🌶️ OpenAI Built a 3nm Chip in 9 Months — And Just Slapped NVIDIA's GB300 in the Face
🤯 Wait — Who Is NVIDIA's Biggest Customer?
Let's start with a counterintuitive question:
Over the past three years, who bought the most NVIDIA chips on the planet?
Not Google. Not Microsoft. It was OpenAI.
Training the GPT series burned through so many GPUs that the whole world took turns calculating their electricity bill. So when on August 25th, OpenAI casually dropped a benchmark report for their own homegrown chip — Silicon Valley collectively did a double take.
💥 The biggest customer just built their own chip — and body-slammed their biggest supplier.
The chip is called Jalapeño. 🌶️
Cute name. Absolutely not cute results.
⚡ Section 1 — From Blueprint to Silicon: 9 Months Flat
Let's talk about how this chip came to be — because the timeline alone is jaw-dropping.
Jalapeño was co-designed by OpenAI and Broadcom, then manufactured by TSMC on their 3nm node.
Here's the stat that broke the internet:
🚀 From architecture design to successful tape-out: just 9 months.
To put that in perspective — a typical chip in the industry takes 2–3 years from concept to mass production. NVIDIA's GB300 was the result of thousands of engineers and years of accumulated R&D.
And OpenAI — an AI "outsider" in the chip world — did it in nine months.
Even more surgical is its purpose: Jalapeño is a pure inference ASIC. No training. No graphics. Every transistor is laser-focused on one thing: running large language models as fast and efficiently as possible.
Translation: "I'm not trying to win everything. I just want to be the world's best at answering questions."
💡 Insight: OpenAI just proved to the entire semiconductor industry that building a cutting-edge chip doesn't have to take forever.
📊 Section 2 — 1.7× Efficiency. Lower Latency. What Does That Actually Mean?
Now for the spicy part — the benchmark results.
In OpenAI's published tests, Jalapeño went head-to-head with NVIDIA's current flagship GB300 and won on two critical metrics:
- ✅ 1.7× better energy efficiency
- ✅ Lower response latency
Don't underestimate those numbers. For AI companies, they're everything:
- 1.7× efficiency = same electricity bill → nearly 2× the inference requests handled
- Lower latency = faster answers to users → better experience → competitive edge in a world where milliseconds matter
And since Jalapeño is purpose-built, its cost structure is naturally cheaper than a general-purpose GPU.
OpenAI has already said the chip will be deployed in production as early as later this year. Analyst firm SemiAnalysis is already calling it: OpenAI could reach profitability as early as Q3.
A chip isn't a PR stunt. It's a spreadsheet.
💡 Insight: 1.7× efficiency isn't a spec sheet number — it's an economics revolution. Every watt saved is a vote against NVIDIA's pricing power.
🤔 Section 3 — Why Did OpenAI Bother Making Their Own Chip?
You might be thinking: why take on such an enormous risk when you can just... buy chips?
Three words: forced, and calculated.
① Cost. OpenAI's compute bill is astronomical. NVIDIA's pricing power is a sword hanging over their neck. Building their own chip means taking back control of their most critical resource.
② Supply scarcity. The global AI arms race has created GPU queues that stretch to infinity. If you make your own, you get exactly as many as you need.
③ Full-stack ambition. This is the biggest one. Chip + Model + Application = a true vertical stack. OpenAI doesn't want to be dependent on anyone. They want to own every layer.
Sound familiar? Think of Apple — once reliant on Intel, now running entirely on their own A-series and M-series chips. Nobody chokes Apple's supply chain anymore.
OpenAI is walking the exact same road.
💡 Insight: Building Jalapeño isn't showing off. It's OpenAI's Declaration of Independence from NVIDIA.
🌏 Section 4 — NVIDIA Fights Back. But China's AI Was Already on Its Own Path.
Of course, NVIDIA isn't going down without a fight.
That same week, at Hot Chips 2026, NVIDIA unveiled first benchmarks for its next-gen rack system Vera Rubin NVL72:
- 🔥 Running DeepSeek V4 Pro: throughput per megawatt up to 30× better than GB300
- 💰 Token cost reduced by up to 35×
Here's the irony that nobody missed:
🐉 NVIDIA's new throne was validated by a Chinese model — DeepSeek.
And then came the even bigger plot twist.
On August 26th — the exact same day OpenAI showed off Jalapeño — China's Ministry of Industry and Information Technology announced at a State Council press conference: the 15th Five-Year Plan will support R&D of high-end training chips, and push breakthroughs in brain-inspired intelligence and world models.
By the end of June this year, China's intelligent computing capacity had reached 2,185 EFLOPS, with 70+ high-bandwidth compute corridors built and nearly 200 AI key standards successfully developed.
Even earlier, China debuted its first AI chip using software-defined + 3D near-memory computing architecture — achieving 520 TFLOPS on a 14nm process, innovating around advanced node restrictions through architectural ingenuity.
Huawei Ascend, Cambricon, and a wave of domestic chip players are all pushing on multiple fronts simultaneously.
💪 While US companies are just starting to internalize chip development, China's AI ecosystem has spent years turning "alternative path" into muscle memory.
🎯 Final Take: This Isn't a Tech Story — It's a Power Story
OpenAI building Jalapeño looks like a chip launch. But it's really a story about compute decolonization.
The chip buyers are afraid of being choked. The chip sellers are afraid of being bypassed. Nobody in this game is a spectator.
For China's AI industry, the real lesson isn't "wow, OpenAI is impressive." It's this:
🃏 When even the giants start hoarding their own supply chains, the hand we're holding — architectural innovation + full domestic stack — has never been more worth playing.
OpenAI built a chip in 9 months from scratch. China is sitting on 2,185 EFLOPS of compute and an entire domestic supply chain.
There's no excuse not to find our own rhythm. 🚀