Subscribe
Share
Share
Embed
Kimi K3's Open Weights Land, Hardware Gate Included
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Monday, July twenty-seventh.
In today's briefing, Moonshot's Kimi K3 lands its full open weights right on schedule, a Hacker News investigation exposes a black market reselling stolen access to the major model APIs, and Nvidia moves from renting out chips to bankrolling OpenAI's own data center buildout.
First up, today in the big model news;
Kimi
Moonshot published Kimi K3's full weights on Hugging Face today, hitting the July twenty-seventh date it committed to ten days ago. The model runs two point eight trillion parameters, with eight hundred ninety-six experts and sixteen active per token, under a modified MIT license and a one million token context window, built on a new Kimi Delta Attention and Gated MLA architecture that lands third on Artificial Analysis's index, just behind Fable 5 and GPT-5.6 Sol. The weights ship natively in a format called MXFP4, which needs Blackwell or MI400 class hardware to run as designed, so self-hosting stays limited to the same well capitalized buyers export controls were built to target. That resolves the question that's hung over Moonshot for the past week: no Entity List sanctions materialized, and the release shipped on schedule anyway. Open no longer means accessible. The hardware is the real gate on who runs frontier open weights now, not the license.
In other news, a widely read Hacker News teardown mapped a multi billion RMB market reselling access to OpenAI, Anthropic, and Google APIs, in some cases at prices up to ninety seven point eight percent below list. Card and account merchants supply bulk registered accounts and stolen virtual cards, pool operators aggregate hundreds of accounts behind a single API key, and Chinese language storefronts resell that access to retail customers, with the top ten relays pulling three point six million monthly visits. It's the same mechanism Anthropic accused Alibaba of running against Fable 5 back in June, twenty eight point eight million exchanges through twenty five thousand fraudulent accounts, except this piece shows it isn't one covert operation, it's a standing commercial layer. That undercuts the premise access control regimes rest on: identity checks and rate limits only gate access if the identity behind an account is real, and identity here is being sold in bulk. If your product relies on per account limits to gate abuse, assume some fraction of your heaviest traffic is pooled behind a single payer, not many genuine users.
In compute economics today, Nvidia is negotiating to backstop roughly two hundred fifty billion dollars of financing for a ten gigawatt OpenAI data center that SoftBank is building in Ohio, plus separate financing worth up to three hundred fifty billion dollars for the chips that go inside it, a five hundred billion dollar campus in total, and OpenAI's first step toward owning infrastructure rather than renting it from Microsoft, Amazon, and Oracle. It's the compute landlord pattern this briefing has tracked since May, xAI hosting Anthropic, sovereign deals in Korea and Japan, but inverted: instead of a landlord renting out capacity, the chip vendor is underwriting its own customer's balance sheet so that customer can build capacity it will fill entirely with that vendor's chips. That's the same circularity investors flagged in Nvidia's earlier OpenAI equity stake, now showing up as debt guarantees, and it lands two weeks after a separate finding that off balance sheet AI infrastructure debt at five major tech firms already totals roughly one point six five trillion dollars. Watch whether this financing gets disclosed as a lease the way Meta's does, or whether a chip vendor's debt guarantee ends up treated differently from ordinary vendor financing.
That's the briefing. Have a great day.