0

OpenAI is matching chips to workloads instead of picking one vendor

In sum – what we know:

• Training versus inference – Friar positions Nvidia as the platform for training, while custom and competing silicon increasingly targets inference, the workload that dominates operating spend at ChatGPT’s scale.
• Jalapeño and MI455 – Broadcom’s custom inference ASIC is running engineering samples in OpenAI’s lab, and AMD’s MI455 has become, in Friar’s words, “a very viable alternative in many cases.”
• Near-term Nvidia – Demand for Nvidia’s training systems likely holds, but a multi-vendor OpenAI can negotiate harder on price and shift share in the fastest-growing workload category.

OpenAI is slowly but surely moving away from Nvidia, and it’s doing so in public. Speaking with Jim Cramer on CNBC’s Mad Money, OpenAI CFO Sarah Friar laid out the company’s multi-vendor compute strategy in unusually plain terms. “Just as we have a strategy for diversifying out our revenue streams, we also have a strategy for diversifying our supply chain, and that is just good CFO risk mitigation,” she said.

The logic is straightforward. “You never know when someone’s supply chain is going to get gummed up, so you need to have multiple providers,” Friar said, pointing to the risk that “TSMC can’t provide all the wafers” as one concrete example. The payoff, according to Friar, is “better lower latency, better reliability, better pricing, because now I have some leverage with my supply chain.” That leverage point is worth sitting with — a buyer at OpenAI’s scale with credible alternatives negotiates very differently than one with nowhere else to go.

Of course, none of this amounts to OpenAI walking away from Nvidia, necessarily — at least not in the near term. Friar called Nvidia “an incredible platform of accelerators for training,” and nothing in the interview suggested the company plans to replace it completely. That said, Nvidia is no longer the only viable choice across every workload, and OpenAI is now matching chips to tasks rather than defaulting to one platform for everything.

Training and inference require different chips

The split between training and inference is essentially what makes diversification possible. Training demands highly scalable, programmable accelerator systems — the kind of work Nvidia’s GPU platforms have dominated. Inference is running the trained model to produce an answer, a token, or a code suggestion for a user. It happens continuously once a model ships, and at ChatGPT’s scale it can come to dominate operating spend.

Friar positioned Nvidia as particularly strong on the training side, citing “Astra, 100,000 GPUs in Stargate, Texas” as evidence of continued large-scale deployments. Custom and competing silicon, meanwhile, is increasingly suited to select inference workloads, where a chip built around a known set of models can be more efficient than a general-purpose accelerator. A custom inference chip doesn’t displace the Nvidia systems training the next frontier model — it carves off a different, fast-growing slice of the workload.

The Broadcom-built Jalapeño chip targets inference workloads

The clearest expression of that strategy is Jalapeño, the chip OpenAI and Broadcom have built together. It was designed from the start for large-language-model inference rather than adapted from a general-purpose AI accelerator, and Friar leaned on exactly that point. The chip is “very focused on inferencing” and “set up exactly for our models,” she said. Broadcom contributes the silicon implementation and networking technology, while Celestica handles board, rack, and system integration.

OpenAI says engineering samples are already running ML workloads in its lab at production target frequency and power, including GPT-5.3-Codex-Spark. The companies claim early testing shows performance per watt “substantially better than current state-of-the-art.” Against general-purpose GPU inference, an ASIC tailored to OpenAI’s own models and serving stack should win on efficiency.

The deployment plan calls for initial rollout by the end of 2026, scaling to gigawatt levels over multiple generations with data-center partners including Microsoft. That’s a forward-looking roadmap, not evidence that Jalapeño is carrying production inference at scale today.

AMD’s MI455 becomes a “viable alternative” for OpenAI

On the merchant-silicon side, Friar said AMD’s MI455 accelerators “have become a very viable alternative in many cases.” That builds on the multi-year chip-supply agreement the two companies previously announced. Reuters reported that the deal covers hundreds of thousands of chips and includes a warrant allowing OpenAI to acquire up to 10% of AMD at $0.01 per share, subject to conditions — an unusual structure that ties the two companies’ fortunes together.

For AMD, which has spent years trying to establish a durable data-center AI business in Nvidia’s shadow, having OpenAI publicly call its accelerators viable is about as high-profile a reference customer as exists.

Friar also mentioned Cerebras “and so on” without detailing volumes or terms. Reuters reported in February that OpenAI had explored inference alternatives with AMD, Cerebras, and Groq amid concerns about aspects of Nvidia’s chips — reporting sourced to unnamed people and not confirmed by OpenAI, so it should carry that caveat.

The diversification extends past chip designers, too. According to reports, OpenAI is deepening ties with Samsung Electronics through joint research on next-generation chips alongside memory demand, potentially strengthening its links to South Korea’s semiconductor supply chain. That’s distinct from sourcing accelerators — it’s about the longer-term chip-development and manufacturing ecosystem.

Market implications across the chip supply chain

For Nvidia, the near-term picture probably doesn’t change much. Demand for its training systems could stay enormous even as OpenAI routes some inference elsewhere, and Friar’s Stargate reference suggests it will. The longer-term risk is share loss in inference, which happens to be the fastest-growing workload category — and a multi-vendor OpenAI can negotiate harder on price regardless of volume, which puts quiet pressure on margins.

The other players each get something different out of this. AMD gets its reference customer. Broadcom gets a showcase for custom ASIC design, networking, and integration, though it also takes on concentration risk if OpenAI’s deployment schedule slips. And foundries, memory makers, and packaging houses see demand rise across the board, since more custom silicon means more reliance on advanced manufacturing capacity rather than less.

Friar drew a line from cheaper inference hardware to cheaper model access, a priority as customers weigh proprietary AI services against open-weight and Chinese models. The strategy hangs together on paper. What’s genuinely unresolved is the mix — OpenAI hasn’t disclosed a vendor-by-vendor breakdown, Jalapeño’s efficiency claims await real benchmark data, and Nvidia still sits at the heart of the company’s highest-end training infrastructure.