OpenAI's Jalapeno Chip: How It Stacks Up Against Nvidia Blackwell
OpenAI's first custom AI chip, Jalapeno, reportedly beats Nvidia Blackwell on performance-per-watt for LLM inference, per SemiAnalysis lab testing.
What is OpenAI’s Jalapeno chip?
Jalapeno is OpenAI’s first in-house AI chip, an application-specific integrated circuit (ASIC) built for one job: running inference for large language models. Unlike a GPU, which handles a wide range of workloads, Jalapeno is narrowly optimized to serve chatbot-style outputs as fast and efficiently as possible. According to testing by the analysis firm SemiAnalysis, which got hands-on access to the hardware, Jalapeno beats Nvidia’s Blackwell chips and Google’s TPUs on performance-per-watt, the metric that measures how much useful AI computation a chip produces for every watt of electricity it consumes.
TL;DR
- Jalapeno is an ASIC, not a general-purpose GPU, meaning it’s built to do one thing (LLM inference) rather than handle training, graphics, or other workloads.
- SemiAnalysis independently tested the chip and reported that it outperforms Nvidia Blackwell on performance-per-watt across most tested scenarios, including both low-latency and high-throughput cases.
- At low concurrency, Jalapeno hit over 700 tokens per second per user running the DeepSeek R1 model, a notable result given OpenAI’s team had no prior in-house kernel support for that model’s attention mechanism.
- The chip skips Nvidia’s CUDA ecosystem entirely, relying instead on hand-tuned, assembly-level kernels and a custom language called Gluon rather than a broad developer-friendly platform.
- OpenAI used its own AI models, including a scaled version of Codex, to write and optimize chip kernels, cutting a process that normally takes specialist engineers months down to a much faster automated workflow.
- The project moved from design to tape-out in about nine months, first announced in October 2025 and formally unveiled in June 2026, with small-volume deployment in OpenAI’s data centers starting this year and a larger ramp planned for 2027.
- SemiAnalysis notes its results are preliminary, based on in-person lab verification rather than the firm’s full standard benchmark suite, so the comparison, while credible, isn’t yet exhaustive.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How does Jalapeno compare to Nvidia Blackwell and Google TPUs?
SemiAnalysis, a firm known for detailed chip and datacenter analysis, ran its own tests on Jalapeno rather than relying on OpenAI’s marketing claims. The headline finding: Jalapeno beat Blackwell on performance-per-watt in nearly every scenario tested, without being specifically tuned for any single point on the performance curve. That matters because chipmakers often optimize for one narrow use case (say, maximum throughput) at the expense of others (like response speed for a single user). SemiAnalysis reported Jalapeno held its advantage whether the workload favored fast, low-concurrency responses or high-volume, many-user throughput.
One specific data point stood out: at concurrency one, running DeepSeek’s R1 model, Jalapeno delivered over 700 tokens per second per user. That’s a meaningful number for anyone thinking about interactive chatbot latency, where users notice lag in real time.
It’s worth noting SemiAnalysis flagged caveats. The firm verified results it observed in person but has not yet run its full preferred benchmark suite on the chip. So this is a strong early signal, not a final, comprehensive verdict.
Why does the DeepSeek R1 test matter so much?
The R1 result is more interesting than it first appears because of a technical wrinkle. DeepSeek’s models use an architecture called multi-head latent attention (MLA), an unusual design that most other AI labs, including OpenAI, don’t use. That meant OpenAI had no existing internal kernel code optimized for MLA when the R1 test came up.
A kernel is a small, highly specific piece of code that tells a chip exactly how to execute the math behind an AI model, things like matrix multiplication or attention calculations. Kernels are notoriously difficult to write well, and a poorly optimized kernel can cut a chip’s real-world performance in half even if the underlying hardware and model are excellent. Writing a good kernel is specialist work traditionally done by a small pool of highly paid engineers.
Faced with a model architecture it had no kernel for, OpenAI reportedly used its Codex coding model to write a functional, efficient MLA kernel for Jalapeno quickly. That the chip still turned in a strong benchmark despite lacking a hand-built, purpose-specific kernel is arguably as notable as the raw performance-per-watt numbers.
Does Jalapeno threaten Nvidia’s CUDA moat?
This is the part of the story with the biggest long-term implications. Nvidia’s dominance in AI hardware isn’t just about chip design, it’s about CUDA, the programming platform Nvidia has built since roughly 2007. CUDA has two decades of libraries, tools, documentation, and trained engineers behind it. That ecosystem is often called the “CUDA moat” because it makes Nvidia chips the path of least resistance: if you already know CUDA, or your team already has code written in it, switching to a competitor’s chip means rebuilding a huge amount of software infrastructure from scratch.
AMD has tried to compete on hardware but has struggled to match Nvidia’s software ecosystem. Google poured resources into building similar tooling around its TPUs. Both efforts required trying to replicate CUDA’s scale and developer-friendliness.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
OpenAI reportedly skipped that approach entirely. Instead of building a broad, human-friendly software platform, the company writes Jalapeno’s kernels in an extremely low-level style, described as similar to assembly code, with some individual kernels running to roughly 3,000 lines, backed by correctness checks and a custom sanitizer. Early kernel development involved humans in the loop, but the process is reportedly shifting toward a more automated, scaled-up internal version of Codex.
The programming language OpenAI built for this is called Gluon, paired with something SemiAnalysis describes as a “linear layout” algebra system OpenAI invented for organizing chip operations. The bet here is a reversal of 20 years of conventional wisdom: instead of building a massive, accessible ecosystem for thousands of human engineers, OpenAI built a narrow, demanding language paired with an AI model willing and able to write in it. If AI can generate competitive kernel code on demand, the scarcity of expert kernel engineers, and the moat that scarcity protects, matters much less.
SemiAnalysis put it directly: OpenAI’s own models, including GPT-5.2-Codex, currently run on Nvidia GPUs, were used to help design and program a chip that could challenge Nvidia’s software advantage. In effect, Nvidia’s hardware is being used to help build its potential successor.
Is Jalapeno actually ready to compete at scale?
Not yet, at least not in volume. OpenAI’s chip program was first announced in October 2025, and Jalapeno was formally unveiled in June 2026, reportedly moving from initial design to tape-out in about nine months, a fast timeline for custom silicon. OpenAI plans to deploy small volumes of Jalapeno chips in its own data centers this year, with a larger ramp-up expected in 2027.
That timeline matters for context. Amazon, Google, Microsoft, and Meta have all built or attempted custom AI training and inference chips with varying degrees of success. Google’s TPUs are the most mature alternative to Nvidia GPUs currently in wide production use. OpenAI’s approach is notable because it arrives after the company already built and trained frontier models on Nvidia hardware, meaning it can use those same AI models to accelerate its own chip software development, a feedback loop some in the industry compare to a slow-building flywheel that speeds up over time.
Frequently Asked Questions
What does “performance per watt” mean and why does it matter?
It measures how much useful computational work (in this case, AI token generation) a chip produces for each watt of electricity consumed. Since power availability is one of the biggest constraints on scaling AI datacenters, a chip that does more work per watt lets a company serve more users or run more models without needing proportionally more electricity or cooling infrastructure.
Is Jalapeno a training chip or an inference chip?
Based on the available reporting, Jalapeno is designed specifically for inference, meaning it generates outputs from already-trained models rather than training new models from scratch. It’s an ASIC purpose-built for that narrower task, not a general-purpose replacement for GPUs used in training.
Can Jalapeno run models other than OpenAI’s own?
Yes. SemiAnalysis’s testing reportedly included open-source models such as DeepSeek’s R1, not just OpenAI’s proprietary models, suggesting Jalapeno is designed as a general-purpose inference chip rather than one locked to a single model family.
How did OpenAI build a competitive chip on its first try?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
The reporting points to heavy use of AI itself in the design and software process, including OpenAI’s Codex model writing low-level kernel code that would traditionally require specialist human engineers. This let the company skip building a broad developer ecosystem like Nvidia’s CUDA and instead rely on AI-generated, hand-tuned code for a narrower, harder-to-use programming language called Gluon.
Are these benchmark results final?
No. SemiAnalysis has said its published results come from in-person lab verification rather than its complete standard benchmark suite. The numbers are credible early signals from an independent third party, but a fuller comparison across more workloads and conditions hasn’t been published yet.