OpenAI Wants AI to Build More Than Software
OpenAI has spent years teaching artificial intelligence to write, analyze, research and code. Now, the company wants its models to help design the physical machinery that makes all that intelligence possible.
Speaking at Goldman Sachs’ Communacopia + Technology Conference, OpenAI Chief Financial Officer Sarah Friar said the company is pushing its technology into specialized fields such as semiconductor design, life sciences and financial services.
The most intriguing example is already sitting in a laboratory.
OpenAI used its own AI models while developing Jalapeño, the company’s first custom inference processor. Engineers completed the chip’s design and reached manufacturing tape-out in approximately nine months, according to Reuters.
Tape-out means the design has been finalized and sent for manufacturing. It does not mean finished chips are already pouring out of factories like spicy silicon nachos. Testing, production, system integration and deployment still follow.
Even so, nine months represents an unusually compressed development cycle for an advanced processor. OpenAI believes its models helped engineers accelerate parts of the design and optimization process.
The company’s larger message is simple: AI can do more than consume computing power. It can help create the next generation of computing infrastructure.
That creates a fascinating loop. Better AI helps engineers build better chips. Those chips then run newer AI systems. The machines are helping improve the machines—and the flywheel has started spinning.
Meet Jalapeño, OpenAI’s First Intelligence Processor
OpenAI officially unveiled Jalapeño with Broadcom in June 2026. The companies describe it as an “Intelligence Processor” designed specifically for large language model inference.
Inference is the stage during which a trained AI model responds to users. Whenever ChatGPT answers a question, a coding agent reviews software or an enterprise assistant completes a task, processors must run the model and generate the result.
Training builds the brain. Inference keeps answering the phone.
OpenAI designed Jalapeño around the workloads it sees across ChatGPT, Codex, its developer API and future agent-based products. Unlike a general-purpose accelerator adapted to many computing tasks, the chip started as a blank-sheet design focused on modern language models.
According to OpenAI’s announcement, its engineers optimized the architecture around model kernels, memory movement, networking and real-world serving patterns. The goal is to combine high throughput with low latency instead of heavily sacrificing one to improve the other.
Broadcom contributed chip implementation, connectivity and its Tomahawk networking technology. Celestica is supporting boards, racks and complete server systems. Taiwan Semiconductor Manufacturing Company is handling manufacturing, Reuters reported.
This is therefore not a story about OpenAI wandering into a garage and returning nine months later holding a homemade processor.
It is a coordinated hardware program involving several deeply experienced technology and manufacturing partners.
Why Nine Months Is Such a Big Deal
Advanced chip development normally moves on a much slower calendar.
Engineers must define the architecture, write hardware designs, verify their behavior, optimize power consumption, resolve timing problems, prepare physical layouts and confirm that the design can actually be manufactured. Each stage can uncover problems that send teams back several steps.
Industry analysis from Futurum Group notes that a frontier-class accelerator reaching tape-out in 18 to 24 months would normally count as fast. OpenAI and Broadcom say Jalapeño completed the journey in nine.
OpenAI attributes that speed to close cooperation between its hardware and software teams, Broadcom’s semiconductor expertise and the use of AI during design and optimization.
However, “AI-designed” needs careful interpretation.
The models did not independently dream up a processor, negotiate with a factory and send everyone a calendar invitation. Human engineers remained responsible for the architecture, engineering decisions and final validation.
AI tools helped them run large numbers of verification and optimization jobs. They could examine logs, identify problem areas, adjust parameters and test alternative designs while engineers focused on higher-level decisions.
That is less theatrical than an autonomous robot chip designer. It may also be more useful.
By multiplying the amount of experimentation a human team can conduct, AI can compress some of the slowest parts of semiconductor development without removing people from the process.
The Chip Is Built for the Inference Explosion
The AI industry often focuses on training enormous models. Training grabs headlines because it consumes gigantic clusters, months of work and mountain ranges of electricity.
Inference may become the larger long-term business.
A model trains periodically, but it can answer billions of requests afterward. Every chatbot response, generated image, coding suggestion and autonomous-agent action adds to the inference bill.
Agentic systems make the problem even tougher. A conventional chatbot may generate one answer. An AI agent might perform dozens of connected steps—searching databases, using applications, checking its work and revising its plan. Small delays accumulate quickly.
Jalapeño was designed for that environment.
OpenAI says the processor balances computing, memory and networking resources while reducing unnecessary data movement. Moving information between components consumes time and power, so keeping more of a workload inside one connected system can improve both speed and efficiency.
The processor is also intended to handle the two main phases of language-model inference. The first processes the user’s prompt and context. The second generates the response token by token.
Hardware often performs better at one phase than the other. OpenAI says Jalapeño aims to perform both efficiently, allowing the same infrastructure to respond to changing workloads.
If that design works at scale, users may notice faster responses and more responsive agents. OpenAI, meanwhile, could serve more requests without increasing power use at the same rate.
OpenAI Publishes Its First Performance Results
In August, OpenAI released its first detailed Jalapeño performance results.
The company tested the chip with three publicly available models: GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T. That selection was designed to demonstrate that the processor could run models developed outside OpenAI rather than functioning only with the company’s internal systems.
According to OpenAI’s benchmark report, Jalapeño delivered between 1.5 and 1.9 times more AI work per watt at peak throughput than the comparison systems. OpenAI also reported between 1.7 and 3.6 times lower end-to-end latency.
For highly interactive workloads, the company measured performance improvements ranging from 2.1 to 4.1 times.
Jalapeño carries a published power rating of 700 watts. OpenAI says its measured sustained consumption stayed at or below 550 watts during the tested workloads.
Those are impressive numbers. They are also company-produced results and should be treated accordingly. OpenAI used the public InferenceX benchmarking framework, but independent testing will provide a clearer picture once outside organizations gain access to comparable systems.
Hardware performance can change dramatically depending on software, model architecture, batch sizes, memory configurations and the comparison point selected.
For now, the results provide evidence that Jalapeño is functioning silicon—not merely a slide deck seasoned with a spicy name.
AI Helped Design the Chip—and Now Programs It
The most interesting part of Jalapeño may not be its raw speed. It is the role AI played throughout development.
OpenAI says earlier model generations helped engineers design and bring up the chip. Newer models are now assisting with optimization and programming.
Hardware engineering involves many repetitive but demanding processes. Teams must run verification tests, inspect failures, tune tool settings, adjust layouts and repeatedly check whether a change improves speed, power efficiency or chip area without breaking something else.
AI agents can run many of these experiments in parallel. They can review logs, identify patterns and suggest which parameters deserve another attempt. Human engineers still supervise the work, judge tradeoffs and approve changes.
This is a powerful use case because semiconductor projects contain enormous search spaces. Thousands of reasonable design choices may exist, but engineers have limited time to test them.
AI does not need to replace the engineer to create value. It only needs to help the engineer explore more possibilities before the clock runs out.
Jalapeño turns that idea into a wonderfully recursive arrangement. OpenAI’s models helped create hardware that will run future models. Those future models could then help improve the next processor.
OpenAI says Jalapeño is only the first product in a multigenerational computing platform. The real test will be whether AI assistance continues shortening development cycles as the company tackles more complicated designs.
One fast chip is a breakthrough. Repeating the trick would make it a strategy.
Broadcom Supplies the Industrial Muscle
OpenAI may own the architectural vision, but Broadcom supplies much of the industrial machinery needed to transform that vision into deployable hardware.
The company has decades of experience developing custom processors, networking equipment and advanced semiconductor systems. That background matters because designing an accelerator involves more than arranging computing units on a diagram.
A modern AI platform must connect processors, high-bandwidth memory and entire racks of servers. It must move data quickly, manage heat and power, achieve acceptable manufacturing yields and run reliably inside enormous data centers.
Broadcom’s Tomahawk networking technology will help connect Jalapeño systems at scale. Celestica will contribute board, rack and server integration. TSMC will manufacture the processor.
This supporting cast explains why the nine-month schedule should not be interpreted as proof that any AI company can suddenly create a competitive chip before its next office lease expires.
OpenAI had detailed knowledge of its workloads. Broadcom brought mature semiconductor intellectual property and implementation experience. Celestica supplied system-level expertise. TSMC provided advanced manufacturing.
Futurum argues that this combination makes Jalapeño’s timeline impressive but difficult to copy. AI accelerated the work, yet it operated inside an unusually strong engineering and supply-chain structure.
That is the practical lesson. AI works best when paired with deep expertise, high-quality tools and people who understand exactly which problems need solving.
Even brilliant automation needs somewhere solid to stand.
From Chip Experiment to Specialized AI Business
Friar’s latest comments push the story beyond OpenAI’s internal processor.
The company now sees chip design as one of several specialized markets where its models could produce measurable business value. Life sciences and financial services are also high on the list.
Generic chatbots can handle a broad range of tasks, but companies increasingly want systems adapted to their workflows, terminology, data and technical standards. A semiconductor engineer needs different tools from a pharmaceutical researcher or financial analyst.
OpenAI appears to be moving toward that more specialized layer.
The Jalapeño project gives it a powerful case study. OpenAI did not merely ask an AI model to explain semiconductor engineering. It integrated AI into a real development program that produced working samples and reached manufacturing tape-out.
That experience could help the company create tools for other chip designers. AI agents might assist with verification, analyze timing failures, optimize layouts or coordinate complicated engineering workflows.
The opportunity extends beyond advanced AI processors. Automotive chips, networking equipment, consumer electronics and industrial systems all require extensive design and validation work.
However, specialized AI demands reliability. A slightly awkward chatbot answer may be annoying. A mistake embedded in a manufactured processor can cost millions of dollars and months of lost time.
OpenAI will need strong evaluation, human oversight and transparent audit trails if it wants engineers to trust AI with decisions that eventually become physical silicon.
The Economics Are Just as Important as the Engineering
OpenAI’s hardware push also tackles a blunt business problem: inference is expensive.
The company serves a rapidly growing number of consumers, developers and enterprises. Every interaction requires computing resources, electricity, networking and data-center capacity.
Reuters reported that OpenAI’s enterprise revenue rose 32% between June and July, while its overall annualized revenue increased 20% during the same period. Friar said the enterprise and consumer sides of the business had reached roughly equal proportions by midyear—earlier than expected.
More business customers mean more sustained workloads. They also demand predictable costs and clear returns.
Custom chips can give OpenAI greater control over the economics of serving its models. If Jalapeño completes more work using less power, the company can potentially process more requests from the same infrastructure investment.
OpenAI is also experimenting with outcome-based pricing, according to Reuters. Instead of charging only for the amount of computing consumed, it may price certain services according to the business result produced.
That approach makes efficient infrastructure especially valuable. If OpenAI receives a fixed payment for completing a task, every reduction in computing cost can improve its margin.
The company has already cut the price of its lower-cost Luna model by 80%, which Friar said helped increase usage approximately tenfold.
Cheaper AI tends to attract more activity. Efficiency lowers the bill, demand races in and the servers discover that retirement has been postponed again.
A New Front in the AI Chip Competition
Jalapeño enters a market dominated by Nvidia, whose GPUs became the standard engines for training and running advanced AI.
OpenAI is not claiming that it will suddenly stop using Nvidia hardware. Demand for AI infrastructure remains enormous, and custom processors can supplement merchant chips rather than replace them.
Still, Jalapeño gives OpenAI another option.
Google has its Tensor Processing Units. Amazon develops Trainium and Inferentia. Meta has built its own AI accelerators. Microsoft also invests heavily in custom silicon. OpenAI is joining a broader movement among large technology companies to design chips around their own workloads.
The logic is straightforward. General-purpose GPUs offer flexibility and a mature software ecosystem, but companies operating at tremendous scale may benefit from hardware tuned to their specific applications.
OpenAI says Jalapeño is flexible enough to run outside models, yet it remains optimized around the company’s knowledge of language-model inference. That specialization could deliver better performance per watt for the tasks OpenAI performs most frequently.
Broadcom CEO Hock Tan previously told Reuters that Jalapeño could compete with Nvidia’s Blackwell processors and Google’s TPUs. Independent benchmarks will eventually test that assessment.
Competition should benefit customers. Nvidia, custom-chip developers and emerging accelerator companies now have stronger incentives to improve efficiency, lower prices and expand capacity.
The AI chip market is not becoming quieter. It is adding more instruments to the band.
Tape-Out Is a Milestone, Not the Finish Line
Jalapeño has reached an important stage, but several challenges remain before it can operate across gigawatt-scale data centers.
Manufactured chips must pass validation. Production yields must reach sustainable levels. Software tools need to mature. Engineers must integrate processors with memory, networking, boards, racks, cooling and power systems.
A chip that performs brilliantly in a controlled laboratory can encounter very different conditions inside a massive production cluster.
OpenAI and Broadcom have said initial deployment is planned by the end of 2026, followed by expansion across multiple generations. Microsoft and other infrastructure partners are expected to participate in future large-scale deployments.
The company must also navigate shortages and supply constraints. Advanced AI processors require high-bandwidth memory, sophisticated packaging and access to limited manufacturing capacity. Strong architecture cannot magically summon more production lines.
OpenAI’s published benchmark results reduce some uncertainty because working samples are already running demanding models. However, deployment will determine whether the performance, efficiency and reliability survive at commercial scale.
The distinction matters. Tape-out proves that engineers finished the design. Working samples prove that silicon exists. Production proves that the company can manufacture it consistently. Large-scale deployment proves that customers can depend on it.
Jalapeño has cleared the early gates remarkably quickly. Several harder ones remain.
Fortunately for the chip, it has a name that already sounds comfortable under pressure.
AI Is Starting to Reshape Its Own Supply Chain
OpenAI’s expansion into AI-powered chip design offers a glimpse of something larger than one processor.
For years, artificial intelligence depended on advances made elsewhere in the technology industry. Semiconductor companies designed faster chips. Cloud providers assembled data centers. AI laboratories used that infrastructure to train better models.
Now, advanced models are feeding improvements back into the supply chain.
They can help engineers test designs, optimize layouts, analyze failures and program new hardware. The resulting chips may reduce the cost of running future models. Those models can then assist with even more ambitious engineering projects.
This does not remove human expertise. It amplifies it.
Jalapeño succeeded because experienced engineers, specialized AI systems and major industrial partners worked together. OpenAI understood its workloads. Broadcom knew how to implement and connect complex silicon. Celestica handled system integration. TSMC supplied manufacturing.
The nine-month tape-out demonstrates how quickly that combination can move.
OpenAI now wants to package what it learned and carry it into chip design, medicine, finance and other specialized fields. Whether it succeeds will depend on accuracy, economics and the ability to turn impressive demonstrations into reliable products.
But the direction is clear. AI is no longer sitting politely at the software layer.
It is moving down the stack—into servers, networks, factories and silicon.
Jalapeño may be OpenAI’s first custom chip, but its bigger achievement could be proving that AI can help build the physical foundation for its own future.
Sources
- Reuters: OpenAI offers AI for chip design and highlights Jalapeño development
- OpenAI and Broadcom unveil the Jalapeño inference processor
- OpenAI publishes Jalapeño’s first performance results
- Reuters: OpenAI unveils its custom chip designed with Broadcom
- Broadcom: Official Jalapeño announcement
- Futurum Group: Did AI break traditional chip-design timelines?
The Kingy Brief
Follow The Kingy Brief.
One consequential launch, one pricing, limit, or shutdown change, one hands-on test, one exact prompt or Test Pack, and one try / watch / skip verdict.
Free · Choose your subjects · Double opt-in · Unsubscribe anytime