GT Axis Daily, July 22, 2026

It launched a trio of new Gemini models, led by Gemini 3.6 Flash, the "workhorse" it uses for the high-volume, everyday tasks. The new version does the same work using 17 percent fewer words behind the scenes, which is the part that costs money, so Google cut the price. It also quietly added a security-only version for governments and teased that training has begun on Gemini 4, the next big flagship. The plain read: the race to the smartest model is turning into a race to the cheapest good-enough model.

Moonshot AI unveiled Kimi K3, a model roughly 75 percent larger than the previous open record holder, and promised to give the whole thing away starting July 27. "Open" here means any company can download it and run it on their own computers instead of renting access from a big provider. Why it matters: every time a top-tier free model drops, it drags down what everyone else can charge for the paid versions.

The company broke ground on a facility in Sturgeon County, Alberta, its first in Canada and one of the largest private investments in Canadian history. This is a one-gigawatt site, meaning it will pull as much electricity as a small city, and Meta signed a decade-plus power deal to feed it. The point for leaders: the biggest AI companies are now competing on real estate and electricity, not just software.

The White House is close to a voluntary deal with OpenAI, Google, and Anthropic that would give federal agencies up to 30 days to review the national security risks of a major new model before the public gets it. Voluntary means no law forces it yet, but when the three biggest labs all sign the same page, that page becomes the de facto standard. An announcement is expected before August.

GitHub, the site where most of the world's software actually gets built, launched a standalone Copilot app that runs AI "agents" in safe, walled-off sandboxes. An agent is AI that doesn't just suggest a line of code, it takes the whole task and runs it start to finish while you watch. The sandbox is the guardrail, it lets the AI work without touching anything it shouldn't. For non-coders, this is the shift from an assistant you babysit to one you can hand a job.

The new model, Muse Image, is Meta's first homegrown system for creating and editing images, and it plugs straight into Instagram, WhatsApp, and Facebook's advertising machine. It also does "agentic" work, meaning it can search the web, run its own code, and rethink a draft before handing it back. The business logic: billions of ad images a year is a lot of money to send to a competitor, so Meta brought it in-house.

Quantum computers use the strange rules of physics to attack problems that stump today's machines, think new drugs or unbreakable encryption. The research momentum has been real all year, but the stocks ran up faster than the reality and are now giving a big chunk of it back. This is the sector's pattern: violent swings on feeling, not fundamentals.

At its Advancing AI event today, AMD showed Helios, a single rack packed with 72 of its newest chips built to train and run AI at massive scale, and named Microsoft, Meta, Oracle, and OpenAI as early customers. A "rack" here is one refrigerator-sized unit of computing that data centers buy by the thousands. The story: for years NVIDIA had almost no real competition at this level, and now it does.

National Grid Ventures is taking a 35 percent stake in Joulent, a firm that builds dedicated power for the giant electricity users, mostly data centers, that can't wait years for the normal grid to catch up. The unglamorous truth of the AI boom: the buildings are ready before the electricity is, and whoever solves the power gap gets paid.

They expanded a financing deal from 5 billion to 25 billion dollars to build on-site power plants for AI data centers, the kind that can be set up fast instead of waiting on the grid. Bloom makes fuel cells, boxes that generate electricity right where it's needed. The signal: the money now assumes grid delays are permanent and is routing around them.

An AI infrastructure consortium completed its purchase of Aligned Data Centers, one of the largest such deals on record, a clean signal that the biggest pools of capital now see data centers as core infrastructure rather than a tech side bet.