Sent to 20 readers
Race to AGI Daily Digest - Friday, June 26, 2026
Share:
TLDR
OpenAI’s Jalapeño custom chip, built with Broadcom and TSMC, shows frontier labs are moving compute in‑house for cheaper, faster inference.
IBM’s proposed 0.7nm nanostack architecture promises up to a 6x jump in AI compute density over current nodes.
Anthropic’s allegation that Alibaba ran a large‑scale Claude distillation attack pushes model cloning and IP theft into the open.
Patronus AI’s $50M raise for digital world models underscores evaluation and safety as a first‑class layer in the AI stack.
The Full Story
Following Monday’s talent moves, Tuesday’s GUI agents, Wednesday’s Doubao benchmark shock, and Thursday’s Qualcomm–Modular deal, Friday zooms all the way down to the silicon.
OpenAI just pulled the curtain back on Jalapeño, a custom chip built with Broadcom -> and TSMC -> and headed for Microsoft’s -> cloud. The goal is simple: cheaper, faster inference for ChatGPT‑class workloads. That’s our debt‑fueled compute arms race turning into a design race, as labs stop waiting for general‑purpose GPUs and start rolling their own. You can dig into the specifics here: Jalapeño custom chip rundown -> and the OpenAI profile ->.
Right alongside that, IBM -> is talking about 0.7nm "nanostack" chips that promise up to a 6x jump in AI compute. If Qualcomm’s Modular buy was about owning the stack from software up, IBM’s move is about stretching the physics from silicon down. Check the roadmap details here: IBM 0.7nm nanostack explainer ->. No surprise that ASML is the lone green chip in a red market today.
Here’s the thing: as compute gets more specialized, the fights over what runs on it get sharper. Anthropic says Alibaba -> ran a huge Claude distillation attack, effectively trying to clone a frontier model instead of licensing access. That hits our "AI security and access controls" and "licensing the training data and models" storylines at once. You can follow the allegations here: Claude distillation dispute -> and revisit Anthropic’s role ->.
On the safety side, Patronus AI raised $50M from backers like Greenfield Partners ->, Lightspeed ->, Notable Capital ->, and Samsung -> to build "digital world models" that stress‑test AI systems before they hit users: Patronus digital world models round ->. Think of it as a simulated universe where agents can break things safely.
And because no week is complete without a content fight, Google DeepMind’s film deal with A24 has set off a creative‑industry backlash. That loops right back to Tuesday’s Getty–OpenAI deal: if labs don’t want to face Anthropic‑style distillation claims, they’ll need Hollywood‑grade licensing too. The film story is here: DeepMind–A24 AI film deal backlash -> and the Google profile ->.
So by Friday, the arc is pretty clear: the race to AGI is turning into a vertical game. Custom chips, new process nodes, synthetic worlds, and hard‑fought licenses are stacking up under the models we see on the surface.
Get This Delivered Daily
Join thousands of AI professionals who start their day with Race to AGI.