America Builds It — China Sets the Price

FROM THE DESK OF STEVEN FELDMAN

OPEN POSITION

The largest capital cycle in corporate history is aimed at a product two rivals have decided to make free. One of them is American.

In the last letter I said I would come back to the AI capital cycle, because it carries a revenue exposure that is only now coming into relief. Here it is.

OpenRouter, which Stripe agreed in August to acquire for a reported $7.5 billion, is the marketplace developers use to reach all the different AI models through a single connection, which makes it the closest thing anyone has to an AI model leaderboard. This February, Chinese-built AI models surpassed their American counterparts for the first time, processing 4.12 trillion tokens against 2.94 trillion. By July, Chinese models held the top spots with 60 percent of routed traffic. By contrast, Meta's Llama had fallen below one percent — a number that explains a great deal about what Meta did next.

The Financial Times recently ran a piece under the headline "The next China shock will come from open-source AI," arguing that countries adopting Chinese models will absorb Chinese standards and governance along with them. I suppose that is something to be concerned about in the long run. But as an investor, I am more concerned about the short run threat to the income statements and stock price of the American AI behemoths.

THE PLAYBOOK -- We Have Seen This Movie Before

China exported roughly 700,000 cars in 2019. In 2025 the number rose to ~7.1 million, a tenfold move in six years.  Solar panels and batteries followed the same arc. The Chinese playbook has not changed: subsidize scale, drive unit cost below the level at which the rest of the world's producers can survive, then export until each country's incumbents are begging their governments for protective tariffs.

But AI has no port, no roll-on roll-off vessel, no dealers. Just a software model to download. We’re gonna need something other than a tariff.

Washington has already reached for the TikTok playbook. Federal agencies, including the Pentagon, are barred from using DeepSeek on federal systems and devices.  Recently, US lawmakers opened a probe into how widely American companies are using Chinese models. There is no prohibition on that use, and the answers will not be comforting: the company Cursor built its Composer 2 model on Moonshot's Kimi, and Alibaba's Qwen has passed a billion downloads.

But TikTok is a service. It has a company behind it, servers to subpoena, an app store listing to remove — chokepoints, in other words. The Chinese have created what is called an “open-weight model” which means the company publishes the model's actual parameters — the trained guts of the thing — for anyone to download, run on their own machines, and modify. Not metered access through somebody's connection. The file itself. Which is why the price is not really the point. Once the weights are out, they cannot be recalled, repriced, or sanctioned back behind a paywall. There is no later in which China decides to start charging.

DeepSeek’s open weight model has already been downloaded onto millions of machines, forked into thousands of derivatives under licenses that permit commercial use, and stripped of any connection to the company that trained it. You can ban a vendor. You cannot un-publish a file.

THE PRICE -- What Free Does to a Revenue Line

The distinction that matters is open versus closed. A closed model is rented: you pay per token for metered access through the vendor's servers, and the weights never leave. That is the business all of this capital is meant to earn back. An open-weight model is the file, and because the file is public, nobody has exclusive supply of it. Dozens of hosts serve the identical model and compete only on infrastructure efficiency, which drags the price of the open tier down toward the cost of the compute underneath it. That is not a rival undercutting on price. It is a floor. And the floor is permanent, because a published file cannot be withdrawn, and each new release resets it lower.

Which means the closed labs are not really selling intelligence. They are selling the gap between their model and the best free one. That gap is now measured in months.

The hosted tier is where that floor shows up in somebody's invoice. DeepSeek's V4-Pro is priced at a fraction of the comparable American frontier model — frontier meaning the handful of most capable systems at the top of the market, the ones all this capital is being spent to build — and it stayed a fraction even after DeepSeek raised its own prices sharply in August.  The Chinese phone company Xiaomi cut pricing on MiMo, its own family of open models, as much as 99 percent in May. These models, launched in April 2025, are now among the most heavily used in the world.

Note what developers are actually optimizing for. They are not always seeking the models with the highest intelligence.  Instead, they are choosing blended cost per token and specific competence in coding and long context. The American labs can hold the intelligence crown and still lose the volume based on cost or sub-specialties.

THE AMERICAN ANSWER -- When You Cannot Win, Commoditize

In August, Mark Zuckerberg published a 6,500-word essay and announced Meta would release the weights for Muse Spark 1.2, its most capable model, plus a new Glimmer family of models built to run on laptops. It was framed in nationalism:  American open source must lead, and Washington should ease the friction that holds it back. Llama, remember, is now below one percent of routed traffic. Muse Spark is the answer to that.

But this is Mark Zuckerberg, and he only cares about Meta. The truth is that it should have been framed by his financial statements.  Meta has guided to $115 to $135 billion of capital expenditure this year, but it will not be a closed-model leader. So it commoditizes the layer it is losing and defends the layer it owns, which is devices, distribution, and attention. This poses another threat to the closed model pricing power.

THE ARITHMETIC -- $725 Billion Against a Floor of Zero

Microsoft, Amazon, Alphabet, and Meta are on track for roughly $725 billion of capital expenditure in 2026, up about 77 percent from around $410 billion in 2025. Dell'Oro Group, which tracks data center and telecom infrastructure spending, puts total global data center capex above $1 trillion this year. Goldman's baseline has annual AI capex at $765 billion in 2026 rising to $1.6 trillion by 2031. Capital intensity at the big four now runs somewhere between 45 and 57 percent of revenue, and Morgan Stanley and JPMorgan have both suggested the sector may need to issue $1.5 trillion of new debt to fund it.

Run the depreciation. Spread one year's $725 billion over a five to six year useful life and you have added $120 to $145 billion of annual depreciation from a single vintage — before power, before financing cost, before the next vintage lands on top of it. That charge is fixed the moment the concrete cures. The revenue meant to cover it is set in a market where the marginal competitor prices at a fraction and publishes the weights for nothing. Note the chain. The capital is spent by the infrastructure owners. The revenue that has to justify it is collected by the model vendors renting that infrastructure. If model pricing breaks, it breaks one step upstream of the balance sheets carrying the depreciation.

That gap is the whole letter. Malinvestment is not capital spent on a bad idea. It is capital committed against a price that someone else has decided will not exist.

THE WITNESS -- Even the Optimist Has Turned

Bill Gates recently published a long essay titled “The turbulent AI era is here. The choices we make now are critical.” Gates remains an optimist about what AI can do in medicine, education and energy. But the tone has changed. He believes many jobs will disappear permanently, that cognitive work will be automated faster than displaced workers can adjust, and that governments are nowhere near prepared for the transition — much the same conclusion I reached in June, arriving from the investment side rather than the policy one.

For this letter, though, the most important passage is not about jobs. Gates writes that if someone had a credible plan for slowing AI globally, he would probably support it. Then he explains why that will not happen: the geopolitical and economic incentives are simply too powerful to stop.

That is the capital cycle in a sentence. Microsoft cannot stop because Google might not. Google cannot stop because Meta might not. And none of them can stop because China certainly will not. The spending does not continue because everyone has done the arithmetic and likes the answer. It continues because no participant can afford to be the one who stopped. Malinvestment on this scale does not require stupidity. It only requires that everyone be trapped.

THE COUNTERARGUMENT -- Where the Skeptics Have a Point

Open weights are not free to run — serving, tuning, and security carry cost. Regulated industries in the West are unlikely to route sensitive workloads to models of Chinese origin, and the governance concern the FT raises cuts both ways: it is also a compliance moat for American vendors. Export controls tightened again on May 31, when BIS extended licensing to any China-parented buyer regardless of where the subsidiary sits, closing the offshore compute channel. China's constraint is memory, and it is binding. The clearest evidence sits with DeepSeek itself. In August it scrapped flat pricing for peak and off-peak rates, taking V4-Pro output from $0.87 to $3.96 per million tokens at peak. The disruptor is not immune to the cost of serving.

All true. None of it restores pricing power in the developer and startup layer, which is where defaults get set and where this decade's incumbents are currently being chosen. Compliance protects the enterprise contract. It does not protect the marginal token. And DeepSeek's price rise does not touch the weights, which remain published and free to run on hardware somebody already owns. Hosted rates are a business decision. The floor is not.

Trillions of dollars of fixed cost, a competitor that has made the output free as a matter of state policy, and a domestic champion that just joined them. All three are the same trade, held by roughly the same shareholders, financed increasingly by the same debt.

CODA — WHAT THIS MEANS FOR INVESTORS -- You Do Not Have to Pick the Winner

There are two ways I want to own the other side of this trade.

Gold is the anchor. Not because AI consumes gold, but because this cycle is increasingly financed with debt against returns that are far from assured. Gold hedges the financial side of the experiment: rising leverage, compressed prospective returns, and the policy response that follows if a meaningful share of this capital turns out to be uneconomic.

Real assets do something different. They are exposure to the physical side of the same experiment. Copper, uranium, power generation and energy infrastructure do not require me to know whether OpenAI, Google, Meta or DeepSeek wins. They require the buildout to continue.

That is the asymmetry I like. The buildout does not have to fail for today's AI valuations to disappoint. Returns merely have to come in below what current prices assume. And whether the software margin is captured in Seattle or Hangzhou, the data centers still need electricity, the copper still has to be mined and strung between substations, and the uranium still has to fuel reactors. The physical inputs are indifferent to who wins the model war. The equity claims are not.

So the positioning is straightforward: gold as the anchor, selected real assets as the beneficiaries of the buildout, funded by less dependence on concentrated mega-cap AI exposure. High conviction on direction, moderate on timing. This is a multi-year structural view, not a call on the next earnings season.

What would change my mind: two consecutive quarters in which the large labs show inference revenue growing faster than depreciation and interest on the same asset base. That would tell me pricing power survived commoditization, and that I am early rather than right.

The spending is committed. The price is not.

Steven Feldman

Steven Feldman is the co-founder & CEO of GBI. This newsletter is for informational purposes only and does not constitute investment advice.