Saturday, July 25 July 25, 2026
China's Kimi K3 triggers a Silicon Valley civil war over open-weight AI access, Etched hits a $10.3B valuation on $300M to challenge Nvidia in inference chips, and Washington's AI policy fault line deepens — your July 25th Morning AI briefing.
Good morning. It's Saturday, July 25th, 2026. Today's briefing is really about one fault line running through every major story — who gets to use the world's most powerful AI models, and who decides. Let's get into it.
The biggest story right now is China's Kimi K3. Developed by Moonshot AI, Kimi K3 is an open-weight model that reportedly narrows the gap with American frontier systems in a meaningful way. And the reaction in Silicon Valley has been intense. An OpenAI executive's public comments about K3 set off a fresh debate about whether the U.S. should restrict access to Chinese open-weight models entirely. Here's the twist: it's not just advocates and academics pushing back. Nvidia, Microsoft, Meta, and a coalition of startup founders are all urging the Trump administration to keep those models accessible. Their argument? A blanket ban would kill small companies and consolidate power with the handful of large labs that can afford to build their own frontier models. This is the "Kimi Effect" — a Chinese open-source release forcing Washington and Silicon Valley to pick sides on open-weight AI.
The same debate is playing out in policy. Trump's AI Executive Order, Executive Order 14409, gives AI labs 30 days to voluntarily submit frontier models for cybersecurity testing through a program called CAISI. It's a softer approach than an outright ban, but it signals that the White House is actively thinking about who controls access to the most capable systems. Meanwhile, DeepSeek's CEO Liang Wenfeng dropped an interesting counterpoint — he made the case that export controls on AI chips are actually the primary thing holding Chinese AI back, not talent or data. In other words, the chips matter more than the models.
Which brings us to this week's biggest infrastructure bet. Etched, the AI inference chip startup founded by three Harvard dropouts, closed a $300 million Series C round this week at a $10.3 billion valuation. Sequoia led the round. a16z participated. SK Hynix joined as a strategic investor. The company has hit $1 billion in chip orders and is now in initial production. Etched's chip, called Sohu, is purpose-built for transformer inference — not training, just running models as fast and cheaply as possible. The valuation doubled in seven months. That kind of growth in specialized inference hardware is a direct signal: the market believes inference costs will define competitiveness in the next phase of AI deployment.
On the product side, Google quietly released Gemini 3.6 Flash this week — a model update that delivers 17 percent token savings over its predecessor, along with a lighter variant called 3.5 Flash-Lite and a specialized Cyber variant designed for security work. No major fanfare, but that Cyber variant is worth watching for enterprise security teams. Separately, Alphabet is reportedly developing a new custom chip designed to run Gemini six to ten times more efficiently — a move that could significantly cut Google's inference costs and give it a structural advantage if it lands on schedule.
In funding, Reid Hoffman-backed AI research lab Prentis — barely three months old — is reportedly in talks to raise $100 million at a $1 billion unicorn valuation. Three months old. That is how compressed the fundraising cycle has become for credentialed AI research labs. And Sam Altman's biometric identity startup World raised $52.5 million through a crypto token sale, funding its eyeball-scanning digital identity project.
On the regulation front, Massachusetts debated an AI guardrails bill this week with language covering scenarios from cyberattacks to loss of control. It's part of a broader $575 million economic development package. And a school district in Houston — Katy ISD — banned generative AI tools for grades K through 6, with restrictions on older students as well.
Here is today's business idea. With inference costs becoming the key battleground, there is a clear opening for a specialized AI infrastructure brokerage — a platform that helps mid-size enterprises benchmark, select, and switch between inference providers like Etched, Nvidia, and cloud APIs based on real-time pricing and performance data. Think of it as a rate-comparison engine for AI compute, subscription-priced to the procurement teams spending six figures a year on API costs. The timing couldn't be better.
That's your Morning AI briefing for July 25th. The open-weight fault line is the story to watch. Have a great Saturday.