AI News Summary — Week 26.2, 2026

This week in AI: OpenAI builds its own chip, Google keeps bleeding talent to Anthropic, and the Trump administration puts the brakes on GPT-5.6 — but with a lighter touch than the Anthropic ban.

OpenAI and Broadcom Unveil "Jalapeño" — A Custom AI Inference Chip

On June 24, OpenAI pulled back the curtain on Jalapeño, its first custom-designed AI accelerator — built in partnership with Broadcom. This is a big deal.

The chip is designed from the ground up for LLM inference, not adapted from general-purpose AI accelerators. Early testing shows it delivers "performance per watt substantially better than current state-of-the-art." That's the kind of efficiency claim that makes NVIDIA nervous.

Key details:

Nine-month development cycle — reportedly the fastest ASIC development ever for a high-performance chip. OpenAI used its own models to accelerate parts of the design process.
Multi-generation roadmap — deployment starts end of 2026, scaling to gigawatt data centers with Microsoft and other partners.
Full-stack play — OpenAI is now building products, models, and the chips underneath them. The goal: lower inference costs, faster responses, more affordable access.

Greg Brockman put it simply: "By designing more of the stack ourselves, we can serve more intelligence with greater efficiency."

For an industry dependent on NVIDIA's CUDA ecosystem and struggling with inference costs, this is OpenAI's bid to control its own destiny.

Google Loses Two More Gemini Researchers to Anthropic

If last week felt like a hemorrhage, this week made it worse. According to reports on June 24, two more leading Google researchers are heading to Anthropic:

Jonas Adler — worked on Google's AI coding effort
Alexander Pritzel — led pretraining work

This comes on top of last week's departures: Noam Shazeer (co-author of the Transformer paper, now at OpenAI) and John Jumper (Nobel laureate, now at Anthropic). That's four senior researchers in a matter of days.

The market noticed. Alphabet stock took another hit this week, adding to the 6% plunge from Jumper's departure last week.

Why Talent Keeps Fleeing

It's not just money, though that's part of it. Anthropic and OpenAI are both approaching IPOs, offering pre-IPO equity that a $2 trillion company can't easily match.

But compute access is the other half. Inside Google, researchers have reportedly had to queue for TPU access as external demand soaks up capacity. A researcher who can't get hardware to run experiments is a researcher with a reason to listen when a rival calls.

The pattern is clear: Anthropic is winning the talent war, one departure at a time.

Trump Administration Asks OpenAI to Delay GPT-5.6

In a move that echoes the Anthropic ban but with a lighter touch, The Information reported this week that the Trump administration asked OpenAI to stagger the release of its next major model, GPT-5.6.

Under the arrangement:

GPT-5.6 will launch in limited preview — only accessible to a group of enterprise customers initially
The administration will approve customer access on a case-by-case basis during the preview period

This is notably more favorable than the treatment Anthropic received. Earlier this month, the government issued an export control directive requiring Anthropic to suspend access to its Mythos 5 and Fable 5 models entirely — including to the company's own non-US citizen employees.

The contrast raises questions about whether the administration's approach is actually about security, or more about which companies it favors. OpenAI's "speed wins" messaging and support for an American AI exports program may have bought it goodwill that Anthropic didn't get.

Anthropic Accuses Alibaba of Largest-Ever Claude Cloning Attack

In the latest chapter of the US-China AI rivalry, Anthropic is accusing Alibaba of launching the largest distillation attack ever attempted.

The numbers: Between April 22 and June 5, operators "affiliated with Alibaba and Alibaba Qwen" generated 28.8 million exchanges with Claude through nearly 25,000 fraudulent accounts — mining capabilities in agentic reasoning, software engineering, and long-horizon tasks.

Anthropic sent a letter to Senators on June 10 (ahead of a Senate committee hearing) calling for:

Updated antitrust laws to let AI firms share info about Chinese attack tactics
More export controls on chips to hamstring Chinese compute access
Laws penalizing Chinese labs' distillation behavior

Meanwhile, Alibaba is suing the Trump administration, claiming it was blacklisted as a "Chinese military company" with "no basis in fact or law."

The timing is notable: the Alibaba attacks happened after Trump's April memo warning that cloning attempts were "unacceptable."

What's Hot This Week

Jalapeño chip: OpenAI + Broadcom unveil custom LLM inference accelerator, deployment by end of 2026
Talent drain continues: Two more Gemini researchers (Adler, Pritzel) heading to Anthropic
GPT-5.6 delayed: Trump admin asks OpenAI for staggered release with case-by-case approval
Alibaba attack: 28.8M Claude exchanges through 25K fake accounts — largest distillation attack ever
Legal spat: Alibaba sues US Department of Defense over "Chinese military" designation

That's a wrap for Week 26.2. Stay tuned!

Sources: The Verge, OpenAI, Ars Technica, The Next Web, Bloomberg, Reuters