OpenAI Tapes Out Its Own Inference Chip With Broadcom and Calls It Jalapeño

01OpenAI taped out its own inference chip with Broadcom, and named it Jalapeño

For years, OpenAI bought its silicon from someone else. Every token its models generated ran on GPUs designed for a thousand other jobs. On Wednesday the company put its name on a chip of its own.

It is called Jalapeño, and OpenAI built it with Broadcom. The two companies describe it as a custom processor for one task: running large language models after they are trained. OpenAI calls it an inference chip, and the distinction matters. Training a model is a one-time cost. Inference is the bill that arrives every time a user hits enter, and it never stops.

Jalapeño is an ASIC, an application-specific integrated circuit. A GPU is a general tool that happens to be good at AI math. An ASIC is the opposite philosophy. It does fewer things and does them on hardware shaped around the work itself. OpenAI and Broadcom say Jalapeño is tuned for the math that LLM inference actually runs, and built to serve both current models and the larger ones OpenAI says are coming.

That framing tells you what the chip is for. OpenAI is not trying to sell silicon. It is trying to lower the cost of answering its own users. The company says Jalapeño targets performance, efficiency, and scale across its AI systems, which is corporate language for the same underlying problem: the servers that run inference are the most expensive thing OpenAI operates, and until now it did not control the part that mattered most.

The partnership structure follows a path other large buyers have taken. Broadcom designs custom accelerators for hyperscalers that want their own chips without building a semiconductor team from scratch. OpenAI supplies the requirements and the workload knowledge. Broadcom supplies the design and manufacturing relationships. The result carries OpenAI's name and OpenAI's specific demands.

What OpenAI has not said is as notable as what it has. There is no public timeline for volume production, no yield figures, no claim about how many GPUs Jalapeño might replace. The announcement reveals a chip and its purpose, and stops there. For a company that has spent the past year signaling it wanted off its dependence on outside silicon, the reveal is the first physical proof that the plan produced hardware.

02The thing eating Accenture's AI budget, per leaked audio: turning PDFs into slides

Accenture employees are burning a large share of the firm's AI tokens on converting PDFs into presentation slides, according to leaked internal audio reported by 404 Media. The company is now scrambling to spend less. That granular detail, low-value automation quietly draining a token budget, is the ground-level version of a number that keeps getting worse in macro estimates.

The blog post "AI's Affordability Crisis" collects those estimates in one place. Platforms spend $8 to $14 to generate $1 in revenue, the post reports, citing a spread of analyses. Independent writer Ed Zitron, working from a test by the pro-AI research firm SemiAnalysis, calculated that Anthropic subsidizes enterprise customers by up to 40 times and OpenAI by up to 70 times. SemiAnalysis ran random long-horizon coding tasks until they hit each subscription tier's limit.

The doubt is not new. The post traces it to Sequoia Capital's David Cahn, whose September 2023 note "AI's $200B Question" asked where the revenue to justify the spending would come from. Nine months later Cahn re-ran the math as "AI's $600B Question." His estimated revenue gap had tripled in under a year.

Science-fiction author and journalist Cory Doctorow closes the loop from the economics side. In an Ars Technica interview tied to his new book, "The Reverse Centaur's Guide to Life After AI," he discusses how to burst the AI bubble by striking at its roots.

The common thread across the three is unit economics, not vendor headlines. Cost pressure that used to live on platform income statements is now reaching the people typing the prompts. The subsidy figures explain why an enterprise can run up a bill that looks irrational: a buyer paying $20 or $100 a month may be consuming many multiples of that in compute. When the discount narrows, the PDF-to-slides habit stops being free.

What enterprises do next is the variable. The leaked Accenture audio suggests the first lever is usage policy: identify which tasks waste tokens, then throttle them before the contract repricing arrives.

03The Investor Backing Both OpenAI and Anthropic Just Called xAI "a Complete Train Wreck"

Reid Hoffman has skin in this fight. The LinkedIn co-founder put money into both Anthropic and OpenAI, spent a decade on Microsoft's board, and has built and funded enough companies to be hard to dismiss. So when he sat down with Rana el Kaliouby on her Pioneers of AI podcast and said Elon Musk's foundation-model bet was failing, it landed as a verdict rather than a jab.

"SpaceX isn't an AI company," Hoffman said. He described SpaceX's strategy as "buying your way into relevance." On Musk's other venture, he quoted Musk back at himself: xAI, he said, "is, as Elon himself has described, it's a complete train wreck for its kind of building of foundational models." He noted all of its founders have left and it's on its "third restart."

That exodus is documented. By May 2026, all 11 of xAI's original co-founders had departed, a cascade that gathered force in February when Tony Wu, called one of the most operationally central founders, resigned. Musk restructured the teams. The departures continued anyway. xAI's Grok models have drawn persistent criticism for trailing Anthropic and OpenAI on benchmarks.

Then Hoffman turned to the company he funds. He raised alarms about Washington's handling of Anthropic, the lab that has staked its identity on safety. Wired reported this week that at high-stakes White House meetings, Anthropic's cofounder Dario Amodei — labeled a "weirdo" by one official, per the report — has been replaced as the company's point of contact by cofounder Tom Brown.

So the two forces sit side by side. The lab selling caution and alignment is being managed out of the room in Washington. The rival Hoffman calls a train wreck, run by a founder with no such reservations, faces no such freeze. Hoffman, holding stakes on the OpenAI side of the line too, is making the case that the market and the government are reading the field wrong.