On June 24, 2026, Broadcom helped unveil the most talked-about chip of the year. Its AI revenue is on track to more than double year over year. Its backlog for custom AI chips sits somewhere north of $70 billion. And its stock is still trading roughly a quarter below the high it hit earlier this year.
That gap is worth sitting with, because it’s not really a story about Broadcom’s stock price. It’s a story about how the entire AI chip industry is quietly restructuring itself, and why the company at the center of that restructuring rarely makes the headlines.
What actually happened
OpenAI and Broadcom announced Jalapeño, OpenAI’s first custom-built chip, which Broadcom’s leadership team delivered to Sam Altman and Greg Brockman as a physical sample. It was designed in about nine months, which both companies call the fastest chip development cycle ever completed at this scale. Broadcom CEO Hock Tan told reporters that early testing shows the chip running at roughly half the cost per token of a typical AI GPU, with performance in the neighborhood of Nvidia’s current top-end hardware.
That claim is still unverified by anyone outside the two companies building the chip. No independent lab has benchmarked it yet. Treat it as a target Broadcom and OpenAI are confident they can hit, not a confirmed result.
But the specific number matters less than what the chip actually is. Jalapeño is not a faster GPU. It’s a different category of hardware entirely, and understanding that difference is the key to understanding everything else in this story.
GPU vs. ASIC, in plain terms
A GPU is a general-purpose tool. It was originally built to render video game graphics, and it turned out to be very good at the kind of math used to train and run AI models. That flexibility is its strength: the same Nvidia chip can train a model, run inference, power scientific research, or render a movie. But flexibility has a cost. A lot of the chip’s silicon is doing work that any one specific task doesn’t actually need.
An ASIC, short for application-specific integrated circuit, is the opposite bet. It’s built to do exactly one job and nothing else. Jalapeño only runs already-trained AI models. It cannot train new ones. In exchange for that narrowness, it can be tuned so precisely for that one job that it needs less power and less silicon to do it, which is where the cost savings come from.
If you’ve followed cryptocurrency mining, you’ve already seen this exact trade-off play out. Early Bitcoin miners used ordinary GPUs, because GPUs could do the math and nothing better existed yet. Once Bitcoin ASICs arrived, purpose-built for exactly one hashing algorithm, GPU mining became uneconomical almost overnight. The ASIC didn’t need to be clever. It just needed to be extremely good at one fixed, repetitive task, at a scale where efficiency mattered more than flexibility.
AI inference is starting to look like that same setup. Once a company is answering hundreds of millions of nearly identical requests a day, the flexibility of a GPU stops being worth its cost. That is the entire logic behind Jalapeño, and it’s the same logic behind every other chip in this story.
Everyone did the same math
Here is the part that rarely makes the headline: OpenAI is the fifth company to reach this conclusion, not the first.
Google has been building its own inference chips, called TPUs, for the better part of a decade, and Anthropic runs a large share of its workloads on them. Amazon builds Trainium, and both OpenAI and Anthropic have committed to renting large blocks of it through AWS. Microsoft builds Maia for its own Azure infrastructure. Meta builds MTIA, though it scrapped a more ambitious in-house training chip earlier this year after deciding Nvidia’s software ecosystem was still too far ahead to replace for that job.
Five of the largest AI companies in the world independently arrived at the same decision, roughly on the same timeline, for the same reason: once you’re serving inference at massive scale, renting a general-purpose GPU stops making financial sense.
None of these chips are for sale. If you’re not one of these five companies, none of this hardware exists for you. Nvidia still sells to literally everyone else, which is a real limit on how far this trend can spread.
Where Broadcom actually sits
This is the detail that turns “OpenAI built a chip” into a much bigger story. Broadcom isn’t only OpenAI’s partner on Jalapeño. It’s also the design and manufacturing partner behind Google’s TPU program, under a supply relationship that reportedly runs for years. And it’s the primary partner behind Meta’s MTIA chips too.
Put plainly, Broadcom is currently the co-designer for three of the five companies racing to reduce their dependence on Nvidia, including two that are direct competitors to each other. It doesn’t need to bet on which AI company wins the race. It gets paid by most of the runners.
Marvell fills the same role for the other two. It’s the design partner behind Amazon’s Trainium and Microsoft’s Maia. Between them, Broadcom and Marvell are estimated to control roughly 95 percent of the entire custom AI chip design market. Every one of these chips, regardless of whose name is on it, is also manufactured at TSMC and depends on memory supplied by SK Hynix, Samsung, or Micron, all three of which are reportedly sold out of AI memory capacity through the rest of 2026.
That’s the real shape of this industry right now. It isn’t five companies competing to replace Nvidia. It’s five companies competing, and a much smaller handful of suppliers who get paid no matter who wins.
Nvidia isn’t losing. It’s hedging.
None of this means Nvidia is in trouble. It still holds somewhere around 80 to 90 percent of the overall AI chip market, and its share of training workloads, where models actually get built, remains above 90 percent. Nvidia’s CUDA software platform is deeply embedded in how AI researchers work, and that kind of switching cost doesn’t disappear because a new chip shows up.
What is shrinking is Nvidia’s grip on inference specifically, the part of the business where these new custom chips are aimed. Estimates vary widely, with some analysts projecting Nvidia’s inference share could fall toward 20 to 30 percent by 2028, and others arguing the shift is overstated. Nobody outside these companies knows for certain yet.
Nvidia’s own response tells you it’s taking the threat seriously without panicking. In March, it invested $2 billion directly into Marvell, tying itself into the very custom silicon ecosystem it’s supposedly competing against. That’s not the move of a company that thinks its position is collapsing. It’s the move of a company making sure it stays in the room either way.
So why is the stock still down
Which brings the story back to where it started. Broadcom’s AI business is genuinely booming, and Jalapeño is real proof that its design relationships extend across the entire hyperscaler landscape. But the stock still hasn’t fully recovered from a sharp selloff earlier this year, and it barely moved on the day Jalapeño was unveiled.
That’s not the market doubting the thesis. Jalapeño won’t reach meaningful production until late 2026, with the real revenue ramp expected across 2027 and 2028. Investors are waiting for Broadcom to convert its backlog into actual shipped revenue, not for another announcement. Its next earnings report, expected in early September, is the number the market is actually watching.
None of this is a recommendation to buy or sell anything, and it isn’t financial advice. It’s a reason to watch a company that most casual AI coverage never mentions. The chip war headlines will keep going to Nvidia, OpenAI, and whoever announces the next accelerator. The company quietly getting paid by almost everyone in that race is easy to miss precisely because it never needs to win.
