GT Axis Daily, July 7, 2026

A model called GLM 5.2, built by the Chinese firm Z.ai, came within one percentage point of Anthropic's best on a closely watched test of AI that can carry out multi-step tasks on its own, and it runs at about a fifth of the cost. "Open-weight" is the key phrase, meaning you can download it and run it yourself instead of paying to rent access through someone else's service. US companies, feeling the sting of rising AI bills, are adopting it fast: on one popular developer platform, daily usage grew roughly 27 times over in its first full week. Why it matters: the thing everyone assumed America had locked up, the best models, is turning into a price fight, and the low-cost option is foreign.

It released two new "realtime" models, gpt-realtime-2.1 and a smaller mini version, built for voice assistants that answer with almost no lag. The upgrade cuts the delay before a response by at least a quarter and handles noisy audio and interruptions better. For a non-technical read: this is the plumbing that makes an AI phone agent or live translator feel like a real conversation instead of a walkie-talkie. Why it matters: OpenAI is competing on speed and price at the developer layer, the same efficiency race pulling buyers toward cheaper options everywhere.

Gemini 3.5 Pro, the model meant to go head-to-head with the best from OpenAI and Anthropic, slipped to July 17 after early testers flagged weak reasoning, shaky coding, and problems handling long multi-step tasks. Google scrapped the old design and is rebuilding it from the ground up. The cheaper, faster sibling, Gemini 3.5 Flash, already shipped and is holding up well. Why it matters: in a market where buyers are getting price-sensitive and foreign models are closing the gap, the last thing the perceived laggard can afford is to keep missing its own launch dates.

Combined capital spending from Amazon, Google, Meta, and Microsoft is on track for as much as $725 billion this year, nearly double last year, most of it going into data centers and the computers inside them. Capital spending is simply the money a company sinks into long-lived physical assets, the buildings and machines it expects to use for years. For months investors flinched at these numbers. Today they leaned in, bidding the whole group up. Why it matters: the mega-caps are betting a staggering sum that AI demand keeps rising, and the market's mood on that bet swings day to day, which is exactly why the steady names still outperform the speculative ones over time.

Oak Ridge National Laboratory, Cleveland Clinic, and IBM ran the first-known calculations of a material used to make fuel for fusion power, the clean, nearly limitless energy source scientists have chased for decades. The specific bottleneck: producing enough tritium, a rare ingredient most fusion machines need, has long blocked the whole field. Doing this math on a quantum machine is an early step toward solving it. Why it matters: this is the kind of concrete, useful result quantum computing has promised for years and rarely delivered, a sign the technology is inching from science project toward practical tool.

The phone-chip giant is reportedly in talks to acquire Tenstorrent, a startup that designs AI chips on RISC-V, an open, license-free chip blueprint anyone can build on instead of paying for a proprietary one. The reported price is $8 billion to $10 billion, though reports suggest the deal has cooled and may not close. The prize would be a real seat at the AI hardware table now dominated by NVIDIA. Why it matters: everyone wants a piece of the chips that run AI, and buying a credible design team is the fast lane in, if the deal actually happens.

US data-center power demand is set to jump from 31 gigawatts in 2025 to 41 this year and 66 in 2027, roughly doubling in two years, according to Goldman Sachs. A gigawatt is about enough to power a mid-sized city, so this is city-scale demand landing on the grid fast. The catch: in the regions carrying most of the buildout, the wait just to plug a new site into the power grid now runs four to seven years. Why it matters: money is no longer the scarce ingredient in AI, power is, and that hands quiet pricing power to whoever controls electricity and the land near it.