Every outlet that covered yesterday’s Pentagon announcement covered it as a product launch: the military now has its own ChatGPT and its own Grok. Not one of them counted the days on the order that makes the other half of the story permanent. The Defense Department memo that started Anthropic’s removal from military systems was signed on March 6 and gave commanders 180 days. That window closes on September 2, which is tomorrow. The timing is the story, and it did not appear in a single writeup I read this morning.
Four things worth your attention today. Three of them are about the terms under which you are allowed to use a model, and only one is about a model getting better at anything.
The Pentagon Started the Last Week of Its Claude Countdown
The Defense Department added ChatGPT Mil and Grok for Government to GenAI.mil on Monday, the secure portal it launched last year with Google’s Gemini for Government as the only option. DefenseScoop reported the same day that the platform has onboarded more than 1.7 million unique users out of roughly 3 million department personnel. Set aside what you think about the customer: that is one of the largest frontier-model deployments anybody has actually stood up, and it happened in about nine months.
Claude is not on it, and the reason is not technical. CBS News reported in March on an internal memo signed by Defense Department CIO Kirsten Davies ordering commanders to remove Anthropic’s products “within 180 days,” and requiring defense contractors to stop using them on department work. Count forward from the March 6 signature and you land on September 2. Anthropic laid out its side in a statement published in late February: the department wanted Claude approved for any lawful use, and Anthropic refused to drop two specific restrictions, one on mass domestic surveillance and one on fully autonomous weapons.
These threats do not change our position: we cannot in good conscience accede to their request.
The department designated Anthropic a supply chain risk, a label Anthropic is contesting in court. Here is where I land, and it is not the comfortable position: Anthropic was right on the merits and it is going to lose the account anyway, and the useful lesson for the rest of us has nothing to do with defense policy. A usage policy is now a procurement variable. Your model vendor’s willingness to say no to a large buyer determines whether that vendor is available to you next quarter, and the vendor most likely to hold a line on principle is also the one most likely to get removed from a category of customer entirely. If you are picking between providers for something you plan to run for years, that risk belongs in the evaluation next to latency and price. Our Claude versus ChatGPT comparison weighs the capability side; this is the axis it does not cover, and until this year nobody needed it to.
Triage: matters. If you sell software into any federal supply chain, the contractor certification requirement in that memo reaches you, not just the department.
Apple’s New CEO Inherits a Billion-Dollar Rental Agreement
John Ternus becomes Apple’s chief executive today, the transition Apple announced back in April when it said Tim Cook would move to executive chairman. Ternus ran hardware engineering. He is inheriting a software problem.
Bloomberg’s Mark Gurman reported last November that Apple settled on paying Google roughly $1 billion a year for a custom 1.2 trillion parameter Gemini model to power the rebuilt Siri, running on Apple’s own Private Cloud Compute servers. The detail I keep coming back to is the one in the same reporting: Apple evaluated Anthropic for the job and the number came back around $1.5 billion a year.
That is the most expensive admission in tech that building your own frontier model is not automatically the right call. The company with the largest cash pile in corporate history, the one with the deepest possible incentive to own the layer, ran the numbers and rented. If you are sitting in a planning meeting where someone insists you have to train your own model to be taken seriously, that argument now has to beat Apple’s arithmetic. It usually cannot.
Triage: matters, and it is quietly the most useful permission slip in this roundup.
Clay Is Raising at $7 Billion, and Your Credit Balance Is Downstream of That
Axios reported yesterday that Clay, the AI sales and marketing data platform, is raising a round led by Wellington Management at a $7 billion pre-money valuation. The company was marked at $5 billion in an employee tender announced in January, and closed a $100 million Series C at $3.1 billion in August of last year. That is a bit more than double in twelve months.
Now the part that is my read rather than reported fact, so treat it as a prediction and not a number: companies that mark up that fast have to grow into the mark, and credit-metered go-to-market tools grow into valuations by changing what a credit buys. Nobody announces that as a price increase. Enrichment waterfalls get re-weighted, a provider that used to cost one credit costs three, the generous free tier on some lookup quietly becomes metered. If your outbound sequence is built on Clay credits and your cost per booked meeting assumes today’s rates, that assumption has a shelf life. Run the number at double and see whether the motion still works.
Triage: breaks your stack, on a delay.
“Open Weights” Now Has a Revenue Ceiling
This one has been building for a month and nobody has framed it plainly, so here it is. The open-weight Chinese models that a lot of teams standardized on this year are quietly acquiring commercial terms.
Moonshot published the Kimi K3 license alongside the weights in late July, and it reads like a permissive license right up until the revenue clause: anyone operating the model as a service has to negotiate a separate commercial agreement once revenue across the licensee and its affiliates passes $20 million over any twelve months. Reuters reported that such an agreement can carry a revenue share of up to 30 percent, and that Alibaba plans to attach similar terms to the open-weight release of Qwen3.8-Max. Alibaba has not settled on a rate. Meanwhile Qwen3.8-Flash-Next shipped on August 26 billed as an early preview of the architecture Qwen4 will be built on, 125 billion parameters with about 6 billion active per token.
For most people reading this, nothing changes. Internal use is free, fine-tuning is free, and if you are under the threshold you are under it. The change matters if you resell inference, because “open weights” is drifting toward “source-available with a revenue ceiling,” and the ceiling is low enough that a successful small API business hits it. We covered MiniMax’s open-weight release a few weeks back in the same spirit of everybody assuming open meant unencumbered. Read the license before you benchmark, not after you have built the product on it.
Triage: breaks your stack, but only if you are selling what you serve.
What I’m Watching
The Pentagon clock expires tomorrow and I expect no announcement at all, because a deadline passing quietly is what a removal order looks like when it works. Apple’s first Ternus quarter will tell us whether the Gemini deal is a bridge or a destination; Gurman’s reporting says Apple wants its own models handling those tasks by 2027, and the gap between wanting and shipping is where Siri has lived for a decade. And if you take one thing from today, take the licensing story, because it is the only one on this list that can change your cost structure without anybody sending you an email about it.