The Bifurcation: Why AI Capital Is Splitting Into Two Incompatible Stacks
Anthropic's confidential trillion-dollar IPO filing and DeepSeek's $7.4B Tencent-CATL round on the same day aren't competing data points — they're the moment the AI capital market officially forked into two non-fungible systems.
Two financings landed inside the same news cycle, and the temptation is to read them as a single story about "AI valuations going parabolic." That reading is wrong. Anthropic's confidential S-1 at roughly $965B and DeepSeek's $7.4B round led by Tencent and CATL at a $59B mark are not points on the same curve. They are the first clearly visible artifacts of a capital bifurcation: AI funding has split into two incompatible stacks that no longer price the same way, hedge the same risks, or answer to the same regulators.
The Anthropic stack is dollar-denominated, public-market-bound, and underwritten by hyperscaler compute contracts. A trillion-dollar IPO target is only legible if you accept the premise that frontier model revenue compounds against a fixed Nvidia-and-cloud cost base, and that U.S. capital markets will keep treating that compounding as a software-margin story rather than a utility-margin one. The S-1 itself is the bet: that the SEC, public shareholders, and a syndicate of investment banks will collectively agree to price intelligence the way they once priced search.
The DeepSeek stack answers a different question entirely. A six-fold valuation jump in two months, with a battery manufacturer as co-lead, is not a software financing in any traditional sense. CATL is on the cap table because Chinese AI scale is now being underwritten as an industrial-policy asset — energy supply, domestic silicon substitutes, and sovereign inference capacity priced into the same instrument. The $59B mark is a strategic number, not a comparable-multiples number, and trying to translate it into Anthropic's framework produces nonsense in both directions.
What makes this a bifurcation rather than a divergence is that the two stacks are becoming actively non-fungible. An LP allocating to Anthropic at IPO cannot meaningfully hedge with a DeepSeek position; the export-control regime and the U.S. outbound-investment rules ensure the two books cannot be netted. Founders raising at frontier scale increasingly have to pick a stack at the seed stage, because cross-stack cap tables are becoming a regulatory liability rather than a diversification benefit. This is the same fork that happened to semiconductors between 2019 and 2024, compressed into eighteen months.
The second-order effect is on everyone downstream. Application-layer companies that assumed they could shop frontier model APIs against each other on price are about to discover that the two stacks have different cost-of-capital, different latency-to-monetization pressure, and different willingness to subsidize inference. Anthropic-stack pricing will be disciplined by public-market gross-margin expectations the moment the IPO prices. DeepSeek-stack pricing will be disciplined by whatever industrial logic CATL and Tencent are actually optimizing for, which is almost certainly not quarterly software margins. Buyers who modeled a single converging API price curve are modeling a market that no longer exists.
The honest read on June 3 is that the AI industry just stopped having one capital market and started having two, and the two are going to drift further apart, not closer together. The interesting trades for the next twelve months are not "which lab wins" but "which stack does a given customer, regulator, or talent pool actually live inside" — because that answer now determines pricing, hiring, and product roadmap more than any benchmark score. The bifurcation is the story. The headline numbers are just the first time it became impossible to ignore.
DeepSeek $7.4B round (this cycle) → · Anthropic confidential IPO filing (this cycle) →