OpenAI presents Jalapeño chip results and accelerates the race for AI infrastructure

OpenAI presents Jalapeño chip results and accelerates the race for AI infrastructure

OpenAI has released new results for Jalapeño, its first custom chip for artificial intelligence, reinforcing a trend that is expected to shape the coming years of the industry: major AI companies do not want to rely only on traditional GPUs to scale their products.

Jalapeño is an ASIC chip, developed in partnership with Broadcom and designed specifically for AI inference — the stage in which an already trained model processes a user request, generates a response, performs a task or powers an agent. Reuters had already reported in June 2026 that the chip was created to accelerate OpenAI’s infrastructure and reduce costs compared with dependence on Nvidia GPUs.

Now, according to results released by OpenAI and reported by The Verge, Jalapeño delivered 1.5 to 1.9 times more AI work per watt in tests with models such as GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T, in addition to 1.7 to 3.6 times lower end-to-end latency compared with the best results recorded on Nvidia GB200 or GB300 systems.

In practice, this means OpenAI is trying to solve one of the biggest bottlenecks in generative artificial intelligence: how to deliver faster, cheaper and more scalable responses.

For end users, this may appear as faster ChatGPT responses, more responsive agents and greater stability during periods of high demand. For companies and developers, the advance may affect API costs, latency, scalability and availability of AI-based applications.

What happened?

OpenAI presented benchmarks for Jalapeño, comparing the chip’s performance with reference systems currently used for AI inference.

According to The Verge, the company used the InferenceX platform to measure Jalapeño’s performance across different models. The comparison included results from systems with Nvidia GB200 and GB300, two of Nvidia’s most advanced chip systems for AI workloads.

OpenAI says Jalapeño delivers more work per watt and lower latency. This combination matters because AI systems usually involve a trade-off between throughput and latency.

Throughput is the amount of work a system can process within a given period. Latency is the response time perceived by the user.

In generative AI applications, both points are critical. A platform needs to handle millions of requests, but it also needs to respond quickly enough to maintain a good experience.

According to Richard Ho, OpenAI’s vice president of hardware, Jalapeño attempts to offer the “best of both worlds”: greater processing capacity and lower response time.

What is Jalapeño?

Jalapeño is OpenAI’s first custom chip for artificial intelligence.

It was developed with Broadcom and built for a specific purpose: running already trained AI models more efficiently.

This differentiates it from traditional GPUs, which are more flexible processors used across multiple stages of the AI lifecycle, including training, inference and other parallel computing workloads.

Jalapeño is an ASIC, short for Application-Specific Integrated Circuit.

In practice, an ASIC is a chip designed to solve a specific type of problem. It is not as flexible as a GPU, but it can be much more efficient for the task it was designed to perform.

In OpenAI’s case, that task is inference for language models and generative AI systems.

Reuters reported that OpenAI designed the chip together with Broadcom, that Canadian manufacturer Celestica will be responsible for the server systems, and that the chips will be used only by OpenAI itself.

Inference is not training

This point is essential.

Jalapeño was not created to train new models from scratch. It was designed for inference.

Training is the stage in which a model learns patterns from large volumes of data. This phase requires enormous computing capacity, memory, communication between chips and long processing periods.

Inference is the stage in which the already trained model is used to answer a question, generate code, summarize a text, analyze an image or execute an action.

When someone uses ChatGPT, Codex or an AI agent, they are triggering an inference workload.

That is where Jalapeño comes in.

This means OpenAI will still depend on other suppliers for several parts of its infrastructure. Axios highlighted that the chip was not designed for training and that OpenAI still depends on Nvidia and other partners for its broader computing strategy.

Why does OpenAI want its own chip?

The short answer is: cost, scale and control.

Generative artificial intelligence depends on extremely expensive infrastructure. Running large models for millions of users requires data centers, energy, high-speed networks, specialized memory and advanced chips.

Over the last few years, Nvidia has become the main supplier of GPUs for AI. This placed the company in a dominant market position and created enormous dependence for AI labs, cloud providers and companies training or operating advanced models.

For OpenAI, relying only on third-party GPUs creates three main problems.

1. Inference cost

Every response generated by a model has a cost.

Even when the user does not notice it, there is spending on processing, memory, energy, cooling, networking and data center operations.

The more users rely on ChatGPT, APIs and AI agents, the greater the pressure on infrastructure.

If a custom chip can deliver more work per watt, the operational cost per response may decrease.

This does not automatically mean API prices will fall. But it means OpenAI may improve its margins, increase capacity and offer more complex products without raising costs at the same pace.

2. Hardware availability

Demand for AI chips remains high.

Companies such as OpenAI, Google, Microsoft, Amazon, Meta, xAI, Anthropic and cloud providers compete for the same supply chain: GPUs, HBM, networking equipment, energy and data center capacity.

By developing its own chip, OpenAI creates an additional alternative to meet part of its demand.

This does not eliminate dependence on Nvidia, but it reduces pressure on a single source of hardware.

Reuters reported that labs such as OpenAI and Anthropic face difficulty obtaining enough computing power to run increasingly advanced chatbots and coding applications.

3. Control over the architecture

When a company uses generic hardware, it needs to adapt its models and systems to what the supplier offers.

With its own chip, OpenAI can optimize the hardware for the real behavior of its models, products and data centers.

This allows deeper adjustments in:

• latency;
• energy consumption;
• memory usage;
• communication between servers;
• execution of large models;
• agent support;
• cost per token;
• performance predictability.

This verticalization is similar to what other companies have already done.

Google has TPUs. Amazon has chips such as Trainium and Inferentia. Microsoft develops its own family of AI chips. Meta also invests in its own accelerators.

The difference is that OpenAI, until recently, was seen mainly as a model and product company. With Jalapeño, it moves into a deeper layer of infrastructure.

Why does this matter for developers?

For developers, the news may seem distant, but it is not.

Most developers will not program directly for Jalapeño. They also should not expect to buy the chip, since OpenAI does not plan to sell it to third parties, according to Axios.

Even so, custom chips can directly affect the experience of those building AI applications.

Faster APIs

If OpenAI can reduce inference latency, applications that depend on real-time responses may improve.

This is relevant for:

• chatbots;
• coding copilots;
• autonomous agents;
• customer support;
• voice assistants;
• document analysis;
• search tools;
• educational applications;
• enterprise automations.

In many products, a difference of seconds changes the user experience.

An agent that takes too long to respond feels stuck. An assistant that responds quickly feels more reliable.

More responsive agents

AI agents require several successive calls to the model.

An agent may need to interpret a request, consult tools, call APIs, analyze results, review its own answer and execute an action.

If each stage has high latency, the experience becomes slow.

The latency reduction promised by Jalapeño may be especially important for this type of application, because agents depend on inference chains.

Greater stability at scale

When demand increases, AI platforms may face queues, usage limits or performance degradation.

A more efficient chip may help OpenAI serve more requests with the same amount of energy and infrastructure.

This matters for companies using AI in production, because availability and predictability are essential.

Possible impact on pricing

There is no guarantee of price reduction.

But if inference cost falls, OpenAI gains room to create more competitive plans, increase limits, offer more powerful models or improve the cost-benefit ratio of its APIs.

For developers and startups, this can make AI applications more economically viable.

Why does this matter for companies?

For companies, the news reinforces that AI is not just software.

Adopting AI at scale depends on an entire chain: models, chips, data centers, energy, networks, security, APIs and governance.

When OpenAI develops its own chip, it is trying to gain more control over that chain.

This can bring impacts in several areas.

Less dependence on external bottlenecks

Companies that use OpenAI products depend on OpenAI’s ability to keep its services available.

If the company can increase capacity with its own chips, it may reduce some infrastructure bottlenecks.

This is important for organizations integrating AI into critical processes such as customer support, software development, legal analysis, technical support and internal operations.

AI more integrated into workflows

The lower the latency, the more natural it becomes to use AI in workflows.

A coding tool that responds quickly can be used continuously. A slow assistant becomes something consulted only occasionally.

Inference efficiency can determine whether AI will be merely a complementary tool or a constant layer inside enterprise systems.

Competition among providers

If OpenAI improves its infrastructure, other providers need to respond.

Google, Anthropic, Meta, Amazon, Microsoft and xAI also compete on performance, cost and capacity.

This competition tends to benefit companies that consume AI, because it pressures the market to deliver better models, faster APIs and more competitive prices.

The impact on Nvidia

Nvidia remains in an extremely strong position.

Jalapeño does not replace GPUs for training, and OpenAI itself indicates that it will continue working with partners such as Nvidia. The Verge reported that Richard Ho said OpenAI’s computing strategy still includes strong partners and that Jalapeño will not replace the company’s entire chip stack.

Even so, the signal is important.

When giant customers start developing their own chips, it shows that the market is looking for alternatives.

Nvidia may remain dominant, but the inference segment may become increasingly contested.

This is especially relevant because, as models mature, operating cost can become as important as training cost.

Training a model is expensive, but serving that model to hundreds of millions of users is also expensive.

If inference becomes the largest cost center, chips specialized in that stage gain strategic value.

The role of Broadcom

Broadcom is a central piece of this strategy.

Reuters reported that OpenAI designed Jalapeño with Broadcom and that the chip was sent to TSMC for manufacturing. The publication also highlighted that companies such as Meta, Amazon and Google turned to companies like Broadcom and Marvell to develop their own chips because this type of technical knowledge is difficult to replicate internally.

This shows that the AI chip race is not only a dispute between Nvidia and AI labs.

It involves a complex chain with:

• chip designers;
• manufacturers such as TSMC;
• HBM memory suppliers;
• server integrators;
• data center providers;
• networking companies;
• energy suppliers;
• cloud operators.

Generative AI depends on this entire chain to work.

What is still unclear?

Despite the released results, there are still open questions.

Real production scale

OpenAI plans to deploy Jalapeño in small volumes by the end of 2026 and increase capacity in 2027, according to The Verge and Axios.

But it is still unclear how many chips will be used, in which data centers, for which products and what percentage of OpenAI’s total workload will be served by this new infrastructure.

This point is decisive.

A chip may perform well in benchmarks, but its real impact depends on deployment scale.

Comparison with real workloads

Benchmarks help, but they do not tell the full story.

Production systems need to deal with irregular traffic, different models, failures, updates, network limits, cooling, load balancing, demand spikes and operational costs.

The real test will be Jalapeño’s performance in production, serving products such as ChatGPT, Codex and OpenAI APIs at large scale.

Continued dependence on Nvidia

Jalapeño is focused on inference.

This means OpenAI will still need GPUs and other accelerators for training, research and specialized workloads.

Axios highlighted that OpenAI itself considers Jalapeño only one part of its computing strategy, alongside Nvidia, AMD and other components.

Therefore, the news should not be read as “OpenAI will abandon Nvidia”. The most likely scenario is a hybrid infrastructure.

Will Jalapeño be sold to other companies?

Based on what has been disclosed so far, no.

Axios reported that OpenAI does not intend to offer the chip to third parties. According to Richard Ho, the company has so much internal demand for capacity that it cannot imagine when it would be able to sell these chips to other customers.

This makes sense.

Jalapeño is not a commercial hardware product. It is a piece of OpenAI’s internal infrastructure.

The goal is not to compete directly with Nvidia by selling chips. The goal is to improve OpenAI’s own ability to operate its models and products.

Why does this signal a new phase of AI?

Until now, much of the generative AI race has been perceived by the public as a model race: who has the smartest chatbot, the best image generator, the most useful agent or the most efficient coding model.

But behind that, there is an even deeper dispute: the race for infrastructure.

Companies that control hardware, data centers and energy may gain a huge competitive advantage.

OpenAI is trying to stop depending only on infrastructure provided by third parties and move toward a more verticalized model.

This may bring the company closer to the strategy of giants such as Google and Amazon, which have been building custom chips for years to optimize their services.

Impact on the AI market

Jalapeño may accelerate some trends.

1. More custom chips

AI companies should continue looking for their own accelerators.

The logic is simple: the higher the inference volume, the more sense it makes to optimize hardware for specific workloads.

2. More pressure on cost per token

Cost per token will become increasingly important.

Companies want better models, but they also want fast and inexpensive responses.

Energy efficiency should become a competitive advantage.

3. More competition in the infrastructure layer

Nvidia will remain strong, but it will face more pressure from internal chips, TPUs, custom ASICs, AMD, Cerebras and other alternatives.

4. More integration between model and hardware

Models may be designed with the hardware they will run on in mind.

This can influence model architecture, quantization formats, context windows, routing, caching and inference techniques.

5. AI as critical infrastructure

The news reinforces that AI is already essential infrastructure for companies.

Just like cloud computing, databases and networks, the AI layer will have its own suppliers, bottlenecks, costs, dependencies and risks.

Possible benefits

Jalapeño may bring important benefits to OpenAI and users of its services.

Among them:

• faster responses;
• greater energy efficiency;
• lower operational cost per inference;
• more capacity to serve demand;
• better support for agents;
• greater control over infrastructure;
• less dependence on a single supplier;
• a foundation for future chip generations.

These benefits, however, depend on OpenAI’s ability to scale production and integrate the chip into its data centers at large scale.

Risks and limitations

There are also risks.

The first is complexity. Creating a custom chip requires high investment, specialized engineering and dependence on a global semiconductor supply chain.

The second is the speed of AI evolution. Models change quickly. A very specific chip may become less suitable if model architectures change significantly.

The third is scale. If OpenAI cannot produce and deploy Jalapeño in sufficient volume, the impact will be limited.

The fourth is dependence on partners. Even with its own chip, the company still depends on Broadcom, TSMC, memory suppliers, server integrators, energy and data centers.

The fifth is competition. Google, Amazon, Microsoft, Meta and other companies are also investing in custom chips. Jalapeño is important, but it does not put OpenAI alone in this race.

Conclusion

Jalapeño’s results show that OpenAI is entering a more mature phase of its infrastructure strategy.

The company does not want only to create advanced models. It wants better control over the costs, latency, availability and scale required to operate AI products used by millions of people and companies.

The chip does not eliminate dependence on Nvidia and does not solve every computing bottleneck. It was also not built to train new models. But it represents an important strategic step: OpenAI is building its own foundation for large-scale inference.

For developers, the impact may appear in faster APIs, more responsive agents and greater stability in AI applications.

For companies, the move reinforces that artificial intelligence is not just a software layer. It is a complete chain of infrastructure, hardware, energy, data, models and products.

The AI race is no longer only about who has the best model. Now it is also about who can operate those models with greater efficiency, lower cost and larger scale.

In this scenario, Jalapeño shows that OpenAI wants to compete not only in model intelligence, but also in the physical rails that make that intelligence possible.

FAQ

What is OpenAI’s Jalapeño chip?

Jalapeño is OpenAI’s first custom chip for artificial intelligence. It was developed in partnership with Broadcom and designed for AI inference, meaning it runs already trained models and responds to user requests.

Does Jalapeño replace Nvidia chips?

Not completely. Jalapeño is focused on inference, while OpenAI still depends on Nvidia and other partners for training, research and other computing workloads. The most likely scenario is a hybrid infrastructure.

Will Jalapeño be sold to other companies?

So far, no. OpenAI has indicated that it intends to use the chip internally because its own demand for computing capacity is very high.

Why is AI inference so important?

Inference is the stage in which a model responds to the user or performs a task. In products such as ChatGPT, Codex and AI agents, inference happens millions of times a day and represents a huge operational cost.

How does this affect developers?

If Jalapeño helps OpenAI reduce latency and increase capacity, developers may benefit from faster APIs, more responsive agents and greater stability in AI-based applications.

Does the Jalapeño chip train AI models?

No. It was designed for inference, not training. Training models still requires large GPU clusters and other specialized accelerators.

Why did OpenAI create its own chip?

OpenAI wants to reduce costs, increase capacity, improve energy efficiency and gain more control over its AI infrastructure. This is strategic for scaling products such as ChatGPT, Codex and enterprise APIs.