NOOPS Weekly — Week of 13 June 2026

There are weeks that add data points and weeks that move the frame. This was the second kind. On the evening of 12 June, the US government reached into the market and switched off a specific frontier model — suspending all foreign access to Anthropic's Fable 5 and Mythos 5, including for foreign-national Anthropic staff inside the United States. A letter landed at 5:21pm ET; the consumer model briefly went dark and fell back to Opus 4.8; the story ran top of Hacker News and across the NYT, BBC and ABC. The biggest IPO in history — SpaceX at a $2 trillion-plus debut — was listing in the same news cycle, and it got drowned out. That magnitude reading is the story under the story: the market now treats model-access policy as bigger news than a trillion-dollar listing.

What looked at first like a discrete safety incident has, over seven days, revealed itself as a regime change. The kill-switch was the trigger; the week's real work was the fallout — sovereignty repricing, open weights taking the intelligence lead, compute hardening into a literal asset class, and valuations detaching from cash flow. Underneath all of it runs the quieter thread we have tracked for months: as generation goes free, value migrates to verification, evaluation and the harness. This week did not contradict that thesis. It supercharged it. We build this wrap around four arcs — the closed-AI watershed and its blast radius; compute becoming something you can trade and own; valuations untethered from the arithmetic; and the maturation of local compute and the harness layer that is quietly the most important story of all.

The closed-AI watershed

Start with the mechanism, because the technical detail is load-bearing. The Fable 5 control did not follow a dramatic breach. As the anatomy emerged — Pliny the Liberator's multi-agent attack using homoglyph obfuscation and an "Opus Loop" feeding restricted fragments to an older unpatched sibling — and as cybersecurity veterans pushed back via freefable.org and Katie Moussouris in The Register ("no jailbreak, just a code prompt"), the verdict consolidated: this was guardrails-as-system-property, surfaced by Amazon researchers and escalated to the White House by Andy Jassy, not to Anthropic. Zvi Moshkowitz's synthesis was unsparing: "the worst possible licensing regime — fully ad-hoc, vibes-based, and based on the whims of people who do not understand how AI works." And by week's end the kill-switch turned out to be reversible: Sacks now says Washington wants controls lifted "as soon as possible" once Fable is patched.

But reversibility is not reassurance — it is the opposite. A hard seizure is a one-time rupture you can price. A switch tied to the security-vulnerability cadence of every frontier release is a recurring, discretionary variable, harder to price, not easier. That is why the premium on full-stack sovereignty hardened rather than fell. "The substitution math just flipped for non-US operators": every closed US token now carries a sovereign-discontinuity cost that was effectively zero a week ago, and the fallout has already widened to India, Korea and the courts, with the EFF filing an amicus brief and SK Telecom caught in the Mythos net.

Two NOOPS theses crystallised here. First, Mark's full-stack thesis: "a frontier model maker has to own the entire stack — probably all the way down to the silicon. Anything else and you have Andy Jassy rug-pulling on you." Applied as a screen, only Google clears it today. The Jassy episode broke Bedrock's neutrality pitch in a single move — and as he put it, no frontier lab will look at AWS the same way again. Second, John's "one-way door": with Washington sparing DeepSeek while freezing its own champion, the asymmetry de-risks the Chinese stack relative to the American one for every non-US buyer — precisely as that stack reaches parity. An ally sitting wholly in the US camp is no longer the obvious posture.

Open weights take the lead — and they're Chinese

The supply response did not wait. GLM-5.2 from Zhipu became the top open-weights model on Artificial Analysis' Intelligence Index at 51, landing on Ollama within days; Kimi K2.7 Code cut thinking-tokens ~30% while improving long-horizon coding. Tunguz's local-coding data put Qwen3.6 within points of Sonnet 4.6 on hardware a developer can own. And the means of production went sovereign too: Huawei claims a 1.6-trillion-parameter training run on domestic Ascend silicon — the part of the pipeline export controls were thought to cap. If it holds, it erodes the cleanest argument that controls preserve a permanent gap.

This is the good-enough thesis and a sovereignty shift arriving in the same instant, and Mark sharpened the punchline into a dare: run GLM-5.2 and GPT-5.5 in an OpenRouter-style fusion harness and get Fable-class results — "so that we can rug pull the White House." The logic builds on OpenRouter's "fusion beats frontier" result, where an ensemble outscored every individual model. You cannot embargo a capability that can be reconstituted from open and commodity parts, so export-control regimes aimed at single frontier models have a shortening half-life. EuroMesh and the DOE's Genesis Mission are the institutional expressions of the same instinct: pool the compute you already own and route around the kill-switch.

Compute becomes an asset class — and valuations lose the plot

The single cleanest signal of the week needed only six words from Mark — "claude, note that as a signal" — when Silicon Data and CME Group launched a GPU price index to underpin a compute futures market, explicitly modelled on WTI crude. John: "this is my inference thesis turned into an asset class." We have argued since March that compute is the underlying commodity of this era; a literal derivatives market forming around GPU-hours is the strongest validation that framing could get. Around it, the capital structure thickens: Nvidia raised $25bn in debt — its first bond sale in five years, and, as John noted, right after returning ~$80bn via buybacks — while TSMC alone funds the boom from cash flow. That contrast is itself an investment map. And RAMageddon arrived on schedule: DRAM and HBM pulled up the stack toward accelerators, a 15% smartphone contraction, Tim Cook telegraphing pricier iPhones. Compute scarcity is now landing on the device in your pocket.

But the same scarcity is also fuelling something less rational. SpaceX priced the largest IPO on record, ran to a ~$2.1 trillion cap, then — Cursor for $60bn folded in — became, by the SMH's measure, the planet's most expensive stock. John was blunt: "this one really is irrational." The arithmetic is the tell — ~$12bn of revenue at an Amazon-scale valuation, on Memphis compute built to train models the company largely isn't training, now rented out of compute desperation. Mark's line is the week's neatest framing: "an AI franchise on revenue that is really a landlord renting idle floorspace." Oracle posted enormous cloud growth days earlier and got a cold reception; SpaceX got euphoria. That gap is the sentiment marker — and the discipline, as John argued reframing OpenAI's "billions in losses," is to separate durable training-capex from genuine operating loss, a discipline the market is conspicuously not exercising.

The harness is the binding constraint

Beneath the geopolitics, the most investable thread compounded quietly. The week's engineering essays — htmx's "Code is cheap," Charity Majors' "return to discipline," Jane Street on the verification bottleneck, Ishmeet Bindra's "reviews are expensive, rewrites are cheap" — converge: once generation is free, the scarce activity is verifying and governing what's produced. John's refinement matters: this is not a return to old practice but a genuinely new paradigm, and his Smalltalk analogy warns the matching disciplines may take years to appear. The missing primitive surfaced too — Fulcrum's Inverse Rubric Optimization showed capability and evaluable capability are different curves, and OpenRouter's battle-royale ("EVALS ARE HARD," the cheapest model winning 27x on cost-per-win) showed leaderboards are weak predictors of real agentic performance. Measuring agents is the binding constraint; the human-approval gate "cannot hold against agentic speed," so value migrates to whoever automates that gate safely. That the harness layer is now M&A-grade (SpaceX paying $60bn for Cursor) and that Polypore is rebuilding the IDE as swappable panels both confirm: the battleground has moved above the model.

And the substrate is coming home. GLM-5.2 quantised to 2-bit at ~82% accuracy; Vicki Boykis declaring local models "good now"; co/core building a peer-shared inference marketplace on an OpenAI-compatible API; Subquadratic's ~1,000x attention-compute cut. Local inference is good enough — for the right job — and every closed-access shock raises the option value of weights you actually possess.

Looking ahead

The de-escalation is coming — John and Mark both expect Fable's controls lifted once it's patched. But the de-escalation is not the story. The story is what the episode taught the market in seven days: that frontier access is a discretionary regulatory variable, that a datacentre's value can be revoked by policy, that the only durable defence is owning the stack or routing around it. Watch three tells. First, whether the fusion rug-pull gets demonstrated — a public Fable-class result from open-plus-commodity parts would reprice single-sovereign lock-in overnight. Second, the GPU futures curve: who builds the WTI of compute, and what financing and arbitrage businesses grow on top. Third, the quiet compounders — evaluation harnesses, verification tooling, local-inference substrate — the layers a deflationary model market does not erode. The kill-switch was the headline. The harness is the thesis. 13 June was, as Mark put it, the day everything changed; the week that followed showed us in which direction.