Nine months from first RTL work to tape-out is fast. Whether it’s fast enough to matter is a different question entirely.
Key Takeaways
- OpenAI and Broadcom unveiled Jalapeño, a custom AI inference chip, completing tape-out in nine months — the fastest ASIC development cycle the companies say has been achieved in high-performance semiconductors.
- The announcement landed the same day Micron reported record quarterly revenue of $41.46 billion, driven largely by AI data center demand for high-bandwidth memory.
- The 2mind read: chip diversification looks real, but a tape-out is a milestone, not proof the chip performs at scale in production.
- As of 2 October 2026, the performance figures come from OpenAI, with SemiAnalysis present for the benchmark runs. Initial deployment is planned for late 2026, with volume in 2027.
What happened
The table below separates what OpenAI and Broadcom said at launch, what has been disclosed since, and what is still missing. It is a snapshot of the sources listed at the end, not of OpenAI’s internal progress.
| Item | Status as of 2 Oct 2026 | Who says so |
|---|---|---|
| Tape-out | Done. Nine months is OpenAI’s figure for initial design (RTL work began February 2025) to tape-out (November 2025). SemiAnalysis counts about 16 months from initial team hiring. | OpenAI and Broadcom via MLQ; Tom’s Hardware; SemiAnalysis |
| Working silicon | First silicon arrived in May 2026. At launch, engineering samples were running GPT-5.3-Codex-Spark in OpenAI’s lab at target frequency and power. | Tom’s Hardware (Hot Chips); MLQ |
| Specifications | Not disclosed at launch. Shown at Hot Chips on 25 August: 13.4 petaflops of MXFP4 compute (SemiAnalysis ties this figure to the B0 stepping; the benchmarks ran on A0), 216 GiB of HBM4, 15.4 TB/s of memory bandwidth, 700 W package. | Tom’s Hardware; ServeTheHome |
| Performance numbers | Published by OpenAI, using SemiAnalysis’s InferenceX benchmark. SemiAnalysis verified the InferenceX runs in person in OpenAI’s lab but says all numbers were provided by OpenAI. It reports accuracy (GSM8k) on par with Nvidia chips. | OpenAI via ServeTheHome and MLQ; SemiAnalysis |
| Independent reproduction | None yet. OpenAI chose the systems, models and operating conditions, and the hardware is at engineering-sample stage. | MLQ; SemiAnalysis |
| Yield and cost | No yield figure appears in the sources listed below. A saving of about 50% versus typical AI GPUs is a reported target, not a measured result. | MLQ |
| Deployment | Initial deployment planned for late 2026. Volume production ramps in 2027. Initial capacity is limited, and Axios reports OpenAI does not expect to sell the chip externally. | Companies via Tom’s Hardware; The Register; SemiAnalysis; Axios via MLQ |
Microsoft appears in the deployment story in two ways. Broadcom’s CEO named it as a data-center partner starting in 2026 (Tom’s Hardware). MLQ separately reports that Microsoft is expected to take about 40% of the initial production run; we could not confirm that figure in the other sources.
On June 24, OpenAI and Broadcom unveiled Jalapeño, a custom accelerator built specifically for large language model inference.
The tape-out — the point where chip design is handed off to manufacturing — was completed just nine months after development began, which the two companies called the fastest ASIC development cycle ever achieved in high-performance semiconductors. It is OpenAI’s first in-house silicon.
The announcement landed the same day Micron reported record quarterly revenue of $41.46 billion for the quarter ended May 28, driven by surging demand for high-bandwidth memory (HBM) from AI data centers.
Two announcements on the same day do not prove a link. But both point the same way: AI demand is now reaching parts of the hardware stack beyond GPUs.
Micron’s $41.46 billion quarter is worth sitting with. Memory chips have historically been one of the more cyclical corners of the semiconductor industry, prone to oversupply and price crashes.
A record quarter driven specifically by HBM demand suggests the AI buildout has moved past the GPU headline story. It’s now stressing the memory supply chain directly, which is a different kind of signal than one company’s earnings beat.
What the August benchmark numbers show, and what they leave out
At Hot Chips on 25 August, OpenAI showed results from InferenceX, a public benchmark run by SemiAnalysis. MLQ reported OpenAI’s absolute values for two cases. The ratios in the table below are our division of those values, and they closely match OpenAI’s rounded ratios.
| Test (as reported) | Jalapeño vs baseline | Ratio |
|---|---|---|
| GPT-OSS 120B vs Nvidia GB200 | 85,448 vs 44,960 tokens/s per kW; latency 1.03 s vs 1.80 s | 1.90x throughput per kW; 1.75x lower latency |
| DeepSeek R1 vs GB300 | 19,641 vs 11,781 tokens/s per kW; latency 1.65 s vs 5.99 s | 1.67x; 3.63x lower latency |
| DeepSeek R1, single-token Jalapeño vs multi-token GB300 | Ratios only, no absolute figures reported | 1.5x per kW; 2.2x lower latency |
| Kimi K2.5 vs GB300 | Ratios only, no absolute figures reported | 1.5x per kW; 3.4x lower latency |
The larger headline figures, 53.7x and 104.3x more throughput, compare systems at a very low-latency operating point where the Nvidia systems produced little total throughput. MLQ notes they do not represent raw chip speed.
The per-kilowatt ratios also depend on the power rating used. ServeTheHome says runs were normalized to package power: 700 W for Jalapeño, 1,200 W for GB200 and 1,400 W for GB300. Multiplying OpenAI’s GPT-OSS figures by those ratings gives about 59,800 versus 54,000 tokens per second per package, roughly 1.1x.
For DeepSeek R1 against GB300 the same arithmetic gives about 13,700 versus 16,500, roughly 0.83x. This is our calculation, assuming the ratings are the denominators; it is not a reported result. It suggests the gain is mainly lower energy per token, not more speed per chip.
SemiAnalysis says Vera Rubin, not Blackwell, is the fairer comparison. Its own estimate puts Jalapeño and Rubin close on output tokens per dollar, with Jalapeño running without speculative decoding. It also says the results come from first-stepping silicon, with a revised version in the fab.
A six-question checklist for any custom-chip benchmark
- Who ran it? OpenAI selected the systems, models and conditions. SemiAnalysis verified the runs in person but did not run the full InferenceX suite or see AgentX results.
- Is the decoding mode the same? Jalapeño used single-token prediction; some Nvidia baselines used multi-token prediction.
- Which Nvidia generation? The baselines are Blackwell-generation GB200 and GB300, not Vera Rubin.
- Which workload? Single-turn 8k-input, 1k-output tests, not multi-turn agentic traffic.
- Which power figure? Rated package power (700 W, 1,200 W, 1,400 W), not a full-system cost.
- Can anyone else buy it? Axios reports OpenAI does not expect to sell it externally, and volume is planned for 2027.
Applied to Jalapeño today, the first two answers are ‘the vendor’ and ‘not matched’, and the last is ‘no’. That does not make the numbers wrong. It means they are a strong early claim, not an industry verdict.
The two lenses
A nine-month tape-out buys OpenAI a design that’s ready for manufacturing, not a chip already running production inference traffic. That gap is where custom silicon projects usually prove themselves or quietly stall.
At launch, neither company had published yield rates or benchmark results (OpenAI later published its own first benchmarks in August 2026, not independently reproduced). Until independent numbers exist, ‘fastest ASIC cycle ever’ describes a development timeline, not a finished product’s performance.
Lens one: This is a healthy diversification of AI infrastructure.
For years, the AI industry’s compute stack has been uncomfortably concentrated around a single supplier. OpenAI spending nine months to tape out a custom chip signals that the largest AI labs are serious about building their own silicon roadmaps.
Custom inference chips, tuned to a specific model architecture, can deliver meaningfully better performance-per-watt than general-purpose GPUs for deployment workloads.
If OpenAI’s chip performs as intended, it reduces both cost and dependency — and that’s a structural improvement for the broader ecosystem, not just for one company.
Micron’s HBM numbers reinforce this: the memory layer of AI infrastructure is already being stress-tested at scale. Demand is real, not speculative.
Lens two: Vertical integration at this scale concentrates power, not just compute.
When the world’s most influential AI lab controls its own chips, its own models, and its own deployment infrastructure, the competitive moat deepens in ways that are hard to reverse.
Smaller AI companies and researchers who depend on shared cloud infrastructure may find themselves further behind — not because the technology is unavailable, but because the cost and latency advantages increasingly favor those who own the full stack.
There’s also the question of validation. A tape-out is a milestone, not a product.
Nine months is fast, but inference chips live or die on yield rates, thermal performance, and real-world benchmark results against production workloads. At launch we had not seen those numbers, and the benchmarks OpenAI later published are its own.
Why it matters
For readers, the practical question is how soon this changes anything. The companies target initial deployment in late 2026. The Register and SemiAnalysis describe volume production in 2027. Axios, via MLQ, reports that initial capacity is limited.
OpenAI says it will keep deploying Nvidia and other accelerators, and Jalapeño is built for inference, not training (MLQ). Chip lifetimes also shape the economics; see our piece on AI chip depreciation.
Most directly affected, per the sources: Nvidia, whose Blackwell systems were the benchmark baseline; Microsoft and other deployment partners; and OpenAI itself. Outside labs are less affected near term, since OpenAI does not expect to sell the chip externally (Axios, via MLQ).
What to watch: results against Vera Rubin, multi-turn agentic benchmarks, whether 2027 production volumes match the plan, and how far Anthropic gets with its reported in-house chip effort.
The race for AI compute is no longer just about who trains the best model. It’s about who controls the hardware it runs on.
Market context from the same two days, not evidence about the chip: gold traded at $3,972 on June 24, and Bitcoin touched $58,131 on June 25. CoinDesk had reported in May that AI and chip stocks had outpaced bitcoin.
What would change our view
If independent benchmark results land roughly in line with Nvidia’s inference chips rather than meaningfully ahead, the ‘structural improvement for the ecosystem’ reading in Lens one gets harder to defend.
It would look more like a cost play for OpenAI alone than a shift for the industry. We’d also revisit this if Anthropic or Google DeepMind announce comparable custom silicon within two quarters.
One trigger has partly happened. On 7 August, Tom’s Hardware reported that Anthropic is co-designing its own inference chips, with Samsung reported, not confirmed, as a manufacturing partner. Anthropic has not announced a chip, a tape-out or a deployment date.
This is not investment advice. It is a reading of public sources, and company-supplied performance claims can change as independent testing arrives.
FAQ
Q. Is Jalapeño already running in OpenAI’s products?
A. Not yet at scale. The companies confirmed the tape-out milestone — the design handoff to manufacturing — and said deployment begins in late 2026; OpenAI first published its own benchmark results in August 2026, which independent parties have not reproduced.
Q. Does this mean OpenAI is moving away from Nvidia?
A. Not entirely. A custom inference chip narrows OpenAI’s reliance on Nvidia GPUs for specific workloads, but general-purpose GPUs remain central to training, where the economics differ.
Q. Why did OpenAI build its own chip instead of buying more GPUs?
A. OpenAI president Greg Brockman said: ‘By designing more of the stack ourselves, we can serve more intelligence with greater efficiency’ (MLQ). SemiAnalysis adds its own view that OpenAI is limited by data-center power, so tokens per megawatt matter most.
Q. Who measured the Jalapeño benchmarks?
A. OpenAI supplied the numbers, using SemiAnalysis’s InferenceX benchmark. SemiAnalysis verified the InferenceX runs in person but says all numbers were provided by OpenAI. No fully independent test exists yet.
Sources
- MLQ News, OpenAI and Broadcom unveil Jalapeño, a custom inference chip (24 June 2026)
- StockTitan, Micron reports record results for its third quarter (24 June 2026)
- Yahoo Finance, Gold breaks below $4,000 for the first time in 2026 (24 June 2026)
- Forbes, Bitcoin hits 21-month low (25 June 2026)
- CoinDesk, Bitcoin slips to 13th largest asset as the AI semiconductor boom accelerates (27 May 2026)
- OpenAI Jalapeno Custom AI ASIC at Hot Chips 2026 — ServeTheHome, 25 August 2026
- OpenAI Jalapeño: Better Than Nvidia Blackwell — SemiAnalysis InferenceX, 25 August 2026
- OpenAI’s Jalapeño shows strong inference gains, but its Nvidia comparison has limits — MLQ News, August 2026
- Broadcom and OpenAI unveil custom-built Jalapeño inference processor — Tom’s Hardware, 24 June 2026
- Hot Chips 2026: OpenAI’s Jalapeño AI ASIC unpacked — Tom’s Hardware, 27 August 2026
- OpenAI’s upcoming Jalapeño chip looks like it’ll be an inference beast — The Register, 25 August 2026
- Anthropic co-designing custom AI inference chips to bypass costly Nvidia GPUs — Tom’s Hardware, 7 August 2026
Correction, October 2, 2026. An earlier version said OpenAI and Microsoft had published neither performance benchmarks nor a deployment timeline for the Jalapeño chip.
At launch they said deployment would begin in late 2026, and OpenAI published its own first benchmarks at Hot Chips in August 2026; those figures come from OpenAI and are not independently reproduced.
Leave a Reply