The frontier became strategic territory this week — and territory gets fenced. Meta closed its frontier models outright; NVIDIA is buying the open estate; espionage and containment proposals arrived at the labs; and the frontier bifurcation spotted in Episode 12 hardened into policy. Meanwhile the accessible side of the fence kept getting cheaper, faster and more reliable: the frontier recedes, the floor rises.
Meta Superintelligence Labs released Muse Wonder and Wonder Full (codename Watermelon) — Wonder Full lands #2 on Artificial Analysis' intelligence index, three points clear of Gemini 4 Argon, GPT-6 Astra and Claude Fable 5.1, passing Opus 5.5 on legal work, long coding and memory. The signal is the licence: both models are closed. The lab that made open weights a competitive weapon has fenced its frontier — Wonder Full gated behind $50/$200 monthly Muse Code plans, with API prices promised below the incumbents. The same week brought Muse Spark Open, reportedly weaker than its predecessor's point release: the open promise kept, with the B-team. Then the counter-beat: NVIDIA — already $800M into Reflection AI — is reportedly exploring a full acqui-hire that would top the $20B Groq purchase as its largest ever, after buying Hugging Face, Poolside and Essential AI. Open weights as public-good counterweight, or open weights as hardware moat? That ambiguity is the signal — and the watch-for is the next Llama licence. If that's closed too, the open-frontier era ended not by argument but by exit.
Andrej Karpathy revisited his April essay on how people judge AI and named the structure: the sharp funnel. ~5,000 people with true frontier access inside the labs at the narrow end — weeks-long agent swarms, open mathematics problems — against ~6B people who've hardly used AI and 1–2B running free chatbots. He reckons 0.2% of the population truly grasps frontier capability, and the gap is widening (4.8M views). His explanation: models excel where answers are checkable and lucrative — code, mathematics — so the narrow end feels the curve going vertical while everyone else underestimates. Same shape in money, per Ramp via the AI Daily Brief: 80% of OpenAI and Anthropic enterprise revenue from 1% of customers; top-1% payers averaging $903/month against a $25 median; 2.2% of US households paying for AI while 1.2B use ChatGPT weekly. The entire CapEx ramp — $825B in combined 2026 guidance, approaching $1T in 2027 — rests on that sliver of whales. And the rankings driving the narrative run on unstandardised accounting: Anthropic's $65B run rate "surges past OpenAI — depending on how you count." Two days, two lenses, one pyramid. Watch for the first policy proposal citing the funnel's shape.
Federal prosecutors charged an OpenAI engineer with planting a hidden recording device in the company's executive meeting room — found September 30, tucked behind a TV. A chatbot account linked to the spouse allegedly asked how to listen in remotely, encrypt a memory card and hide a homemade recorder; OpenAI's own models helped crack the case. The US Attorney called it a matter of national security, and Pragmatic Engineer added the detail the wire stories missed: the accused worked on the Agents SDK — the tooling layer that decides what autonomous agents can do. The week stacked four containment vectors around it: physical (Ukrainian drones struck Yandex data centres, after Iran's Gulf strikes on Amazon facilities), doctrinal (Nadella's assume-breach essay — treat AI as already compromised, 10M views), the espionage case itself, and institutional (400 scientists across 34 countries — Hinton and Bengio among them — proposing frontier models be isolated from external networks: air-gapping as policy language, connectivity restraint rather than capability restraint). Meanwhile diplomats met in the Reykjavik room where Reagan and Gorbachev sat in 1986 — the informality before the formality is how treaty processes start. Frontier labs are becoming strategic installations, and those get fences, clearances and guards.
The economics on everyone's side of the fence kept rotating — and the rotation is the tell. Anthropic is serving Opus 5.5 at a verified 285 tokens/second; OpenAI's docs suggest a 280 theoretical max for GPT-6-1 Sol at a 50% price premium; and Amazon returned to the Decart table at a $7B valuation — a 75% premium one month after walking away — for a stack claiming 8× faster inference at one-eighth the cost. Microsoft's Maia 300 claims 40% more AI output per unit of electricity: the binding constraint has shifted to power, and the marketing now says so out loud. And the quiet one: Modal crossed $1B ARR, with sandboxes alone a $300M run-rate product (~3B spun up, scaling to a million concurrent) while Perplexity SPACE disclosed a <0.0002% error rate — competing on reliability, not benchmark scores. Advantage is being redefined away from intelligence peaks toward throughput, isolation and trust — and the sandbox layer is becoming the agentic era's trust boundary: where your agent's blast radius gets contained. Watch Factory's agent sandboxes inside the customer's own cloud — the sovereignty story of 2027.
The fence is rising on three fronts at once. Market: the frontier's best weights are now closed, by the company that made openness a weapon. Experience: the funnel says most people's AI is a free chatbot while five thousand insiders run weeks-long swarms. State: the security establishment is arriving with clearances and containment proposals.
But the fence has a weakness. Everything on the accessible side — workhorse models, sandboxes, eval tooling, the open weights that still exist — keeps getting cheaper, faster and more reliable on a metric the fence can't revoke: capability per pound, per second, per joule. The frontier recedes; the floor rises. Both were true this week.
The practical reading for a sovereign individual: know which side of the fence you're building on. Weights you can run and inspect — the floor rising in your favour. Agent execution inside your own perimeter — the trust boundary under your control. And the public-good stakes are sharp: if open weights end up as merely a hardware moat, the sovereign option quietly closes. The next Llama licence, the Reflection deal terms, whether isolation language reaches legislative text — those are the tripwires.