One-Line Summary

China’s Moonshot puts a 2.8T open‑weight model into play while U.S. labs jockey for compute and Google slows Gemini — a week where access and cost, not just raw power, shape AI choices.

Big Tech

Xi pushes ‘AI for all’ message at Shanghai summit

China’s President Xi Jinping uses his first appearance at the World AI Conference in Shanghai to call for inclusive global collaboration on artificial intelligence, saying, “AI development should not be a solo performance by a single country, but a symphony of international cooperation,” while stressing that safety risks must be contained. ¹

For business leaders, the speech signals China’s emphasis on low‑cost, widely available AI and comes as Chinese labs rapidly advance open‑weight systems. The message aligns with models like Kimi K3 that prioritize access and affordability alongside capability. ¹

Meta weighs leasing compute to Anthropic amid its capacity spree

The New York Times reports Meta is in talks to lease computing power from its AI data centers to Anthropic in a deal that could reach $10 billion over two years, with monthly payments and an option to exit early — a potential step toward a new AI infrastructure business for Meta. ²

Anthropic, meanwhile, formalizes massive capacity on AWS, committing more than $100 billion over ten years to Amazon technologies and securing up to 5 gigawatts of new capacity to train and run Claude, including nearly 1GW of Trainium2 and Trainium3 by the end of 2026. Anthropic also says it already uses over one million Trainium2 chips across training and serving. ³

The startup adds capacity through a deal with SpaceX for more than 300 megawatts (over 220,000 Nvidia GPUs) at its Colossus 1 data center, and notes other strategic compute arrangements including a $30 billion Azure capacity commitment and a 5GW agreement with Google and Broadcom beginning to come online in 2027. ⁴

Google’s Gemini 3.5 delay fuels concern about AI pace

The Los Angeles Times reports Google is months behind schedule on Gemini 3.5 Pro, with internal friction and a push to improve coding capabilities contributing to delays as rivals ship stronger models; Google says it is shipping quickly across models, testing 3.5 Pro, and engaging with the U.S. government on safety. ⁵

CNBC adds that the Gemini delay heightens concerns Google is falling behind OpenAI, Anthropic, and Meta ahead of Alphabet earnings, intensifying pressure on its AI roadmap. ⁶

Industry & Biz

Moonshot unveils 2.8T open-weight Kimi K3 with 1M context

Moonshot AI, the Beijing startup behind the Kimi assistant, unveils Kimi K3 — a 2.8 trillion‑parameter system it describes as the world’s largest open‑weight AI model, built for advanced reasoning, long‑horizon coding, and knowledge work, with a 1‑million‑token context window. The company says K3 competes with top frontier systems on select tasks. ⁷

Moonshot positions K3 as a Mixture‑of‑Experts model and plans to release full weights, enabling organizations to download and run it. According to Emergent’s launch brief, weights are scheduled for public release by July 27, 2026, and API pricing lists input at $0.30 per million tokens for cache‑hit, $3.00 per million for cache‑miss, and $15.00 per million for output. ⁸

Early third‑party evaluations cited by Reuters show Kimi K3 scoring strongly: Arena.ai ranks it first for web interface building; Vals AI places it second overall behind Fable 5 and ahead of GPT‑5.6 Sol; and Artificial Analysis reports performance comparable to OpenAI’s GPT‑5.5 and Anthropic’s Claude Opus 4.8 on complex, multi‑step tasks. ⁷

Axios frames K3 as part of a broader shift toward cheaper, customizable intelligence. It reports companies increasingly choose open‑weight models for routine workloads and notes that, on marketplace OpenRouter, Chinese open‑weight models occupy the top five spots by weekly token usage — pressure that could reshape pricing and deployment decisions for many teams. ⁹

What This Means for You

For most teams, capability that is “good enough,” customizable, and cost‑controlled beats paying premiums for every task. Axios reports businesses are already routing routine coding, summarization, and customer support to cheaper open‑weight systems, keeping premium models for the hardest problems — a pattern Kimi K3’s approach may reinforce. ⁹

Open‑weight access matters beyond cost: it allows self‑hosting and fine‑tuning under your own controls. Emergent’s brief notes Moonshot plans to release K3’s weights by July 27, 2026, which could make frontier‑adjacent performance easier to integrate into workflows with strict data governance. ⁸

Compute supply is a bottleneck you can feel. Anthropic’s SpaceX deal added more than 300 MW (over 220,000 GPUs) and enabled higher usage limits for Claude Code and its API — a reminder that vendor capacity expansions can translate directly into fewer throttles for your team. ⁴

If your stack leans on Gemini, build in a multi‑model fallback. Given reporting on Gemini 3.5’s delays, pilot alternatives side‑by‑side for key jobs (coding, structured data, RAG) so delivery doesn’t stall if one provider lags. ⁵

Action Items

1. Test Kimi K3 on a real task: Use the Kimi app or API to run a document‑heavy workflow (e.g., summarize a 30‑page report plus follow‑up Q&A) and compare output quality and iteration speed to your current model.
2. Do a 30‑minute cost check: Estimate your typical token mix and compare against Kimi’s published rates (cache‑hit/miss input and output) to spot where switching could cut monthly spend without hurting quality.
3. Set a model‑selection rubric: Route routine tasks (summaries, extraction, baseline code) to lower‑cost models by default; reserve premium models for the 1–2 toughest workflows. Document this in your team’s runbook.
4. Lift concurrency where available: If you use Claude Code or the Claude API, retry high‑throughput jobs that previously hit caps — recent compute additions raised usage limits for paying tiers.