GT Axis Daily, July 18, 2026

GPT-5.6 launched in three tiers instead of one: Sol (the flagship, strongest at coding and complex reasoning), Terra (a balanced everyday option), and Luna (the cheap, fast version). Splitting into tiers lets a business match the model to the job instead of overpaying for horsepower it doesn't need on simple tasks. Sol runs $5 per million input tokens (the chunks of text an AI reads) and $30 per million output tokens, meaningfully cheaper than routing everything through one do-it-all model.

Thinking Machines Lab, founded by OpenAI's former chief technology officer, released its first model, called Inkling, as "open-weight," meaning anyone can download the whole thing and run or modify it themselves instead of renting access. It's a large model, 975 billion total parameters, though it only draws on a fraction (about 41 billion) for any given task, which keeps it fast, and it reasons across text, images, and audio. The company admits it isn't the strongest model available today, open or closed. The real bet is customization: businesses fine-tune it themselves rather than accept whatever OpenAI or Google decides to ship.

Moonshot AI's Kimi K3 has 2.8 trillion parameters (the internal knobs a model tunes during training, roughly speaking, more capacity to learn patterns), making it the biggest open-weight model available anywhere. It handles a 1-million-token context window (close to 750,000 words held in memory at once) and reads images natively. Moonshot claims it beats Anthropic's Opus 4.8 and OpenAI's GPT-5.6 Sol on several benchmarks, though it still trails the very best closed models. The gap between Chinese and American frontier AI, which used to be measured in years, is now measured in weeks.

Gemini 3.5 Pro, the flagship model Google demoed back in May, is now delayed by several months because it fell short of internal expectations, particularly on coding and long, multi-step reasoning tasks. Alphabet shares dropped roughly 4% in a single day on the news, erasing about $200 billion in market value. The lesson for any leader watching from outside AI: promising a capability at a conference is cheap, shipping it on time against fast-moving competitors is the actual test.

Apple filed suit accusing OpenAI of systematically stealing trade secrets, alleging OpenAI's hardware chief, a former Apple vice president, coached job candidates to bring physical Apple parts to interviews and directed departing employees on how to evade Apple's security checks on the way out. More than 400 former Apple employees now work at OpenAI. This is a hard reversal from the two companies' 2024 partnership, and it signals Apple sees OpenAI as building competing hardware, not just competing software.

Microsoft is preparing to launch Project Perception, a tool that mixes its own models with OpenAI's and Anthropic's, automatically routing each security task, bug-finding, vulnerability analysis, incident response, to whichever model handles it best and cheapest. It's positioned as a lower-cost rival to Anthropic's own security tool, Mythos. The bigger idea: instead of betting the farm on one AI provider, Microsoft is building a broker that shops around, exactly the kind of infrastructure move that doesn't make headlines but ends up everywhere.

Anthropic gave Claude Enterprise admins new tools: usage and cost analytics broken down by team and person, plus a self-service flow for HIPAA readiness (the federal healthcare privacy law), letting a company's own admin review the compliance agreement and turn on health-data-safe settings without waiting on a sales negotiation. For any business that touches patient data, this cuts a weeks-long compliance process down to a few clicks.

The EU's Chips Joint Undertaking and quantum computer maker Pasqal launched Q-PLANET, a six-year, €50 million program spanning 28 organizations across 11 countries. The goal is manufacturing quantum computing components, lasers, atom chips, vapor cells, the way ordinary computer chips get mass-produced: standardized and repeatable, instead of one-off lab builds. Quantum computers are still mostly science-fair projects. Making the parts mass-manufacturable is the unglamorous step that turns a lab demo into an actual industry.

TSMC, the Taiwanese foundry that manufactures chips for NVIDIA, Apple, and nearly everyone else in AI, reported second-quarter profit up 77% year over year to about $22 billion, and raised its 2026 spending plan to $60-64 billion. It's also adding another $100 billion to its Arizona operations, bringing total US investment to $265 billion, mostly for advanced 2-nanometer (a measure of how small and efficient a chip's circuits are, smaller generally means faster and more power-efficient) production and packaging. Demand for AI chips isn't cooling off. The world's biggest chipmaker keeps needing more factories, not fewer.

NVIDIA is partnering with Japanese industrial giants Fujitsu, Kawasaki Heavy Industries, FANUC, and Yaskawa, who together build a huge share of the world's factory robots, plus more than 20 other Japanese firms, to build shared AI models for robotics using NVIDIA's Cosmos platform. A related deal includes 27,500 of NVIDIA's newest Rubin chips to power an AI factory in Japan. For anyone outside tech, this is AI moving off screens and into machines that physically build things: the training happens in simulation, then gets deployed to real factory floors.

Chevron signed a 20-year deal to build a dedicated 2.67-gigawatt power plant, enough electricity for roughly two million homes, in West Texas solely to supply a new Microsoft AI data center campus in Pecos. It's "behind-the-meter" power, meaning it's built specifically for this one customer rather than feeding the public grid. First power is targeted for 2028. AI data centers have gotten so power-hungry that the biggest tech companies are now signing multi-decade energy contracts directly with oil and gas majors instead of waiting on utilities.

Brookfield expanded its financing partnership with Bloom Energy, whose fuel cells generate power on-site instead of waiting on grid connections, from $5 billion to $25 billion. The point is speed: getting new AI data centers powered up fast is now more valuable than getting the cheapest power, because the bottleneck for AI growth isn't chips anymore, it's electricity.

3M and Microsoft announced a partnership to scale up 3M's cooling and materials technology, branded EBO, for AI data centers, where managing heat from densely packed chips is a growing engineering problem. It's a reminder that the AI buildout is pulling in unexpected industrial suppliers, not just chipmakers and cloud companies. Anyone who solves a physical bottleneck, power, cooling, materials, has a seat at the table now.