The Omniscient Bulletin · 2026-07-17
The Omniscient Bulletin — July 17, 2026
Moonshot puts an open model at the frontier, TSMC adds $100 billion in Arizona, and the EU cracks open Google
Thursday belonged to the powers that shape AI, states and giants alike. China's Moonshot released an open model that reaches the frontier, and hours later Bloomberg reported Google's own flagship is stuck, held up by the very coding skills its rivals have raced ahead on. TSMC answered the AI boom with another $100 billion for Arizona and a record quarter. In Brussels regulators forced Google to share its search data and open Android to rival assistants, while in Shanghai China gathered 29 countries to launch an AI body of its own. A big funding round and a quiet advance in how agents log in round out the day. Here's what happened.
Frontier
China's Moonshot puts an open model at the frontier with Kimi K3
Moonshot AI, the Chinese lab behind the Kimi assistant, released Kimi K3 on Thursday, a 2.8-trillion-parameter mixture-of-experts model it calls its most capable yet. It went live through Kimi's app and API the same day, with the full weights promised as a free download by July 27. On independent evaluations it lands near the top of open models and trades blows with the US frontier, one ranking placing it just behind the best American models. At $3 per million input tokens it is priced like a Western flagship, not a budget challenger. Another sign the open-weight gap is closing.
Compute
TSMC pours another $100 billion into Arizona as its AI quarter breaks records
TSMC used a record quarter on Thursday to make its biggest American bet yet. On its earnings call the chipmaker said it will invest another $100 billion in Arizona for at least four more fabs running its cutting-edge 2-nanometer process, lifting its total US commitment to about $265 billion. Second-quarter revenue was $40.2 billion and profit jumped 77% from a year earlier, both driven by AI demand, and TSMC raised its 2026 capital budget to as much as $64 billion. It is the clearest signal yet that the money behind the buildout is still accelerating.
Policy
Brussels orders Google to open Android and share search data with AI rivals
The European Commission ordered Google on Thursday to pry open two of its most valuable moats. Under the Digital Markets Act, Google must share anonymized search and ranking data with rival search engines and AI chatbots, and give competing assistants like ChatGPT and Claude the same deep Android access it reserves for Gemini. The obligations, issued as binding orders rather than fines, phase in from January 2027. Google's top lawyer objected that they risk undermining privacy and security. It is one of the sharpest moves yet to loosen Google's grip on how people reach AI.
Industry
Fireworks raises $1.5 billion, betting companies want their own tuned models
Fireworks AI, a platform that helps companies fine-tune and run open models on their own data, raised a $1.5 billion round on Thursday at a $17.5 billion valuation. Atreides Management, Index Ventures and TCV led it, with Nvidia among the backers. The company, founded in 2022 by former Meta engineers, says it now clears more than $1 billion in annualized revenue, five times a year ago, and serves over 40 trillion tokens a day. Most of that traffic runs on models tuned to customers' own data. It is a wager that the future is many specialized models, not one giant brain.
Policy
China gathers 29 countries to launch its own global AI body
On the eve of Shanghai's World AI Conference, 29 countries signed an agreement on Thursday to establish the World AI Cooperation Organization, a China-led body based in Shanghai. Foreign Minister Wang Yi signed for China, and UN Secretary-General Antonio Guterres attended. Founding members include Russia, Pakistan, Indonesia and Brazil, among others. Framed as a forum for the Global South, it is Beijing's bid to offer an alternative to US and European AI governance, complete with cheaper models and a seat at the table. The lines of an AI bloc are hardening.
Agents
1Password lets Claude log in for you without ever seeing the password
1Password released an integration on Thursday that lets Anthropic's Claude sign in to websites on your behalf without the model ever seeing your credentials. When an agent reaches a login, the password and any one-time code are injected into the page after you approve with biometric authentication, and never enter Claude's memory or Anthropic's systems. A new agentic mode locks the vault down to only the credentials approved for the task at hand. It is an early answer to a real problem: how to let AI agents act for you online without handing them the keys to everything.
Frontier
Google's next flagship model is late, and coding is the reason
Google is months behind on Gemini 3.5 Pro, its next flagship, because the model still is not good enough at coding, Bloomberg reported. Chief executive Sundar Pichai had said in May that it would ship in June; instead engineers are scrambling to close a gap with Anthropic and OpenAI, whose models now beat Gemini at the tasks developers care about. A recent attempt to fix it by refreshing the training data disappointed, and layers of internal review slowed things further. For the company that helped invent the technology behind modern AI, it is an uncomfortable place to be.