The most valuable businesses in the world rarely make their money selling you something once. They make it by collecting a small fee every single time you use the thing they sold you.
That is the quiet genius of a toll booth. You can own the car, buy the gas, and drive the open road, and you will still hand over a few coins every time you cross the bridge.
For three years, Nvidia (NVDA) has run the only toll booth that matters in artificial intelligence (AI). Its graphics processing units, or GPUs, sit underneath nearly every chatbot reply, every generated image, and every line of code a machine writes for someone.
The company controls roughly 90% of the chips that power AI data centers, according to Fortune, and in October 2025, it became the first company in history worth $5 trillion.
So every time you ask ChatGPT to draft an email, a sliver of that toll quietly flows to a chipmaker in Santa Clara, California.
On June 24, the company paying the most tolls decided to build its own road.
OpenAI unveiled Jalapeño, its first custom chip, designed in-house and manufactured with Broadcom (AVGO), and it is pointed at one specific job Nvidia has owned outright.
What OpenAI built with Broadcom in nine months
Jalapeño is designed for inference, the work of running an AI model after it has already been trained, according to Yahoo Finance. Training teaches the model. Inference is the model answering you.
The distinction sounds technical, but it is where the money increasingly lives. The chip showed better performance per watt than current state-of-the-art processors in early testing, a direct swipe at Nvidia, the same report noted.
More AI:
- OpenAl, Cerebras deal was supposed to be good news
- Microsoft’s CEO sends the Al industry a strong warning
- Nvidia pours cold water on Al fears
What caught my attention was not the chip itself but how quickly it was created. OpenAI designed the silicon, leaned on its own models to accelerate the work, and went from early schematics to fabrication-ready in nine months, a cycle the industry normally measures in years, according to VentureBeat.
The chip is “part of our long-term full-stack infrastructure strategy to make compute more abundant,” OpenAI president Greg Brockman said in a statement.
OpenAI is not the only one trying this. Amazon (AMZN), Google (GOOGL), Microsoft (MSFT), and Meta (META) all design their own AI chips, and Amazon and Google now rent that silicon to outside customers, according to Yahoo Finance.
Broadcom’s chief executive framed the larger lesson bluntly. Companies that want to lead in AI “cannot, should not rely on some other third-party GPU to do it for you, because it’s such a key part,” said Broadcom CEO Hock Tan, as Axios reported.
Inference is the slice of Nvidia that matters most
Here is the part Wall Street is actually chewing on. Training a frontier model is an expensive, occasional event. Inference happens billions of times a day, forever, every time a person opens a chatbot.
That makes it the steady, recurring toll, and the fastest-growing slice of AI spending as these tools reach more people.
OpenAI has been hunting for inference alternatives since 2025, frustrated that Nvidia’s chips were not fast enough for products like its Codex coding assistant, according to CNBC.
It has since spread its bets across AMD (AMD), Cerebras (CBRS), and Broadcom, and analysts had already revamped Broadcom’s price target on the OpenAI deal months ago, TheStreet reported.
Related: The SpaceX–OpenAI IPO boom: What investors must do now
The timing is not an accident. As OpenAI prepares for a widely expected initial public offering (IPO), cheaper inference is one of the few levers it has to narrow steep losses, since it runs its models far more often than it trains them, the VentureBeat report indicated.
None of this is a sure thing. Similar in-house chip efforts at Microsoft and Meta ran into delays or failed to match Nvidia’s performance, and analysts say custom silicon does not threaten Nvidia’s dominance in the near term, according to NBC News.
Even so, Nvidia’s share of the broader AI chip market has slipped to roughly 70% as custom silicon gains ground, Tom’s Hardware noted.
Nvidia is not blinking. The company says its chips still power the vast majority of OpenAI’s inference work and offer the best performance per dollar, and Sam Altman has said OpenAI hopes to remain “a gigantic customer for a very long time,” according to CNBC.
What Jalapeño means for the AI stock trade
When I ran Nvidia’s roughly $4.9 trillion market value against the International Monetary Fund‘s 2026 gross domestic product (GDP) tables, the company landed somewhere odd.
The popular line that Nvidia is worth more than every economy on Earth except two is now outdated. After a recent pullback from its $5 trillion peak, it sits behind three.
- The United States leads all economies at about $32.4 trillion in output, according to the IMF.
- China ranks second near $20.9 trillion, the IMF confirms.
- Germany is third at roughly $5.45 trillion, just ahead of Nvidia, the IMF notes.
- Nvidia’s market value sits near $4.9 trillion as of June 2026, according to CompaniesMarketCap.
- Japan’s entire economy runs about $4.4 trillion, smaller than a single chipmaker, according to the IMF.
That valuation rests almost entirely on one belief — that AI spending keeps compounding and that Nvidia keeps its cut. Jalapeño is the first real crack in the second half of that bet.
Broadcom is the quieter winner. It already sits on a $73 billion AI backlog and is targeting $100 billion in annual AI chip revenue by 2027, according to the Tom’s Hardware analysis. Its stock rose more than 1% on the Jalapeño news, the Yahoo Finance report also noted.
The one job in the headline is inference, the toll you pay every time the machine answers a question. OpenAI is not trying to beat Nvidia at training, the hard and costly work of building a model from nothing. It is trying to stop renting the booth on the part that runs billions of times a day.
For anyone holding an S&P 500index fund, this is not abstract. A startling share of that index now rides on one company keeping a grip it has held since the AI boom began. OpenAI just showed that grip can be pried loose at the edges, and it did it in nine months.
Watch two things from here:
- Whether OpenAI’s chips actually reach live data centers by the end of 2026, as promised
- Whether Nvidia’s next earnings call offers any hint that inference demand is starting to leak
The toll booth is not closing. For the first time, though, somebody just built a road around it.
Related: OpenAI makes IPO decision amid Anthropic, SpaceX fervor
