AI Daily Brief: China's Kimi K3 Storms the Frontier, TSMC Bets Another $100B on the AI Buildout
The largest open-weight model ever released just landed from Beijing, Google's rebuilt Gemini 3.5 Pro is due today, and the big labs keep shopping for their own silicon. Plus: TSMC's blockbuster quarter, and why Indian IT stocks chose this week to rally.
Two things happened in the last 24 hours that, read together, describe where AI is heading: the frontier got more crowded, and the factories underneath it got more expensive.
The largest open model ever released is Chinese
Moonshot AI shipped Kimi K3 on Wednesday, a 2.8-trillion-parameter mixture-of-experts model that Fortune says pushes Chinese AI into frontier territory. On GDPval-AA v2, a benchmark built around real work across 44 occupations, it placed third overall, behind only the top tiers of Claude Fable 5 and GPT-5.6 Sol and ahead of Claude Opus 4.8. In blind Arena testing, developers preferred it to every leading US model for front-end coding. The full weights go public on July 27, and the API is priced at a fraction of what the closed frontier charges. The gap between "open" and "frontier" has been narrowing for two years; this is the week it stopped being a gap you could comfortably plan a business around.
Google's rebuilt Gemini 3.5 Pro is due today
Google is widely expected to make Gemini 3.5 Pro generally available today, after scrapping the original base model and restarting pretraining when early testers flagged structural problems. The reported spec sheet is aggressive: a 2-million-token context window and a Deep Think reasoning mode aimed at long, multi-step work. Worth stressing: Google has confirmed none of this, and as of earlier this week there was no model card or API listing in public documentation. If it ships on schedule, it lands into a far tougher week than Google planned, with Kimi K3 having just reset expectations for what you can get without paying frontier prices.
TSMC's quarter says the buildout is accelerating, not cooling
The company that manufactures nearly every chip in this story posted second-quarter revenue up 34% to $40.2 billion, raised its 2026 capital spending guidance to as much as $64 billion, and committed another $100 billion to Arizona, taking its planned US investment to $265 billion. Full-year revenue is expected to grow more than 40% on orders from Nvidia, Apple and AMD. Capex is the most honest signal in this industry: forecasts can be talked up, but foundries do not spend $64 billion on capacity for demand they don't believe in.
Everyone wants their own silicon
The other chip story this month is who is trying to need Nvidia less. Anthropic is in talks with Samsung about a custom AI chip built on its 2-nanometer process, weeks after OpenAI showed off Jalapeno, the custom chip it designed with Broadcom. Nobody in this race expects to replace Nvidia soon. What they want is negotiating leverage and a hedge, and Samsung wants exactly the frontier-scale foundry business that TSMC's earnings just showed everyone it is missing.
What it means for India
The IT complex had its best session in months on Friday, with the Nifty IT index up as much as 2.3% and Infosys, Tech Mahindra and HCL Tech each gaining over 3% as June-quarter results came in without major disappointments. The more durable India angle in today's stories is Kimi K3's July 27 weights release: IndiaAI's model builders, from Sarvam to BharatGen, get frontier-class open weights to build on, running on the mission's 38,000-plus subsidised GPUs. And TSMC's capex surge is a rising tide for the packaging-and-testing ecosystem India is assembling through Micron's and Kaynes' new facilities. India is not at the frontier of this week's news, but it is positioned downstream of almost all of it.
Markets and AI money
The odd couple of the week: TSMC delivered one of the best quarters in semiconductor history and its shares fell 4% anyway, with the selloff spreading to memory names, SK Hynix dropping over 11% in Seoul. That is less a verdict on AI demand than a market that had already priced in perfection and is nervous about memory supply. Indian equities read the same news differently, with the Sensex up over 1% as IT led the advance.
| Metric | Move |
|---|---|
| TSMC Q2 revenue | $40.2B, +34% YoY |
| TSMC 2026 capex guide | $60-64B, raised from $52-56B |
| TSMC shares (Thu) | -4% |
| SK Hynix (Thu, Seoul) | -11% |
| Sensex (Fri) | 78,005, +1.06% |
| Nifty IT (Fri) | +2.3% intraday |
The frontier is getting cheaper to access and more expensive to build. Both halves of that sentence are good news for somebody.
Keep reading
AI Daily Brief: Google's New Flash Models Land as AMD and Anthropic Circle a Chip Deal
Google quietly shipped three new Gemini models overnight, Anthropic locked up 3.5 gigawatts of TPU capacity, and AMD kicked off its biggest AI event of the year with Anthropic rumors swirling. Below, what the compute land grab means for India, and how the chip trade is pricing it all in.
AI Daily Brief: Nvidia Ships Its Groq Chip and a New Model of Its Own, Qualcomm Bets $14 Billion Against It
Nvidia's first Groq-derived inference chip heads into production on Samsung's line just as Samsung's chairman prepares to meet Jensen Huang, Qualcomm commits $14 billion to not need Nvidia at all, and Nvidia's own research lab quietly shipped a new open model that decodes six times faster than the competition. Plus what the chip realignment means for India, and the day's market moves.