OpenAI Drops the Chips Behind Its $38.5 Billion Loss to Build Its Own

01OpenAI Stops Buying the Chips That Drove Its $38.5 Billion Loss and Starts Building Them

OpenAI spent its way into a $38.5 billion loss buying other people's silicon to run its models. Now it wants to make its own.

The company, with Broadcom, introduced a custom processor called Jalapeño this week. It is not a training chip. OpenAI says Jalapeño is built for inference: the work of answering a user's prompt, the part of the business that scales with every query and never stops once a model ships. The Verge reports the chip is an ASIC, an application-specific integrated circuit designed to do one job rather than the general-purpose math that GPUs handle. OpenAI says it is tuned for current and future large language models.

The choice of inference over training is the whole story. Training a model is a one-time capital event. Inference is the recurring bill, and that bill has become the pressure point for the entire industry. SemiAnalysis, described as a pro-AI semiconductor shop, ran long-horizon coding tasks until they hit the usage limits on OpenAI and Anthropic subscription tiers. Writer Ed Zitron, working from those numbers, calculated that OpenAI could be subsidizing some customers by as much as 70 times the price they pay. Estimates of the gap run from $8 to $14 in cost for every $1 of revenue.

A company losing money on every token has two levers: charge more, or make the token cheaper to produce. Jalapeño is a bet on the second. Owning the silicon lets OpenAI tune the hardware to its own models and skip the margin a chip vendor adds on top.

The plan still runs through Broadcom, which handles the design and manufacturing partnership, so OpenAI trades one dependency for another rather than escaping supply chains entirely. The company gives no shipping date, no benchmark figures, and no per-query cost comparison against the GPUs it runs today. It says the chip improves performance, efficiency, and scale. Whether those gains close a 70x subsidy gap is a number OpenAI has not put on the table.

What is on the table is the direction. The largest buyer of AI inference hardware has decided that buying is the problem.

02An Anthropic Cofounder Got Called a 'Weirdo' at the White House. The Company Sent Someone Else.

The Trump White House has stopped dealing with Anthropic CEO Dario Amodei. At high-stakes meetings, one official called the cofounder a "weirdo," according to Wired, and the company swapped in cofounder Tom Brown.

That substitution reads as a concession. Amodei spent years positioning Anthropic as the safety-first lab, the one willing to slow down where rivals sprint. Inside an administration that prizes speed and loyalty, the posture wore thin.

The friction is now open. MIT Technology Review recapped the company's running feud with the US government. The dispute traces to April, when Anthropic said it had built a model called Mythos. The outlet laid out three things to watch as the standoff escalates.

Investors are starting to talk. Reid Hoffman, a backer of both Anthropic and OpenAI, raised alarms about how the government has handled Anthropic's pulled models. He made the remarks on the Pioneers of AI podcast, a venue where he rarely defends one portfolio company against Washington in public.

Hoffman used the same appearance to call xAI "a complete train wreck" and to say SpaceX "isn't an AI company." Those lines drew the headlines. The quieter point was his unease about a government willing to yank a private lab's models, a precedent that reaches every frontier developer.

The two sides now sit across a table. Anthropic argues that caution is its product. The White House, going by the "weirdo" line and the personnel swap, would rather deal with Brown.

What that buys Anthropic is unclear. Brown holds the cofounder title but not Amodei's public identity as the industry's safety conscience. Whether a friendlier face changes the administration's treatment of the company's models is the question the next round of meetings will answer.

03The internet checks that a human is acting. AI agents broke that in three places at once

Cloudflare now lets an AI agent deploy code before it has an account. Run wrangler deploy --temporary and a Worker goes live for 60 minutes, with no signup, no OAuth flow, no token to paste. Claim the account inside that window and it becomes yours. Ignore it and it expires on its own. The company says background agent sessions have no human to click through a browser, so the browser step had to go.

Anthropic moved the other direction in the same stretch. It began requiring identity verification for certain Claude capabilities, routing people through partner Persona. You supply a government photo ID and a live selfie. The stated purpose is confirming who is using the tool before access is granted.

A third signal came from research, not a product. A widely shared writeup reframes prompt injection as role confusion. The argument runs like this: a model receives one continuous string holding system prompts, user messages, tool outputs, and its own prior replies. It has no reliable way to tell whose words are whose. Edit the string and you edit the model's reality.

Stack the three and one assumption sits under all of them. Access control begins by confirming the actor's identity and role. Each layer presumes a person is acting and the user is the role issuing orders. Agents void that presumption. Cloudflare hands the agent its own identity so it can ship. The human, on Anthropic's side, must now prove a separate one. The model itself cannot keep the roles straight to begin with.

For anyone deploying agents, the bill arrives at all three layers at once. An autonomous agent spins up accounts and infrastructure no person registered, which muddies ownership, billing, and cleanup. That same agent reads tool output that, by the role-confusion account, can pose as the user's own instructions. Identity stops being a one-time login and becomes a standing risk surface the developer has to manage.