OpenAI teams up with Broadcom to design its own AI chip, codenamed Jalapeño, to reduce its dependence on Nvidia, which controls nearly the entire AI chip market. The main reason is control over the supply chain — no more waiting in line to buy GPUs — and the ability to design chips tailored specifically to its own workloads, rather than the general-purpose chips Nvidia sells. If it succeeds, this is a signal that Big Tech is starting to shift from being “buyers” to becoming “makers” of AI chips themselves, which would shake up the balance of power across the entire AI infrastructure market.
What the chip everyone in AI is watching looks like
Jalapeño isn’t a chip sold to the general public yet, so there are no numeric specs to compare the way you would with a typical GPU (cores, TDP, bus width) — nothing like the graphics cards we’re used to.
What can be confirmed clearly is the production structure: OpenAI teamed up with Broadcom to handle the design, while actual manufacturing goes to TSMC — the same formula Google uses for its TPUs and Amazon uses for Trainium.
Put simply, OpenAI isn’t competing with Nvidia by selling general-purpose chips to others. It’s competing by having its own chip that no longer depends on anyone else.
When ChatGPT broke down because of a GPU shortage
Anyone who uses ChatGPT often has probably run into rate limits or unusually slow responses at certain times — that’s the result of GPU shortages.
OpenAI’s compute demand is growing faster than Nvidia can manufacture chips to keep up. Order queues are long, and the cost per chip keeps climbing because every AI company is fighting over the same line.
When demand outstrips supply, what follows is service disruption, users getting throttled, and engineering teams constantly firefighting.
This is a critical turning point — when the infrastructure an entire business depends on ends up in the hands of a single supplier, the risk is too high to ignore. Having your own chip, then, isn’t just about cost — it’s about business survival.
Where Jalapeño fits into OpenAI’s hardware plan
Jalapeño doesn’t replace the existing deals with Nvidia, AMD, Oracle, or Microsoft — it’s an additional layer OpenAI is adding so it doesn’t have to rely entirely on GPUs from a single supplier.
In simple terms, OpenAI will keep buying GPUs from Nvidia for most of its heavy workloads, while running its own custom chip in parallel for specific tasks (such as some inference workloads) where it can control cost and supply on its own.
Designing a chip in-house takes enormous time and capital — no company does it lightly. But after the lesson learned from the GPU shortage, spreading this risk has become a necessity, not an option.
From 100% reliance on Nvidia to having its own alternative
Previously, OpenAI had almost no other choice — it had to queue up for Nvidia GPUs mainly through cloud partners, with prices fluctuating with the market and nearly all the negotiating power sitting with Nvidia. This was a risk a company at this scale had been carrying for a long time.
With Jalapeño added to the mix, the situation changes considerably. It’s not about ditching Nvidia — it’s about having a “second card” to negotiate with and to spread out supply chain risk.
| Factor | Before: fully reliant on Nvidia | Now: supplemented with Jalapeño |
|---|---|---|
| GPU/chip source | Through cloud partners, queued by market availability | Has its own in-house designed chip running in parallel |
| Price negotiating power | Sits almost entirely with Nvidia | Some risk now spread out |
| Speed of deploying heavy workloads | Fast, system ready to use | Fast, system ready to use |
What this chip will actually change for AI users
If Jalapeño works out as planned, the first thing ChatGPT users will notice is that peak-traffic periods — like evenings or the start of a school term — won’t feel as laggy as before, because compute no longer has to wait in line behind a single GPU vendor.
On the pricing side, API costs also have a chance to trend downward over the long run, since the cost per token won’t be tied solely to whatever price Nvidia sets — though this part is still a matter for the future, with no confirmed figures right now.
More importantly, there’s stability during traffic surges. The more OpenAI plans to roll out new compute-hungry features — like video generation or long-running agents — the more it needs its own backup compute. Otherwise, whenever Nvidia can’t manufacture fast enough, the entire platform’s service could stall immediately.
Jalapeño compared to other custom AI chips from Big Tech
Honestly, in this game OpenAI is arriving quite late compared to its peers. Google has been building its own TPUs for a long time now, and Gemini currently runs primarily on TPUs, relying on Nvidia the least among this group. Amazon also has Trainium and Inferentia, which it has offered AWS customers as GPU alternatives for a while now. Microsoft has Maia, but it’s still mixed heavily with Nvidia — which is roughly the position OpenAI (tied to Microsoft) currently occupies.
| Factor | OpenAI Jalapeño | Google TPU | Amazon Trainium/Inferentia | Microsoft Maia |
|---|---|---|---|---|
| Project start point | Just started | Started earliest | Started a while ago | Started after TPU |
| Current reliance on Nvidia | Still heavily reliant | Least reliant | Mixed | Still heavily reliant |
In short, Jalapeño is still the newest kid in this group — it has a long way to go before it catches up with the others.
The pros and cons of OpenAI building its own chip
Building your own chip isn’t just for show — it’s purely a matter of long-term cost calculation. Cut out the margin you’d otherwise pay Nvidia, and put that money into more compute instead. But the reality is that Nvidia built CUDA over the course of a decade, while Jalapeño is just starting from zero. That gap can’t be closed in a year or two.
Pros
- +Reduces long-term cost — no longer paying Nvidia's full margin
- +More negotiating leverage with Nvidia when signing GPU purchase contracts
- +Full control over its own hardware roadmap, designed specifically for its own workloads
Cons
- −Will take several years before an in-house chip can genuinely replace Nvidia at large scale
- −Supply chain risk now shifts to dependence on Broadcom and TSMC
- −Still lacks a software stack mature enough to rival CUDA, which the entire industry's developers are already used to
The real cost of breaking away from Nvidia
The chip’s price tag is the smallest part of the bill. The real cost is hidden in the software team having to rewrite drivers and compilers from scratch, instead of using CUDA, which the whole industry already knows by heart. That’s years of work, not just a small patch.
Another factor is manufacturing yield — a new chip from a new production line carries a higher risk of defect rates than a production line Nvidia has been running for a decade. The more complex the chip (5nm and below), the more painful it gets when yields aren’t stable.
Crucially, OpenAI isn’t actually becoming “independent” — it’s just shifting the bottleneck from Nvidia to Broadcom instead. Design, packaging, and manufacturing queues at TSMC still depend on a small handful of players, just with a different party now holding the stronger hand.
The real turning point isn’t today — it’s next year
Jalapeño is still genuinely in the design phase — there’s no mass production to point to yet. What’s worth watching is when the TSMC production queue actually starts moving, and what the first-run yield looks like.
Nvidia itself isn’t sitting idle either — every custom chip Big Tech announces is a signal that major customers are spreading their risk away from a single GPU vendor. If OpenAI pulls this off, Google, Amazon, and Meta — which already have their own custom silicon — will likely accelerate their efforts too. And the question that follows is whether Nvidia will respond with pricing, or with a faster new roadmap.
Worth following next: Jalapeño’s actual production timeline, how Nvidia’s stock behaves around progress announcements, and which Big Tech player will be next to announce its own custom chip.