DeepSeek

Phase A — Understand the business

Lens 1 · Company Overview

DeepSeek (Hangzhou DeepSeek Artificial Intelligence Co.) is a frontier AI lab founded July 2023 by Liang Wenfeng, spun out of and bankrolled by his quant hedge fund High-Flyer. The business model is unusual and is the single most important thing to understand about the name: DeepSeek gives the product away. It open-weights its frontier models under the maximally permissive MIT license (V4-Pro, V4-Flash, V3.2, R1 all downloadable on Hugging Face, commercial use / fine-tuning / distillation permitted with no royalty or revenue share) and runs a free consumer chatbot.

Revenue, such as it is, comes from (1) a usage-metered API (where most commercial demand lands) and (2) enterprise relationships. Open weights are explicitly "a lever, not the business" — the strategic bet is that broad adoption beats licensing revenue. There is no disclosed ARR; on public evidence the API is a thin, low-margin commodity line priced near cost (see Lens 5/7).

• Products: the DeepSeek-V series (general/MoE foundation models — latest V4-Pro 1.6T-param MoE, 49B active, and V4-Flash 284B / 13B active, both 1M-token context, released April 24, 2026) and the DeepSeek-R reasoning series (R1 shipped Jan 2025; R2 still unreleased as of June 2026 — Reuters reported Liang was "not satisfied with its performance").
• Customers: developers self-hosting the open weights; API consumers; and, increasingly, Chinese cloud giants (ByteDance, Tencent, Alibaba) deploying V4 on domestic silicon, plus Southeast-Asian sovereign/enterprise adopters (Singapore's OCBC runs 30+ internal tools on DeepSeek+Qwen; Indonesia's Indosat; Malaysia's sovereign-AI stack).
• Suppliers: compute. Historically NVIDIA (H800/H100, plus a pre-control A100 stockpile — see Lens 2); increasingly Huawei Ascend silicon, on which DeepSeek now runs its own infrastructure.
• Competitors: the global frontier (OpenAI, Anthropic, Google DeepMind, xAI, Meta) and the Chinese cohort (Alibaba's Qwen, Moonshot/Kimi, MiniMax).
• Contract structure: API is pure pay-as-you-go, no take-or-pay, no lock-in (MIT weights mean a customer can leave for self-hosting at zero switching cost — a moat problem, not a feature; see Lens 3).

Lens 2 · Supply Chain

The AI stack is compute → model → product. DeepSeek sits at the frontier-labs node, and its supply chain is defined almost entirely by one variable: access to advanced accelerators under US export controls.

Upstream → DeepSeek → end use, named:

• Accelerators (the chokepoint). Original training ran on NVIDIA H800 (the export-throttled China SKU). SemiAnalysis reported DeepSeek's actual fleet is far larger than the headline run implies — on the order of ~50,000 Hopper GPUs (≈10k H800 + ≈10k H100) plus a pre-control A100 stockpile (10k–50k). The strategic pivot of 2026: DeepSeek now runs on Huawei Ascend (the 950-series), and V4 was optimized for Ascend 950 — the first time a frontier-class model validated Huawei silicon at scale.
• Foundry / fabrication. Huawei Ascend is fabbed domestically (SMIC), and that node is itself supply-constrained: US controls on advanced chipmaking equipment mean China "can barely keep up." Huawei targets ~750k units of the 950PR in 2026, mass production from April, full shipments H2 2026.
• DeepSeek (the lab) — trains and serves the models.
• Downstream buyers / distribution. Chinese hyperscalers (ByteDance, Tencent, Alibaba) — whose V4-driven demand reopened Huawei orders "within days" of launch; Hugging Face (open-weight distribution); the DeepSeek API and app; SE-Asian sovereign deployments.

Chokepoints / single-source dependencies:

The second-order effect the KB flagged as a gap — "export-control 2nd order effects on Chinese labs" — is now the whole story: controls intended to slow China's frontier instead forced DeepSeek onto Huawei and collapsed NVIDIA's China share to ~0%.

Lens 3 · Competitive Advantages (moats)

DeepSeek's edge is real on engineering, thin on economics.

What's genuinely durable:

• Compute-efficiency IP. The core competence — born of High-Flyer's quant DNA and sharpened by having to do more with throttled chips — is squeezing frontier capability out of less compute (MoE architecture, training/inference optimization). Liang's own framing is "computing-power arbitrage through algorithmic innovation". This is the moat that produced the cost shock (Lens 8).
• Mindshare / brand. R1 made DeepSeek the most famous AI name in the world for a week (Lens 8). In open-weight land, that gravitational pull drives downloads, fine-tunes, and ecosystem — the "TikTok of LLMs" framing.
• Sovereign-stack lock-in (emerging). By co-optimizing with Huawei Ascend, DeepSeek is becoming the default model of the non-NVIDIA, non-US AI stack — China + sympathetic sovereigns. That is a structural, policy-reinforced moat no Western lab can contest.

Where the moat is weak — and this is the crux of the bear case:

• MIT license = near-zero switching cost. Anyone can take the weights and self-host. DeepSeek captures none of that usage. Its own framing concedes weights are "a lever, not the business" — but the API it does monetize is a commodity racing to the price floor (Lens 7).
• Bargaining power is inverted. It needs compute (NVIDIA/Huawei) far more than they need it; and its customers can walk for free. The only counterparty that needs it is the Chinese state, which now controls it.
• The 3-6 month lag is structural, not closing. DeepSeek itself claims V4-Pro "trails state-of-the-art closed models by only 3 to 6 months". As a fast-follower, it is a price-destroyer for the frontier, not a frontier-setter.

Lens 4 · Segments

segments.csv is header-only — DeepSeek discloses no segment revenue, geography split, or product mix. No `` segment data exists. On a qualitative web basis the economic activity splits into: (a) open-weight distribution (zero direct revenue, maximal reach), (b) API usage (the monetized sliver, commodity-priced), and (c) strategic/sovereign deployments on Huawei hardware (value accrues to Huawei and the state more than to DeepSeek's P&L). Geographically the centre of gravity is China + Southeast Asia, structurally walled off from the US/allied market by bans (Lens 10). Revenue by segment: n/a — private, not disclosed.

---

Phase B — Measure performance

Lens 5 · Funding & Valuation Trajectory (+private overlay — replaces "Earnings Result")

There is no earnings print. The performance signal for a private lab is the funding/valuation trajectory, and DeepSeek's is singular.

• Self-funded era (2023 → early 2025): built entirely on High-Flyer capital and Liang's own resources; took no VC money through early 2025 — deliberately.
• Maiden external round (June 2026): >¥50bn (~$7.4B) raised at a valuation >$50B.
• The structure is the story. Capital went not into DeepSeek directly but into a limited partnership managed by Liang, stripping outside investors of governance and locking capital for five years. The single exception is China's National AI Industry Investment Fund, which invested directly and retained voting rights and no lock-up. Backers: Liang himself ¥20bn ($3B, the largest single backer), Tencent ¥10bn ($1.4B), CATL ~¥5bn.
• Read-through: investors get economic exposure with no control and the state holds the only votes — i.e. DeepSeek is, in governance terms, state-steered. OpenAI has publicly branded it "state-controlled".

Burn / unit economics: undisclosed. The famous training-cost figures are: V3 final run ~$5.576M (512 H800 × theoretical $2/GPU-hr) and, in a rare 2026 disclosure, R1 training ~$294K. Both are final-run compute only — they exclude prior research, ablations, the ~50k-GPU fleet, and the A100 stockpile. Treat them as marketing-grade lower bounds, not true cost of development.

Lens 6 · Founder Signal & Sentiment Trend (+private overlay — replaces "Earnings Calls")

No earnings calls. The substitute is founder communication + ecosystem sentiment. Liang is famously low-profile; the recurring themes across his rare interviews and the secondary literature are consistent and worth tracking as a "tone" baseline:

• AGI as the goal, framed as "curiosity-driven exploration, academic research" rather than profit-seeking.
• China must be a "contributor, not just a beneficiary" of global tech — a nationalist-research framing that aligns neatly with the state's interest (and helps explain the funding structure).
• "Arbitrage philosophy" carried from quant trading into compute — find the inefficiency, exploit it algorithmically.

Sentiment trajectory of the name (the closest analogue to call-tone): peak-euphoria Jan-Feb 2025 (the R1 shock), maturing through 2025 into "credible, cheap, fast-follower" status, and by 2026 hardening into a geopolitically polarized read — celebrated in China/SE-Asia as the sovereign-stack champion, treated in the US as a security threat and IP-laundering suspect (Lens 10). The free-and-open posture that won goodwill in early 2025 is now read in the West through a national-security lens.

Lens 7 · Cap Table & Secondary Marks (+private overlay — replaces "Comps")

Syndicate quality: this is not a crossover-fund / Fidelity-T.Rowe-Coatue IPO-proximity cap table. It is a domestic-strategic + state syndicate — founder, Tencent, CATL, and the National AI Industry Investment Fund as the sole voting holder. The presence of a state fund with exclusive votes is the opposite of an IPO-readiness tell; it signals the company is being positioned as national strategic infrastructure, not a float candidate.

Valuation comps (private peers, all ``, unaudited):

The single most important comp fact: DeepSeek is valued at ~5-6% of OpenAI/Anthropic despite shipping a model those labs concede is only 3-6 months behind, and despite resetting global API pricing. The market is pricing DeepSeek not on capability but on (a) its inability to monetize (open weights + commodity API) and (b) its un-investability to Western capital (state control, bans). The gulf is "China's AI pragmatism vs Silicon Valley's valuation machine".

API price marks (the closest thing to a "secondary mark" on the product): V4-Flash $0.14 in / $0.28 out per 1M tokens; R1 $0.55 in / $2.19 out; V3.2 $0.028 in (cache hit) / $0.28 (cache miss) / $0.42 out. Versus GPT-5.5 ($5 in / $30 out) and Claude Opus 4.7 ($5 in / $25 out), DeepSeek is ~11-36x cheaper on input and 35-100x cheaper on output.

Lens 8 · Stock-Price Catalysts (adapted — funding/product events that moved the market)

DeepSeek has no stock, but it has moved other people's stocks more than almost any private company in history. The pattern reveals what the market reacts to:

• Jan 27, 2025 — the R1 shock. R1's release (frontier-comparable reasoning, "trained for a fraction of the cost") triggered an AI-complex selloff: NVIDIA −17% in a day, ~$600B market-cap erased — a record single-day loss. Investors feared a collapse in data-center chip demand.
• The Jevons-paradox counter-narrative (Feb 2025): Nadella and others argued cheaper intelligence increases total compute demand; SemiAnalysis noted DeepSeek's efficiency had already induced demand with "tangible effects to H100/H200 pricing." The market re-rated within weeks — Jevons won the framing.
• April 24, 2026 — V4 launch → Huawei demand spike. V4 (Ascend-optimized) reopened ByteDance/Tencent/Alibaba chip orders to Huawei "within days," and is a key driver of Huawei's guide to ~$12B AI-chip revenue in 2026 (+60% YoY) as NVIDIA's China share went to ~0%.
• June 2026 — the $7.4B / $50B round. Validated DeepSeek as a national champion and crystallized the state-control governance read.

What the market actually reacts to for this name: (1) cost/efficiency surprises (the entire thesis), (2) which silicon it endorses (it now moves Huawei's order book), and (3) the geopolitical frame (state control, bans). It does not react to revenue — there is none to react to.

---

Phase C — Judge people & books

Lens 9 · Management

Liang Wenfeng (founder & CEO). Born 1985, Wuchuan, Guangdong; BEng/MEng from Zhejiang University. The archetype is unambiguously founder-operator with a quant-research soul, not a professional manager.

1. Track record (quantified): co-founded High-Flyer (2013, age ~28-30); built it into one of China's largest quant funds, >$100B AUM by 2021. Critically, he used the fund's profits to build the GPU cluster before there was an AI company — a contrarian capital-allocation bet that became DeepSeek's founding asset.
2. Tenure & skin in the game: founder since inception (2023). In the 2026 round he personally put in ~$3B — the single largest check — and the LP structure routes all outside capital through a vehicle he manages. Skin in the game and control are both maximal.
3. Capital-allocation history: exceptional and idiosyncratic — reinvested trading profits into research compute ahead of demand; gives the product away to win mindshare; deliberately refused VC for ~18 months to preserve mission and independence. This is a builder optimizing for capability and influence, not near-term return.
4. Red flags: the dominant governance flag is state alignment — the National AI Fund's exclusive voting rights make the state, not minority investors, the controlling external voice; OpenAI's "state-controlled" label is a real (if self-interested) characterization. Plus the unresolved IP-provenance question (distillation — Lens 10). No evidence of self-dealing or promotional accounting, but the absence of any audited disclosure is itself the risk.
5. Founder vs professional manager: pure founder-researcher. At this stage that is a strength (vision, speed, willingness to make the contrarian compute bet) — but it concentrates key-person and governance risk to an extreme degree.

Lens 10 · Forensic Red Flags + Regulatory Findings

There is no audited financial statement to forensically analyze — which is itself the headline forensic flag: every cost and capability claim is self-reported and, where checkable, understated (the $5.6M/$294K training figures vs the ~50k-GPU reality — Lens 5). A skeptic should assume the public cost narrative is engineered for maximum shock value.

SEC enforcement: none possible — DeepSeek has no CIK and is not an SEC registrant. No EDGAR LR or AAER search applies.

Non-SEC / regulatory & legal findings (web, per Stage-1 guidance):

• IP / distillation — the central allegation. OpenAI and Microsoft have investigated whether DeepSeek distilled outputs from ChatGPT to train R1/V3; in Feb 2026 OpenAI warned Congress DeepSeek is "illegally distilling" US models, alleging employees used third-party routers and ~automated access to conceal origin. Anthropic separately reported "industrial-scale" distillation campaigns by three Chinese labs — DeepSeek, Moonshot, MiniMax — generating >16M exchanges via ~24,000 fake accounts against Claude. Unproven in court, but credible and corroborated across multiple US labs — a genuine legal/reputational overhang and the bear case's sharpest edge.
• Government bans / national-security actions. Banned or restricted on government devices by Italy, Australia, Taiwan, South Korea, India, the Czech Republic, and ≥17 US states; US federal "No DeepSeek on Government Devices Act" introduced; NASA and New York State barred employee use. This structurally caps the addressable Western market.
• Security posture. A Cisco study found DeepSeek failed to block a single harmful prompt in testing (vs GPT-4o ~86%, Gemini ~64%). Open weights let users strip safety mechanisms — a real exploitation surface and a regulatory target.
• Censorship / data: documented topic censorship and data-handling concerns cited by multiple governments as ban rationale.

Verdict (Lens 10): No securities-fraud findings are possible for a non-filer, but the forensic picture is opaque-by-design financials + a credible, multi-source IP-provenance allegation + a broad and widening ban map. For a private name these are the material findings; all ``, unaudited.

---

Phase D — Project & stress-test

Lens 11 · IPO-Readiness & Path-to-Tradeable (+private overlay — replaces "Forward Projection")

No private-watch.json entry exists in this research root, so this is reconstructed web-only. The honest assessment: DeepSeek is structurally far from a tradeable security, and the 2026 round pushed it further away, not closer.

• Stage: first-ever external round just closed (Jun 2026) at >$50B.
• IPO readiness: low / actively disfavored. The deal was engineered to avoid the normal investor-governance path — 5-year lock-ups, an LP vehicle controlled by the founder, and state-only voting rights. That is the architecture of a national strategic asset, not a pre-IPO float. A US listing is essentially foreclosed by the bans and the state-control characterization; even a domestic (HK/STAR) listing is not signaled.
• Catalysts that would unlock tradeability: (1) a domestic China listing decision (no public signal); (2) resolution of the distillation litigation overhang; (3) a credible recurring-revenue story to justify the valuation. None are near.
• What to actually track instead (since the equity isn't accessible): DeepSeek is best treated as a read-through indicator, not a position — its model cadence moves NVIDIA (negatively, on efficiency fears) and Huawei / the China semi complex (positively, on endorsement), and it sets the price floor for the entire LLM API market, pressuring the gross margins of every closed-source lab.

No Brier forecast logged (breadth/watchlist loop; and the binary that matters — a DeepSeek listing — has no near-term resolution date).

Lens 12 · Bull vs Bear

Bull case. DeepSeek is the most capital-efficient frontier lab on earth and the anchor tenant of the entire non-US AI stack. It ships open-weight models 3-6 months behind the closed leaders at ~1/50th the cost, has the mindshare to dominate the global open-weight ecosystem, and — by validating Huawei Ascend at frontier scale — has made itself indispensable to China's ~$12B-and-growing sovereign-compute buildout and to SE-Asian sovereigns who want AI without US dependency. With the state as patron, it has effectively unlimited runway and political air cover. If "good-enough, open, and 50x cheaper" is where the bulk of global inference demand actually lands, DeepSeek wins the volume war even while the West wins the valuation war.

Bear case (risks that could permanently impair).

1. It cannot monetize what it creates. MIT weights → zero capture of self-hosted usage; the API is a commodity racing to ~$0. A $50B valuation on a near-zero-margin, give-it-away model is the definition of value created ≠ value captured. The bull case is an influence thesis, not a return thesis.
2. The IP-provenance overhang. If the distillation allegations harden into enforcement/legal findings (US labs are coordinating the case), DeepSeek's models — and the credibility of its cost narrative — are tainted, and Western enterprise adoption (already capped by bans) goes to zero.
3. Permanent fast-follower / commodity status. It sets the price floor but not the frontier. If frontier capability (agents, multimodal, reasoning depth) keeps mattering more than price, DeepSeek is structurally a margin-destroyer for others rather than a value-compounder for itself.

Pre-mortem (it's late 2027, the thesis broke). Most likely: a US/allied legal finding on distillation + a tightening of the open-weight export/usage regime walls DeepSeek almost entirely out of the Western and SE-Asian enterprise market; meanwhile the closed frontier's agentic capability pulls far enough ahead that "50x cheaper but 9-12 months behind" stops being good enough, and the API price floor stops mattering because the work has moved up-stack. DeepSeek remains a celebrated domestic champion but never becomes an investable, value-capturing business.

Are the marks too high? At ~$50B, relative to OpenAI/Anthropic at ~$850-965B, DeepSeek looks almost cheap on a capability-per-dollar basis — but that discount is correct: it prices the monetization gap and the un-investability. The valuation is rational, not a bargain.

Contrarian view (what the market refuses to see). The consensus treats DeepSeek as "China's OpenAI." It isn't — it's China's AWS-of-models-as-public-good: a state-subsidized commoditizer whose output (cheap, open intelligence) is the product and whose job is to deny the West pricing power and deny NVIDIA the China market. Viewed that way, DeepSeek is already succeeding at its actual mandate — which is precisely why it will never be a great equity.

Lens 13 · Devil's Advocate (short-seller)

If this were shortable, here's the dismantling. The whole edifice rests on a cost claim that is, on the lab's own architecture, understated by one to two orders of magnitude — $294K/$5.6M "training cost" against a ~50k-GPU fleet and a pre-control A100 hoard. Strip the marketing and DeepSeek is a well-funded fast-follower whose moat is a license that guarantees it captures none of its own value and whose largest "customer" relationship is a state that took the only votes. Revenue is concentrated in a commodity API priced at the floor, so any margin is illusory; the moment a rival (Qwen, Moonshot, Llama-successor) ships an equally-good open model — and they are right behind — DeepSeek's mindshare premium evaporates because there is nothing to switch away from. The most dangerous competitor bulls underrate is Alibaba's Qwen, which has comparable openness plus a real cloud business to monetize it through. The capital-allocation "red flag" isn't fraud — it's that the founder structured the round to entrench control and please the state, aligning the company with a sovereign agenda over commercial value. For today's $50B to make sense, you must believe DeepSeek converts open-weight mindshare into durable, high-margin revenue — and there is no evidence and no mechanism for that. Knock 20-30% off the (already nonexistent) growth-in-monetization and the value isn't impaired by 20-30%; it was never financial to begin with. The single scenario that permanently impairs it: a distillation finding that simultaneously taints the models' provenance and validates every Western ban — plausibility moderate and rising, given three US labs are now publicly building that case.

Lens 14 · Management Questions (ordered by information value)

1. Open weights mean you capture none of self-hosted usage and your API is priced at the floor — what is the actual path to durable, high-margin revenue, and what is current ARR and gross margin on the API line?
2. The National AI Industry Investment Fund holds the only voting rights. In concrete terms, what decisions does the state control, and how do you reconcile that with calling DeepSeek "curiosity-driven academic research"?
3. US labs allege industrial-scale distillation of their models (OpenAI to Congress; Anthropic's 16M-exchange claim). What is your provenance trail for R1/V3/V4 training data, and would you submit to independent audit?
4. SemiAnalysis puts your real fleet at ~50k Hopper GPUs plus an A100 stockpile. What was the true all-in development cost of V4, including research, ablations, and amortized infrastructure — not the final-run number?
5. You've made Huawei Ascend your training/inference base. What are the real performance and yield gaps vs NVIDIA Hopper/Blackwell, and how exposed are you to domestic fab-capacity constraints?
6. With ≥17 US states, the US federal government, Italy, Australia, Taiwan, South Korea, India and others banning you, what is your realistic addressable market outside China + sympathetic sovereigns?
7. Why MIT license rather than a source-available/commercial-use-restricted license that would let you monetize large deployers? What did you model as the trade-off?
8. R2 has slipped because you were "not satisfied." What capability bar must it clear, and what does the delay say about the reasoning frontier vs your fast-follower cadence?
9. The Cisco safety result (zero harmful prompts blocked) is a regulatory and reputational liability. What is your safety/alignment investment, and how do you square it with shipping strip-able open weights?
10. Five-year investor lock-ups and a founder-controlled LP vehicle — is a public listing (anywhere) part of the plan, or is DeepSeek explicitly being built as permanent strategic infrastructure?
11. Qwen, Moonshot, and MiniMax are right behind you on open models, and Qwen has Alibaba's cloud to monetize through. What is your durable advantage over a peer with the same openness and a distribution/billing engine?
12. How dependent is the business on continued state subsidy (compute, capital, political cover), and what does DeepSeek's economics look like without it?
13. What is your inference-serving gross margin at current API prices, and what happens to it as you scale serving on Ascend?
14. Which single capability — agents, multimodality, long-horizon reasoning — do you believe will matter more than price over the next 24 months, and are you resourced to lead or only to follow on it?
15. If a Western jurisdiction issues an adverse distillation/IP ruling, what is your contingency for the models' commercial usability outside China?

---