Open Weights, Closed Seas

There is a useful, ugly symmetry to the headlines this week: one set of actors is trying to make powerful things cheap and widely available; another is making the world expensive by threatening the pipes that carry oil. Both moves are lessons in leverage. In AI, leverage looks like model size, token efficiency and inference costs. In the Gulf, leverage is geography — a few miles of water that translate into global price flows.

If you want a single throughline, here it is: scarcity creates power. Whoever controls the scarce input — whether it's frontier model capability or safe, open sea lanes — sets the terms. But scarcity can be manufactured, eroded, or commoditized. Ben Thompson argues that Chinese open-weight models are attempting to erode scarcity in AI capability and let market economics do the rest (and not incidentally to give domestic firms lower-cost inputs) Stratechery. Meanwhile, missiles and militias hunt tankers, and insurance premiums and spot prices tell the rest of us how scarcity gets priced into everyday life Yahoo.

This newsletter will do two things you do not get from cable: connect the dots across these theatres, and spare you the moralizing that adds nothing to your situational awareness. Expect numbers, a little gallows humor, and the occasional unpleasant truth: technology that lowers the cost of doing something useful will be used — for productivity, for competition, and yes, for war.

Ben Thompson reframes the Chinese open-weights story as pure economics: the front-loaded R&D costs are separate from the real ongoing cost that matters — cost of goods sold (COGS) for inference Stratechery. The upshot is delightful in its mundanity: weights themselves can be free, but running them at scale is not. That means the war will be won by whoever optimizes the marginal math — memory footprint, inference efficiency, and token economy — not merely by who shouts "bigger" loudest.

Concrete pricing Thompson cites is instructive: Kimi K3 at about $3 per million input tokens and $15 per million output tokens, versus some Western counterparts like Sol at $5/$30. He also notes the parameter counts being thrown around — Qwen 3.8 Max (2.4T), Moonshot AI (2.8T) — but stresses that raw parameters are theatre; what matters is the real work per dollar. As he puts it, tokens are not a commodity; intelligence is becoming a commodity-like output, governed by footprint and efficiency.

"Tokens are not a commodity" — the claim is part provocation, part taxonomy. The practical conclusion is that profitability will increasingly be about cost structure rather than price power.

Two strategic consequences follow. First, China’s push for open weights — encouraged from the top down — is a deliberate attempt to reduce barriers to experimentation and commoditize the complements (tools, fine-tuning pipelines, deployment know‑how). That will spur global innovation but also raise cybersecurity and competitive risks. Second, distillation — using strong Western models as teachers to train local variants — gives Chinese engineers a recurring structural edge if Western labs insist on using ToS and enforcement as gatekeeping. Thompson urges policy nudges: favor fair use, loosen distillation restrictions for U.S. firms, and avoid letting frontier labs set global norms Stratechery.

Practical note for product people and policy wonks: the battle isn't just about who owns the weights — it's about who controls the cheap, reliable inference layer. Whoever wins that layer sets the baseline for everything built on top.

Google announced an efficiency offensive: Gemini 3.6 Flash, 3.5 Flash‑Lite, and 3.5 Flash Cyber. The pitch is explicitly operational — lower latency, better token efficiency, and stable behaviour for agentic workloads — with headline pricing for 3.6 Flash of $1.50 per million input tokens and $7.50 per million output tokens Google Blog.

Two bells ring here. First, margin pressure from more efficient models matters. If Google can shave 17% (or much more on some benchmarks) off output tokens and improve throughput, it changes the enterprise calculus: cheaper inference lowers the floor price at which automation is profitable. Second, Google’s productization acknowledges an ugly truth of modern conflict: AI is now a utility of statecraft. The 3.5 Flash Cyber variant — exposed via the CodeMender programme and targeted to governments and "trusted partners" — is a reminder that not all capability is being commercialized equally. Restricted access to security‑focused models creates asymmetries: one country's defenders get specialized, low‑cost cyber tooling while others make do with commodity models.

Google also embeds a Frontier Safety guardrail, promising better resistance to misuse in areas like chemical, biological, radiological, nuclear, and cyber operations while trying not to overblock legitimate uses. Fine balance — or bureaucratic theatre — depending on how the safety gates are operated in practice. In short: incumbents are moving from a capability arms race to an efficiency arms race, while also trying to stitch a safety moat. Which matters, because in kinetic or gray‑zone conflict, speed and cost of decisions are as consequential as accuracy Google Blog.

Qwen's latest image model is not about making art for art's sake. Qwen‑Image‑3.0 is explicitly engineered for utility: up to 4.5k tokens of input, micro‑detail fidelity (10px text, pores, hair), layout mastery (newspaper PDFs, storyboards) and multilingual, interface‑accurate rendering Qwen. The marketing word is Real (实), and they mean it: this is about deployable output, not just Instagram aesthetics.

Why that matters beyond designers and publishers: tools that reliably produce complex, authentic visual artifacts change the cost of producing believable documents, mockups, and translated interfaces. There are plenty of innocent, productive uses — automated UI mockups, rapid storyboarding for small studios, educational materials at scale — but the same reliability makes disinformation cheaper and more convincing. A bad actor who can generate authentic‑looking government forms, forged headlines, or satellite‑style imagery (or at least convincing composites) is a tactical problem in any contested sea lane or domestic theatre.

To be blunt: as image models graduate from "good enough for social media" to dependable for operational workflows, the friction for deception falls. That is not an argument for censorship; it is an argument for accounting. Military planners and corporate security teams must treat high‑fidelity synthetic content as part of the threat landscape — alongside mines, missiles and malware Qwen.

The other front is less academic. A tanker attacked in the Strait of Hormuz forced crew abandonment as the U.S. conducted its 10th consecutive night of strikes against Iran, and the region is bristling with diplomatic and military risk Yahoo. The Strait remains a genuine global bottleneck — roughly 20% of the world’s crude and natural gas transits nearby — so disruptions show up quickly in prices: Brent above $88 and U.S. pump prices nudging toward $4 a gallon.

Compounding the danger, Yemen’s Houthis declared a naval blockade of Saudi Arabia, framing it as an "eye for an eye" and threatening Bab el‑Mandeb — another critical corridor Yahoo. Analysts warn that a full closure of Bab el‑Mandeb could shave about 7% off global oil supply, and between Gulf and Red Sea disruptions the market is already pricing non‑trivial risk premia. Casualty counts and wounded figures underscored the human toll — Iran reported dozens killed and hundreds wounded in recent strikes, and U.S. officials say nearly 100 service members have been injured since early July, most returning to duty Yahoo.

Do not romanticize blockades. They are simple, ancient instruments with modern consequences: insurance spikes, shipping reroutes that cost time and fuel, and price shocks that sap political support for governments. Diplomacy is still possible — Pakistan‑brokered talks and ad hoc mediation continue — but the battlefield logic is brutal: raising the marginal cost of moving oil is an effective lever because the world refuses to run on political virtue alone.

And yes, this ties back to tech. Commoditized, cheap AI speeds up information operations, automates open‑source intelligence, and lowers the barrier to orchestrating psychological pressure against shipping companies, insurers and ports. On the other hand, better, cheaper cyber‑focused models (hello, 3.5 Flash Cyber) accelerate both defence and attack in the same window. The result is a feedback loop: kinetic pressure raises prices; AI lowers the cost of doing the informational, financial and asymmetric parts of that pressure. Whoever adjusts fastest wins the short game; whoever rebuilds the resilience wins the long one.

There is a false moral binary doing rounds in policy circles: cage models and you stop misuse; open models and you invite chaos. Reality, as always, is more boring and more useful. Openness lowers friction: it democratizes innovation, forces competition, and drives down the marginal cost of useful AI. It also makes some kinds of misuse cheaper. The right question is a different one: where do you want friction, and who pays for it?

The geopolitics here are plain. China’s open‑weight strategy aims to reprice AI inputs for domestic and friendly firms; Western incumbents are pressuring for safety guardrails that can also serve as competitive moats. Google is attempting the middle path — product efficiency plus curated safety — and Qwen is turning images into infrastructure. Meanwhile, missiles, militias and blockades make energy and shipping a political lever again. Those threads are not independent. Cheap models accelerate everything from logistics optimization to propaganda; chokepoints and violence monetize scarcity.

So where does that leave you, the reader who would rather understand than panic? Two modest prescriptions:

• Treat infrastructure as the battlefield: that means ports, undersea cables, insurance markets and the inference layer of AI. Hardening any one of those corridors has outsized returns.
• Insist on realistic policy: favor openness that creates observable, measurable competition in inference costs while building reciprocal, enforceable norms around distillation, export controls and access to cyber‑specific models.

If you want a soundbite for friends: the coming decade will be decided not by who owns the biggest model, but by who runs it cheapest, defends the pipes, and keeps the markets predictable enough to plan investments. And if that sentence makes you feel a little queasy — good. Economics is often grim, and history likes winners who understand the ledger.