GT Axis Daily, July 20, 2026

Moonshot AI's Kimi K3 has 2.8 trillion parameters, meaning it's built from roughly 2.8 trillion adjustable settings that shape how it reasons, the largest of any model whose underlying code and weights anyone can download and run themselves. It doesn't beat the top US models on every benchmark, but it closes the gap fast and does it at a fraction of the price. The signal for leaders: the US lead in frontier AI is shrinking, and it's shrinking on cost as much as capability.

GPT-5.6 ships in three tiers, Sol for heavy lifting like coding and research, Terra for everyday balanced work, and Luna for fast, cheap tasks, so businesses can match the model's horsepower to the job instead of overpaying for one-size-fits-all. It's a pricing and positioning move as much as a technical one.

Gemini 3.5 Pro was supposed to ship weeks ago, but internal testing showed it struggling with "long-horizon coding," meaning tasks that require the AI to hold a plan together across many steps without losing the thread. Google keeps retraining it and keeps missing its own bar. For a company racing OpenAI and Anthropic, a stumble on the flagship model is the story competitors will use against them.

The lawsuit claims a former Apple engineer downloaded confidential files on unreleased products before joining OpenAI, and that OpenAI's hardware chief, a former top Apple product lead, used inside knowledge while recruiting. The backdrop matters: Apple just passed NVIDIA to become the world's most valuable company, and this lawsuit reads like a company protecting its next act, the AI hardware device OpenAI is rumored to be building with former Apple design chief Jony Ive's team.

President Xi Jinping unveiled the World AI Cooperation Organization at a Shanghai summit, a new international body meant to set shared rules for AI safety and access, separate from anything the US is building. Founding members include Brazil, Russia, Indonesia, and South Africa. Translation: there are now two competing tracks for how the world governs AI, and most of the Global South just picked a side.

The Model Context Protocol, MCP, is the connective tissue that lets AI assistants plug into your calendar, files, or business software and take action instead of just talking. Its biggest rewrite yet finalizes July 28, adding a formal deprecation policy so nothing breaks without a year's notice, but the rebuild itself carries breaking changes now. For any business building AI agents in-house or through a vendor, this is the plumbing update that decides whether your integrations still work in two weeks.

GPT-5.6 Sol is now available in GitHub Copilot, meaning millions of working developers can call on OpenAI's flagship reasoning model without leaving their code editor. More reliable AI coding tools translate into fewer bugs shipped and less time spent double-checking machine-written code.

Shares crashed roughly 25% in a single session after IBM warned its hardware and software business fell short of targets, blaming a severe memory chip shortage that pushed IBM's own customers to spend on physical servers and storage instead of IBM's software and consulting. The CEO's own words: "we did not adapt and move quickly enough." For any leader watching the AI boom from the outside, this is the clearest proof yet that physical component shortages, not model quality, are what actually derail AI-era business plans.

SoftBank triggered a contractual right to sell its remaining 10% stake, and Hyundai is buying it for roughly $325 million, ending five years of shared ownership. Hyundai's plan: fuse Boston Dynamics' robots with its own car-manufacturing scale to build what it calls an end-to-end AI robotics supply chain, from R&D to factory floor.

A new consortium backed by SoftBank, Sony, NEC, and Honda is standing up 140 megawatts of computing power, enough electricity for roughly 100,000 homes, running on NVIDIA's newest chip platform. It's Japan's government and biggest companies pooling resources to build sovereign AI capacity instead of renting it from American cloud giants.

It's the largest capital raise in tech industry history, split between open-market stock sales and a private placement, with Berkshire Hathaway agreeing to put in $10 billion of it. The size of the check is the story: this is what it now costs a company Google's size to stay competitive in AI data centers.

Brookfield, one of the world's largest infrastructure investors, expanded its partnership with fuel-cell maker Bloom Energy fivefold, from $5 billion to $25 billion, to build and finance fast, onsite power for AI data centers. Data centers now need electricity faster than power grids can deliver it, so this money buys AI companies power plants of their own rather than waiting in line for the grid.