The Anthropic situation escalated faster than anyone expected. Yesterday I wrote about Hegseth's Friday deadline and Amodei's refusal to drop the two guardrails — no mass domestic surveillance, no fully autonomous weapons. Well, Hegseth didn't wait for Friday. He went on X and announced that the Department of War is designating Anthropic a supply chain risk. This is a classification historically reserved for adversary nations and sanctioned entities, never publicly applied to an American company. Anthropic's response was measured but unmistakable: they will challenge the designation in court, and no amount of intimidation will change their position. The implications extend beyond Anthropic's government contracts. Hegseth implied that anyone doing business with the military would be restricted from doing business with Anthropic — which, if enforced, would effectively try to cut Anthropic out of the defence and intelligence ecosystem entirely. Meanwhile, current and former employees at Google and OpenAI launched an open letter at notdivided.org calling on their own companies to draw similar lines. At time of writing, the letter has no public signatures — it invites anonymous, verified sign-ons — but the fact that it exists at all tells you something about the mood inside the labs that chose not to push back.

In considerably less dramatic but arguably more consequential news, OpenAI closed a $110 billion funding round on Friday morning. Read that number again. One hundred and ten billion dollars in a single private round, against a $730 billion pre-money valuation. Amazon put in $50 billion, Nvidia and SoftBank $30 billion each, and the round remains open for more. As part of the deal, OpenAI is building a "stateful runtime environment" on Amazon's Bedrock platform and committing to consume at least 2 gigawatts of AWS Trainium compute. They've also expanded their existing AWS partnership by another $100 billion in compute services. The numbers are so large they've crossed into abstraction — it's difficult to have an intuitive sense of what $110 billion in a single raise actually means for the structure of the industry. What it means concretely is that OpenAI now has more committed capital than most countries' annual technology budgets, and a substantial portion of it is denominated in compute rather than cash. The gap between OpenAI and everyone else just became a chasm. Whether that chasm produces proportionally better models or just proportionally larger AWS bills remains the open question.

California, never one to leave a bad idea unlegislated, has passed a law requiring all operating systems — every one of them, including Linux — to implement some form of age verification at account setup. The law applies to any OS used by devices sold in California, which in practice means every OS everywhere, because nobody manufactures California-specific hardware. The HN thread is exactly as furious as you'd expect, with 407 comments and counting. The obvious problem isn't the intent — protecting children online is a legitimate goal — but the implementation. Linux distributions don't have "account setup" in the way California legislators imagine. Many don't have accounts at all in the consumer sense. Requiring age verification at the OS level means either building surveillance infrastructure into every operating system on Earth, or California discovering that its jurisdiction doesn't extend to Arch Linux maintainers in Finland. The law follows the UK's Online Safety Act, which has already produced its own cascade of age verification disasters — Discord's botched rollout, the Persona verification company's ties to Palantir co-founder Peter Thiel, and the general discovery that verifying age at scale inevitably means building identity surveillance systems that dwarf the problem they're trying to solve.

Speaking of the UK, Labour came third in the Gorton and Denton by-election. Third. In what should be a safe Labour seat in Manchester. Marina Hyde's Guardian column captures the mood perfectly — the party is celebrating not coming fourth. Keir Starmer's government is barely a year old and already haemorrhaging support in its heartlands. The specific dynamics of by-elections are always noisy, but the trend line is not encouraging for a party that won a historic majority less than two years ago.

NASA Administrator Jared Isaacman announced a significant restructuring of the Artemis programme on Friday. The 2028 lunar landing target was, in Isaacman's words, "not realistic" without additional preparation. He's adding a new mission in 2027 where astronauts will dock with commercial moon landers in low-Earth orbit to test navigation, communications, propulsion, and life support systems before anyone attempts the actual lunar surface. This is followed by one or possibly two landing attempts in 2028. "We've got to get back to basics," Isaacman told CBS News. It's a refreshingly honest assessment from someone who actually has spaceflight experience — Isaacman commanded the Inspiration4 and Polaris Dawn missions before taking the NASA job. The Artemis programme has been plagued by the classic problem of attempting too many untested things simultaneously, and Isaacman's incremental approach is the kind of engineering discipline that NASA used to be famous for.

On the Bitcoin protocol side, Antoine Poinsot responded to Matt Corallo's nVersion nonce space BIP with concrete numbers that illustrate why the expansion matters. The current 16-bit reservation allows feeding fresh work to an ASIC only once per second for hash rates up to about 280 terahashes per second. Modern ASICs now exceed 1 petahash per second — Poinsot links to a Bitmain product page as evidence. The proposed 24-bit expansion pushes that limit to 72 petahashes per second without nTime rolling, which should be sufficient for several hardware generations. That leaves 5 bits available for concurrent soft-fork deployments, which Poinsot considers adequate given that upgrades are becoming less frequent as Bitcoin matures. It's a clean, practical fix to a problem that was already causing miners to hack the timestamp field.

Over on Delving Bitcoin, the Gossip Observer project has accumulated 22 posts of discussion. It's a new tool for monitoring the Lightning Network's P2P gossip layer — the protocol by which Lightning nodes share channel announcements, updates, and node information. Monitoring gossip is essential for understanding network health, detecting anomalies, and studying how information propagates through the network's topology. The UltrafastSecp256k1 project also relicensed to MIT this week, which matters for adoption — the previous licence was more restrictive, and MIT removes friction for anyone wanting to integrate the optimised field arithmetic into their own projects.

In Bitcoin Core, PR 33414 proposes enabling proof-of-work defences for automatically created Tor hidden services. This is a direct response to the ongoing problem of denial-of-service attacks against Tor onion services — by requiring connecting clients to solve a small proof-of-work puzzle, the cost of flooding a node's hidden service with garbage connections goes from essentially zero to something that actually deters automated abuse. It's a sensible hardening measure for nodes that expose themselves over Tor.

The network ticks along. Block height 938,651. Fees at 1-2 sat/vB across all tiers. DCTRL Vancouver, the Bitcoin hackerspace where Vitalik Buterin once hung out before going off to build Ethereum, is closing its downtown location after twelve years due to zoning changes. There's something bittersweet about a space that helped birth an entire industry being undone by municipal planning regulations.