When AI Companies Start Building Chips: OpenAI's Jalapeño Is a Big Bet
> Source: easy-learn-ai / commit 71dca12 > Date: June 25, 2026
Introduction: A Strange Loop
In 2023, OpenAI released ChatGPT and taught the world what a "large language model" is.
Three years later, OpenAI did something that seems unrelated: together with Broadcom, it built a chip named "Jalapeño" — the Mexican chili pepper.
The name itself is telling. No obscure codename, just a pepper: spicy, direct, down-to-earth.
But behind this chip lies a deeper question: as training and inference costs keep climbing, must companies that sell models also build their own chips?
Chapter 1: Why Chips Became AI Companies' Lifeline
AI models are, at their core, massive amounts of math. Every ChatGPT query triggers billions to tens of billions of operations — not on general-purpose CPUs, which are versatile but slow, but on GPUs, whose parallel compute turned NVIDIA into the oil giant of this era.
The problem: the oil giant calls the shots.
It's Not Just Chinese Companies Being Squeezed
NVIDIA GPUs are in chronic shortage. An H100 officially costs tens of thousands of dollars — double that on the gray market — and delivery is entirely at NVIDIA's discretion.
OpenAI's compute spending is astronomical: an estimated over $3 billion in 2025 alone.
More critically, OpenAI's commercial lifeline sits in NVIDIA's hands:
- If NVIDIA raises prices, OpenAI must accept.
- If NVIDIA cuts supply, OpenAI grinds to a halt.
- If NVIDIA ships better chips to competitors first, OpenAI can only watch.
- "First" — OpenAI's first self-designed chip, a milestone.
- "Inference chip" — not a training chip.
- "With Broadcom" — the veteran ASIC design house brings the expertise OpenAI lacks.
- Google built TPU, mostly for its own use
- Amazon acquired Annapurna Labs for in-house chips
- Microsoft co-designed the Maia chip with AMD
- Tesla built Dojo for autonomous driving training
- OpenAI now has Jalapeño
- Training chips: NVIDIA still leads, with Google TPU and AMD MI series eroding share
- Inference chips: fragmentation and ever-deeper specialization as companies build their own
- Edge chips: AI in phones, cars, and IoT led by Qualcomm, Apple, and Huawei
- OpenAI official announcement: https://x.com/OpenAI/status/2069770172802773292
- gdb on energy efficiency: https://x.com/gdb/status/2069809298612621629
- Discussion of 9-month tape-out: https://x.com/kimmonismus/status/2069795647956373632
- Community spec estimates: https://x.com/scaling01/status/2069867464716939413
This isn't a business partnership — it's dependence. Building its own chip isn't a question of "want" for OpenAI; it's a question of "must."
Chapter 2: What Is Jalapeño?
According to OpenAI's announcement, Jalapeño is OpenAI's first in-house inference chip, designed in partnership with Broadcom.
Key phrases:
Training vs. Inference: Two Different Battlefields
Training teaches a model from scratch: massive data, compute, and time — a GPT-4-class run can cost tens of millions of dollars and take weeks. Training chips demand extreme compute density and memory bandwidth.
Inference is serving trained models to users in real time — every ChatGPT reply is an inference. Inference chips prioritize low latency, high energy efficiency, and low cost.
Choosing inference for Jalapeño is shrewd:
1. The market is bigger — training is episodic; inference runs 24/7 for every user query. 2. The technical barrier is lower — competing head-on with NVIDIA's CUDA ecosystem for training is nearly impossible; inference offers more design freedom. 3. ROI is more direct — every dollar saved on inference goes straight to profit.
Why the Name Jalapeño?
OpenAI hasn't explained it. But the name signals an attitude: not lofty like "Orion" or "Apollo," not cold and technical like "A100" or "H100." A jalapeño is an everyday ingredient — but spicy and essential. Perhaps the message: this chip isn't for showing off; it goes into every ChatGPT reply, like pepper in every Mexican dish — never the star, but indispensable.
Chapter 3: Why Call It a "Big Bet"?
Building chips sounds cool but is extremely risky.
Hardware Has No "Iterate and Patch"
Software ships 1.0 today, 1.0.1 tomorrow, 1.1 next week. Hardware is different: a single tape-out costs tens of millions of dollars, and a design flaw means the money is gone. Iteration cycles are measured in months and years. Jalapeño reportedly took 9 months from design to tape-out — fast for the chip industry, glacial compared to software.
The Ecosystem Is the Moat
NVIDIA's real moat isn't the chip — it's CUDA. Engineers worldwide write models in CUDA, which runs only on NVIDIA silicon. That's a moat of habit, not just technology. Jalapeño must be compatible with existing models and frameworks, or it's an expensive brick. And CUDA compatibility means navigating around NVIDIA's patents.
Economies of Scale
NVIDIA ships millions of GPUs a year, amortizing R&D. Jalapeño will initially be used only internally, at low volume and high unit cost. The investment pays off only when OpenAI's inference volume makes the chip's total cost of ownership lower than buying NVIDIA GPUs. That takes time, scale — and luck.
Chapter 4: The Bigger Picture — Vertical Integration
Jalapeño isn't an isolated event. It's part of the industry's vertical-integration trend.
Traditionally: chip companies (NVIDIA, AMD, Intel) make chips; cloud companies (AWS, Azure, GCP) build data centers; AI companies (OpenAI, Anthropic, Google) build models; app companies build products. But every layer takes a margin and is a bottleneck, so everyone is moving up and down the stack:
When AI is the core competitive asset, compute is the lifeline — better held in your own hands.
The Chinese Path
In China, the logic is even more extreme. With NVIDIA's H100/H200 under export controls, companies including Huawei, Alibaba, Baidu, Biren, Moore Threads, MetaX, and Iluvatar CoreX are all building AI chips. Per easy-learn-ai's tally, seven Chinese companies are shipping or preparing to ship chips benchmarked against H100/H200. But hardware specs are one thing; software stacks and mass production are another. China's AI chip ecosystem is going through a painful 0-to-1 phase — and unlike OpenAI, Chinese companies have no fallback.
Chapter 5: What Happens After Jalapeño?
Short Term: Not Much Changes Immediately
Production is limited; Jalapeño will serve ChatGPT, Codex, and API services while NVIDIA GPUs remain the workhorse. Large-scale deployment typically takes 6–12 months as OpenAI validates stability, efficiency, and cost.
Medium Term: A Changed Cost Structure
If Jalapeño hits its design targets, OpenAI's inference costs fall significantly — potentially lower subscription prices or bigger free tiers, higher margins, and pressure on rivals like Anthropic and Google to accelerate their own chips.
Long Term: A Reshuffled Landscape
If OpenAI proves the in-house inference chip works, others will follow:
The AI chip market moves from "one dominant player" toward a warring-states era.
Epilogue: The Pepper Metaphor
A jalapeño isn't the hottest pepper (that's the Carolina Reaper) or the most common (the bell pepper). But it's spicy, fragrant, and versatile.
Maybe OpenAI is saying: we don't aim to be the hottest chip — we aim to be the most indispensable one. Like Mexican food: you can skip the steak, but not the jalapeño.