Custom Silicon, Ford's Reversal, and Computer Use

OpenAI and Broadcom unveiled Jalapeño on 24 June, OpenAI's first custom chip. It is an inference processor, built to run trained models rather than train them, and OpenAI says it targets better performance per watt and lower running costs than the general-purpose GPUs it depends on today. The launch lands as rivals Amazon and Google push their own silicon to cut their reliance on Nvidia.

The detail: Per OpenAI and Broadcom's joint announcement, OpenAI designed the accelerator while Broadcom supplied the silicon and networking and Celestica builds the boards, racks and systems. OpenAI says it went from design to tape-out in nine months, which it calls the fastest ASIC development cycle yet in advanced semiconductors, and that engineering samples are already running its own workloads in the lab, including a model it names GPT-5.3-Codex-Spark. Early testing shows performance per watt "substantially better" than current hardware, though OpenAI says final performance is still being measured, with a technical report to follow. Broadcom's chief executive, Hock Tan, said the platform will deploy at gigawatt scale with Microsoft and other partners. The design work was itself sped up using OpenAI's models, and Tom's Hardware describes the part as a reticle-sized ASIC.

Related: The move extends a pattern Pulse24 has tracked, of large buyers building around Nvidia rather than only buying from it. The pressure behind it is supply: days earlier, Google restricted Meta's access to Gemini over compute capacity, and Micron warned the chip shortage will run past 2027. The caveat is execution. Jalapeño is not in production, OpenAI targets first deployment only at the end of 2026, and custom-silicon programmes often slip between tape-out and volume.

Ford said it has brought in about 350 veteran engineers over the past three years, many of them former staff or hires from suppliers and internally called "gray beards," to fix quality problems that followed an aggressive move to automate design and inspection with AI. As Bloomberg first reported, chief operating officer Kumar Galhotra said Ford had been "relying more and more on automated quality systems" with disappointing results, and the specialists now "hunt for failure points before a part ever reaches the plant floor." Ford credits the reversal with helping it become the top mainstream brand in the 2026 J.D. Power Initial Quality Study, its first time in 16 years.

The detail: Ford's vice-president of vehicle hardware engineering, Charles Poon, said the company had "mistakenly" believed that "just introducing artificial intelligence" and feeding it Ford's existing design requirements "would produce a high-quality product." The flaw, he said, was that experienced staff left before their judgment was captured in the training data, so the automated tools amplified weak inputs rather than catching design flaws. Ford scored 152 problems per 100 vehicles in the J.D. Power study, better than the 193 that left it 14th a year earlier, and says the wider fix should cut costs by about $1bn this year. That is Ford's own account, not an independent audit, and it sits against a rougher record: Ford has issued 51 recalls in 2026 covering more than 11 million vehicles, the most of any US carmaker.

Related: Ford is not dropping AI; it is using people to improve the data the models learn from, and it has added a 40-person software quality team and more than 100,000 automated tests. The episode echoes arguments that the hard part of AI work is now review, not generation, and warnings about "verification debt" as unreviewed machine output piles up. The counterpoint is that automation is working in other settings: Stripe's agents merge more than 1,000 pull requests a week with no human-written code, in a domain where a bad change is cheaper to catch than a shipped vehicle.

Google added "computer use" to Gemini 3.5 Flash, released on 24 June as a public preview. Rather than routing between a reasoning model and a separate agent model, one Flash model can now see a screen and act on it across browser, mobile and desktop, through the Gemini API and Google's Enterprise Agent Platform.

The detail: The model pairs the capability with two optional enterprise safeguards: it asks for confirmation before sensitive or irreversible actions, and it halts automatically if it detects an indirect prompt injection, backed by what Google calls targeted adversarial training. Google is candid that this is not enough on its own, framing it as "defence in depth" and telling developers to add sandboxing, human-in-the-loop checks and strict access controls. On the OSWorld-Verified benchmark for screen control, Google reports 78.4 against GPT-5.5's 78.7, a 0.3-point gap, but every score on that board is reported by the model's own maker, with no independent verification as of late June. To lower the barrier to trying it, Google shipped a GitHub reference implementation and a Browserbase-hosted demo alongside the API. The feature is the native successor to the standalone Gemini 2.5 Computer Use model from October 2025.

Related: It lands in a busy month for agent plumbing. Last week Cloudflare and NewCore gave agents their own identities; this week Okta extended its Cross App Access framework to more than 25 partners, citing a finding that only 13% of firms could stop a rogue agent. The safeguards matter because the attacks are real: an amateur hacker used Claude and OpenAI agents to compromise 14 companies, and the Five Eyes alliance warned that AI cyber capability is "months, not years" away. Investor money is following the gap: Patronus AI raised $50m to build systems that stress-test agents before deployment.

---