revenue was a record $8.5 billion in the third quarter, up 97% from a year earlier, according to its earnings release. In our transcript, Su gave no revenue figure for the OpenAI host CPU win.
How Each Side Reads These Figures
For readers who see AMD gaining from OpenAI’s chip plans either way, Su placed AMD silicon in both: Instinct GPUs in the 6 gigawatt deal, and EPYC host CPUs for OpenAI’s first custom accelerator platforms.
For readers who see limits, the custom programme is about 1.7 times the AMD GPU deal by our arithmetic, in our transcript Su gave no units, revenue or dates for the host CPU win. Beyond the analyst questions quoted above, we cite no analysts’ views.
As we read the two announcements, OpenAI’s largest stated chip programme is the one it designs itself, and Su’s remark places an AMD part beside the first platforms of that kind.
As we read the call, Su answered the ASIC question on total cost of ownership and fungibility, while the host CPU line shows AMD also selling into systems built around custom accelerators.
Lisa Su, AMD Q3 2026 earnings call (our transcript)
“EPYC’s leadershipperformance and mature ecosystem, proven at scale, have made it the host CPU of choice for custom accelerators”
Three Implications
FOR OPENAI WATCHERS OpenAI’s announced chip programmes total 16 gigawatts by our arithmetic: 10 with Broadcom, targeted to complete by end of 2029, and 6 with AMD, starting with 1 gigawatt in the second half of 2026.
FOR AMD SHAREHOLDERS Su said AMD won major head-node deployments with several of the largest model developers and expects server revenue to double in 2027; in our transcript she gave no revenue figure for the OpenAI host CPU win.
FOR BROADCOM WATCHERS The OpenAI-Broadcom announcement describes Broadcom Ethernet, PCIe and optical connectivity in the racks and does not mention CPUs; Su did not name Broadcom.
The framework states: “The May 2026 Map of AI argued that the seven-layer AI stack is now governed from the bottom — energy, silicon, compute capacity.”
As we read it, OpenAI’s two programmes are silicon-layer commitments measured in gigawatts, and Su’s host CPU line shows a chip supplier placing parts on both sides of a lab’s choice between its own accelerators and a vendor’s GPUs. The framework was published in May 2026 and does not discuss this host CPU win.
In our transcript of the call, Su did not say whether the platforms she described are the Broadcom-built accelerators, how many processors OpenAI is buying, or when they ship. We set the two side by side only because the Broadcom collaboration is the custom accelerator programme OpenAI has announced; in the call transcript and the releases we read, neither AMD nor OpenAI links the host CPU win to the Broadcom programme.
Gigawatts here are the companies’ announced targets, not installed capacity, and the Broadcom term sheet is not a final contract as described. Quotations come from our own machine transcript of the call audio; it is not an official AMD transcript. We did not contact AMD, OpenAI or Broadcom.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
OpenAI’s announced custom accelerator programme with Broadcom covers 10 gigawatts, about 1.7 times its 6 gigawatt AMD GPU deal by our arithmetic. Lisa Su said on 3 November that OpenAI chose AMD’s Turin EPYC processors as host CPUs for its first custom accelerator platforms, giving AMD a stated role beside OpenAI’s own chips as well as its GPUs.
A note on sourcing. Gigawatt figures and timings come from the OpenAI-Broadcom announcement of 13 October 2025 and AMD’s release of 6 October 2025. Quotations from Lisa Su and analysts come from our own machine transcript (faster-whisper) of AMD’s third-quarter 2026 earnings call of 3 November, from the live webcast audio; it is not an official transcript. The 1.7 times ratio and the 16 gigawatt total are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
