TSMC buys $500 million of Anthropic and OpenAI stock

Two $500 million checks, both signed at the same board meeting, both for equity rather than capacity. The amounts are almost irrelevant. The direction of the money is not.

TSMC's board approved up to $500 million of Anthropic stock and the same again in OpenAI on Tuesday, buying equity in the two AI labs whose accelerators it fabricates — a $1 billion loop that the foundry's own filing records in a few lines and explains not at all.

The terms sit in TSMC's Form 6-K. The Anthropic leg buys 6,756,756 non-voting Class N shares at the IPO's $74 midpoint through a subsidiary, TSMC Partners. It is not a standalone deal: it sits inside the $26 billion concurrent private placement that also carries Google at about $10 billion, Nvidia at about $10 billion, AMD at roughly $2.5 billion, and Broadcom, Micron and Microsoft at about $1 billion each, according to Anthropic's amended S-1. TSMC is the smallest of the seven participants. The shares carry no underwriter lock-up, are restricted under Rule 144, and convert to Class A only when transferred to an unaffiliated third party. Closing waits on the IPO, which is expected to price on November 18 and trade from November 19 under the ticker ANT.

The OpenAI leg is a different animal. It buys preferred stock at roughly $992 a share, up to 504,000 shares, which TSMC's own disclosure puts at no more than 0.04% of the company. Reuters reported the plan on Monday as a bet on long-term AI demand, noting the chipmaker would benefit "not only from supplying AI infrastructure, but also from owning stakes in some of the industry's most influential developers" — that is the wire's framing, not a TSMC statement.

What the filing does not contain is a single sentence of rationale. There is no C.C. Wei quote, no strategic explanation, no mention of securing capacity or of reading demand signals. The board also approved $40.93 billion in capital appropriations and a NT$9 dividend at the same meeting, which is worth keeping straight because the numbers get conflated in the coverage: the two stakes are a rounding error against the capex, and the capex is the actual commitment.

Put the size in perspective. One billion dollars combined is about 2.4% of that capex line, and the $500 million Anthropic leg is roughly 0.04% of a company valued at up to $1.405 trillion on a fully diluted basis. TSMC is not buying influence with this money. It is buying a seat, and a small one.

The structure is what deserves attention. A foundry is taking equity in the customers of its customers — the labs that rent time on Nvidia GPUs and Google TPUs, which TSMC builds. That is not the usual circular-financing complaint, where a vendor lends money to its own buyers so they can buy more of its product. TSMC is not financing purchases; it is buying shares in companies whose growth it already profits from at the wafer level. The argument that this closes a loop is still fair, because the loop is real in the other direction: the manufacturer whose capacity limits Anthropic's growth is now one of its shareholders. Tech Times, reading the S-1, drew out the awkward symmetry — the filing lists foundry capacity shortfalls among Anthropic's supply risks, and the filing is silent on whether the shareholder gets priority on allocation.

That is the question the deal does not answer, and it is the one that matters more than the dollars. Taiwan's own coverage is blunter than the US wires: Economic Daily's headline called it TSMC joining "AI closed-loop investment," and quoted institutional analysts warning that tying the ecosystem together with equity helps until the industry changes, at which point it adds risk. MRB Partners strategist Phillip Colmar told Business Insider Taiwan that if AI commercialization or borrowing costs disappoint, the revenue everyone in this ecosystem has promised each other may not arrive.

My read: this is a customer-relations instrument dressed as an investment thesis. Five hundred million dollars does not buy TSMC anything it could not have got with a phone call, but it does buy a quarterly flow of information about the two most important demand signals in the industry, and it aligns TSMC with the labs whose ordering patterns decide its advanced-node mix. The capacity-allocation silence is the part to watch — if Anthropic's next gigawatt turns out to be easier to book than a rival's, the small stake will have bought something large. We covered the same mechanism in September, when Nvidia's $10 billion anchor closed the loop on circular AI money — the difference being that Nvidia's check was big enough to move the valuation. TSMC's is not.

Google DeepMind released AlphaGenome 2 on Wednesday, a version of its DNA model that predicts how a genetic variant behaves in one specific cell type rather than averaging across a whole tissue — and on the benchmark it was built for, that change roughly doubles the share of causal variants it can recover.

The numbers from DeepMind's launch post: 107 million single-cell expression profiles across 68 tissues trained the conditioning model, which covers 637 human and 311 mouse cell types and still reads a million-letter DNA sequence as input. The architectural change is that the cell type is fed in at every stage — encoder, transformer and decoder — and DeepMind's 78-page preprint reports that conditioning only at the output layers produced minimal gains. There is also a single-cell encoder that lets a lab condition the frozen model on a few dozen of its own cells without retraining, which is the part working biologists will notice first.

On TenK10K, a single-cell eQTL dataset spanning 28 immune cell types, recall of causal variants at 90% precision rises from 16% to 30%, and effect direction from 23% to 43%, against a baseline of AlphaGenome 1 fine-tuned on single-cell data. On bulk GTEx eQTLs across 49 tissues, causal-variant recall goes from 17% to 26%. Nature's news piece leads with the case that made the point: a team that spent eighteen months of experiments confirming an ETS2 regulatory link in inflammatory bowel disease watched the model reach the same answer in about thirty minutes. The same researcher, James Lee at the Crick Institute, had been one of AlphaGenome 1's sharper critics precisely because it could not resolve cell type specificity.

The caveats are in the paper rather than the blog. AlphaGenome 2 still recovers only a minority of fine-mapped eQTLs and putative causal GWAS variants at high precision, still misses effects outside accessible chromatin and at very distal elements, and still handles post-transcriptional mechanisms poorly. Its condition embeddings are learned from observational data, so they are correlational; on AG1's bulk splicing, chromatin and polyadenylation benchmarks it performs comparably to its predecessor rather than better. And the licensing moved in the opposite direction from the science: free for non-commercial research through an API, commercial access behind a private preview, and no model weights at launch — a reversal from AlphaGenome 1, which shipped weights with its Nature paper. We covered the mapped genome when DeepMind's AlphaGenome Atlas predicted all 9 billion DNA swaps; that atlas stays on the older model, because one forward pass per cell type is too expensive to repeat across the whole genome.

What to watch: TSMC's two filings close after Anthropic's November 18 pricing, and the allocation question will surface in the first supply-crunch quarter. On AlphaGenome, watch for the weights — an API-only release is a different kind of scientific contribution than an open one, and the genomics community will say so.

Is a 0.04% stake worth a shareholder's seat at the table, or is TSMC just buying a better phone call? Write to us via our contact page.