GT Axis Daily, July 14, 2026

GPT-5.6 came out in three flavors: Sol for the hardest work, Terra for everyday tasks, and Luna for cheap and fast. The pitch is agents, meaning software that carries out a whole task on its own instead of handing you a draft, and OpenAI shipped a "ChatGPT Work" version aimed at exactly that. The number that matters for a budget owner is price: Terra does roughly what the last flagship did at half the cost per unit of work.

Seedream 5.0 Pro can split one generated image into ten or more separate transparent layers, so a designer can open it in Photoshop and move things around instead of starting over. It also renders readable text in about fourteen languages, which is the boring feature that actually makes it usable for real marketing work. TikTok's parent now has a design tool with a billion-user distribution channel attached.

IBM told investors its quarter came up short, roughly $17.2 billion against an expected $17.9 billion, and the stock fell about 24%. CEO Arvind Krishna's explanation is the real headline: in the last weeks of June, customers moved money out of software and into servers, storage, and memory, racing to lock in hardware before shortages and price hikes hit. Translation for a leader: the AI compute crunch is no longer an infrastructure story, it is now a line item competing with your software renewals.

Apple's federal complaint accuses OpenAI of systematically taking trade secrets, alleging it coached departing Apple staff on dodging exit security and told job candidates to bring actual Apple hardware parts to their interviews. More than 400 former Apple employees now work at OpenAI, including its hardware chief, a former Apple VP. Two companies that were partners eighteen months ago are now in court, and the reason is that OpenAI decided to build devices.

Google told Meta it could not supply as much Gemini capacity as Meta wanted to buy, and Meta has since told staff to be more careful with their usage. Sit with that. Google is spending north of $180 billion this year on infrastructure, owns some of the largest compute in the world, and still cannot serve a paying customer the size of Meta. Compute, not talent and not ideas, is the binding constraint in AI right now.

Copilot agents in VS Code can navigate a live web page, click, type, read the error console, and take screenshots, all inside the editor. The old loop was: AI writes code, developer switches to a browser, tests it manually, copies the error back into chat. That round trip is now gone. It's on by default, and the agent cannot touch a tab until you explicitly share it.

June revenue came in at roughly $13.8 billion, up 68% from a year ago, and first-half revenue is up 36%. Two details matter more than the headline. First, its most advanced manufacturing and its advanced packaging (the step that stitches chips together into an AI accelerator) are sold out through year-end. Second, TSMC has told Apple, NVIDIA, and Qualcomm that leading-edge prices are going up 5% to 10%. When the world's only real chip factory is full and raising rates, everybody downstream eats it.

Pasqal and the EU's Chips Joint Undertaking launched Q-PLANET, a pilot production line for neutral-atom quantum processors, pulling in 37 partners across 12 countries and running through 2031. The interesting choice: they're publishing open manufacturing blueprints so smaller companies can build on the same standards instead of each reinventing the factory. That's the same playbook that made regular chips cheap.

The custom data-center chip, codenamed "Iris," is slated to start manufacturing in September. Every large AI buyer is now trying to build its own silicon, for one reason: NVIDIA's margin is their cost. Meta getting cut off from Google's Gemini capacity, then racing its own chip to the fab, is the same story told twice.

The Hyperion campus in Richland Parish, Louisiana is now planned at 5 gigawatts, roughly ten million square feet, up from a $27 billion plan announced last October. For scale, 5 gigawatts is about what five large cities draw. The local detail worth knowing: parish teachers' annual bonuses jumped from $10,000 to over $50,000 on the tax revenue, which is exactly the political argument these projects will be won or lost on.

It expanded its compute partnership with Google and Broadcom for multiple gigawatts of capacity, most of it US-based. Separately it's in early talks with Samsung about a custom chip on a 2-nanometer process (smaller number, denser and more efficient chip), and it's reportedly preparing IPO paperwork for as soon as October. Nothing is built yet, but the shape is clear: secure the compute, own the silicon, then sell the story to public markets.

Local opposition, permitting fights, and power constraints have stalled projects at that scale in 2026. That is the counterweight to every gigawatt headline above. Compute is the bottleneck, and the bottleneck behind the bottleneck is electricity and a county board that has to approve it.