GT Axis Daily, June 27, 2026

Z.ai (the company formerly known as Zhipu) put out GLM-5.2, and it lands within a single percentage point of Anthropic's best model on a closely watched test of "agentic" work, meaning AI that can plan a task, write code, check itself, and loop until it's done. The catch for the American labs: GLM-5.2 runs at roughly a fifth of the cost and is free to download and run on your own servers. When a free model gets that close to the paid leaders, it puts a ceiling on what everyone else can charge.

DiffusionGemma is an open model that builds whole blocks of text at once instead of one word at a time, which makes it up to four times faster on the right hardware. Think of it like a Polaroid developing: it starts blurry and sharpens across a few passes, rather than typing left to right. The tradeoff is quality, so Google steers it toward speed-critical jobs like live editing and autocomplete, not your most important writing. It's free and open, which is the real tell about where Google wants developers building.

Two more researchers seen internally as key to Google's Gemini model are heading to Anthropic, the latest in a steady walk of top talent out the door. This follows a Nobel laureate leaving for Anthropic and a star researcher going to OpenAI earlier in the year. In a field where a handful of people can swing who builds the best model, this is not a rounding error, and investors treated it that way.

While most of Big Tech rebounded Friday, Nvidia and Alphabet stayed red. The read is simple: these are the two names most exposed to the question hanging over the whole sector, namely whether the enormous money going into AI chips and buildouts pays off fast enough. When the market rewards everyone except your two biggest AI spenders, it's telling you what it's worried about.

Teams that once raced to use as many AI "tokens" as possible (tokens are just the units AI providers bill by, like minutes on an old phone plan) are now getting blindsided by the invoices. Some firms burned through an annual AI budget in four months. A wave of new tools is showing up to track and cap that spend, and the expense-management company Ramp just hit a $44 billion valuation partly on the back of helping businesses rein it in. The story underneath: AI went from "use it freely" to "show me the return," fast.

The shift has a name in the industry, from "tokenmaxxing" (spend whatever it takes) to efficiency (get the same job done for less). One startup, Lindy, moved all of its traffic off Anthropic's models to DeepSeek, a cheaper Chinese open-weight option, and expects to save millions within months. Others, like Uber, slapped monthly limits on AI use after blowing through budgets. For leaders, the lesson is that the "best" model and the "right" model are no longer the same thing.

President Trump signed executive orders this week launching a national push to build a working quantum computer at a Department of Energy lab, plus a mandate to harden government systems against future quantum-powered hacking. Quantum machines use the strange rules of physics to crack problems regular computers can't, think new materials, drug discovery, or breaking today's encryption. Federal backing does for this sector what it did for early space and chips: it de-risks years of expensive research and tells private money it's safe to follow.

Researchers at Duke and IonQ connected three quantum processors sitting a couple of meters apart and got them to share a single fragile quantum state, a first for this kind of setup. In plain terms, instead of building one impossibly large quantum computer, you wire several smaller ones together like servers in a data center. That modular approach is widely seen as the realistic path to machines big enough to be useful.

The robotics company Figure says its California plant now rolls a finished humanoid robot off the line every hour, a 24-fold jump in under four months, with more than 350 built so far. The point isn't a flashy demo, it's manufacturing: getting from "we built one" to "we build them on a schedule" is the hard part that decides who actually ships. The first line is built to make up to 12,000 a year.

Stocks tied to AI infrastructure, the chips, the warehouses, the power, dropped two days running as investors worried that soaring costs could eat into profits down the line. Chip names took the brunt: Marvell fell 5.2%, Arm nearly 4%, Intel 3.4%, and Japan's SoftBank, a huge AI infrastructure backer, slid 5.6%. Nothing broke, and demand is still real. This was the market catching its breath and asking the question it had been ignoring: at these prices, when does the spending pay for itself?

The company behind the Claude models is putting $50 billion into US data centers, starting in Texas and New York, with construction creating thousands of jobs as sites come online through the year. For a model maker to pour that into physical buildings is a statement: owning your own computing is now a competitive weapon, not an expense to outsource. It also deepens the bet that demand for AI computing keeps climbing.