AINews A custom Broadcom-built ASIC for LLM inference puts OpenAI on the same vertical-integration path Google and Amazon paved years ago.
Priya Nair OpenAI just taped out its first piece of silicon, and the name tells you the team has a sense of humor: Jalapeño. Built with Broadcom, it's an inference-only accelerator, not a general-purpose GPU, and OpenAI is branding it an "Intelligence Processor." Strip away the marketing and what you've got is the most predictable move in modern AI infrastructure. The frontier labs that can afford it eventually build their own chips, because the math on renting Nvidia at scale stops working.
The real story here isn't a single chip. It's that the economics of running models, not training them, are now the thing that decides who survives. Jalapeño is OpenAI admitting that inference cost is the constraint, and that it would rather own the bottom of the stack than keep paying GPU margins to serve every ChatGPT token.
This is the TPU playbook, run faster #
None of this is new. Google has been shipping custom TPUs since 2015, and Amazon built Trainium and Inferentia for the same reason: a hyperscaler running one dominant workload can design silicon that does that one thing cheaper than a flexible GPU ever will. The trade-off is well understood. An ASIC like Jalapeño is less flexible than Nvidia's hardware, but it's cheaper to run and can be tuned for a narrow set of tasks. If your task is serving transformer inference all day, every day, that's a good trade.
What's genuinely notable is the timeline. OpenAI says it went from design to tape-out in nine months, which it calls the fastest high-performance ASIC development cycle it's aware of, with its own models accelerating parts of the work. Greg Brockman told CNBC the degree of that acceleration "was very surprising to us." Take the self-congratulation with salt, but nine months for a from-scratch accelerator is genuinely quick for an industry where two to three years is normal. The partnership was only made public in October, after 18 months of quiet collaboration.
Broadcom is the unsung half of this. It's been the picks-and-shovels winner of the custom-silicon boom, doing the same job for multiple frontier labs and hyperscalers. CEO Hock Tan described demand from his six big customers as "simply insatiable," and said it runs through 2028. Celestica handles boards, racks, and system integration, and Broadcom's own Tomahawk networking silicon ties the racks together. OpenAI designs the chip; the partners industrialize it.
Why inference, and why now #
The choice to start with inference rather than training is deliberate and correct. Pre-training is bursty, experimental, and benefits from the flexibility of GPUs, so OpenAI will keep buying Nvidia for that. Inference is the opposite: a relentless, predictable, always-on flood of tokens that scales directly with how many people are using your product. Every fraction of a cent you shave off a generated token multiplies across billions of requests.
That's where performance-per-watt matters more than raw speed. Power and cooling are the binding constraints on data center buildout now, not silicon supply alone, and OpenAI is targeting deployments measured in gigawatts (the broader plan reportedly aims at 10 GW of power). OpenAI says early tests show performance-per-watt "substantially better" than current state-of-the-art hardware, with engineering samples already running real workloads in the lab, including a GPT-5.3-Codex-Spark coding model that currently runs on Cerebras inference hardware.
Here's the honest caveat: those are self-reported numbers, not finalized, with a technical report promised later. We don't know what Jalapeño was benchmarked against, on which models, or under what conditions. Every chip vendor cherry-picks. "Better perf-per-watt than the alternatives" is the most reliable sentence in the entire semiconductor industry, and it's true of almost nothing in practice until independent testing shows up. Treat the claim as direction, not fact.
What it actually means if you build on OpenAI #
For working developers, the chip itself is invisible. You will never pip install
Jalapeño or write a kernel for it. It lives behind the API. What changes is the stuff downstream of the silicon, and that's where to pay attention.
Inference pricing. This is the whole point. Custom silicon is how you cut the cost of serving a token without cutting the model. If Jalapeño delivers, expect that to show up over time as cheaper per-token rates or fatter rate limits on inference-heavy models, especially the coding and agentic ones OpenAI is clearly optimizing for. The chip was explicitly pitched around low operating cost on real-time coding models. If you're runningCodex-style agents that burn tokens in loops, that's the workload being targeted.Capacity and availability. Brockman flatly said OpenAI "cannot get compute fast enough." Owning a chip line is a hedge against Nvidia allocation queues. For developers, more total inference capacity means fewer capacity-related 429s and less throttling during demand spikes. That's arguably a bigger near-term win than price.Model-hardware coupling. The flip side of a narrow ASIC is that it's tuned to the models OpenAI ships, not yours. This deepens platform lock-in in a subtle way. As OpenAI co-designs models and silicon, the cost and latency advantages accrue to staying inside OpenAI's stack. Self-hosting an open-weights model on commodity GPUs won't get the same efficiency curve. If portability matters to you, factor that in.
The practical decision for most teams doesn't change today. Jalapeño won't be in production at scale until late 2026 at the earliest. Hock Tan's own framing was small prototype deployment late this year, a real ramp in 2027, and "full tilt" in the first half of 2028. So the developer-facing effects (cheaper inference, more headroom) are a 2027 story, not something to plan around this quarter.
The part worth watching skeptically #
Two things deserve a raised eyebrow. First, one report says Broadcom demanded Microsoft guarantee it will buy 40 percent of the first batch of chips. That's a single, uncorroborated claim, so don't bank on the number, but if it's even directionally true it tells you something real: OpenAI's custom-silicon bet is being de-risked by committing a chunk of output to a partner before the chip has proven itself. Vertical integration narratives love to skip that part.
Second, OpenAI is hedging hard. It already has deals for Amazon's Trainium, AMD accelerators, and Cerebras inference systems alongside its enormous Nvidia footprint. Jalapeño is one supplier in a portfolio, not a replacement for any of them. That's the smart play, but it also means "OpenAI builds its own chip" is less a declaration of independence from Nvidia and more a diversification of an inference bill that's growing faster than anyone can build power plants to feed it.
The strategic logic is sound and the precedent is proven. Google and Amazon both showed that owning inference silicon pays off if your volume is high enough, and OpenAI's volume qualifies. Whether Jalapeño specifically delivers is a question for the technical report and for 2027 production numbers, not for a launch-day press release. For now: real shift in strategy, genuine implications for what you'll pay to run AI in two years, and a performance claim that hasn't earned your trust yet.
Sources & further reading #
Priya Nair· AI & Developer Experience Writer Priya covers AI frameworks, developer productivity tooling, and the startup ecosystem across South and Southeast Asia, bringing a researcher's rigour and a practitioner's empathy to every story. She is deeply sceptical of benchmarks and asks hard questions so her readers don't have to.
Discussion 0 #
No comments yet
Be the first to weigh in.