DeepSeek nears $12B funding round

Bloomberg reported on October 6 (picked up by Reuters) that DeepSeek is close to securing at least 80 billion yuan — roughly $12 billion — in its latest funding round, well past the $8 billion target it set in August. Tencent and battery giant CATL are participating, the proceeds are aimed primarily at compute infrastructure, and multiple reports point to a planned IPO in early 2027. In the space of a year, the lab famous for efficient training has moved from “does not need money” to a super-round with its target raised 50% inside two months.

Key points

Amount: at least 80 billion yuan (about $12 billion), beyond the roughly $8 billion target set in August
Investors: Tencent and CATL participating; earlier reporting placed China’s National AI Industry Investment Fund in a prior round
Use of proceeds: primarily compute infrastructure
Timetable: a listing in early 2027, per multiple reports
Caveat: Bloomberg cites people familiar; terms are unsigned and the figure can move

Read it against the Anthropic ledger

Two weeks ago, Anthropic’s leaked filing supplied the US-side sample: $4.6 billion revenue, an $8 billion operating loss, 47% channel concentration, a two-trillion-dollar narrative. DeepSeek’s ~$12 billion round is the other end of the same “raise to buy compute, IPO to refinance” structure: an order of magnitude smaller, backed by industrial capital (Tencent) rather than dollar funds, listing in Hong Kong rather than Nasdaq. The two ledgers together are the first complete picture of the 2026 global AI capital market — one story, two valuation languages, two regulatory environments.

What the investor mix signals

Tencent brings distribution and cloud; CATL is new money from the “AI is a power consumer” narrative — industrial capital arriving with expectations of bound application scenarios, not just financial return. The same week, reports surfaced that Moonshot AI closed a pre-IPO round at a reported $50 billion valuation: the listing race among China’s frontier labs has moved from rumor to price discovery. For DeepSeek itself, the IPO timetable also explains the round’s rhythm — a final private raise before a listing is usually the last rehearsal of pricing and governance.

Discount the numbers until the filing

“Close to securing” and “reportedly” mean no signed terms and a movable figure — the previous round’s $8 billion target also began as a rumor, and the final size could land above or below today’s report. The checkable anchors are two: Bloomberg and Reuters corroborating each other, and the early-2027 IPO timetable; within six months a prospectus will replace every rumor. For the industry, the signal worth registering is that Chinese AI compute investment has now entered a race of the same order as Dell, Apollo and Jera’s $140 billion program — denominated in a different currency, and facing the same memory and power inflation that raises the price of “buying compute” month by month — which may be why the round’s target grew 50% in two months. If the early-2027 window holds, DeepSeek would become the first Chinese frontier lab to carry its unit economics into a public market, and every number in this round — the raise, the investors, the compute bill — will be re-audited in a prospectus.

The efficiency route’s capital paradox

DeepSeek made its name by “doing more with less compute,” and is now raising a fortune to buy compute — contradictory on the surface, identical underneath. When inference rather than training becomes the dominant cost center, efficiency advantages amplify scale pressure: cheaper tokens mean more tokens, and the Jevons paradox is replaying on the inference side. Whoever converts an efficiency edge into an infrastructure edge gets to tell the most complete unit-economics story in the 2027 listing window — that is what the $12 billion is actually buying. The alternative reading — that the round signals inference demand outgrowing even DeepSeek’s efficiency — is the one its competitors will prefer to tell.