OpenAI is finally making chips.
When many people saw this news, their first reaction was: Nvidia is in trouble.
But what I see is precisely the opposite.
The most significant meaning of the first chip, Jalapeño, is not that it's coming directly for Nvidia.
This is the first time OpenAI has publicly admitted it's not satisfied with being just a model company.
What it wants to control is the entire process of producing intelligence.
From models, to chips. From data centers, to energy. From training, to inference. From producing Tokens, to selling Tokens.
Jalapeño appears to be a chip on the surface, but it's actually more like a roadmap.
OpenAI has finally laid its ambition on the table.
I. The Model Gap is Shrinking, the Compute Gap is Widening
Since the explosion of large models, almost all attention in the AI industry has been on the models.
The industry was shocked by GPT-4, then Claude caught up, Gemini caught up, DeepSeek delivered high cost-performance, Meta pushed open source. Every release, everyone focuses on the same set of things: parameters, leaderboards, coding ability, math ability, long context, multimodal capabilities.
Models are, of course, important. But a change has already occurred: the window of model leadership is getting shorter. Today, a model is just released, and within months, the open-source community, competitors, and cloud providers catch up. Performance gaps still exist, but they are increasingly difficult to constitute a long-term moat on their own.
The things that truly create differentiation are moving to a deeper, more foundational level. Compute supply, inference costs, system throughput, networking capabilities, data center construction, energy acquisition. These aren't as flashy as model launches, nor do they go viral immediately. But they determine whether an AI company can run long-term.
Jensen Huang recently said something: Nvidia systems might not have the lowest purchase price, but they can generate the lowest-cost Tokens, the highest Token throughput, and ultimately bring the highest revenue.
Huang's statement was direct. The industry has long complained that Nvidia is expensive. Huang didn't argue about the purchase price; he reframed the problem in another dimension: don't look at how much the machines cost to buy, look at the production cost per Token.
This is the new ledger for the AI era. Servers and GPUs are not the ultimate unit; the Token is.
OpenAI happens to be at the very center of this problem.
ChatGPT handles a massive volume of requests daily, Codex consumes even more inference steps, and in the future, there are Agents, video generation, robotics, long reasoning chains. The more useful the model, the greater the Token consumption. The more successful the product, the thicker the inference bill.
The brutal part is here: the more users OpenAI has, the more money Nvidia makes. The stronger OpenAI's products, the heavier the underlying compute tax.
If every Token has to pass through an external hardware platform paying a toll, it's hard for OpenAI to have a complete moat. It can have the strongest model, a super entry point, a developer ecosystem. But the core production cost is always in someone else's hands.
This is the essence of Jalapeño. OpenAI has started building its own Token factory.
II. GPT Begins Designing GPT
The most underestimated detail about the Jalapeño chip is the nine-month tape-out.
Traditional high-performance ASIC projects typically have cycles of 18 to 36 months. Advanced processes are even more troublesome—architecture, verification, physical implementation, packaging, software stack, debugging—any hiccup can rapidly escalate costs. OpenAI and Broadcom compressed the cycle to nine months.
This cannot be understood as the chip industry suddenly becoming simple. OpenAI did not spontaneously grow a semiconductor supply chain. Broadcom has deep experience in custom chips and network infrastructure; Celestica handles boards, racks, and systems engineering.
What OpenAI truly contributed is something scarcer: it knows how future models will run.
Many chip companies building AI accelerators face the challenge of guessing the workload. Model architectures change, inference methods change, service patterns change. Once a chip is taped out, it's not as easy to roll back in the physical world as it is in the software world.
OpenAI doesn't have to rely entirely on guesswork. Running ChatGPT, Codex, and APIs daily, it knows which kernels are used most, which memory transfers are most wasteful, which network bottlenecks most affect cluster efficiency, which latencies directly hurt product experience. It also knows how future Agent products will consume inference resources.
This experience was once just backend engineering knowledge; now it's being written into the chip architecture.
A crucial statement in OpenAI's official press release: OpenAI used its own models to accelerate parts of the design and optimization process. It also said that models provided to users are helping improve the infrastructure that will run future models.
GPT has started participating in designing the machines for the next generation of GPT.
For decades, the chip chain was: first design the chip, the chip runs the software, the software runs the AI. Now, the chain is turning back: AI helps humans design chips, which then run the next generation of AI.
Once this loop is established, nine months might just be the beginning. The future could be six months, three months, or even more frequent iterations.
The chip industry had its own rhythm, the model industry had its own rhythm. The former was slow, the latter fast. Jalapeño is pulling these two rhythms together.
If this step succeeds, OpenAI's flywheel will become formidable. Better models help design better chips, better chips lower the running cost of the next model generation, lower costs support more users and products, more users and products generate more real workload data, which in turn defines the next generation of chips.
This is the cycle OpenAI truly wants.
III. Cutting the Inference Tax, Controlling Cash Flow
Jalapeño is not a training chip; it targets large language model inference. This is key.
Training is like building an aircraft carrier. A huge one-time investment, requiring extremely strong general-purpose capability, and constant adaptation to new models, architectures, and experiments. The training market still heavily depends on Nvidia—not just the GPUs, but the entire platform: CUDA, networking, systems, software libraries, developer ecosystem.
Inference is more like a fleet of taxis. Running daily, hourly, by the minute. Every time a user asks a question, an API responds, an Agent takes a step forward, inference happens. It cares more about low latency, low cost, high throughput, high utilization.
Training burns big money in phases; inference burns daily cash flow.
This is also the most painful problem for AI companies as they enter the commercialization stage. GPT training is expensive once, but inference happens every day. The Agent era will further amplify this problem—one task may involve dozens or even hundreds of model calls. Long context, chain-of-thought reasoning, multimodal generation, code execution—all continue to push Token consumption higher.
Jalapeño is precisely targeting this inference tax. It's more like OpenAI's own TPU. Google, Amazon, Meta, Microsoft have all taken similar routes—as long as the workload is sufficiently large, custom ASICs make economic sense for high cost-effectiveness.
OpenAI now meets these conditions. Real requests, a product roadmap, a model team, industry partners like Broadcom, and immense cost pressure.
Jalapeño doesn't need to be sold externally to prove its value. As long as it makes ChatGPT answers cheaper, makes Codex run faster, and makes API margins higher, it's meaningful.
OpenAI also mentioned that Jalapeño will reduce data transfers, balance compute, memory, and network resources, bringing actual utilization closer to theoretical peaks. Compute is expensive often because it's not fully utilized—GPUs waiting for networks, memory transfers slowing down computation, poor scheduling causing idle time—all waste eventually turns into electricity bills and capital expenditure.
The purchase price is only the first layer; system efficiency is the final account.
IV. OpenAI is Looking More and More Like Apple
Many interpret Jalapeño as OpenAI challenging Nvidia, but I think OpenAI doesn't want to become the next Nvidia; it's more like emulating Apple.
Apple's greatest strength has never been any single point. The iPhone is strong, iOS is strong, the A-series and M-series chips are strong, the App Store is strong. But the truly difficult thing to compete against is how all these things are placed within the same closed loop.
Chips are optimized for the system, the system is optimized for applications, and the application experience in turn defines the next generation of chips. This closed loop allows Apple to deliver experiences under the same battery, same size, and same thermal constraints that others find hard to replicate.
OpenAI is building something similar. The model is the intelligence kernel, ChatGPT is the super entry point, Codex is the development tool, API is the ecosystem distribution layer, Jalapeño is the custom chip, and data centers are the AI factories.
Over the past two years, OpenAI CEO Altman has repeatedly discussed chips, energy, nuclear fusion, data centers. Looking back now, he might not have been chasing trends at all; he has stopped planning OpenAI in the way an AI startup would.
If Nvidia sells shovels, then OpenAI wants to own the mine.
Nvidia wants to be the factory equipment supplier for all AI companies, selling GPUs, networking, systems, software ecosystems, AI factory solutions—its ideal customers are every company that needs to produce Tokens.
OpenAI wants to build a factory for itself, selling not the equipment, but the final, generated intelligence.
In the short term, OpenAI still depends on Nvidia. Training and general-purpose computing still require the GPU platform, and Jalapeño likely won't cover all workloads quickly. It will probably first enter OpenAI's most certain, largest-scale, highest-optimization-return inference scenarios.
In the long term, cracks have appeared. When model companies start having their own chip roadmaps, Nvidia's customers are no longer just customers. They also become another type of player in the AI infrastructure landscape.
Words Beyond the Layout
Over the past two decades, the most important asset on the internet was traffic. Whoever controlled the users, controlled the value.
Today, new rules are emerging in the AI era.
Models are becoming more like traffic, while compute is becoming more like land.
Models will iterate, products will change, leaderboards will keep refreshing. But those factories that produce intelligence—chips, networks, data centers, energy—will increasingly concentrate in the hands of a few players.
GPT designing GPT looks like just another tape-out.
But what it truly announces is this:
OpenAI is no longer satisfied with being the smartest company; it wants to be the company that controls the production of intelligence.
This article is from WeChat public account:Layout Beyond, author: Huahua
This article is from WeChat public account:Layout Beyond, author: Huahua, title image from: AI-generated
