Jalapeño: How OpenAI Designed Its First AI Chip Using Its Own LLMs – and What This Means for the Industry

A Chip Designed with AI – OpenAI's Jalapeño Sets New Standards

At the end of August 2026, OpenAI officially unveiled its first proprietary AI accelerator chip: Jalapeño. The technical specifications alone are impressive – up to 13.4 petaflops of 4-bit compute performance, 232 gigabytes of next-generation high-bandwidth memory, and a memory bandwidth of 15.4 terabytes per second. Benchmarks show that Jalapeño can reduce end-to-end latency – the time between entering a prompt and the last generated token – by a factor of 3.6 compared to Nvidia's GB300, while consuming less energy. Yet perhaps the more significant story behind this chip is not the result, but the journey: OpenAI relied heavily on its own Large Language Models when designing its chip.

From Concept to Silicon in Under 20 Months

What traditionally takes years in the semiconductor industry, OpenAI accomplished within a remarkably short timeframe. Less than 20 months passed from the first architectural concept to the finished first silicon. Particularly striking: only nine months elapsed between the first RTL code (Register-Transfer Level, the logical description of chip functionality) and the so-called tapeout – the moment when the finished design goes into production. In an industry where complex chip projects routinely take five to seven years, this represents an extraordinary pace.

This pace was achieved by a team averaging fewer than 100 people – a remarkably minimal headcount for such an undertaking. Add to this the partnership with Broadcom, which handled the physical implementation from the gate level onward, while OpenAI was responsible for the end-to-end system design, including the inference accelerator, the memory hierarchy, and the network architecture.

"The models give our engineers superpowers. Our engineers are still the driving force, they are still the final authority. But they can do things much faster and explore many more paths." – Richard Ho, VP Hardware at OpenAI

How LLMs Changed the Chip Design Process

Automation in chip design is not a new concept – Electronic Design Automation (EDA) has existed for decades. What sets Large Language Models apart from previous automation approaches is their ability to understand language and code semantically, rather than merely processing them syntactically. This makes them particularly valuable for those phases of the design process that still reside in the "linguistic domain" of the problem – that is, where specifications, architecture documents, and code interact closely.

OpenAI built its frontend workflow around the open-source framework XLS (Accelerated Hardware Synthesis), originally developed at Google. XLS enables chip designers to work in higher-level programming languages such as C++ or the Rust-inspired domain-specific language DSLX – the tool then automatically translates these into Verilog, the classic hardware description language. Because this code is structured like software, LLMs can work with it particularly effectively: they iterate rapidly, suggest optimizations, and support engineers in exploring many parallel design paths.

This approach – using LLMs for high-level synthesis – is regarded by experts such as Andrew Kahng, Distinguished Professor at UC San Diego, as especially promising. The ability to iterate quickly is considered a central advantage: the more design cycles a team can complete per day, the faster it converges on an optimal result.

Software Optimization through AI: From 0.31% to Realistic Target Values

Also noteworthy is how OpenAI proceeded after receiving the first physical chips in May 2026. Rather than pursuing classical software engineering, the team directed its internal AI models toward developing benchmark software. For a kernel benchmark focused on Multi-Head Latent Attention (a critical computation block in modern language models), performance rose from an initial 0.31 percent of theoretical compute capacity to significantly higher values – a leap that demonstrates the ability of LLMs to take on complex optimization tasks on the software side as well.

For Dr. Maik Bunzel, founder and managing director of mabucon.eu, this approach illustrates a principle that reaches far beyond the semiconductor industry: "What OpenAI is demonstrating here is the fundamental principle behind every successful AI agent deployment: the AI does not take control – rather, it dramatically extends the human expert's capacity to act. The engineer remains the final authority – but the speed at which they can test hypotheses and run iterations multiplies."

What Does This Mean for Companies Beyond the Chip Industry?

The Jalapeño chip is a semiconductor product, but what is truly relevant about this story is the workflow principle. OpenAI has demonstrated how LLM-powered agents can be deployed within a highly specialized, technical design process – with measurable results in development speed and quality. This pattern can be transferred to many other industries:

• Product development and engineering: Iteration cycles can be significantly shortened through AI-assisted design and testing processes.
• Software development: Agentic Coding – AI agents that autonomously generate, test, and optimize code – follows exactly the same pattern as OpenAI's XLS workflow.
• Compliance and documentation: Wherever specifications need to be translated into actionable outputs, LLMs can serve as translators and accelerators.
• Research and development: The parallel exploration of many solution paths – which is resource-intensive for humans – becomes routine through AI agents.

The decisive lesson is not that AI replaces people, but that it scales output per expert. A team of fewer than 100 people realizing a chip design project with AI support that would traditionally require several hundred specialists – that is the real signal.

Skepticism is warranted – but the momentum is real

Experts from the semiconductor industry point out that Broadcom, as an experienced partner, contributed significantly to the project's success and rapid timeline. A team without comparable infrastructure and partnerships would have faced considerably greater challenges. Whether Jalapeño's benchmark figures will be fully confirmed in real production environments also remains to be seen once the broad rollout takes place.

At the same time, experts such as Ravi Krishna from chip design startup Verkor.io are convinced that timelines will shorten further as LLM capabilities continue to grow. This points to a kind of self-reinforcing cycle: better AI models accelerate hardware development, while better hardware in turn enables more powerful AI models.

Dr. Maik Bunzel, founder and CEO of mabucon.eu, sees this development as an important reference point for companies: "The question is no longer whether AI will be integrated into critical development processes, but how quickly companies build the right workflows to benefit from it. OpenAI demonstrated this with Jalapeño for the chip industry – but the principle applies equally to logistics, financial services, or the process industry."

Outlook: The chip as a blueprint for AI-driven development

Jalapeño is more than a product – it is proof that LLM-powered Agentic Workflows function in highly complex, domain-specific environments. The chip itself will play a significant role in reducing OpenAI's inference costs and decreasing its dependence on Nvidia. But the methodology behind its creation has potentially greater implications than the product itself.

For companies, this leads to a clear strategic conclusion: those who begin now to integrate AI agents into iterative domain-specific processes – combining the strengths of LLMs (language understanding, code generation, rapid iteration) with human expert knowledge – will be able to realize substantial competitive advantages in the years ahead. The window in which this transformation can be approached selectively and in a controlled manner is open. But it will not remain open indefinitely.