GT Axis Daily, June 25, 2026

Anthropic, the maker of Claude, publicly accused Chinese giant Alibaba of running thousands of fraudulent accounts to pull off "distillation," which is using one company's finished AI to train your own cheaper knockoff, like photographing every page of a textbook instead of writing one. Why it matters: distillation is the quiet threat to every frontier lab's moat, because the expensive part is building the first great model and the cheap part is imitating it. This is one of the first times a top US lab has named a Chinese rival and called it theft outright, and it lands just as Washington tightens the rules on who can touch the best models.

The new "Claude Tag" lets anyone in a Slack channel type @Claude to hand it a task, and one shared Claude works out in the open where the whole team can see what it is doing and pick up where a colleague left off. It holds context across the channel, can plug into a team's tools, data, and code, and replaces Anthropic's older one-off Claude in Slack app. Why it matters: most AI assistants are one-person, one-window tools. Dropping a single shared assistant into the place teams already talk turns it from a personal gadget into shared infrastructure, and shared infrastructure is a lot harder to rip out. It runs on Claude Opus 4.8 and is in beta for Enterprise and Team customers.

The ChatGPT maker, working with Broadcom, unveiled its first custom processor, called Jalapeño, built specifically for "inference," the run-the-model-for-users step that eats most of an AI company's compute bill. They took it from blank page to factory-ready in nine months, which they call the fastest chip development ever, and used OpenAI's own models to speed up the design work. Why it matters: every big AI player is trying to escape Nvidia's premium, and a purpose-built chip that does more work per watt of power is how they claw back margin. First chips are due to go live by the end of 2026.

Researchers at Duke and IonQ linked three separate quantum modules into one entangled system, the first time three remote "nodes" have been wired together this way. In plain terms, today's quantum computers hit a ceiling on how many qubits fit in a single box, and networking several boxes into one machine is how the field gets past that, the same way data centers stitch thousands of servers into one. They hit it at 84 to 88% fidelity, meaning accuracy, without the shortcuts that usually inflate these results. Why it matters: it nudges "modular" quantum from theory toward a real path to scale.

The company reported $41.5 billion in revenue for its May quarter, up 346% from a year ago, with gross margin of 84.9%, meaning it keeps roughly 85 cents of every sales dollar before operating costs, a level chipmakers almost never touch. It guided next quarter to $50 billion and about 86% margins, and called out HBM4, its newest high-bandwidth memory, already shipping in volume to its lead customer. Why it matters: memory used to swing from glut to shortage every couple of years, and Micron just signed multi-year "Strategic Customer Agreements" that lock in demand, turning the most cyclical corner of chips into something that reads like a subscription.