Moonshot AI unveiled Kimi K3 in Beijing on Thursday, July 16. It is the largest open weight model shipped to date, a 2.8 trillion parameter mixture of experts that activates roughly 16 of 896 experts per token, with full weights scheduled to drop July 27. Moonshot’s own benchmarks put K3 competitive with Anthropic’s Claude Fable 5 and ahead of Opus 4.8 and OpenAI’s GPT-5.6 Sol, at $15 per million output tokens against Fable 5’s $50 per million.

Markets read it as a second DeepSeek moment. TSMC fell roughly 7 percent on Friday. SoftBank, often treated as a proxy for OpenAI exposure, fell 9 percent. z.ai, a competing Chinese lab, lost close to 30 percent in Hong Kong trading. Nvidia dropped 1.2 percent and briefly forfeited its title as the world’s most valuable company to Apple. Meta lost 2.4 percent. The Nasdaq 100 was down 1.0 percent by early afternoon Eastern (Fortune, July 17, 2026).

The pricing logic is familiar: if Chinese frontier models can be trained and served cheaply, the argument for hyperscaler capex weakens, and with it the demand for advanced packaging, accelerators, and the physical inputs behind them. That thesis fell apart under its own foundry the day before it was traded.

What TSMC actually said on Wednesday

TSMC reported Q2 revenue of $40.20 billion and gross margin of 67.7 percent, both above its own guidance. Operating margin came in at 60.3 percent against a 56.5 to 58.5 percent range. Net income hit NT$706.56 billion, up 77.4 percent year on year, the company’s fifth consecutive record quarter. High performance computing, where AI accelerators book, was 66 percent of revenue and rose 20 percent quarter on quarter. Full year 2026 capital expenditure guidance moved up to $60 to $64 billion. Arizona commitment climbed to $265 billion with another $100 billion committed on the call (Investing.com, July 16, 2026; Hardware Busters, July 16, 2026).

If Chinese open weights had already broken the AI capex thesis, the foundry’s own book was the wrong place to look for the break. CoWoS advanced packaging is still the binding constraint on the AI wafer line. Nvidia’s H2 orders route through it. Broadcom and Google TPU route through it. The queue does not clear on a benchmark leaderboard, and TSMC just committed another $100 billion of Arizona capex on the assumption it will not.

Where the compute footprint is actually moving

The more consequential detail sits inside China’s own buildout. Meituan open sourced LongCat 2.0 on June 30, a 1.6 trillion parameter mixture of experts model that the company says was trained end to end on domestic Chinese ASIC superpods, the first time a frontier scale model has completed full pre training on non Nvidia, non Google TPU hardware (SiliconANGLE, June 30, 2026). Analysts flagged Huawei Ascend as the likely vendor, though Meituan did not confirm it. OpenRouter’s five most used models this week are all Chinese: Tencent, Xiaomi, DeepSeek, MiniMax, and z.ai (Fortune, July 17, 2026).

On Friday, Xi Jinping delivered the opening keynote at the World Artificial Intelligence Conference in Shanghai and reaffirmed that China’s frontier stack will stay open source. “We must seize this rare historic opportunity, encourage open source, openness, cooperation, and sharing,” he told the room, adding that the country should “oppose the practice of overstretching the concept of national security in the field of AI” (Fortune, July 17, 2026; Quartz, July 17, 2026).

The night before, 29 countries signed the founding agreement of the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai. Brazil signed on as a founding member, alongside Indonesia, Malaysia, South Africa, Senegal, Russia, and Pakistan among the 29 signatories (Al Jazeera, July 17, 2026; Caixin, July 17, 2026). UN Secretary General António Guterres attended the ceremony.

For the materials read, the substitution matters. If a growing share of Chinese frontier training and inference runs on Ascend, Cambricon, or Alibaba T-Head silicon, the gallium, germanium, rare earth magnets, and copper the AI stack needs do not vanish. They route through Chinese domestic supply chains that are already the most concentrated in the world. Gallium output sits at roughly 95 percent China per the USGS Mineral Commodity Summaries 2025. Germanium is 60 percent China. Rare earth separation is above 70 percent China (USGS MCS 2025; Tantalum SOV50 concentration series). The Kimi K3 story is not a demand cut. It is a geography cut.

Where Brazil sits

Brazil’s absence from the frontier model race has framed our AI improvement coverage all year. But the WAICO signature this week and the timing of the Serra Verde acquisition put the country in a different position than it looked on Monday.

USA Rare Earth’s $2.83 billion takeover of Serra Verde was announced April 19 and is expected to close in Q3 2026, which is now. Serra Verde’s Pela Ema mine is the only producing operation outside Asia that ships all four magnetic rare earths at commercial scale, and its Phase 1 output is already under a 15 year offtake to a US government special purpose vehicle with contractual price floors (USA Rare Earth 8-K, April 20, 2026). The Alto Paranaíba stake by a TDK backed battery partner and the CBMM-Nyobolt Series C round this desk covered in June are the other two working links.

On the materials side, Brazil is now a founding WAICO member and the anchor supplier of the West’s diversification magnet chain at the same time. That posture is unusual. It is the strongest bidding position Brasília has held since the R$15 billion Sovereign Plan was released in May.

What it means

The Friday selloff priced a benchmark. It did not price the wafer book, the packaging line, the gas turbine queue, or the concentration of the upstream inputs. TSMC’s Q2 refuted the efficiency story at the foundry level. Xi’s WAICO speech and Meituan’s Chinese silicon disclosure refuted the demand story at the geography level. What actually happened is the AI compute footprint began to bifurcate along silicon geography, and the materials bill bifurcates with it. The West’s diversification supply, which is the pitch behind SDX and Serra Verde and CBMM, is the answer to that bifurcation, not to a single Chinese model release.

What to watch

• July 27, 2026. Kimi K3 weights drop publicly. Watch the inference migration to Chinese hosted endpoints, and the gallium premium in the run up to Beijing’s November 27 suspension expiry.
• Q3 2026 close for USA Rare Earth and Serra Verde. Watch the shareholder vote and any residual Chinese response to a US buyer taking commercial control of a producing Brazilian heavy rare earth mine, given Beijing already put MP Materials and USA Rare Earth on its export control list on June 22.
• Whether the WAICO signatories, including Brazil, use the new Shanghai framework for any materials procurement pooling or bilateral offtake announcements in the next 90 days. That would move SOV50 and SDX at the constituent level.