Why Is DeepSeek Raising Another $7 Billion Just Months After Its First Round?

Six pain points: what teams actually need to clear after the headlines

The funding story is loud. For engineering and product teams, the decision blockers are more specific:

1. Numbers are not signed: the ~$7B target and ~$70B pre-money are still in talks — do not treat them as closed terms in budgets or vendor diligence packs.
2. Compute is the real bill: analysts estimate that for every 10 billion yuan raised, roughly 7 billion yuan goes to chips, data centers, bandwidth, and liquid cooling — the cadence tracks compute buildout, not valuation slogans.
3. Voting rights are unusual: most Round-1 outside capital entered via an LP controlled by Liang Wenfeng — no vote, five-year lock-up. The sole exception: China's National AI Industry Investment Fund.
4. P/S is extreme: implied ~140–150x versus dealmaker estimates of ~65x for OpenAI and ~21x for Anthropic — an options bet, not a cash-flow multiple.
5. Leaks already stalled the deal: talks paused in late July after closed-door investor remarks circulated online; governance and comms pressure rose with the check size.
6. Retail investors are not in the room: both rounds are institutional. Individuals wait for a STAR Market listing — or focus on model routing and local inference costs they can control today.

Timeline: from "no fundraising" to a $70 billion valuation in four months

Numbers at a glance: Round 1 closed vs Round 2 in talks

Inside the deal: why the math — and the voting rights — don't add up cleanly

1. The real bill is compute, not headlines

Shortly after closing Round 1, DeepSeek said it would double headcount across data-center and AI-agent teams, and Reuters reported it was hiring chip-design engineers for in-house AI inference chips. Industry analysts estimate that for every 10 billion yuan raised, roughly 7 billion yuan goes into compute — chips, data centers, bandwidth, liquid cooling. The fundraising cadence is a race to keep pace with that buildout, not a valuation vanity exercise. Chip context: DeepSeek custom AI chip analysis.

2. Most investors don't get a vote

In Round 1, most outside capital flowed through a limited partnership controlled by Liang Wenfeng — meaning those investors received no voting rights and a five-year lock-up. The one exception: China's National AI Industry Investment Fund, which invested directly with voting rights and no lock-up. The structure keeps Liang's control near 84% and has drawn governance scrutiny — including from Forbes — about founder control versus state influence.

3. A 148x P/S is an options bet, not a cash-flow multiple

At a ~$70 billion pre-money valuation against $400–500 million ARR, implied P/S sits around 140–150x — far above OpenAI's ~65x and Anthropic's ~21x. One dealmaker's assessment, as translated in Chinese coverage: pricing a foundation-model company is fundamentally an options bet, not a cash-flow valuation. Investors are pricing the chance DeepSeek becomes infrastructure-level in China's compute ecosystem and enterprise agent market — not today's revenue run-rate.

Peer comparison: China's foundation-model valuation race

DeepSeek and Moonshot — still private — both carry P/S multiples around 140–150x, well above what listed peers Zhipu and MiniMax trade at. Private-market investors are, for now, paying a steeper premium for the two labs that have not yet faced public-market scrutiny.

Controversy: leaked remarks, missing votes, and bubble warnings

1. Leaked closed-door remarks stalled the deal: the late-July pause was reportedly driven by Liang's frustration that first-round investor-meeting comments spread online — a reminder that a larger investor base makes a low profile harder to keep.
2. Voting-rights structure draws scrutiny: most external investors have no vote and a five-year lock-up; only the National AI Industry Investment Fund gets direct voting rights with no lock-up. Forbes and others have raised governance and state-influence questions; DeepSeek has not publicly settled them.
3. Valuation-to-revenue gap remains untested: 140–150x P/S is extreme even versus high-growth SaaS (often 30–50x). Whether it holds depends on converting technical lead into scaled enterprise revenue after a STAR Market listing — still unproven by public markets.

Why it matters: STAR Market rules, compute self-reliance, and the global AI race

• STAR Market rules loosened for AI: on June 17, 2026, the Shanghai Stock Exchange said at the Lujiazui Forum that the STAR Market's fifth listing standard would cover AI — no requirement for profit or large revenue if technology is strong enough. That backdrop makes a late-2026 filing and 2027 listing target plausible.
• From "three no's" to capital center stage: DeepSeek spent years funded only by Liang's quant fund High-Flyer under a no-fundraising / no-IPO / no-commercialization stance. Round 1 ended that policy — a sector marker alongside Hong Kong listings for Zhipu and MiniMax and Moonshot's rapid raise cadence.
• Global re-pricing of frontier labs: OpenAI was reportedly valued at $300 billion in 2025; Anthropic's valuation reportedly surpassed OpenAI's by June 2026. Paying a premium for a globally competitive Chinese lab is partly a bet on technical pace plus market scale.
• Compute self-reliance is the subtext: reports of in-house inference chips and owned data centers mirror the broader domestic-compute strategy — and explain why modest ARR still requires fast, large raises.

Six-step runbook: how technical teams decide under funding noise

1. Separate "in talks" from "closed": put Round-2 numbers on a rumor board; do not write them into confirmed financing terms or vendor diligence conclusions.
2. Stress-test ARR and P/S: recompute at $400–500M ARR and 140–150x P/S; if you depend on DeepSeek APIs, keep alternate routes (Flash / Pro / competitors).
3. Track compute and chip narratives: large shares of capital go to data centers and custom silicon — inference cost curves can still swing hard. Keep multi-vendor and local-inference plans.
4. Treat governance as diligence: no-vote LPs, five-year lock-ups, and the state-fund exception matter for long-term control and compliance if you are an enterprise buyer.
5. Put the IPO calendar on the risk board: end-2026 filing and 2027 listing are targets, not guarantees; regulation and markets can rewrite the timeline.
6. Budget production agents and local inference separately: funding headlines do not replace 24/7 nodes or high-memory Macs. Local V4 Flash quantized runs: ds4 local vs cloud rental; official Flash review: V4 Flash 0731 benchmarks.

Three hard data points worth quoting

• $70B vs >$350B yuan post-money: Round-2 pre-money is about +43% versus Round 1's post-money (Caijing via dealmakers; not officially confirmed)
• 140–150x P/S: implied at $400–500M ARR; vs OpenAI ~65x and Anthropic ~21x (dealmaker estimates)
• 7 of 10: analyst framing that roughly 7 billion yuan of every 10 billion yuan raised goes to compute (chips, data centers, bandwidth, liquid cooling)

Closing: valuation races make noise — stable compute is the engineering reality

If Round 2 closes, DeepSeek will reset Chinese foundation-model fundraising records and push "options-style" valuations further into the open. For most technical teams, the operable questions remain API cost curves, multi-model routing, and whether agents can run reliably in production.

Beyond the funding feed, three structural bottlenecks still show up:

• Laptop sleep interrupts: lid close or network switches kill long sessions; spent tokens are not refunded;
• High-memory local inference is hard to keep always on: 96GB / 128GB / 256GB purchase TCO is steep; short projects fit elastic rental better;
• No 24/7 control plane: OpenClaw Gateway and multi-model fallback need a dedicated node — not a personal notebook.

If you are wiring DeepSeek APIs or local quantized inference into production agent workflows, MACCOME Mac cloud hosts give real macOS, SSH handoff, and isolated environments suited to always-on agents. Public plans: Mac mini cloud rental rates.

Sources: Caijing (via Sina Finance and Wall Street CN), The Standard HK, Gate News, ChainCatcher, Forbes, SCMP, Caixin Global, Reuters, Bloomberg, CIW on the cap-table structure, DeepSeek API docs / TechCrunch / Hugging Face on V4, and 36Kr / TMTPost peer comparisons. Most figures are anonymous-sourced media reports, not official disclosures — verify before citing as settled fact.