Taiwan Semiconductor Manufacturing Co.’s board approved on Nov 10 an investment of up to $1 billion in Anthropic and OpenAI, with no more than $500 million going to either company.
The same meeting approved a capital budget of about $40.93 billion for advanced process technology and capacity, set the third-quarter dividend at NT$9 a share, and agreed to establish a company that will invest in the chip ecosystem. The decisions were taken together, and reading them together is more informative than reading any one of them alone.
The investment is small next to the capital budget, roughly one fortieth of it, which is the first thing to understand about the move. TSMC is not buying a stake in order to make money on two startups. It is buying a seat at a table where the design of future data centres is settled, and it is paying a price that is trivial relative to the cost of being absent from that conversation.
The customers are asking for something specific. An executive said at a Nikkei forum that demand from customers for advanced process technology was more than the company could feed. That is a sentence about scarcity, and scarcity is what gives a supplier the ability to choose which demand it serves. Taking equity in two of the largest buyers of leading-edge logic is one way to make that choice easier to defend.
There is a version of this that looks like circular financing, and investors will ask about it. A supplier invests in a customer, the customer spends the money on the supplier’s product, and the supplier records revenue that partly funded the investment. That loop is not new in the industry, and it is not necessarily improper, but it does make the reported growth slightly harder to interpret.
The scale here keeps the loop small. Anthropic and OpenAI buy computing, most of it through cloud providers who buy chips, and $500 million does not move a revenue line for a company spending tens of billions on capacity. What it moves is the relationship. A supplier with equity in a customer gets a call before a design decision, and design decisions at the leading edge are worth more than the equity.
The dividend and the capital budget say something about the internal view. A company that raises its capacity spending by that much and still pays NT$9 a quarter is telling shareholders that the cash generation supports both. That combination has been the core of the investment case for two years, and the board is choosing to keep it rather than to conserve.
The ecosystem investment company is the part that will take longer to understand, because the announcement did not say what it will buy. The plausible reading is packaging, substrates, memory interfaces and the other links that have become the binding constraint on how many accelerators can be built. TSMC’s own advanced packaging has been the tightest part of the chain, and a vehicle that invests in the suppliers around it is a way to add capacity without adding it inside the company.
For Anthropic and OpenAI, a strategic investor of this kind is worth more than the money. Both companies are competing for allocation of leading-edge capacity against buyers with longer relationships and larger commitments, and both have argued that their growth depends on securing it. A foundry on the share register is an argument in a negotiation that neither can otherwise make.
There is also a defensive logic that neither company will state. Advanced chips are the subject of export controls, and the largest buyers of them are increasingly the subject of policy attention in Washington. A supplier with equity in its customers has an interest in the customers remaining in good standing, which is a form of insurance that does not appear on a balance sheet.
The risks run in the other direction too. A foundry that owns part of two model developers has a conflict when a third asks for capacity, and a startup that is not on the list will notice. Neutrality has been one of TSMC’s most valuable assets, since it manufactures for everyone who competes with everyone else. Equity stakes are cheap. Neutrality, once spent, is not easily bought back.
None of this is announced as a strategy. The board approved a number, a budget and a dividend on the same day, and the company framed the investment as a limit rather than as a target. Two stakes of up to a half billion dollars each, in companies whose valuations are measured in the hundreds of billions, is a small position in financial terms.
The larger signal is who is buying whom. For thirty years the capital flowed from the software companies to the factories, and the factories competed for the business. The direction has partly reversed. The factory now holds shares in the customers, the customers depend on the factory for the scarcest input in the industry, and the dividend is set by a board that can see both sides of the ledger.
What that arrangement produces over the next two years will be decided in allocation meetings rather than in press releases. The stake buys a voice. Whether it buys capacity, and at what price, is the question the customers will be asking, and the one the company has not had to answer yet.
