1. Yet another DeepSeek moment?

On July 17 China's Moonshot released Kimi K3, a 2.8-trillion-parameter model that tops the leaderboards and comes close to Fable 5 at roughly a third the token price ($3/$15 per million). The vibes are more mixed than the scores. Early hands-on testing finds it strong on benchmarks but weaker on the agentic tool-calling real work depends on, with a single token-hungry "max" reasoning mode that made one trivial task cost 25 cents. And "open" is aspirational for now: the weights are only promised by July 27, and self-hosting a 2.8T model is anything but free. Still, a Chinese model reaching the Western frontier at a fraction of the price, again, is the signal that rattled markets.

  • "Near-free" is a headline, not a usage bill. A third the sticker price, token-hungry reasoning, and enterprise-scale infrastructure to self-host add up fast.
  • The strategic point is the cadence, not this one model: China keeps landing frontier-adjacent models faster and cheaper, and Western pricing power keeps eroding at the margin.

2. China launches rival AI order

That same week, at Shanghai's World AI Conference, Xi Jinping established the World Artificial Intelligence Cooperation Organisation: Shanghai-headquartered, 29 founding nations, routed through BRICS and the Shanghai Cooperation Organisation. He pitched China as the partner that ships affordable, open models to the developing world. The striking part is the asymmetry. There is no matching Western pole. The US, once the world's convener, is restricting access rather than organizing: export controls, a G7 coalition still only proposed, Commerce cutting foreign users off from top US models. China is building distribution and governance while Washington guards the gate.

  • "Rival order" flatters the contest: one side is standing up an institution and courting the Global South; the other has controls and a proposal. On coordination, the US is not leading, it is largely alone.
  • Cheap open weights are the on-ramp. The competition for the developing world's AI stack is, for now, being run mostly by one player.

3. American open-weight leader?

Two days before Kimi K3, Mira Murati's Thinking Machines open-sourced Inkling, a 975B-parameter model that debuted as the leading US open-weights release but, by the lab's own admission, not the best overall. It leads US labs and trails China's frontier, with a high hallucination rate that limits accuracy-critical use. What matters strategically is how it was built: pre-trained from scratch on 45 trillion tokens, not distilled from a frontier rival the way several Chinese open models appear to be. Reception on quality was lukewarm, but the West now has a genuine, independently-trained open-weight contender, not just closed APIs.

  • An independently-trained US open model is worth more long-term than a higher benchmark score built on distillation.
  • Controllability and provenance (who trained it, on what) are becoming procurement criteria alongside raw capability.

4. Record chip earnings, falling chip stocks

The picks-and-shovels are printing money. TSMC posted record Q2 revenue of $40.2B, profit up 77%, and pledged another $100B for Arizona, and ASML raised its 2026 guidance a second time. The stocks fell anyway. But the chip index was already about 20% off its peak before Kimi landed, and the Friday rout was overdetermined: Netflix, rate fears, valuation doubts, and Kimi all triggers, not one cause. Record fundamentals, a nervous tape, and a market voting on returns it can no longer take for granted.

  • Pinning the selloff on Kimi is tidy but hard to prove. The deeper worry predates it: whether ~$700B of hyperscaler capex earns its return.
  • Supercycle or peak? Demand signals (ASML's second hike) and price signals (the rout) point opposite ways.

5. "We must act now": economists break glass

On July 13, over 200 economists and AI researchers, including 16 Nobel laureates, signed a Stanford statement led by Erik Brynjolfsson urging governments to prepare now for AI's economic disruption: steam, electricity, and computers gave societies decades to adapt; AI may give us only a few years. The ask is not "stop AI" but build the institutions, reskilling, safety nets, and complements-not-substitutes design, before the shock, not after.

  • Even the economists concede they are driving in the fog on productivity and jobs.
  • For boards, workforce transition just moved from HR footnote to a governance question with a countdown.

More AI stories of the week