The AI Margin Squeeze - inevitable?

Sat, 25 Jul 2026 10:45:49 -0500

Recent updates 🗞️

🇪🇸 Spain wins 2026 Fifa World Cup
🚀 SpaceX stock drops below IPO price
🛢️ Iran US conflict resumes; Crude back up to $90/barrel
📚 Reading “The Algorithm” by John McNeill
🎥 Watching “Your Friends & Neighbors S2”
YTD Portfolio Performance: +12.31% YTD

I’ve been thinking a lot lately about a deceptively simple question: what is a token actually worth?

Not philosophically - commercially. If you’re OpenAI or Anthropic, a token is your product. If you’re one of the thousands of companies now building on top of these models, a token is your input cost. And if you’re an investor watching a trillion dollars of data center capex roll down the pipeline, the price of a token is the variable that decides whether this is the greatest infrastructure investment in history or the greatest capital bonfire since the dotcom fiber overbuild.

Benedict Evans published an essay this month — “Ways to think about token pricing” — that frames this better than anything I’ve read. And then, a week later, a Chinese lab called Moonshot AI released Kimi K3 and gave us a live experiment in his thesis. I think the two events together tell us more than either does alone.

Evans’ argument: every road leads to commodity

Evans’ core claim is uncomfortable for anyone holding AI lab exposure: every dynamic we can currently see points toward foundation models becoming low-margin commodity infrastructure, with the value captured by whoever builds on top.

His reasoning runs on four questions:

How many use cases actually need the frontier? Plenty of work runs fine on a small, cheap, “good enough” model. The expensive top-right of the price/performance curve only matters if enough buyers have an ROI for it.
Does the frontier keep moving? If capability gains slow while efficiency gains continue, the frontier premium collapses.
Will frontier competition stay fierce? Today, a mid-single-digit number of labs use mostly the same science and mostly the same data to get mostly the same results. No network effect, no winner-takes-all mechanism — yet.
Who captures the value of the high-end use cases? Even where you need the big model, does it capture the value, or does it end up as a component inside someone else’s product? As Evans puts it, every SaaS company is a “database wrapper” — and Oracle didn’t capture SaaS.

His most striking comparison is mobile data. Cellular traffic grew by orders of magnitude over 20 years into a trillion-dollar revenue industry with $200 billion in annual capex — and the carrier stocks went nowhere. All the value was captured further up the stack. Selling tokens, like selling bits, is an opaque unit of marginal cost that maps to nothing a customer actually values.

As a techie-turned-MBA, this is the part that grabbed me: it’s marketing myopia in reverse. The labs think they’re selling intelligence. The market may decide they’re selling bandwidth.

Then Kimi K3 showed up

On July 16, Moonshot AI released Kimi K3: a 2.8 trillion parameter mixture-of-experts model with a 1M-token context window. Its self-reported benchmarks beat Claude Opus 4.8 and GPT-5.5 — the models sitting just behind the American frontier — while trailing only Claude Fable 5 and GPT-5.6 Sol. On Artificial Analysis’ long-horizon knowledge work eval, it scored an Elo behind only Fable 5. It took the top spot on Arena’s frontend-code leaderboard outright. And Moonshot has promised open weights by July 27.

Read that again through Evans’ framework. A Chinese lab, using mostly the same science and presumably cheaper compute, has replicated near-frontier capability within months of the leaders — and is about to hand the weights to anyone with a GPU cluster. If you wanted a single data point for the commoditization thesis, this is it. The moat, whatever it is, is measured in months.

But here’s what caught me off guard: Moonshot raised prices.

Kimi K3 launched at $3 per million input tokens and $15 per million output tokens — the same price as Anthropic’s Sonnet tier, and the most expensive model ever released by a Chinese lab. Their previous model, K2.6, cost $0.95/$4. The supposed commoditizers just tripled their prices.

So which is it — commodity or premium?

I think both readings are true, and the tension between them is exactly Evans’ point about the supply crunch.

The bull case reading: near-frontier intelligence has real pricing power right now. Moonshot looked at demand — driven overwhelmingly by AI coding, the one use case with undeniable product-market fit — and concluded buyers would pay Sonnet prices for Sonnet-class capability regardless of the flag on the lab. When the marginal buyer is a CFO funding software development, intelligence is not a commodity. It’s priced on value, not cost. That’s the first thing they teach you in a pricing strategy course, and Moonshot clearly took the class.

The bear case reading: this is what a supply crunch looks like from the inside. Evans is explicit that today’s pricing is transitory — capacity is constrained, one use case is soaking up all the inference in the world, and everyone can name their price. The test isn’t whether Moonshot can charge $15 today. It’s whether anyone can charge $15 in 2028, when the trillion dollars of capex has landed, inference efficiency has compounded, and there are open-weight K3-class models running in every cloud. K3’s own open-weights release on July 27 starts that clock.

The contrarian take: Kimi K3’s price hike isn’t evidence against commoditization — it’s the last, best moment to charge premium prices before the shakeout. Charge while the crunch lasts.

The variable nobody can model: politics

There’s one more wrinkle that makes this genuinely hard, and Evans flags it: Beijing is reportedly considering curbing overseas access to China’s top AI models — potentially including open-source releases. Some voices in Washington have floated similar ideas from the other direction.

Think about what that means. The strongest force pushing AI toward commoditization is the steady flow of near-frontier open weights out of Chinese labs — DeepSeek, GLM, and now Kimi. That flow is not a law of nature. It’s a policy choice, and it could be switched off by either government. If it is, the surviving closed frontier labs suddenly look a lot less like mobile carriers and a lot more like TSMC: a tiny number of players at a frontier too expensive for anyone else to reach.

So the honest answer to “will AI labs capture value?” might be: ask the export-control lawyers.

Evans ends his essay with the observation that for foundation models to avoid becoming commodity infrastructure, “something needs to happen that we don’t see yet." Charlie Munger would have appreciated that sentence. It’s an admission that the honest state of knowledge is uncertainty — and as he liked to say, “knowing what you don’t know is more useful than being brilliant.”

A Chinese lab just shipped a near-frontier model and tripled its prices in the same week the world’s best tech analyst argued prices must eventually collapse. Both of them might be right. That’s not a paradox — that’s a cycle. And cycles, as always, are where the interesting investing happens.

What are your thoughts — is intelligence a commodity, or the last great premium product? Let me know in the comments below.