Five stories crossed my feed today. Three of them change what you can build. One is geopolitical theater. And one is about to change what you pay. Here’s the operator’s read.
Google Targets Gemini 3.5 Pro Launch After Scrapping and Rebuilding the Base Model
Verdict: Matters.
Google is targeting today for the general availability of Gemini 3.5 Pro, timed to the opening of Shanghai’s World Artificial Intelligence Conference. The kicker: Google DeepMind scrapped the original base model after engineers found structural failures in recursive tool-calling and SVG generation, then rebuilt from scratch.
The rumored specs include a 2-million-token context window, a Deep Think reasoning layer on the $250/month Ultra tier, and API pricing near $1.25 input / $10 output per million tokens. I need to flag this clearly: as of this writing, no model card, no pricing page, and no gemini-3.5-pro listing appear in the public Gemini API docs. Every spec is third-party reporting from unnamed sources. The model exists, Google has it running internally, and it slipped from June to July. But until the API endpoint goes live, treat every number as unconfirmed.
What I’m watching: the recursive tool-calling fix. If that holds, Gemini 3.5 Pro becomes a real option for agentic workflows that currently choke on multi-step tool chains. If it doesn’t, the rebuild was cosmetic.
TSMC Posts Record Quarter: AI Chip Demand Is Accelerating, Not Plateauing
Verdict: Matters.
TSMC just posted Q2 2026 numbers that should end the “AI demand is peaking” narrative. Revenue hit $40.2 billion, up 36% year over year. Net profit surged 77.4% to a record NT$706.6 billion (roughly $22 billion). Gross margin climbed to 67.7%, above the company’s own guidance ceiling. The chip maker raised its full-year 2026 revenue growth outlook to slightly above 40% and bumped capex guidance to $60-64 billion.
High-performance computing now makes up 66% of TSMC’s total revenue, with the 5nm node leading at 33% share and 3nm close behind at 30%. HPC revenue surged 20% quarter over quarter.
Here’s why this matters if you build with AI: TSMC holds roughly 73% of the global advanced foundry market. When they raise capex guidance, they’re telling you GPU and AI accelerator supply is expanding. When margins stay at 67.7%, they’re also telling you nobody’s undercutting them. Your inference costs track these dynamics. The demand curve hasn’t bent.
Kimi K3: China Drops the Largest Open-Weight Model at Near-Frontier Performance
Verdict: Matters.
Moonshot AI released Kimi K3, a 2.8 trillion-parameter open-weight model with a 1-million-token native vision context window. Full weights drop July 27. On GDPval-AA v2, a benchmark measuring real-world tasks across 44 occupations, Kimi K3 scored 1,687, placing it third overall behind Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,747.8).
Two architectural innovations worth knowing: Kimi Delta Attention (a hybrid linear attention mechanism) and Attention Residuals (a drop-in replacement for residual connections that the team claims delivers consistent scaling gains). The model also ships with an always-on “thinking mode” for reasoning tasks.
The builder calculus: a 2.8T open-weight model scoring near frontier closed models changes the self-hosting math. You won’t run 2.8T parameters on a single box, but distilled variants and quantized versions will follow the weights release. If your use case tolerates a small quality gap below Fable/Sol, the inference cost savings of running your own model just got a lot more interesting.
WAICO: 29 Countries Form a China-Led AI Governance Body (Without the US or EU)
Verdict: Marketing.
Twenty-nine countries signed an agreement in Shanghai to establish the World Artificial Intelligence Cooperation Organization, an independent intergovernmental body headquartered in Shanghai. The signing list includes Russia, Pakistan, Indonesia, Kazakhstan, and Laos. Chinese Foreign Minister Wang Yi signed on behalf of China. UN Secretary-General Antonio Guterres attended.
Who’s missing: the United States, the European Union, Japan, South Korea, the UK, and every country with a frontier AI lab.
I’m filing this under marketing because WAICO has no enforcement mechanism, no standards body, and no compute. It’s a geopolitical signal: China is building a parallel governance track for countries that want AI cooperation without US-aligned strings attached. For builders, this changes nothing about what models you can use or what APIs you can call today. Watch it if you ship to WAICO-member markets and need to navigate compliance expectations, but don’t lose sleep over it yet.
DeepSeek Preps IPO, Builds Its Own Chip, and Introduces Surge Pricing
Verdict: Breaks your stack.
Three DeepSeek developments in one week, and they all point the same direction. First: DeepSeek raised $7 billion in the largest single AI fundraising in China’s history, valuing the company at roughly 400 billion yuan. Second: DeepSeek is developing its own AI chip designed specifically for inference, not training. Third: DeepSeek is prepping an IPO filing that could see a debut in 2027.
Meanwhile, DeepSeek V4’s official release is targeting mid-July with a new pricing wrinkle: peak-hour API rates will run double the off-peak price during Beijing business hours (9am-12pm and 2pm-6pm). The old deepseek-chat and deepseek-reasoner model names retire permanently on July 24.
Why this breaks your stack: if you built on DeepSeek’s API for cost arbitrage, surge pricing means your inference bill just got variable. Plan for it or migrate the peak-hour traffic. Longer term, a DeepSeek with its own inference silicon and public-market capital is a company positioning to undercut everyone on price while controlling its own supply chain. That’s the competitive pressure that drives inference costs down for the whole industry, which is great for builders, until your current provider’s margins get squeezed and the free tier disappears.