An AI company that isn’t making money should take lessons from the Hong Kong MTR.

Author: Michael Wenye Li

Edited by: TechFlow

Deep Tide Guide: An AI lab burned several tens of billions of dollars, but no one can explain when (or whether) the money will be earned back. API pricing falls by about 10x every year, open-source keeps chasing closed-source, and training costs only keep piling up. This article steps out of the tech-industry lens and uses Hong Kong’s MTR 45-year business model to provide an extremely thought-provoking answer: don’t try to make money from ticket sales—own the real estate above the stations.

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

There’s a business that looks like this: in the beginning you pour in tens of billions of capital, with zero revenue. The core service price is set close to marginal cost. You create huge value for users, but builders can’t hold on to even a penny. And you still have to keep investing in the next generation of infrastructure.

This isn’t about an AI lab—it’s about a large rail system.

Many people compare railways to the AI industry, and most people’s conclusion is: general technologies have public-good attributes, and commercial viability depends on government subsidies.

I want to challenge that conclusion, because Hong Kong’s MTR actually solves this problem. It’s one of the very few subway systems worldwide that can sustain itself commercially—an listed company, paying dividends, and not receiving government operating subsidies.

The financial structure is the same

MTR’s core rail business has never been able to finance its own expansion. 2018 was the best year before the pandemic, when the transport business EBIT was HK$2 billion. And from 2024 to 2026, the estimated capital expenditure is HK$87.9 billion, almost entirely for rail. The peak rail profits over three years only cover 8% of the capital expenditure. Ticket revenue has never been enough to build the next line; that was never its design intent.

MTR keeps fares at an affordable level through a government fare adjustment mechanism. You can’t set fares at a level that can recoup construction costs—nobody would be able to afford it, and it would run counter to the purpose of public transportation. Each line might cover its own operating costs, but ticket revenue will never be enough to fund the construction of the next line.

The AI API pricing faces the same mirrored problem. Distillation and open-source alternatives make API prices fall at a rate of roughly 10x per year. Any lab whose pricing is above marginal cost will lose volume to competitors. Each model can achieve operating profitability at the inference layer, but margins can never cover the next round of training expenses.

The global standard solution is subsidies. London’s Underground runs on TfL funding. China’s high-speed rail carries debts of more than a trillion, and 94% of the lines aren’t profitable. AI is walking the same road: the CHIPS Act, the Stargate program, sovereign wealth fund investments, Pentagon contracts. The default endgame is relying on subsidies for quasi-public infrastructure.

MTR found another way.

Rail + real estate

When MTR was built in 1979, the designers understood from the start that fares would never recoup construction costs. So they structured the company around a completely different premise: railways make surrounding land appreciate in value, so they must hold the land.

MTR develops residential buildings, office buildings, and shopping malls above and around stations, capturing the value uplift created by its own infrastructure. Property profits then feed back into rail operations and fund the next line. Today, MTR owns 13 shopping malls and manages 47 station-over property development projects, with property contributing the bulk of actual profits.

The logic is straightforward: don’t try to capture value from the rail service itself—own the assets that increase in value because of the railway.

The AI analogy

“When will AI labs make money?” is isomorphic to “When can railways sustain themselves through fares?” The answer is the same: they can’t, and the question is wrong to begin with.

A biotech startup uses advanced models to screen drug compounds, saving two years of clinical trial time. A logistics company uses it to optimize routes, saving $40 million in fuel costs. An independent developer delivered, over a weekend, a project that a five-person team would have taken three months to complete. In each case, the model provider captures only a fraction of a percent of the value via API fees. The provider can’t raise prices because there are four other labs and more than a dozen open-source alternatives offering roughly comparable capabilities. The remaining value flows to users and the broader economy.

That’s how general technology works. Steam engines, electricity, and TCP/IP have never contributed meaningful revenue to their creators.

MTR’s takeaway: stop trying to make fares cover construction costs—go find your “station-over property.”

Four candidate approaches, ranked by defensiveness

First is deployment rights granted by the government. The government authorizes a lab to have exclusive access to national medical records, tax systems, or defense logistics. The domain data the lab accumulates, the depth of system integration, and regulatory qualifications take years to replicate. This is MTR’s mechanism itself: the state grants development rights based on natural monopoly attributes.

Second is reinforcement learning reward data accumulated via pre-training interactions. Billions of interaction signals are used to train the next generation of models. Unlike model weights (which depreciate through distillation), RL data is almost impossible to copy, and it compounds across generations. It can’t be monetized directly, but it’s like a piece of land that appreciates in value—yet to be developed.

Third is upfront deployment-style integration. Instead of selling model interfaces to a consulting firm and letting it capture the productivity surplus, own the entire service delivery layer end-to-end. Like Palantir embedding engineers into government institutions rather than selling software licenses. The lab doesn’t charge API fees to law firms; the lab becomes legal research services itself, pricing based on delivered outcomes rather than consumed tokens. Switching costs stack up continuously as domain data and institutional knowledge accumulate. This is MTR’s shopping malls: monetize the passenger flow generated by rail, rather than raising fares for riders.

Fourth is national data set data hosting. Governments around the world hold large quantities of underutilized datasets (patient records, tax filings). A frontier lab designated as the hosting party gains exclusive access, trains models based on these data, and builds products. But this creates a public-private data monopoly, which requires strict governance architecture: clear usage boundaries, public return of benefits, independent oversight, and truly binding accountability mechanisms.

Redefining the question

Labs that can survive are not those that make APIs profitable; they are those that have already found their “station-over property,” and started building now. APIs are like railways: they will never make enough money. The money is in the assets around the railways that appreciate.

Problems at the policy level follow as well: rather than subsidizing training operations, governments should design institutional mechanisms (deployment-right frameworks, data hosting structures, productivity measurement standards) that enable labs to capture the value surplus created by their own infrastructure.

Finally, there’s a bit of irony. AI policy discussions are dominated by frameworks from China and the US: US free-market labs versus China’s state-supported champion companies. The most referable institutional model may be neither. It might be Hong Kong’s model: a 45-year hybrid public-private structure, commercialized operations, and self-financing achieved through institutional design rather than ideology.