iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
OpenAI is developing Jalapeno, a programmable custom chip aimed first at running (“inference”) large language models. In a Forbes Australia interview published October 13, 2026, OpenAI hardware chief Richard Ho said the company hopes to have some users running some models on it in the first half of 2027, but described that schedule as a stretch goal. The project is being developed with Broadcom as OpenAI tries to improve efficiency and add capacity alongside chips from Nvidia, AMD, Cerebras and other suppliers.
What OpenAI has actually announced
Jalapeno is the reported name of OpenAI’s first custom semiconductor. Ho described it as an inference chip: hardware designed to process requests from trained models rather than primarily to train those models. He said the device is programmable, somewhat more specialized than a general-purpose GPU, and intended for transformer-based systems and large language models.
The design is not described as locked to OpenAI models. Ho said the objective is flexibility as model architectures change, so the chip can adapt instead of supporting only one generation of software. OpenAI has not published independent benchmark results, a per-chip performance rating, a price, or a confirmed full-production date.
Why OpenAI wants its own silicon
Ho said the effort began with a practical question: could general-market suppliers provide everything OpenAI needed as demand for computing grew? After early discussions involving AMD, Nvidia and Intel, the team chose to pursue a blank-slate design.
#1 Best Overall
The strategic goal is control across more of the infrastructure stack—data centers, chips, systems, software and models—so engineers can optimize the boundaries between those layers. Ho identified “intelligence per watt” as the key measure: delivering as much useful model output as possible for the energy consumed. OpenAI’s 2025 announcement likewise said custom hardware could incorporate lessons from its models and products.
That rationale does not yet come with a public savings estimate. Ho said he had not analyzed how much OpenAI might save, and an OpenAI spokesperson told Forbes that the company had no estimate. The reported benefit is therefore a design objective, not a quantified result.
Inference first, not a confirmed training replacement
OpenAI is focusing Jalapeno on inference because it expects the amount of model-serving work to keep rising and wants to reduce the cost of answering requests. Ho said pre-training could become a future direction, but it is not his current focus.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsRank #2
A Reuters report from February 2025 discussed an earlier proposed training-capable chip and an expectation that TSMC mass production might begin in 2026. That earlier report should not be treated as proof of Jalapeno’s architecture, fabrication status or production. The later Forbes interview identifies Jalapeno as an inference chip without confirming those earlier details.
What Broadcom is doing
OpenAI says it is designing accelerators and complete systems with Broadcom. Broadcom’s publicly described role includes the accelerator-development partnership and Ethernet and other connectivity technology needed to link large groups of processors.
Ho said OpenAI selected Broadcom after evaluating potential partners. He pointed to relevant analog intellectual property and existing working relationships among members of his hardware team, including experience with Broadcom during development of Google’s first TPU. The available reporting does not establish that Broadcom alone manufactures Jalapeno or identify the chip’s current foundry.
The partnership is broader than one chip. OpenAI’s October 13, 2025 announcement covered custom accelerator systems and networking, with deployment planned for OpenAI facilities and partner data centers.
The 10-gigawatt figure is a systems target
OpenAI’s 2025 announcement described 10 gigawatts of custom accelerator systems, with deployment targeted to begin in the second half of 2026 and finish by the end of 2029. Those are announced system-level targets. They do not state how many Jalapeno chips are involved, how much power an individual chip uses, or what capacity Jalapeno itself has.
Ho said Jalapeno’s gigawatt capacity had not been publicly shared. Treating the 10-gigawatt collaboration figure as a rating for the chip would confuse a data-center deployment plan with an individual component specification.
Timeline and what each date means
| Date | Development point | What is established |
|---|---|---|
| February 2025 | Reuters reports an earlier training-capable chip proposal and a then-current 2026 TSMC production expectation. | An earlier plan, not confirmation of Jalapeno’s design or manufacturing. |
| October 13, 2025 | OpenAI announces its accelerator and systems collaboration with Broadcom. | A 10-gigawatt system target, with rollout targeted for the second half of 2026 through the end of 2029. |
| October 13, 2026 | Forbes Australia publishes Ho’s interview naming Jalapeno as an inference chip. | Ho says the company hopes to put something into production in the first half of 2027. |
| First half of 2027 | Possible initial use by some users on some models. | Ho calls this a stretch goal dependent on software and other preparation; it is not a confirmed deployment date. |
How Jalapeno fits with Nvidia and other suppliers
Ho described Jalapeno as an addition to OpenAI’s existing fleet, not a replacement for Nvidia. He said the fleet will continue to include Nvidia, AMD, Cerebras and other suppliers because OpenAI’s demand for computing is so large. Custom hardware could give OpenAI another supply source and allow tighter tuning for particular workloads, while purchased accelerators continue to provide broad compatibility and available capacity.
| Question | Purchased general-purpose accelerators | OpenAI custom inference hardware |
|---|---|---|
| Workload flexibility | Designed for many customers and workloads; established software ecosystems are a major advantage. | Intended to remain programmable for changing transformers and language models, but its real-world flexibility has not been independently measured. |
| Efficiency per watt | Varies by model, software and system configuration; no comparison is supplied here. | Improving intelligence per watt is the stated design goal; no public benchmark or savings figure establishes an advantage. |
| Software and system integration | Existing tools and deployed systems can reduce integration work. | OpenAI can co-design hardware, systems and software, but the amount of required preparation remains part of the schedule risk. |
| Supply and maturity | Commercial products already exist from multiple vendors. | Jalapeno’s production status and available capacity are not confirmed. |
| Dependence on external suppliers | Purchasing keeps the buyer dependent on vendor road maps and supply. | Could reduce dependence for selected workloads, while OpenAI still expects to use outside suppliers. |
The reporting supports the rationale for this two-track strategy, not a winner on performance, cost or reliability. No head-to-head test or independently measured Jalapeno result has been published in the sources available for this story.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Development scale and the nine-month claim
Ho said about 100 people worked on the overall effort, with an average of roughly 70 during development. That count includes much more than chip designers: rack and board engineering, power delivery, thermal design, supply-chain work and software are part of the system effort.
Best Value
OpenAI also says the chip took nine months to build. That is a company-reported development duration, not an independently audited measurement, and it does not mean a nine-month path from concept to large-scale customer deployment. The remaining software, systems and operational work is one reason Ho called the first-half 2027 target a stretch.
Will consumers be able to buy or rent Jalapeno?
Nothing in the current reporting describes a retail product, replacement part, development kit or public cloud instance. Ho said serving OpenAI’s internal demand is the priority and that external sale or cloud access is not the current focus. The announced deployment language refers to OpenAI facilities and partner data centers, not a consumer purchasing program.
What remains unknown
- Jalapeno’s individual chip capacity, power draw, process node and foundry are not publicly established by these reports.
- No independent performance, latency, energy-efficiency or cost benchmark has been released.
- OpenAI has not quantified expected savings.
- The financial terms of the Broadcom arrangement were not disclosed.
- The first-half 2027 production and usage plan is a stretch goal, not confirmation that users are already running models on Jalapeno.
- The 10-gigawatt figure applies to the announced accelerator-and-networking systems collaboration, not to Jalapeno as a single chip.
Why the story matters
OpenAI’s move reflects a broader infrastructure pressure: serving ever-larger models requires enormous, continuously available inference capacity. Designing a chip can let a model developer tune compute, networking, power and software together, but it also shifts responsibility for hardware validation, supply chains and deployment from vendors to OpenAI. For now, the evidence shows an ambitious supplement to a multi-vendor fleet, with goals and dates still ahead of verified production.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
