Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

The Commoditization of Intelligence: When Tokens Become a Utility

The best model wins the demo. The cheapest good-enough model wins the invoice, and the invoice is what actually scales.

In July 2026, CNBC reported a number that should have rattled every AI lab with a private valuation: American companies now route up to 46% of their tokens through Chinese open-weight models, according to data from OpenRouter, the service that brokers traffic across dozens of AI providers. A year earlier that figure was around 11%. In the first half of 2025 it was 4.5%. Chinese models have cleared 30% of enterprise token volume every single week since February, and in one February week they out-tokened American models outright for the first time.

Why the stampede? Not patriotism, and not benchmarks. Money. The open models coming out of DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, MiniMax, and Zhipu’s Z.ai run 60% to 90% cheaper than the frontier tier from OpenAI and Anthropic. As of mid-2026, DeepSeek’s V4 Flash costs about $0.14 per million input tokens. OpenAI’s GPT-5.5 lists at $5.00. That’s not a discount. That’s a different planet: the same routine work at roughly one-thirtieth the price.

The comforting story is that these are the cheap seats, fine for bulk drudgery, no threat to the real frontier. That story is wrong. Z.ai’s GLM 5.2, released in June 2026, benchmarks around fifth in the world, a hair behind the American labs, and became the fastest-growing model on Vercel’s gateway, with daily usage up roughly 27-fold in its first week. Near-frontier quality at a rounding-error price is exactly the condition economists have a dry, undramatic word for.

Commoditization.

Every Abundance Story Ends at Marginal Cost

One economic law runs underneath the entire book: when the marginal cost of producing something falls toward zero, its price follows, and every moat built on the old price collapses with it.

We’ve watched this movie in every input humanity ever managed to mass-produce. Aluminum was once more precious than gold. Napoleon III served his best guests on aluminum plates and handed the silver ones to lesser nobles, right up until the Hall-Héroult process turned it into a soda can. Long-distance calls, memory chips, solar watts, cloud storage, internet bandwidth: each began as a premium good with fat margins and a proud brand, and each ended as a metered utility you never think about. Nobody asks which company’s electrons are lighting their kitchen.

Intelligence is now on that curve. As I argue in Unscarcity, the three inputs to everything (energy, labor, and intelligence) are all sliding toward zero marginal cost at once, and intelligence is sliding fastest. Generating a paragraph cost dollars in 2023. By 2026 a full research report costs a fraction of a cent. When large language models get cheap enough to meter in hundredths of a penny, “tokens” stop being a product and start being a dial tone. Genius on tap, priced like tap water.

This is the same logic behind the energy standard: once a thing becomes fungible and abundant, its unit price stops carrying information about which lab made it. A kilowatt-hour is a kilowatt-hour. And increasingly, a competent token is a competent token. The buyer stops asking whose model it came from and starts asking what it costs.

“Good Enough” Is the Most Dangerous Phrase in Business

The frontier labs are not losing the benchmark war. OpenAI shipped GPT-5.6 the same week this story broke, and it’s superb. They’re losing something subtler and more fatal: the boring middle of the market.

Most tokens in the world are not doing frontier work. They’re classifying support tickets, summarizing meetings, extracting fields from invoices, drafting the fourteenth version of a marketing email, and answering the same customer question for the ten-thousandth time. That work doesn’t need a genius. It needs a reliable clerk who costs nothing. And the moment a Chinese open-weight model is good enough for the clerk work, a purchasing manager staring at a bill routes it there without a flicker of guilt.

That’s how commoditization actually kills margins. Not by beating the leader at the top, but by making the top irrelevant for most of the volume. The leader keeps the trophy and loses the business. When your differentiator (raw capability) becomes something four other labs can replicate within months and give away as open weights, your differentiator has stopped differentiating.

“But the Frontier Labs Will Always Stay Ahead”

Maybe. Probably, even, at the very top. But that misunderstands how commoditization works. It doesn’t require the cheap option to catch the leader. It only requires the gap to stop mattering for most of the work. Intel stayed ahead of the commodity x86 clones on raw performance for years and still watched its margins compress, because “fast enough and half the price” is a winning hand in the boring middle where the volume lives. The frontier labs can stay two steps ahead forever and still lose the invoice, because the invoice was never being paid for the last two steps.

There’s a second escape hatch people reach for: switching costs will lock everyone into the American incumbents. But open weights are the anti-lock-in technology. When the model is a downloadable file with a permissive license and an API that mimics the incumbent’s, switching suppliers is a config change, not a migration. That’s why the OpenRouter numbers moved so fast. The routing layer exists precisely to make the supplier interchangeable. The market built its own commoditization machine.

So Where Does the Moat Go?

If capability is no longer defensible, does that mean there’s no money left in AI? Of course not. It means the defensible value moved to a different floor of the building. When the product commoditizes, the margin migrates to whatever hasn’t commoditized yet.

Integration and distribution. The winner isn’t the smartest model; it’s the one already living inside your IDE, your CRM, your spreadsheet, your operating system. Being the default is worth more than being the best. This is the old network-effects game, where the value is the workflow you’re wired into, not the intelligence itself.

Trust, reliability, and accountability. A model you don’t control, running on weights from a lab in another jurisdiction, is cheap precisely because you’re absorbing a risk you can’t see. Enterprises will pay a premium for a supplier who is legally on the hook when the clerk hallucinates a refund policy. In a world where anyone can generate plausible output, verifiable output, output someone will stand behind, becomes the scarce good. That’s the same insight that makes Impact and Civic Standing the load-bearing currencies of the book: when intelligence is free, trust is what stays expensive.

Switching costs and the rentier layer. The compute landlords figured this out early. You don’t need to own the smartest model if you own the warehouse everyone rents to run any model. Own the substrate, tax the traffic.

The pattern is always the same. Abundance at one layer forces the scarcity, and the profit, up or down to the next one.

The Spigot Problem: When Cheap Becomes a Leash

Now for the twist that turns an economics story into a geopolitics story.

If everyone routes their routine cognition to Beijing’s models because they’re cheap, everyone has quietly handed Beijing a lever. In July 2026 that lever became visible from both ends. On July 8, the U.S. House Homeland Security Committee opened a probe into how deeply American business now leans on Chinese models: your customer-service bot and your code assistant running on weights you don’t control. In the same week, reporting surfaced that China is weighing whether to restrict access to its most powerful models, now that American dependence is measurable. The seller counted its buyers.

This is the oldest pattern in resource politics, and it’s the spine of the book’s chapter on the Geopolitics of Abundance. Oil taught us the rule: the barrel doesn’t care who pumped it, but the country holding the spigot cares very much. When the strategic resource shifts, and it’s shifting from oil to compute and the intelligence that compute produces, power flows not to whoever builds the smartest model but to whoever supplies the cheap, plentiful one. Cheap is what spreads. Cheap is also what can be switched off.

Which is why, as I argue in Energy Sovereignty Is Compute Sovereignty, the fight underneath the fight was never about who has the best demo. It’s about who controls the machines and the megawatts that make intelligence pourable in the first place. A commodity you import is a commodity someone else can ration.

The Unscarcity Read

None of this is a surprise if you’ve internalized the framework. The commoditization of intelligence is not a bug in the AI boom; it’s the AI boom working exactly as an abundance engine is supposed to. Intelligence is doing what electricity did, falling from a premium product to invisible infrastructure, becoming part of what the book calls the Foundation: the 90% of goods and services that get so cheap they’re simply there, like gravity.

And as always, the scarcity doesn’t vanish. It relocates. When the tokens themselves cost nothing, the remaining value climbs to the Frontier: judgment about which token to trust, integration into the systems that matter, accountability for outcomes, and control over the physical supply chain of compute and energy. The moat didn’t disappear. It moved to a floor the copycats haven’t reached yet.

The labs racing to build the smartest model are fighting over the trophy. The real game, the one the book is about, is what happens to a civilization when the thing that was scarce and expensive for all of human history becomes a rounding error on an invoice. When genius is on tap, the interesting question stops being how smart and becomes who decides, who’s accountable, and who controls the tap.

The purchasing manager already knows. The best model wins the demo. The affordable one wins the world, right up until the supplier decides to close the valve.

That’s the fork in the road Unscarcity was written to map. When intelligence stops being scarce, everything we built on its scarcity (prices, moats, even the balance of power between nations) has to be rebuilt on something else. The book is the blueprint for what that something else looks like.

Sources

• Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge — CNBC, July 7, 2026 (OpenRouter token-share data)
• Lawmakers probe growing use of Chinese AI models in U.S. companies — CNBC, July 8, 2026
• Chinese AI Models Now Capture Up to 46% of US Enterprise Token Usage — Forkast

Related Articles

• Unscarcity: The Word for When Scarcity Becomes Optional — The core thesis: what happens when energy, labor, and intelligence all approach zero marginal cost.
• Compute Landlords — When the model commoditizes, the margin moves to whoever owns the warehouse.
• Energy Sovereignty Is Compute Sovereignty — Who controls the reactors controls the tokens.
• Network Effects — Why distribution, not raw capability, is the durable moat.
• The AI Talent Paradox — When compute and capital go abundant, the scarcest input is a few hundred minds.
• The Energy Standard — Why fungible, abundant goods stop carrying a brand premium.