Two record-setting open models drop overnight, and Anthropic quietly signs a nineteen-billion-dollar power lease in Kentucky — guess which one most feeds will lead with. This is the AI Daily Briefing. Today: who actually owns the weights when the biggest open model comes out of China, and whether a seventeen-times revenue multiple survives a second look. Hit follow and you won't have to come looking for the next episode. Michael Nuñez, writing in VentureBeat:
Moonshot AI, the Beijing-based artificial intelligence startup backed by Alibaba, on Thursday released Kimi K3— a 2.8-trillion-parameter model that the company says is now the largest open-source AI model in the world, and one that benchmarks show performs neck-and-neck with the most powerful proprietary systems from Anthropic and OpenAI.
Kimi K3 — largest open-source model ever, weights out in the open, from Moonshot in Beijing. The parameter count's the headline, but that's not the part I'm watching. Now the weights themselves come with the provenance problem. Earlier this week it was where the compute physically sits. Now it's: the biggest open model on the planet was trained by a lab, on a stack, in a jurisdiction outside U.S. reach — and you're downloading it into production. And notice what's missing. VentureBeat's got 'rivaling top U.S. systems' in the headline — where's the technical report? Big parameter claim, release-day splash, no paper I can find. Same playbook, just faster. 'Open weights' sounds like transparency until you remember public weights and a documented training process are two completely different things. Right — you can run it without ever knowing what went into it. That's a weird definition of open. Here's Jordan Novet at CNBC:
That’s boosting cloud startup Fireworks, which competes with Amazon and Google to host models that developers can weave into applications. The Nvidia-backed company said Thursday that it has exceeded $1 billion in annualized revenue, five times what it had last year, and it has now raised a $1.5 billion round at a $17.5 billion valuation.
Finally a number I can trust this week. Fireworks at a billion in annualized revenue, five times last year — that's a P&L line, not a slide deck. And it's the story I've been circling all week — the weights are open, the stack running them absolutely is not. Fireworks is the tollbooth. But look at the multiple. Seventeen-and-a-half billion on one billion ARR — that's 17x revenue. I want to know how much of that billion is Cursor or Cursor-like customers versus real diversification. Right, they used to get over half their revenue from Cursor alone. Qiao says they've spread it out — but 'diversified' from one customer to how many? Three? The round doesn't answer that. And the tailwind is CFOs, not engineers. Finance execs got the frontier-model bill and told their teams to go find open alternatives. Fireworks is selling anxiety relief. Which is why the Kimi K3 piece we just hit matters here — the biggest open model ever drops, and somebody's got to run it in production. That somebody bills $1 billion a year. Data Center Frontier, with David Chernicoff:
TeraWulf has secured one of the largest dedicated artificial intelligence infrastructure leases announced to date, signing Anthropic to a 20-year, $19 billion agreement for approximately 401 MW of critical IT capacity at the company’s Justified Data campus in Hawesville, Kentucky.
Twenty years, 401 megawatts, nineteen billion dollars, in Hawesville, Kentucky. This is the most concrete infrastructure commitment Anthropic has ever put its name on, and it comes one day after we walked through Ode as their services play. So put that together: a services arm, Claude as the model, and now a dedicated compute campus locked in through the mid-2040s. That's a very different animal from the safety lab that used to send us position papers. And notice what TeraWulf's doing to get there fast — converting a dead aluminum smelter. The heavy electrical gear is already in the ground, so they skip years of greenfield permitting. That's the only reason late-2027 is even a plausible date. They're also dumping their Texas JV stake for about 530 million dollars to fund stuff they own outright. Twenty-year contracted revenue is beautiful on a deck — but can a bitcoin-miner-turned-landlord actually deliver 401 megawatts of AI-grade capacity on schedule? The Meta move I flagged Monday — own the campus, own the chips, own the weights — now you can trace the same shape at Anthropic. Different actor, same vertical grab. This one's from Unite.AI:
Thinking Machines Lab has released Inkling, its first general-purpose artificial intelligence model, giving developers access to a nearly one-trillion-parameter system that can process text, images, and audio. The July 15, 2026 launch represents the most significant product milestone yet for the heavily funded AI company founded by former OpenAI Chief Technology Officer Mira Murati.
Inkling — Mira Murati's first shipped product. 975 billion parameters total, but here's the part that actually matters: it's a sparse Mixture-of-Experts, so only 41 billion light up per token. So the near-trillion headline is basically a marketing number. I want to know whether 41 billion active parameters fit an enterprise inference budget, or if this is a research flex nobody runs at scale. And it drops the same day as Kimi K3, which we just covered — two enormous open-weight models in one window, and neither one showed up with a technical report I can point you to. But the interesting move is the packaging. Apache 2.0 weights on Hugging Face; fine-tune it on their own Tinker platform. Download it free, then pay to shape it — the weights are open, the stack around them very much isn't. Right, that's the Fireworks trade from one story ago, just wearing a lab coat. Public weights, monetized inference and training layer. The business model is the same. And Murati's positioning it as the customizable foundation — not the most capable model in the world. Which is a smart thing to say when you haven't published the benchmarks to prove otherwise. KWXX writes:
A function of AI, referred to as “deepfake” technology, enables the realistic digital imitation of an individual’s voice, face, likeness and performance. Deepfake technology has been linked to identity theft, fraud, election interference, cyberbullying and non-consensual pornography, causing irreparable damage and lasting harm.
Hawaii's Governor Green signed an AI safety bill yesterday, HB 2137, alongside a dementia care package. And the framing I actually care about: the AI measure got signed in the same ceremony as a bill targeting a disease that hits 31,000 residents — and that number, the state says, doubles by 2050. Right, and here's what I keep watching with state-level AI bills — which problem is this one actually solving: deepfakes, data privacy, or algorithmic decisions in something like Medicaid eligibility? The excerpt cuts off right at HB 2137, so we don't yet know. That's the tell, though. A state with $309 million in annual dementia-related Medicaid costs is exactly where AI shows up in production — eligibility screening, care coordination. If the safety bill doesn't reach that, it mostly targets the scary demo while missing the actual deployment. And it lands the same day we're talking about near-trillion-parameter open models and a $17.5 billion inference company. Fifty states writing fifty different definitions of 'AI protection' while the compute stack consolidates in Kentucky. Guess which one moves faster. The one with a P&L. Hawaii's got three million dollars appropriated for the memory network — Fireworks does that in ARR before lunch. If you follow the policy and power struggles behind AI, check out Anthropic Pentagon Watch — a daily briefing on Anthropic’s fight with the DoD over Claude, military AI use, autonomous weapons, and procurement blacklisting. Find it wherever you listen to podcasts.
What we’re watching next: initial capacity under TeraWulf’s Anthropic lease is expected to be operational by late 2027. If you want to dig deeper on anything we covered today, you’ll find links to every story in the show notes. Follow the threads that caught your ear. That’s AI Daily Briefing for today. This is a Lantern Podcast.