Author: Michael Wenye Li

Translation: Shenchao TechFlow

Shenchao Introduction: AI labs have burned through hundreds of billions of dollars, but no one can clearly say when the money will be earned back. API pricing drops by a factor of ten each year, open source is chasing closed source, and training costs keep piling up. This article steps outside of the tech industry's perspective and provides a highly enlightening answer using the commercial model of Hong Kong's MTR over 45 years: stop thinking about making money from ticket sales and start owning properties above the stations.

They can't make money, and the question itself is wrong

There is a type of business that looks like this: billions in capital invested upfront, not a penny in revenue. The core service pricing approaches marginal cost. It creates tremendous value for users, but the builders can barely keep a dime. They must continually invest in the next generation of infrastructure.

This is not about AI labs; it is about large railway systems.

Many people use railways to draw analogies for the AI industry, and the conclusion for most is: general technologies have public good attributes, and commercial viability relies on government subsidies.

I want to challenge this conclusion. Because Hong Kong's MTR has actually solved this problem. It is one of the few subway systems in the world that is commercially self-sustaining, a publicly traded company that pays dividends and does not take government operational subsidies.

Financial structure is identical

The core railway business of MTR has never been able to self-sustain its expansion. The best year prior to the pandemic, 2018, saw an EBIT of 2 billion HKD from transport operations. However, the estimated capital expenditure for 2024-2026 is 87.9 billion HKD, almost all of which is used for railways. The peak railway profit over three years is only enough to cover 8% of the capital expenditure. Ticket revenue never covers the construction of the next line; that was never its design intention.

MTR keeps ticket prices at an affordable level through government pricing adjustment mechanisms. You cannot set ticket prices to recover construction costs, because then no one can afford them, which contradicts the essence of public transport. Each line might cover its own operating costs, but the ticket revenue can never support the construction of the next line.

The pricing of AI APIs faces a mirrored version of the same problem. Distillation and open-source alternatives are causing API prices to drop at a rate of about ten times per year; any lab pricing above marginal cost will lose volume to competitors. Every model can achieve operational profitability at the inference level, but the profit margins can never support the expenses of the next round of training.

The common solution globally is subsidies. The London Underground relies on TfL funding, while China's high-speed rail carries a trillion-dollar debt, with 94% of its lines not making money. AI is following the same path: the CHIPS Act, the Stargate project, sovereign wealth fund investments, Pentagon contracts. The underlying assumption is that it relies on subsidies for quasi-public infrastructure.

MTR has found another way.

Railways + Properties

When MTR was built in 1979, the designers understood that ticket prices would never recover construction costs. Thus, they structured the company around a completely different premise: railways would increase the value of surrounding land, so they must hold onto the land.

MTR develops residential buildings, office buildings, and shopping centers above and around the stations, capturing the value created by its infrastructure. Property profits are reinvested into railway operations and fund the next line. Today, MTR has 13 shopping centers and manages 47 station property projects, with property contributing a significant portion of actual profits.

The logic is clear: stop trying to capture value from the railway service itself, and instead own the assets that appreciate because of the railways.

AI's Correspondence

“When will AI labs make money?” and “When will railways sustain themselves on ticket prices?” are isomorphic questions. The answer is the same: they cannot, and the question itself is wrong.

A biotech startup uses cutting-edge models to screen drug compounds, saving two years of clinical trial time. A logistics company optimizes its routes with it, saving $40 million in fuel costs. An independent developer delivers a project in one weekend that took a five-person team three months to complete. In each case, the model provider only captures a fraction of the value through API fees. The provider cannot raise prices because there are four other labs and dozens of open-source alternatives offering similar capabilities. The residual value flows to users and the broader economy.

This is how general technologies work. Steam engines, electricity, and TCP/IP have never contributed much income to their creators.

The lesson from MTR: stop trying to have ticket prices cover construction costs; go find your “property.”

Four candidates, ranked by defensiveness

The government-granted deployment rights come first. The government authorizes a lab exclusive access to national medical records, tax systems, or defense logistics. The data accumulated, the depth of system integration, and regulatory qualifications take years to replicate, which is MTR's own mechanism: the state grants development rights based on natural monopoly attributes.

The accumulated reinforcement learning reward data comes second. Billions of interaction signals are used to train the next generation of models. Unlike model weights (which can depreciate due to distillation), RL data is almost impossible to replicate and accumulates with cross-generation compounding. It can't be directly monetized, but it's like a piece of land that appreciates but hasn't been developed yet.

Pre-deployment integration ranks third. Instead of selling the model interface to a consulting company and letting it take the productivity surplus, it's better to own the entire service delivery layer end-to-end. Just like Palantir embedding engineers in government agencies instead of selling software licenses. The lab does not charge law firms API fees; instead, the lab becomes the legal research service itself, pricing based on the results delivered rather than the tokens consumed. Switching costs will continuously accumulate as domain data and institutional knowledge build up. This mirrors MTR's shopping centers: monetizing the passenger flow created by the railways instead of increasing ticket prices for passengers.

National dataset data hosting comes fourth. Governments hold a large number of underutilized datasets (patient records, tax filings). A cutting-edge lab designated as the host gains exclusive access to these data to train models and build products. However, this creates a public-private data monopoly that requires strict governance structures: clear usage boundaries, revenues returning to the public, independent oversight, and truly binding accountability mechanisms.

Redefining the question

Labs that survive are not those that make APIs profitable but those that find their “properties above the station” now and start building. APIs are like railways; they will never make enough money. The money lies in the assets that appreciate around the railways.

Policy issues also arise: rather than subsidizing training operations, the government should design institutional mechanisms (deployment rights frameworks, data hosting structures, productivity measurement standards) that allow labs to capture the residual value created by their infrastructure.

Finally, there is an irony. AI policy discussions are dominated by the US-China framework: US free-market labs versus China's state-supported champion enterprises. The most valuable institutional model for reference may be neither; it could be Hong Kong's model: a 45-year-old public-private hybrid that operates commercially, achieving self-financing through institutional design rather than ideology.