DeepSeek Round 2 restart: 500 млрд ¥ pre-money — cap table, P/S 148×, compute burn
Кому: platform-инженерам и тех, кто маршрутизирует agent pipeline и не смешивает funding headline с signed term sheet. Внутри: таймлайн с caveat на каждую цифру, две data tables (round compare + operating metrics), LP/vote mechanics, peer cap table matrix, controversy flags, 5 шагов изолированной Mac-валидации. Статус Round 2: unconfirmed negotiation — signing target end-Aug 2026.
Содержание
Продукт: V4-Flash-0731 benchmarks. Compute stack: собственный AI chip inference.
4–5 августа 2026 dealers из окружения Caijing сообщили restart Round 2: target ~50 млрд юаней (~$7B) при pre-money ~500 млрд юаней (~$70B) — +43% vs Round 1 post-money (~350 млрд юаней). При close cumulative raise >100 млрд юаней (~$14B) в <5 месяцев. Все Round-2 параметры: negotiation stage, zero official DeepSeek disclosure.
01 · Три failure modes: не путать reported terms с signed close
- Stage mismatch. Amount, valuation, timeline — anonymous dealers via financial media. Internal runbooks должны маркировать «in talks», не «closed». Regulatory filing от DeepSeek на Round 2 — отсутствует.
- P/S arithmetic. ~$70B pre-money / ~$400–500M ARR (API-token heavy, media-sourced) → implied P/S ~140–150×. OpenAI ~65×, Anthropic ~21× по dealer estimates. Это options premium, не DCF на текущий cash flow.
- Capital narrative ≠ runtime SLA. Fundraising cadence tracks data center + chip capex; не гарантирует production-ready agent harness или API stability. Model routing decision — через isolated acceptance, не через valuation headline.
02 · Таймлайн: от High-Flyer-only до $14B cumulative target
- Апрель 2026: registered capital up; Liang Wenfeng direct stake 1% → 34%; controlled entities ~84.29%. Round 1 opens; V4 preview.
- Июнь 2026: Round 1 closed ~50 млрд юаней (~$7.4B), post-money >350 млрд юаней — largest first-round in Chinese AI history. Backers: National AI Industry Investment Fund, Tencent (10B ¥), CATL (5B ¥), JD.com, NetEase, IDG Capital.
- 14–17 июля 2026: STAR IPO prep + Round 2 talks (~$71B / ~480B ¥ pre-money reported); ARR $400–500M first public mention.
- 25–26 июля 2026: talks paused — Bloomberg cites Liang frustration over leaked closed-door investor remarks.
- 4–5 августа 2026: restart per Caijing dealers — still 50B ¥ target, ~500B ¥ pre-money (~$70B); signing end-Aug; both sides want low-profile process.
Caveat на каждую Round-2 цифру: unconfirmed, terms may shift pre-signing.
03 · Таблицы: round compare + operating metrics
| Round 1 (closed) | Round 2 (in talks) | |
|---|---|---|
| Talks | April 2026 | Mid-July, pause, restart Aug 4–5 |
| Close / target | June 2026 | Late August 2026 (planned) |
| Amount | ~50B ¥ (~$7.4B) | Target ~50B ¥ (~$7B) |
| Valuation basis | Post-money >350B ¥ | Pre-money ~500B ¥ (~$70B) |
| Valuation step-up | — | ~+43% vs R1 post-money |
| Cumulative if R2 closes | — | >100B ¥ (~$14B) in <5 months |
| Metric | Value | Provenance |
|---|---|---|
| ARR | ~$400–500M | Mostly API tokens; media, not audited |
| Gross margin | Reported >50% | Unaudited |
| Implied P/S | ~140–150× | vs OpenAI ~65×, Anthropic ~21× |
| MAU | 100M+ (external) | Methodology undisclosed |
Hard numbers · burn & valuation math
- Cumulative target: >100B ¥ (~$14B) raised in <5 months if Round 2 closes at reported terms.
- Step-up: R2 pre-money (~500B ¥) ≈ +43% over R1 post-money (~350B ¥).
- Compute allocation heuristic: ~70% of each 10B ¥ tranche → compute-related spend (dealer estimate).
- P/S band: 140–150× on ~$400–500M ARR — options bet on infrastructure-level outcome, not current revenue multiple.
04 · Deal mechanics: compute bill, LP structure, 148× P/S as option
4.1 Primary burn vector = data center + in-house inference silicon
Post-Round 1: DeepSeek announced 2× headcount in data-center and agent teams. Reuters flagged chip-design hiring for custom inference ASIC path. Fundraising velocity = function of capex runway, not vanity multiple expansion.
4.2 Cap table: LP no-vote + 5yr lock-up (except state fund)
Round 1 structure: most external capital via Liang-controlled LP — zero voting rights, 5-year lock-up. Exception: National AI Industry Investment Fund — direct equity, voting rights, no lock-up. Liang control ~84% — governance scrutiny (Forbes et al.). Round 2 terms on vote/lock — unconfirmed.
4.3 148× P/S = long-dated call on frontier infra
Dealer framing from Chinese coverage: foundation-model valuation = option on regime shift, not trailing revenue multiple. Investors price probability DeepSeek becomes default compute/agent infrastructure layer in China enterprise — payoff asymmetric, downside = full round write-off if technical lead erodes.
05 · Peer matrix: private premium vs listed secondary
| Company | Listing | Valuation / mkt cap | ARR (reported) | Raise pace |
|---|---|---|---|---|
| DeepSeek | Private, STAR IPO prep | ~500B ¥ pre-money (~$70B, in talks) | ~$400–500M | 2 rounds / 4 months, >$14B combined target |
| Moonshot (Kimi) | Private | ~$20B (May 2026); later ~$30B talks | ~$200M | 4 rounds / 6 months, ~$3.9B total |
| Zhipu AI | Listed (HK) | ~350B ¥ mkt cap (May 2026) | Undisclosed | ~8.3B ¥ pre-IPO |
| MiniMax | Listed (HK) | ~210B ¥ mkt cap (May 2026) | Undisclosed | ~11B ¥ pre-IPO |
Private-market bid for two still-private frontier labs > secondary multiples on listed peers — liquidity discount inverted for «winner» narrative.
06 · Controversy flags
- Leaked investor transcript (July) paused talks — larger cap table = harder low-profile process.
- Vote asymmetry: LP no-vote mass + state fund exception — unresolved public governance response from DeepSeek.
- 140–150× P/S vs high-growth SaaS 30–50× — holds only if post-STAR enterprise revenue conversion validates option thesis. Untested.
07 · STAR fifth standard (Jun 17 2026), compute sovereignty, global re-pricing
- STAR Market fifth listing standard expanded to AI Jun 17 2026 — unprofitable/low-revenue AI firms can file on tech strength. Underpins reported DeepSeek filing end-2026 / debut 2027 path.
- End of 5-year «no raise, no IPO, no commercialization» funded solely by Liang's High-Flyer quant fund until Round 1 close June 2026.
- Global backdrop: OpenAI ~$300B (2025); Anthropic reportedly > OpenAI valuation by June 2026. Chinese frontier premiums fit wider frontier re-pricing wave.
- Compute self-reliance subtext: chip + DC buildout behind modest revenue + aggressive raise cadence — see in-house inference chip reporting.
08 · 5 шагов изолированной валидации (Mac platform engineer)
- Provision isolated Mac node — Cursor / Claude Code / OpenClaw only; no personal iCloud, no corp SSO on daily driver.
- DeepSeek API key in node-local env — never sync to production laptop Keychain; document revoke + wipe path.
- Same agent task suite on
deepseek-v4-flashvs Kimi K3 / Qwen3.8-Max API; log latency p95, failure rate, $/task — independent of dealer P/S math. - Hard constraints doc: rate limits, region lock, key rotation schedule, billing alert thresholds — funding headlines don't change these.
- Destroy node post sign-off; output = internal selection memo, not bookmark folder of valuation articles.
09 · FAQ
В: Round 2 закрыт?
О: Нет. Negotiation, target late Aug 2026. Final terms may differ from reports.
В: Зачем снова raise?
О: DC + in-house chip + headcount capex outpacing Round 1 coverage per media.
В: $70B confirmed?
О: Нет. Anonymous Caijing dealers — not official DeepSeek.
В: Больше голосов для инвесторов?
О: Round 1: most LP no vote + 5yr lock; exception National AI Fund.
В: IPO timing?
О: Reported STAR filing end-2026, debut 2027. Private round institutional-only.
10 · Аренда Mac: funding news ≠ production routing plan
Caijing summary на daily driver = Keychain contamination risk (DeepSeek key рядом с prod Anthropic creds), Windows/Linux VPS без native macOS/Xcode sidecar fidelity. Для reproducible agent bake-offs, day billing, real Apple Silicon — isolated bare-metal Mac rental. Тарифы: bare-metal macOS.
11 · Источники
- Caijing reporting, via Sina Finance, Wall Street CN
- The Standard (HK), Gate News, ChainCatcher on restart
- Forbes: «DeepSeek Just Raised $7.4 Billion. Here's The Catch."
- South China Morning Post, Caixin Global, Reuters, Bloomberg
- CIW on Round 1 cap-table / voting structure
- DeepSeek API docs, TechCrunch, Hugging Face DeepSeek-V4
- 36Kr, TMTPost: Moonshot, Zhipu, MiniMax
Данные на 6 августа 2026. Figures predominantly anonymous sources — verify before production decisions.