I’ve been looking back at some of the startup projects I worked on, and at the same time, thinking about what companies are doing through the lens of financial trading. Projects and trades are very different. One deals with users, products, and teams; the other with prices, positions, and risk. But when I review them side by side, I keep coming back to two questions: what is the supply and demand, and when should resources go in or come back out?

I’m not trying to call everything a trade. Startups, career choices, and financial markets have different objects, time horizons, and feedback loops. I’ve just noticed that I tend to ask the same questions in each: is there real supply and demand here? Once I understand that, should I commit resources, how much, and under what conditions should I keep going or pull back?

Supply and demand

When building a product, it is easy to say that users have a need. The harder part is making the supply and demand concrete: who uses it, who pays, who provides the product or service, and how do the two sides meet? A user saying they want something does not mean they will pay for it. Even when someone will pay, that does not mean the provider can keep delivering at a reasonable cost. Market size is only one part of deciding whether a project is worth pursuing. You also need to understand how the exchange happens and who ultimately gets the money.

I first encountered the idea of supply and demand in Wang Huiwen’s Tsinghua Product Course. After working on a few startup projects, I got into the habit of looking at both sides: is there evidence of demand, and can the supply meet it? If it can, who can get the product in front of users, and what will delivery and customer acquisition cost? Focusing only on the product idea can make “I think users need this” sound like proof that a market exists.

Ride-hailing platforms make the two sides easier to see. Passengers need rides; drivers supply vehicles and their time. The platform has to match them by location and time of day. DiDi says in its annual report that its system forecasts changes in supply and demand across regions and time periods, then adjusts driver incentives and dispatch. Matching also takes distance, wait time, and the preferences of both passengers and drivers into account. When there are too few drivers, passengers may wait longer. When there are too few passengers, drivers spend more time without fares. The platform is constantly trying to make the two sides meet.

In 2016, Uber announced that it would combine its China business with DiDi. The company said Uber China was handling more than 150 million trips a month, but both companies had invested billions of dollars in China and were still losing money. DiDi’s registration statement says it acquired Uber China in exchange for shares. Uber China stopped operating independently, and Uber received a stake in DiDi. A large number of trips and heavy investment do not necessarily turn into profit.

The shift from Google Search to AI Mode also shows how demand and supply change together. People have always needed to find information, understand it, and compare their options. Search used to return pages and links for people to sort through. AI Mode now combines multiple searches, synthesized answers, and follow-up questions in one process. In 2026, Google reported that AI Mode had more than one billion monthly active users worldwide and that queries had doubled every quarter since launch (its account of AI Mode usage). These are Google's own figures, and they suggest that people are beginning to use search for longer, more complex questions. How that usage will turn into revenue, whether publishers will lose traffic, and whether the computing costs can be covered are still open questions.

In this AI cycle, I’d separate the model business from the “selling shovels” business. Some companies mainly provide model capabilities; others charge by token or API call; still others provide cloud services, computing capacity, and chips. One company may work across several of these layers. Models get most of the attention, but each call consumes tokens and computing resources. As usage grows, demand for cloud services and chips grows with it. Viewed through supply and demand, selling the shovels can be a very profitable business. NVIDIA’s second-quarter FY2027 results reported 88.3billioninrevenueand88.3 billion in revenue and 62.7 billion in operating income for its Compute and Networking business. When demand surges in one part of the market and supply cannot keep up, suppliers may earn substantial revenue and profit. As supply catches up and alternatives multiply, those margins can change. The next question is where the money is going today, and how long that advantage can last.

Entry and exit

Once the supply and demand are clearer, the next question is when to commit resources and how much to put in. In a startup project, entering might mean hiring a team, spending money, and setting aside time to test the market. Later, you might invest more, change direction, or shut the project down. In financial trading, entry and exit involve decisions about price, position size, and how long to hold. Career choices involve resources too, but their effects may take years, even decades, to become clear. A trade that lasts a few days, a project that runs for years, and a career that spans a lifetime cannot use the same stop-loss or take-profit rules. The time horizon changes how much you can commit and how long you can wait for feedback. In his discussion of investment timing, Dixit writes that under uncertainty, waiting can be valuable because time may bring more information about a project’s prospects. That does not mean waiting is always the right choice (Dixit on investment timing).

Trading is often described as buying low and selling high. In practice, it involves at least two judgments: when to enter and when to exit. You also have to decide how much to put in each time, whether to add to the position, and how much to take out. Short selling requires the same judgments about entry, position size, and exit, just in the opposite direction.

If we reduce those two judgments to coin tosses, each with a 50% chance of being right, and assume the judgments are independent, the chance of getting both right is 25%. If each judgment is right 80% of the time, the chance of getting both right is 64%. This only describes the probability that both judgments match the expected outcome. It is not the market’s actual probability of rising or falling, and it is not the probability of making a profit. Real judgments are not necessarily independent, and they involve more than guessing whether prices will rise or fall. Knight distinguished measurable risk from uncertainty, where the probabilities themselves cannot be reliably estimated. The development of a project, a career, or a new technology cannot necessarily be assigned a 50% or 80% probability.Knight on risk and uncertainty

The win rate alone does not determine whether a trade is profitable. Profit also depends on position size, the potential gain relative to the loss, costs, and discipline when exiting. Even with an 80% accuracy rate, a single large loss or costs that eat away at returns can leave you unprofitable. A lower win rate can still produce a positive expected return if losses stay small and the gains are large enough. Repeated decisions do not guarantee profit. Over time, they need to have positive expected value, with costs and risks kept within what you can bear.

Judgment and decisions

Once I’ve thought through supply and demand, entry, and exit, I still have to ask: what do the facts in front of me support? I try to separate what has already happened, what I infer from it, and what I hope will happen. Whether users have paid, whether suppliers can deliver, and what a price has done are observable facts. Whether demand will grow or a price will keep rising is an inference. Wanting a project to succeed or a market to rise is a desire. A desire can tell me where I want to go. It is not evidence that the market will move in that direction.

Simon wrote that decisions are constrained by information, time, and our ability to calculate. We cannot work out every possibility before acting (Simon on rational choice). A more practical approach, I think, is to write down what supported the decision and what remains unclear, along with what would make me change my mind. When the outcome is known, I can look back and see where my reasoning stopped matching reality.

Before continuing to observe, it helps to know what I am waiting for. Howard’s work on the value of information asks whether new information might change a decision, and whether that change would improve the outcome (Howard on the value of information). If waiting could bring information that changes the choice, it may be worthwhile. If the information will not change what I do, more observation may not improve the decision. The cost of waiting matters too. An opportunity may pass.

“Make a plan, execute the trade, and review what happened.” I now think the point is to settle on a plan before entering, follow it when executing, then review whether the reasoning and actions held up. In a project, you might compare progress with the milestones and decide whether to keep investing, change direction, or leave. In a trade, you can write down the position size, the loss you can bear, and the conditions for exiting. If a review does not change what you do next time, all those notes amount to busywork.

Projects, careers, and trades run on different time horizons and carry different costs. The same standard cannot be copied across all of them. I’ll leave it here for now. These are some scattered thoughts I put down when I had a bit of time.

References

  1. Wang Huiwen, Tsinghua Product Course: course materials (PDF)
  2. DiDi Global Inc., Form 20-F for fiscal year 2024: annual report
  3. Uber, “Uber China Merges with Didi Chuxing,” 2016-08-01: announcement
  4. DiDi Global Inc., Registration Statement on Form F-1/A, 2021: filing
  5. Google, “How AI Mode is changing the way people search in the U.S.,” 2026-05-19: article
  6. NVIDIA, Form 10-Q for the quarter ended 2026-07-26: second-quarter FY2027 results
  7. Avinash K. Dixit, “Investment and Hysteresis,” Journal of Economic Perspectives, 6(1), 1992: article
  8. Herbert A. Simon, “A Behavioral Model of Rational Choice,” The Quarterly Journal of Economics, 69(1), 1955: article
  9. Frank H. Knight, Risk, Uncertainty and Profit, Chapter X, 1921: chapter
  10. Ronald A. Howard, “Information Value Theory,” IEEE Transactions on Systems Science and Cybernetics, 2(1), 1966: article