OpenAI builds its own AI chip with Broadcom

Overview

Updated Jun 25

Every time you ask ChatGPT a question, OpenAI rents the chips that answer it, mostly from Nvidia. On June 24, OpenAI and Broadcom unveiled Jalapeño, OpenAI's first chip of its own, built to do that work for roughly half the cost.

The chip only runs models, it doesn't train them. But running models is where the money goes once hundreds of millions of people use a product daily. Owning that hardware lets OpenAI cut its single largest cost and lean less on Nvidia.

Why it matters

The cost of running AI sets the price of every chatbot, coding tool, and assistant. Cut it in half and the economics of the whole industry shift.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

Do they have plans to deploy them? If so, when?

Yes — Jalapeño goes into initial deployment by end of 2026, with Broadcom and OpenAI committed to scaling to 10 gigawatts of capacity through 2029.

This is when OpenAI actually starts capturing the promised cost savings, shifting from renting Nvidia compute to running on its own silicon.

4 details2 sources1 disagreement

Initial deployment is targeted for late 2026, with the chip designed in just nine months.
The October 2025 deal with Broadcom set a 10-gigawatt deployment target running from late 2026 through 2029.
Data centers for the rollout are being built with Microsoft and other partners.
Jalapeño only handles inference (serving models), so Nvidia GPUs remain in play for training — the deployment plan doesn't replace Nvidia entirely.

Room for disagreement

Custom chip rollouts routinely slip; OpenAI has no prior track record shipping silicon at scale, so whether late 2026 holds is an open question — though neither Broadcom nor OpenAI has publicly flagged delays.

0 AI answers with web search. May be wrong.

What are the specs of the chip? Give me the details!

Jalapeño is a reticle-sized inference ASIC built on TSMC's 3nm process with eight HBM memory sites — purpose-built for LLM serving, not training.

The spec choices (expensive HBM, maximum die size, 3nm node) show OpenAI is prioritizing raw throughput and latency over cheap hardware — a bet that owning the best inference silicon beats renting GPUs at scale.

6 details3 sources2 disagreements

Built on TSMC's 3nm process — the same node as leading-edge GPUs — which OpenAI says is key to hitting close to theoretical peak utilization.
Eight HBM memory stacks surround the central compute die: HBM is far faster than cheaper DRAM alternatives and addresses the memory-bandwidth bottleneck that limits GPU efficiency on LLM workloads.
Reticle-sized die: the largest single die a fab can print, maximizing transistor count per chip.
Uses a systolic array architecture — efficient for the matrix math that dominates inference — developed by a team led by Richard Ho, a former Google chip engineer.
OpenAI and Broadcom have not disclosed peak FLOPS, memory bandwidth numbers, or power draw; a detailed technical paper is expected in coming months.
Designed in nine months — claimed to be the fastest ASIC development cycle ever for a high-performance advanced-node chip.

Room for disagreement

OpenAI calls Jalapeño 'the best inference platform for LLMs' — but Nvidia's Vera Rubin delivers 50 petaflops per GPU at one-tenth the cost per token (per Nvidia's own GTC 2026 claims), and independent benchmarks haven't yet compared the two head-to-head.
Some analysts argue that without disclosed FLOPS or bandwidth numbers, the '50% cost reduction' claim is unverifiable until Jalapeño runs in production at scale.

0 AI answers with web search. May be wrong.

Key Indicators

~50%

Lower serving cost

Reported cut in inference cost versus standard AI GPUs, pending an independent technical report.

9 months

Design to tape-out

Time from start to a finished design ready for manufacturing, sped up using OpenAI's own models.

10 GW

Total deal scale

Custom accelerators OpenAI plans to deploy with Broadcom through 2029, enough to power several large data centers.

~40%

Microsoft's expected share

Portion of the chips Microsoft is reported to plan to buy.

Voices

Curated perspectives — historical figures and your fellow readers.

People Involved

Greg Brockman

President and co-founder, OpenAI

Leading OpenAI's push to design its own hardware

Hock Tan

President and CEO, Broadcom

Manufacturing partner for OpenAI's chip

Jensen Huang

CEO, Nvidia

Faces a major customer building an alternative

Organizations Involved

OpenAI

Artificial intelligence company

Designed its first chip

The maker of ChatGPT, now designing its own processors to lower the cost of running its models.

Broadcom Inc.

Semiconductor company

Builds and ships the chip

The chipmaker turning OpenAI's design into silicon and supplying the networking to connect the chips at data-center scale.

Timeline

September 2025 June 2026

4 events Latest: June 25th, 2026 · 5 months ago

Tap a bar to jump to that date

June 2026
Wall Street reacts: Broadcom gains, Nvidia slips
Latest Market

Broadcom shares rose about 2% the day after the Jalapeño unveil. Nvidia fell roughly 0.75%. JPMorgan kept its Overweight rating on Broadcom with a $580 price target; UBS maintained its Buy rating.
June 25th, 2026
OpenAI and Broadcom unveil Jalapeño
Product

OpenAI reveals its first custom chip, built to serve models for about half the cost of standard GPUs. Designed in nine months, it is set for initial deployment by the end of 2026.
June 24th, 2026
October 2025
OpenAI and Broadcom announce the 10-gigawatt chip plan
Deal

The two say OpenAI will design custom accelerators and Broadcom will build and deploy them, starting in late 2026 and running through 2029. Broadcom shares jumped.
October 13th, 2025
September 2025
OpenAI and Nvidia announce a multi-gigawatt deal
Deal

OpenAI and Nvidia unveil a partnership to deploy large amounts of Nvidia compute, reportedly worth up to $100 billion.
September 22nd, 2025

Scenarios

Jalapeño serves real traffic on schedule

Possible Resolves by Jan 31, 2027

Discussed by: OpenAI and Broadcom announcements, CNBC, TechCrunch

Deployment hits its stated target. By the end of 2026, OpenAI confirms Jalapeño is running production inference for ChatGPT or other products inside its data centers. This is the companies' own timeline, and the October 2025 deal already set the supply chain in motion.

Deployment slips into 2027

Possible Resolves by Jan 31, 2027

Discussed by: Semiconductor analysts, The Register

First-generation custom chips often miss their first date. Manufacturing yields, software readiness, or rack integration push large-scale deployment past late 2026. OpenAI keeps leaning on Nvidia and AMD chips in the meantime while Jalapeño volumes ramp slowly.

Independent testing confirms the ~50% cost cut

Uncertain Resolves by Q2 2027

Discussed by: SemiAnalysis, VentureBeat, OpenAI's promised technical report

OpenAI has said a detailed performance report is coming. If that report or an independent benchmark shows inference costs falling by roughly half versus comparable GPUs, the headline claim holds up. If the gain is smaller, the strategic case weakens.

OpenAI commits to a second-generation chip

Possible Resolves by End of 2027

Discussed by: Broadcom, OpenAI, Bloomberg

The companies call Jalapeño the first step in a multi-generation platform. If OpenAI publicly commits to or unveils a successor chip, it signals the first one worked well enough to keep building in-house silicon rather than retreating to GPUs.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

May 2016

Google builds the TPU (2015)

Google revealed it had quietly built its own AI chip, the Tensor Processing Unit, to run its machine-learning workloads. It had been using the chips in its data centers for about a year before saying so. The goal was the same as OpenAI's: serve models faster and cheaper than off-the-shelf hardware.

Then

Google cut its cost of running AI services and reduced its dependence on outside chip suppliers.

Now

The TPU became a core part of Google Cloud and proved a big AI company could design competitive silicon in-house.

Why this matters now

Jalapeño follows the same playbook a decade later. It shows custom inference chips can pay off, and that the hard part is volume manufacturing, not the first design.

November 2020

Apple drops Intel for its own silicon (2020)

Apple replaced Intel processors in its Mac computers with chips it designed itself, the M1. After years of buying from Intel, Apple decided controlling the chip let it tune hardware and software together for better speed and battery life.

Then

The first M1 Macs beat Intel models on speed and efficiency, and Apple's margins improved.

Now

Apple now designs the chips in nearly all its products, and Intel lost a marquee customer.

Why this matters now

OpenAI's pitch echoes Apple's: design the chip and the model together and you get gains a general-purpose supplier can't match. The risk for Nvidia is losing a top customer the way Intel lost Apple.