Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.（DEEPSEEK） · AI Applications & Foundation Models

DeepSeek: a technically strong AI model company, with valuation already pricing in too much

Jun 9, 2026·~19 min read

Quick ReadPlain-language overview · read this first

This research note is about DeepSeek. Its stance is clear: Avoid. The point is not that the technology is weak, but that investors should not buy it at the current price.

DeepSeek mainly develops large models, the underlying “brain” that powers a wide range of AI applications. Ordinary users can access its website and App for free, while developers and enterprises pay by usage for API access. Its technology is strong and its pricing is exceptionally low. Among AI users in China, it has about 81.6 million weekly users, second only to ByteDance’s Doubao.

The report’s biggest concern is this: DeepSeek has not yet gone public, has no audited financial statements, and outsiders simply cannot see whether it is actually profitable, whether cash has truly been collected, or how much more money expansion will burn. Whether the technology can gain traction is one question. Whether the company can keep earning cash in an intense price war is another, and the latter is impossible to verify right now.

On valuation, the latest reports say its post-financing valuation is as high as 52 billion to 59 billion U.S. dollars. The report’s calculation is that this already pays upfront for the most optimistic scenario over the next ten years, leaving buyers no margin for error. The report argues that a valuation below 10 billion to 18 billion U.S. dollars would be reasonable. At present, it is clearly expensive.

The other key risks are also worth watching closely: industry competition is brutal, the price war has not ended, and pricing power is weak; regulation and chip supply both constrain the company; and the company depends heavily on its founder as an individual. The report’s conclusion is that DeepSeek deserves long-term attention, but now is not the time to buy it.

The above is only a plain-language explanation of this research note and is not investment advice. The stock market involves risk; invest with caution.

Lead
DeepSeek is a frontier Chinese AI model company founded in 2023, building adoption through free Web/App access, metered API monetization, and an open-source model ecosystem. Its latest DeepSeek-V4 model was released in April 2026, with API pricing far below OpenAI and Anthropic, but the company remains a private business with no publicly audited financials and a latest media-reported post-financing valuation of roughly $52 billion to $59 billion. Rating Avoid: the business is technically impressive, but the current valuation appears to have prepaid an optimistic ten-year outcome without enough verifiable cash-flow evidence.

Full report

Conclusion First

Let us put the most important point upfront: if DeepSeek is bought as a business rather than chased as a concept trade, my conclusion is "Avoid." The reason is not that I deny its technical breakthroughs. It is that as of June 9, 2026, DeepSeek remains a private, unlisted company with no public share price, no verifiable audited annual or quarterly reports, and none of the 10-Ks, 10-Qs, prospectuses, or shareholder letters that value investors rely on most. What you can see is strong product capability, fast iteration, very low API pricing, intense market attention, and a string of rapidly revised media-reported financing valuations. What you cannot see is "how much money the company has actually earned, what the quality of its cash flow is, how much capital growth will consume, and what minority shareholders can ultimately receive." For a balanced but conservative investor with a holding period of more than 10 years, that information structure is itself a red flag.

Preliminary Conclusion

Item	Conclusion
Investment rating	Avoid
Core judgment	DeepSeek is a business that is understandable, but cannot yet be priced safely. Its technology and cost efficiency are strong, and the industry opportunity is large, but competition is extremely intense, price wars have become normal, regulatory and compute constraints are heavy, and the company lacks audited financial disclosure. Based on the latest media-reported post-financing valuation of $52 billion to $59 billion, the market is already paying upfront for a very optimistic ten-year outcome.
Does the current price offer a margin of safety?	No / impossible to confirm, but under a conservative framework it should be treated as no.
Suitable investor type	Better suited to professional primary-market investors who can obtain first-hand diligence materials, board or contractual protections, and tolerate long lock-ups plus liquidity discounts; not suitable for ordinary long-term value investors.
Biggest uncertainty	Real revenue and free cash flow are unknown; the current valuation is a negotiated valuation rather than a publicly verifiable market price; China's AI model industry's price wars and regulatory/compute constraints may keep long-term returns depressed.

If I had to summarize it in one Buffett-style sentence: I am willing to study DeepSeek as a business for the long term, but I am not willing to buy a minority stake at today's opaque valuation, which already appears to discount optimistic expectations.

Business Understanding and Industry Structure

DeepSeek's business is not mysterious: it is essentially a foundation model company that acquires users, developers, and enterprise adoption through free Web/App access, paid API services for developers and enterprises, and ecosystem diffusion via open-source model weights and toolchains. Its official website states clearly that users can use DeepSeek for free and can also access the API. The company publishes model, pricing, transparency, and technical report pages. Its latest public models include the DeepSeek-V4 series released on April 24, 2026. The official pricing page shows that the current DeepSeek-V4-Flash API price is about $0.14 per million input tokens and $0.28 per million output tokens, while DeepSeek-V4-Pro is about $0.435 for input and $0.87 for output.

From a "how does it make money" perspective, its customers fall into three groups. The first is ordinary end users, mainly using free Web and App products to create brand reach and data feedback. The second is developers, who call APIs by token and top up balances. The third is enterprise customers, who integrate the model into their own products, customer service, coding, office workflows, or AI agent applications. The official FAQ shows that the platform supports top-ups and that balances do not automatically expire. Official documentation also provides OpenAI-format and Anthropic-format API endpoints, which lowers trial and migration costs for developers and helps customer acquisition. It also means customers do not face much difficulty switching providers: adoption friction is low, and lock-in is low as well.

Can I understand this business? I can understand the business layer, but not enough of its full economics. The business layer is clear: build stronger, cheaper, easier-to-integrate models; expand traffic through free products; monetize compute and model capabilities through APIs and the tool ecosystem. The problem is that long-term value is determined less by "whether the model becomes popular" and more by "whether the model can keep converting into free cash flow under high R&D, high compute, heavy depreciation, and intense price competition." On that front, DeepSeek's public information is severely insufficient.

At the industry level, generative AI is still in a high-growth but far from mature phase. Gartner expects global AI spending to reach $2.52 trillion in 2026, up 44% year over year, including about $26.38 billion for AI Models, about $588.6 billion for AI Services, and about $452.5 billion for AI Software. This shows the market is large enough. It also suggests that the real money may not all sit in "the model itself," and may instead flow to cloud platforms, agent layers, vertical software, data, service integration, and infrastructure. Gartner also points out that AI in 2026 is in the so-called "trough of disillusionment," with many enterprises preferring to buy AI capabilities from existing software vendors rather than making all-in new "moonshot" projects. That is not easy for independent model companies.

The competitive landscape is extremely intense. Stanford HAI's 2026 AI Index shows that as of March 2026, leading model performance had clearly converged: Anthropic, xAI, Google, OpenAI, Alibaba, and DeepSeek were all in the top tier, but DeepSeek's Arena Elo was about 1424, below Anthropic's 1503, OpenAI's 1481, and Alibaba's 1449. Meanwhile, the re-widening gap between open and closed models in 2025 shows that open source does not automatically equal durable leadership. More importantly, the performance gap between Chinese and U.S. models has narrowed sharply, which means DeepSeek no longer has the comfortable window of "others have not caught up yet."

In China, DeepSeek is certainly a first-tier player, but it is not an undisputed champion. Reuters, citing QuestMobile data, said ByteDance's Doubao had about 155 million weekly active users, while DeepSeek had about 81.6 million weekly active users, ranking second. Alibaba's Qwen is also catching up quickly through subsidies and product upgrades. Reuters also noted that while DeepSeek-V4 made clear progress, its relative lead was narrowing against competitors such as Qwen and Kimi. For long-term investors, this means DeepSeek looks more like a strong player in a good industry than a business "sitting on an undisputed monopoly profit pool."

As for "would I be willing to hold this business if the stock market closed for five years," my answer is: if I could buy the whole company at a reasonable valuation and obtain full financial statements and control rights, I would keep studying it; if the opportunity is merely a minority stake at today's rumored high valuation with extremely weak information rights, I would not. This is not a rejection of the business. It is a rejection of the imbalance between "price paid, rights received, and risks assumed."

Scores

Dimension	Score	Reason
Business understandability	3/5	The product and business model are understandable, but real economics and cash-flow structure lack disclosure.
Industry attractiveness	2/5	The industry opportunity is huge, but capital intensity is heavy, price competition is strong, technological substitution is fast, and regulatory and compute constraints are high.

Moat and Management

DeepSeek does have some things that "look like a moat," but most remain technical advantages or engineering-efficiency advantages rather than the kind of "sustainable, price-raising, cash-flow-realizing" moat value investors prefer. Official materials show that the company was founded in 2023, focuses on AGI foundation models, and emphasizes self-developed training frameworks, self-built intelligent computing clusters, and ten-thousand-GPU-class compute resources. Its models and inference tools are released under the MIT license, and its technical reports are publicly published. Reuters has also repeatedly noted that DeepSeek shook the market through low cost, high efficiency, and open weights, prompting rapid responses from domestic and overseas peers.

But a real moat should be tested across ten factors:

Moat dimension	Judgment	Evidence and explanation
Brand advantage	Medium	DeepSeek has built a strong brand in the global technical community and among Chinese AI users, but its domestic consumer user scale still lags Doubao.
Cost advantage	Medium to strong	Official API pricing is meaningfully lower than OpenAI GPT-5.5 and Anthropic Sonnet 4.6, indicating strong engineering and compute-utilization efficiency.
Scale advantage	Medium to weak	User and developer adoption is strong, but compared with Microsoft, Google, Alibaba, and ByteDance, DeepSeek lacks a larger traffic and cloud distribution base.
Network effects	Weak	Model products do not have classic two-sided network effects. The effect is more "data-feedback-iteration," and open source weakens exclusivity.
Switching costs	Weak	Official support for OpenAI/Anthropic-compatible interfaces helps integration, and also helps customers switch.
Channel advantage	Weak	The company does not have super-app entry points such as Alibaba Cloud, WeChat, or Douyin.
Patents, licenses, regulatory barriers	Medium	Compliance thresholds, data localization, and model registration are entry barriers, but regulation also constrains the company.
Data advantage	Medium to weak	The company discloses use of public and licensed data, but has not disclosed clear exclusive data assets.
Corporate culture and operating capability	Medium to strong	Reuters describes it as more like a research lab, with a flat hierarchy, young research team, and high engineering efficiency.
Capital allocation capability	Uncertain	Its long refusal of external financing and support from High-Flyer suggest long-termism, but without audited financials and a return record, ROIC and shareholder-return quality cannot be verified.

The key judgment is: this moat is narrowing, not widening. There are three reasons. First, Stanford HAI shows that leading model performance is converging, and the gap between open and closed models did not shrink one-way but reopened to 3.3% in 2025. Second, Reuters has repeatedly reported that DeepSeek's low-price and open-source strategy forced peers such as Alibaba, Baidu, and ByteDance to cut prices and upgrade frequently. That shows it changed the industry, but does not necessarily mean it created an exclusive profit pool for itself. Third, price competition in China's enterprise AI market is already very intense, and open source can weaken the platform's ability to charge.

On management, what I see is a mix of technology-driven, long-term oriented, but externally under-disclosed traits. Positive evidence includes founder Liang Wenfeng's long-term support for company operations through High-Flyer rather than rushing to an IPO; Reuters said he had publicly expressed that the company was not aiming at a price war, but placed AGI first; in the latest financing round, Liang was also reported to be contributing RMB 20 billion personally, which at least indicates substantial alignment and long-term commitment. At the same time, Reuters describes the company as "more like a research lab than a profit-centered enterprise." That culture is an advantage during periods of technical breakthrough, but it has not yet been proven in the task of "sustaining cash-flow distributions to shareholders."

The negative evidence is equally obvious. DeepSeek provides almost no listed-company-level disclosure: no verifiable segment revenue, no capital expenditure explanation, no equity incentive details, no shareholder letters, and no fixed mechanism for candid discussion of returns on capital and free cash flow. The official public materials focus mainly on products, research, transparency, and legal policies rather than investor relations. For long-term owners, this means management's motives may be good, but there is still not enough verifiable evidence to justify trust.

Scores

Dimension	Score	Reason
Moat strength	2/5	It has technology and cost advantages, but pricing power is weak, switching costs are low, and the industry catches up quickly.
Management and capital allocation	2/5	The founder shows strong signs of long-termism, but the company lacks listed-company-level verifiable disclosure and cannot prove excellent capital allocation.

Financial Quality and Owner Earnings

This section must be direct: most of the key financial metrics you need are currently unavailable publicly. On DeepSeek's official public pages, I can find product entry points, model pricing, technical reports, transparency pages, legal terms, and hiring information. I cannot find audited annual reports, quarterly reports, prospectuses, investor relations pages, 10-Ks, 10-Qs, or public share count and per-share data. Therefore revenue growth, gross margin, operating margin, net income, operating cash flow, free cash flow, ROE, ROIC, net debt/EBITDA, interest coverage, share-count changes, dividends, and buybacks can basically only be marked "unknown" at the public-information level. This is not analytical laziness. It is the boundary of the available evidence.

In that situation, the most useful step is to lay out the publicly verifiable operating anchors:

Item	Publicly verifiable information	Note
Company status	Operated by Hangzhou DeepSeek Artificial Intelligence Co., Ltd.; still in private financing stage and has not disclosed listing plans.	Unlisted, no public share price
Founded	2023.	Too short a history for a 10-year audited lookback
Product format	Free Web/App, paid API, open-source models.	Typical "traffic entry + developer platform" model
Latest major model	DeepSeek-V4, publicly released on April 24, 2026.	Still in rapid iteration
Current official API pricing	V4-Flash: $0.14 input / $0.28 output; V4-Pro: $0.435 input / $0.87 output.	Extremely low pricing
Disclosed revenue/cost anchors	For V3/R1 in 2025, disclosed theoretical daily revenue of $562,000 and theoretical daily inference cost of $87,100; Reuters noted that actual revenue would be significantly lower.	This is the closest public number to a "revenue anchor"
User-scale anchor	About 81.6 million weekly active users in China, below Doubao's 155 million.	Proves product influence, but not monetization quality
Latest valuation anchor	Reuters said on June 3, 2026 that its valuation after the first financing round could reach $52 billion to $59 billion.	Still a negotiated valuation

These anchors lead to a plain but important judgment: DeepSeek is not a mature business already proven to generate stable free cash flow. It is a frontier AI lab/platform in the commercialization-verification stage, with capital needs still rising. Reuters disclosed in March 2025 that under the theoretical maximum load for V3/R1, DeepSeek's annualized revenue could reach about $205.1 million, corresponding to annualized inference cost of about $31.78 million, and a theoretical "inference-layer gross contribution" of about $173.4 million. But the company also emphasized that actual revenue would be significantly lower. Combined with another Reuters report citing experts that the widely circulated "$6 million training cost" referred only to the chip usage cost of the final training run, while the entire development process could require much larger investment, and some industry participants believe total early-stage investment may have exceeded $1 billion, it is hard to view this company as a mature asset with distributable owner earnings.

Worse, the industry's commercialization reality is not forgiving. In an article discussing Alibaba and other Chinese AI companies, Reuters noted that Chinese consumers are clearly more resistant to paying for AI subscriptions, while the enterprise market has already seen margins severely compressed by price wars. AI companies are shifting toward enterprise APIs, but sharp API price cuts and widespread open-source models will weaken long-term pricing power. DeepSeek itself continued to push major price reductions in 2025 and 2026. For questions such as "are profits real cash or accounting profit, and does growth consume more money as it gets larger," my answer is: within the constraints of public information, I would rather treat DeepSeek as a company whose growth requires sustained capital investment, not as one that already becomes more profitable as it grows.

Accordingly, my conservative Owner Earnings estimate is very restrained:

Owner Earnings element	Publicly verifiable status	Conservative judgment
Net income	Unknown	Cannot confirm positive net income
Add back non-cash charges	Unknown	Cannot estimate reliably
Less maintenance capital expenditure	Unknown, but likely not low	Frontier model companies usually require sustained compute and R&D investment
Less changes in working capital	Unknown	Not verifiable
Real distributable cash flow	Unknown	Conservatively treated as roughly 0 or negative

This conclusion may sound cautious, but it is exactly the discipline of a long-term owner: before audited data exists, do not mistake theoretical revenue calculations from a research-oriented growth company for distributable cash flow.

Valuation, Margin of Safety, and Opportunity Cost

First, a note on scope: because DeepSeek is not a listed company, the discussion below is not about a "share-price range," but an overall equity valuation range for the company. There is no publicly verifiable per-share price or share count, so any "per-share target price" would be false precision.

Owner Earnings Discount Method

Without audited financial statements, the honest approach is not to pretend to run a precise DCF, but to use a stress-test DCF. My core assumption is simple: because there are currently no verifiable positive Owner Earnings, I treat 2026 starting Owner Earnings as close to zero. I then examine what level of Owner Earnings the company would need to reach over the next ten years to justify the valuation reported in the media today. This is more reliable than pretending to know current net income.

I use three scenarios, all assuming the company gradually ramps from "current near-zero distributable earnings" to scaled Owner Earnings in year ten:

Scenario	Year-10 Owner Earnings assumption	Discount rate	Terminal growth	Implied intrinsic value estimate
Conservative	$1 billion	12%	3%	About $6 billion to $8 billion
Base	$3 billion	11%	4%	About $18 billion to $30 billion
Bull case	$6 billion	10%	4%	About $45 billion to $60 billion

How should this table be read? It should be read as one simple sentence: if you truly buy DeepSeek today at $52 billion to $59 billion, you are effectively paying for the optimistic scenario in which the company can produce roughly $6 billion of Owner Earnings ten years from now. For a company whose real revenue, profit, and free cash flow are not publicly transparent today, and whose industry faces continuous price wars and enormous capital investment, that premise is not cheap.

Put more sharply: at the $52 billion to $59 billion valuation reported by Reuters, if you assume a perpetuity model, a 10% discount rate, and 4% terminal growth, the market is equivalently saying that DeepSeek should already have roughly $3.1 billion to $3.5 billion of "current stable Owner Earnings." If we acknowledge that it is still ramping today and must deliver gradually over the next ten years, the required year-10 earnings become even higher. For a company whose audited revenue is not even publicly verified, this is almost fully prepaying the best part of the future.

Relative Valuation Method

I will not fabricate PE, PB, EV/EBITDA, or P/FCF here, because DeepSeek has no public net income, book equity, EBITDA, or free cash flow. But two relative-valuation perspectives remain useful.

The first is to use the company's own disclosed "most optimistic public revenue anchor" to reverse-engineer multiples. Reuters cited DeepSeek's disclosure in March 2025 that V3 and R1, under theoretical full load, had daily revenue of about $562,000, equal to annual revenue of about $205 million, while the company also emphasized that actual revenue would be significantly lower. If this optimistic theoretical annual revenue ceiling is compared with the latest $52 billion to $59 billion valuation, the implied EV/Sales is roughly 254x to 288x. If we instead look at the theoretical annualized gross contribution after inference costs of about $173 million, the implied multiple becomes even more extreme. In other words, even using the company's own most optimistic public revenue anchor, this valuation is already extremely "future-heavy."

The second perspective is its price competitiveness and the profit pressure behind it. Official current pricing shows DeepSeek V4-Pro at about $0.435 input / $0.87 output; Anthropic's official Claude Sonnet 4.6 page lists $3 input / $15 output; OpenAI's official API page shows GPT-5.5 at about $5 input / $30 output. This indicates that DeepSeek is highly competitive on unit price, making it a strong cost challenger. For shareholders, however, this is a double-edged sword: cost advantage can drive adoption, but if the industry enters a lasting price war, low price is more like the "entry point" of the moat, not necessarily the "destination" of high returns. In addition, Alibaba Cloud's official documentation shows Qwen3.5-Plus pricing across different token ranges has also fallen to around $0.4 to $0.5 for input and $2.4 to $3 for output, so competitors are catching up quickly.

Asset or Liquidation Value Method

DeepSeek is not a company suited to high-precision liquidation-value analysis, because its core assets are talent, model weights, engineering know-how, brand, developer mindshare, and the data feedback loop, not tangible assets such as land, factories, or inventory that are easy to recover. The assets that could theoretically form an "asset floor" are mainly cash and some compute/equipment. The problem is that we do not even know its net cash and debt.

The only item that can be used, with difficulty, for an "asset floor" discussion is the proposed financing cash reported by Reuters: if the latest roughly $7.4 billion first financing round is completed at a $52 billion to $59 billion post-financing valuation, the new cash raised would represent only about 12.5% to 14.2% of the total valuation. This means that even if the financing closes, most of the current price is still a prepayment for future earning power, not support from hard assets. If the financing does not close, that asset floor becomes even thinner. For conservative investors, this means there is almost no "liquidation protection."

Margin of Safety and Opportunity Cost

After combining the three methods, I arrive at the following range judgment:

Valuation conclusion	Overall company valuation range
Conservative intrinsic value range	$5 billion to $10 billion
Reasonable intrinsic value range	$15 billion to $30 billion
Bull-case intrinsic value range	$45 billion to $60 billion
Ideal buy valuation range	No higher than $10 billion to $18 billion
Acceptable hold valuation range	$18 billion to $30 billion
Clearly overvalued range	Above $40 billion
Current media valuation comparison	$52 billion to $59 billion, close to or above my bull-case upper bound

The meaning behind this table is clear: the current rumored market valuation is not "leaving you a margin of safety"; it is "requiring you to believe in the most optimistic execution outcome." Especially for your balanced but conservative investment objective with a holding period of more than 10 years, the most fragile assumption when buying at this level is not "whether demand will grow," but "whether DeepSeek can truly turn adoption into billions of dollars of Owner Earnings amid years of price wars and heavy capital investment." If that assumption is marked down, returns will be permanently impaired.

Opportunity cost is also unfavorable. In public markets today, you can at least choose highly liquid broad-market indices; for example, the latest SPY price is about $739.22. On the risk-free/high-grade bond side, the latest public data shows the U.S. 10-year Treasury yield at about 4.55%, and China's 10-year government bond yield at about 1.73%. For a private company with no liquidity, no audited statements, no public shareholder-rights structure, and an already high valuation, expected returns should be meaningfully higher than those alternatives to justify the capital allocation. Under my valuation framework above, DeepSeek does not offer that excess compensation at the current valuation.

Risks, Checklist, and Final Recommendation

The most important risk is not short-term price volatility, but permanent capital loss. DeepSeek currently faces six core risks.

The first is competitive risk. Leading model performance is converging quickly. DeepSeek remains first-tier, but no longer has unique performance leadership. In China, it faces a group of strong rivals including Alibaba Qwen, ByteDance Doubao, Moonshot/Kimi, and MiniMax. Overseas, it competes with closed-model giants such as OpenAI, Anthropic, and Google. The scarce capability in this industry is not "being able to build a model," but "being able to preserve high returns after scaled commercialization."

The second is technology substitution and price-war risk. DeepSeek itself was one of the initiators of China's price war. Reuters reported that its V2 once triggered an industry price war, that it cut prices repeatedly in 2025 and 2026, and that the latest V4-Pro made a 75% discount permanent. This is good news for consumers, but not necessarily for shareholders, because it suggests industry value is shifting quickly to customers rather than accumulating for shareholders.

The third is regulatory and data-compliance risk. Reuters summaries show that DeepSeek has faced investigations, restrictions, and bans on government devices in multiple countries because of privacy and security concerns. Its privacy policy also states clearly that data will be processed and stored in China. For international enterprise customers, this can directly affect adoption, deployment scope, and order quality from government or regulated industries.

The fourth is compute and supply-chain risk. Reuters says DeepSeek is constrained by U.S. export controls and cannot freely obtain frontier chips the way U.S. peers can, directly affecting its training and inference expansion path. Its adaptation to Huawei chips is a commendable engineering achievement, but it also shows that its expansion capacity is heavily constrained by supply chains and geopolitics.

The fifth is management and key-person risk. Liang Wenfeng is extremely important. Funding, strategic direction, culture, and external financing are all tightly tied to him. High founder ownership and large follow-on investment can improve alignment, but they also increase key-person dependence. If the founder misjudges, talent losses accelerate, or the organization fails to transition from a research lab to a commercial company, the impact on shareholders would be large. Reuters has already mentioned that DeepSeek has lost talent to competitors.

The sixth is overvaluation risk. I consider this the most realistic risk. The financing valuation Reuters mentioned on April 17, 2026 was still $10 billion; on April 22 it became more than $20 billion; on May 6 it became up to $50 billion; and on June 3 it moved further to $52 billion to $59 billion. In less than two months, valuation expectations almost jumped continuously. That pattern looks more like liquidity and narrative-driven repricing than value gradually verified by disclosed audited financial results. For value investors, this method of valuation re-anchoring should itself be treated as a risk.

The strongest opposing view can be compressed into one sentence: DeepSeek may be a very strong AI company, but between a "strong company" and a "good investment" stand three gates: real cash flow, capital expenditure discipline, and purchase price. None of these three gates has been verified yet. The possible mistake is not underestimating the technology, but overestimating the speed and quality with which technology converts into shareholder cash returns. What bears truly see is an earnings model that has not closed the loop, an industry profit pool being eroded by price wars, and a valuation that has already heavily discounted ten years of optimistic expectations.

If the following facts appear in the future, I would acknowledge that my cautious judgment should be revised upward: first, audited-level financial disclosure appears and proves the company has formed sustained positive operating cash flow and free cash flow; second, enterprise customer revenue grows meaningfully, and unit economics keep expanding even as prices continue to fall; third, DeepSeek reopens a reliable performance gap versus Qwen, Doubao, OpenAI/Anthropic in the Agent era, and converts that gap into sticky commercial contracts; fourth, regulatory pressure remains controllable and international adoption is not clearly suppressed. Conversely, if there are continued sharp price cuts, missing revenue disclosure, compute constraints, high user numbers but weak payment, and later rounds still relying on higher valuations to survive, the current "Avoid" conclusion would be further strengthened.

Investment Checklist

Check item	Conclusion
Can I understand this business?	Pass
Does it have stable long-term demand?	Pass
Does it have a durable moat?	Fail
Does it have pricing power?	Fail
Can it generate stable free cash flow?	Uncertain
Are its returns on capital excellent?	Uncertain
Is management trustworthy?	Uncertain
Is capital allocation rational?	Uncertain
Is the balance sheet sound?	Uncertain
Is valuation below intrinsic value?	Fail
Is the margin of safety sufficient?	Fail
Would I feel comfortable holding it long term?	Fail
What key facts would make me sell?	If already held, sustained price wars, financing dependence, regulatory escalation, key talent loss, or failure to turn cash flow positive should all trigger reassessment
Do I want to buy merely because of market sentiment and news heat?	Most likely requires serious self-examination

Open Questions and Limitations

There are only three key unresolved questions, and each is enough to change the conclusion: real revenue scale, real cash-flow quality, and real shareholder terms and dilution structure. As long as these three items remain unknown, any "very precise" attractive valuation has limited credibility.

Final Judgment

【Final Rating】 Avoid

【One-Sentence Investment Thesis】 DeepSeek is an excellent AI company worth tracking for the long term, but without audited financials, without public information on shareholder rights, and with the latest valuation already near the upper bound of an optimistic scenario, it is not a buying point that fits conservative value-investing discipline.

【Core Bull Case】

Technical and engineering efficiency is outstanding, and the company reshaped industry expectations with low costs and open weights.
Product adoption is broad, with a strong brand and developer mindshare; weekly active users in China rank among the leaders.
Cost competitiveness is strong, with official API pricing meaningfully below several Western frontier models.
Founder long-termism and alignment appear strong. The company long refused external financing, and in this round the founder was reported to have personally committed a large follow-on investment.

【Core Bear Case】

There are no publicly audited financials, so profit, free cash flow, ROIC, and liabilities cannot be verified.
Industry competition and price wars are extremely strong, and pricing power is insufficient.
Regulatory, data-compliance, and geopolitical risks are high, affecting international expansion and enterprise adoption.
The current media-reported valuation of $52 billion to $59 billion already requires the market to believe in an extremely optimistic ten-year execution path.

【Key Assumptions】

DeepSeek can effectively convert technology leadership into enterprise revenue over the next ten years.
Industry price wars will not destroy the model-layer profit pool for the long term.
Compute and supply-chain constraints will not continuously suppress training and inference expansion.
Sufficiently transparent financial disclosure will appear in the future, allowing value to be verified.
Minority shareholder terms, dilution arrangements, and the liquidity path will not meaningfully erode returns.

【Fair Buy Price】 Measured by overall company valuation, I think the ideal buy range should be below $10 billion to $18 billion; $18 billion to $30 billion is barely worth studying; above $40 billion clearly requires near-perfect execution; $52 billion to $59 billion offers no margin of safety.

【Target Holding Period】 Only if future information becomes transparent and valuation is reasonable would it be suitable to hold for more than 10 years; under current conditions, I do not recommend establishing a position.

【Expected Annualized Return】 Using the current media-reported midpoint valuation of about $55.5 billion as the starting point, and roughly applying the ten-year intrinsic-value scenarios above:

Conservative scenario: about -18% to -15%/year
Base scenario: about -10% to -6%/year
Bull-case scenario: about 0% to +4%/year

These results do not mean DeepSeek will deteriorate. They mean that if bought at today's high valuation, shareholder returns may still be insufficient even if the company performs well.

【Maximum Loss Risk】 If commercialization falls short of expectations, the price war persists, regulation escalates, or the next financing round is repriced materially lower, permanent capital loss could reach 60% to 90% or even higher. For minority shareholders in the primary market, an extreme scenario approaching a 100% loss cannot be ruled out.

【Tracking Indicators】 The most important items to track in the future are not trending topics, but the following:

Revenue and revenue mix on an audited basis
Whether operating cash flow and free cash flow turn positive
Enterprise API revenue share and ARPU changes
Gross margin / inference cost trend
Capital expenditure and the intensity of training/inference compute investment
Major customer concentration and renewal status
Stable performance lead versus Qwen, Doubao, OpenAI, and Anthropic
Regulatory events and international availability
Key talent losses
Future financing terms, valuation, and dilution arrangements

【Signals That Would Trigger Reassessment】

The company begins disclosing audited-level financials and shows sustained positive free cash flow
Enterprise paid usage grows clearly faster than price reductions
Regulatory restrictions ease materially
It rebuilds a stronger performance and ecosystem advantage in the Agent era
Or conversely, there are repeated major price cuts, negative regulatory escalation, major talent losses, or a significant down-round valuation

【Final Recommendation】 Put DeepSeek on the "high-quality watchlist," not on the "currently buyable list." From the perspective of a long-term business owner, it is a company worthy of respect. From the perspective of a long-term value investor, respecting a company does not mean buying it at any price, and it certainly does not mean paying for the market's optimism when information is opaque.

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.

Mentioned Tickers

BABA GOOGL

Baillie Framework · Ten Questions for Growth Investing

Hunting ten-year five-baggers among great growth stocks — pressing the upside question: "Can it get much bigger?"

How large is its market ceiling? Is it expanding an existing pie, or creating an entirely new market?6/10

Conclusion: DeepSeek faces an extremely high ceiling, but it is not creating an entirely new market on its own. It is making the existing pie of AI foundation models, APIs, developer tools, and agent workflows larger and cheaper, while trying to control the model-layer gateway. Under Baillie Q1, the market space itself is large enough; the real question is not whether there is TAM, but how much of the profit pool DeepSeek can ultimately capture.

The external market anchor is very large: Gartner expects global AI spending to reach about 2.52 trillion dollars in 2026, but AI Models account for only about 26.38 billion dollars, far smaller than AI infrastructure, services, and software. This is critical: DeepSeek is not a company that directly eats “all AI spending.” It is more like a low-cost challenger in the foundation model/API/open-source ecosystem layer. If enterprises ultimately buy AI capabilities mainly through cloud vendors, incumbent software providers, and system integrators, model companies may drive industry adoption without necessarily taking the largest profit pool.

DeepSeek’s strength is that it does have the ability to “expand the pie.” Official API pricing shows V4-Flash cache miss input at about 0.14 dollars per million tokens and output at about 0.28 dollars per million tokens, while V4-Pro input is about 0.435 dollars and output about 0.87 dollars. It also supports OpenAI/Anthropic-compatible interfaces and a 1M context window; this low pricing and compatibility lower the trial threshold for developers and enterprises. Hugging Face’s technical readout of V4 also confirms that V4-Pro has 1.6T total parameters and 49B active parameters, while V4-Flash has 284B total parameters and 13B active parameters, and both offer a 1M token context window, with positioning tilted toward long-context and agent workload efficiency. This shows DeepSeek is not only taking existing chatbot users; it may also push agent, coding, office automation, and enterprise process scenarios that were previously uneconomic because costs were too high or context windows too short into practical use.

Even so, this is still not “creating a new market from zero.” A more accurate description is: DeepSeek is repricing the existing foundation model market and expanding usable scenarios through low cost, open source, and long-context capabilities. What it creates is new usage density and new application boundaries, not an exclusive new profit pool. Converging model performance, compatible interfaces, open-source substitutes, and price wars will all make it easier for customers to switch and will hand part of the value to the application layer, cloud platforms, and end users.

Under a consistent valuation framework, DeepSeek remains a private company with no public share price; media-reported financing valuation anchors are around more than 52 billion dollars, while the report range is about 52 billion to 59 billion dollars. Axios, citing Bloomberg, also said it was raising about 7.4 billion dollars at a valuation of about 52 billion dollars. So the Q1 answer is moderately positive: the market ceiling is high, and DeepSeek can expand total model usage; but its value-capture ceiling is clearly below the grand figure of “global AI spending,” and the core issue is whether it can convert a low-price technical advantage into enterprise revenue, stickiness, and pricing power.

Jun 9, 2026
Can its revenue at least double over the next five years? Will growth be driven mainly by volume, price, or new businesses?5/10

Conclusion: revenue could double over five years, but it is not a high-confidence bet; the real driver can only mainly come from API/enterprise paid usage volume and penetration of agent scenarios, not from price increases. DeepSeek has no publicly audited revenue baseline, so the idea of “doubling” is easily amplified by a low base: if true commercial revenue today is small, doubling or even multiplying several times is not difficult. But that does not mean it has proved sustainable commercialization, and user enthusiasm cannot be directly converted into revenue growth.

The key constraint is the revenue definition. DeepSeek once disclosed that if V3/R1 were all billed at R1 pricing over 24 hours, theoretical daily revenue would be 562,027 dollars and inference cost 87,072 dollars. But the same official note also admitted that actual revenue was significantly lower because only part of the service was monetized, Web and App remained free, and discounts applied. This makes the point precisely: traffic, tokens, and brand attention are prerequisites for revenue, not revenue itself.

Breaking down growth, volume is the first driver. DeepSeek’s free entry points, open-source ecosystem, OpenAI/Anthropic-compatible interfaces, long context, and tool-calling capabilities lower the adoption threshold for developers; the current official V4-Flash/V4-Pro support 1M context, and API pricing is charged per million tokens, with V4-Flash cache miss input at 0.14 dollars and output at 0.28 dollars, and V4-Pro input at 0.435 dollars and output at 0.87 dollars. This kind of low-price strategy helps pull developer and enterprise workloads in, especially coding, customer service, office automation, data analysis, and long-context agents.

Price is not the main driver and may even be a drag. DeepSeek’s low pricing is itself a customer acquisition weapon, but model-layer competition is intense, and official materials also note that product prices may be adjusted. If Qwen, Doubao, Kimi, OpenAI, and Anthropic continue cutting prices, DeepSeek is more likely to offset lower unit prices with higher call volume than to double revenue through rising ARPU.

New businesses are optional upside, not a proven main engine. Enterprise private deployment, industry agents, toolchains, and partnerships with cloud vendors/hardware ecosystems could move DeepSeek from a “cheap API” toward more stable enterprise revenue; AP’s coverage of V4 also mentioned its strengthened agentic capabilities, free Web/mobile access, and open-model path. But as of now, public information still does not show enterprise contracts, renewal rates, net revenue retention, or segment revenue.

So my view is: the probability that revenue at least doubles over five years is not low, but the investment implication is limited; what truly needs to be verified is whether paid API and enterprise revenue can grow faster than price erosion and eventually settle into auditable revenue and cash flow.

Jun 9, 2026
Five years from now, what will take over as the next growth engine? Does this “second curve” exist today?5/10

Conclusion: the most promising second curve five years from now is “enterprise-grade agentic workflows infrastructure,” not ordinary chatbots themselves. More specifically, it means using long-context models to read codebases, contracts, knowledge bases, and business logs; connecting retrieval, code, spreadsheets, tickets, and internal systems through tool calls; and embedding models into real workflows through enterprise APIs, private/local deployment, and open-source ecosystem services. But today this curve can only be said to exist technically in prototype form, not to have been validated in revenue.

The evidence for a technical second curve is quite clear. DeepSeek’s official API page shows that V4-Flash and V4-Pro both support 1M context, up to 384K output, JSON output, tool calls, and both OpenAI Format and Anthropic Format base URLs. This is not simply “a model that chats better”; it allows the model to ingest long task traces, long documents, and multi-round tool results, making it suitable for long-chain tasks such as coding agents, enterprise knowledge-base agents, compliance review agents, and data-analysis agents. Hugging Face’s introduction to V4 also emphasizes long context and agentic workloads: V4’s innovation is not just benchmarks, but making long-running tool-call tasks more usable and more KV-cache efficient. This shows DeepSeek’s second-curve direction is not another consumer app, but an upgrade from “answering questions” to “completing processes for enterprises.”

Enterprise API is the first commercialization path. Low-price APIs can first attract developers and small and medium-sized enterprises to try the product, while compatibility with OpenAI/Anthropic formats lowers migration friction; but this is also a double-edged sword, because easy integration also means easy replacement. Therefore, if DeepSeek wants API to become a growth engine five years from now, it cannot rely only on cheap tokens. It needs stickier enterprise-layer capabilities: stable SLAs, permission isolation, audit logs, private knowledge-base integration, reliable tool calling, industry templates, and long-term contracts. Public materials today can prove API capability and pricing advantage, but they cannot prove large-scale enterprise renewal revenue or net revenue retention.

Private and local deployment is the second and more important path, especially for finance, manufacturing, government/enterprise, healthcare, and legal customers that are sensitive to data boundaries. The DeepSeek V4-Pro model card shows that its model weights use the MIT License and provide usage paths such as vLLM, SGLang, Docker, and local deployment. This gives DeepSeek a chance to convert open-source influence into revenue from enterprise deployment, inference optimization, dedicated adaptation, and technical support. The problem is that open source itself is not commercialization: MIT weights expand the ecosystem while weakening exclusivity; the money is made by building hosted inference, private deployment, toolchains, enterprise support, and customization services around the open-source model, not by “many model downloads.”

From the market-demand side, the external soil for the second curve exists. Gartner expects global AI spending to be about 2.59 trillion dollars in 2026 and explicitly mentions that enterprises will use AI agents in multiple workflows, with AI Models spending growing sharply year over year in 2026. But this also reminds us that the largest profit pools may not all sit in the foundation model layer. AI Services, AI Software, and AI Infrastructure are all larger, and enterprises more often buy AI through existing software and cloud vendors. If DeepSeek remains only a “cheap model supplier,” the second curve may be captured by cloud vendors, application software, and system integrators; only by becoming the enterprise agent execution layer or a standard component for private deployment can it capture more value.

So my judgment is: this second curve exists today, but what exists is a technical asset and strategic option, not a validated revenue curve. A verifiable revenue second curve would require at least three types of hard evidence: first, enterprise API revenue mix, renewal rates, and large-customer retention; second, private/local deployment contracts, deployment pricing, and renewal status; third, gross margin and free cash flow after inference, training, R&D, and support costs. The central issue in the current report remains that DeepSeek has no publicly audited financials and cannot prove that these businesses have already formed stable cash flow.

Therefore, the growth engine five years from now will likely be a combination of “agentic workflow + enterprise API + private deployment + open-source ecosystem commercialization”; but as of June 9, 2026, it looks more like a set of promising technology options. In the Baillie framework, this is a positive because the second-curve direction is real and the market is large enough; but it is not yet a decisive positive because the commercial loop, pricing power, and revenue quality remain unverified publicly.

Jun 9, 2026
What is its core competitive advantage? Will this moat widen or narrow over the next three to five years?4/10

Conclusion: DeepSeek’s core competitive advantages are “high value-for-money model capability + engineering cost efficiency + open-source brand + developer mindshare,” but this is more like a shallow and fluid technology/distribution moat, not a commercial moat where customers are locked in. Over the next three to five years, the base case is that the moat narrows rather than widens.

Its strongest point is first cost efficiency. The official pricing page shows that DeepSeek V4-Flash / V4-Pro support 1M context, with V4-Pro cache miss input and output priced at 0.435 dollars and 0.87 dollars per million tokens respectively, while V4-Flash is even lower at 0.14 dollars and 0.28 dollars. This proves it does have an advantage in “delivering strong model capability cheaply to developers”: DeepSeek’s official model and pricing page lists the above V4 prices and 1M context. This advantage is attractive to developers, startups, and cost-sensitive enterprises, and it is the core reason the report recognizes its strong technology and engineering efficiency.

The second layer of advantage is open source and brand. DeepSeek-V4’s Hugging Face model page shows that V4-Pro has 1.6T total parameters and 49B active parameters, V4-Flash has 284B total parameters and 13B active parameters, and the model weights use the MIT License: V4 model weights and license are public. This has built a strong “cheap, open, and good enough” perception for DeepSeek among developers. Open source amplifies adoption and makes third-party frameworks, inference services, and private deployment teams more willing to support it.

The problem is that these advantages do not naturally translate into strong lock-in. DeepSeek’s official documentation clearly supports OpenAI/Anthropic-compatible APIs, with the OpenAI format base URL at https://api.deepseek.com and the Anthropic format base URL at https://api.deepseek.com/anthropic: the official documentation says OpenAI/Anthropic SDKs or compatible software can be used for access. This is excellent for customer acquisition, because customers can try it by changing a few lines of configuration; but in reverse, customers can also switch back to OpenAI, Anthropic, Qwen, Kimi, Doubao, or Gemini with similarly low friction. In other words, it lowers the entry barrier and also lowers the exit barrier.

So DeepSeek’s moat is not typical software stickiness, data network effects, or enterprise process lock-in, but “the speed of sustained leadership.” As long as it can keep being faster, cheaper, and better than peers, the moat exists; once model performance converges and prices continue to fall, this moat will be filled in. Stanford HAI’s 2026 AI Index has noted that as of March 2026, Anthropic, xAI, Google, OpenAI, Alibaba, and DeepSeek all sat in the top Arena Elo tier, and competitive pressure was shifting toward cost, reliability, and vertical-scenario performance: frontier model performance is converging. This is a double-edged sword for DeepSeek: it can break through on cost, but it will also be forced to keep competing on cost.

Over three to five years, whether the moat can widen depends less on “whether the next-generation model is very strong again” and more on whether DeepSeek can turn model advantages into enterprise contracts, private deployment, Agent workflows, toolchain ecosystems, compliance capabilities, and real commercial data feedback. If these are achieved, the moat could upgrade from “developers like it” to “enterprises cannot do without it.” But based on current public information, DeepSeek looks more like a powerful model-layer challenger than a platform company that has already proved high switching costs. Base case: the technology and cost advantages will remain, but converging model performance, open-source diffusion, and price wars will continue to compress exclusivity, and the moat will most likely narrow over the next three to five years.

Jun 9, 2026
If its core business is disrupted, does it have the DNA to reinvent itself? How does it treat mistakes and bad news?5/10

Conclusion: DeepSeek has a strong “technical organization reinvention” gene, but it has not yet proved “commercial and governance-level reinvention.” If the core model business is disrupted by stronger models, price wars, or regulation, I believe it can quickly adjust its research direction, architecture, and product form; but if the bad news comes from revenue missing expectations, uncontrolled capital expenditure, valuation markdowns, customer churn, or regulatory penalties, public information is still insufficient to prove that it would disclose this to shareholders in a timely, complete, and candid way like a mature, excellent company.

The positive evidence is clear. DeepSeek is not a conventional AI project incubated from a traditional internet product line, but a research organization that grew out of High-Flyer’s quant, machine learning, and compute resources: AP reported that Liang Wenfeng first built the High-Flyer quant fund and then extended quant-model capabilities into AI; Yicai also verified that High-Flyer Quant announced in 2023 that it would concentrate resources to establish an independent research organization exploring AGI, and responded that the research “has nothing to do with finance”. This shows it has already gone through one core identity shift: from “using AI for financial trading” to “putting compute, talent, and cash flow accumulated in finance into AGI foundation research.” This kind of migration is not a marketing rebrand but a real change in resource-allocation focus.

The second piece of evidence is model iteration speed. Official news shows that DeepSeek-V3 was released in December 2024, using 671B MoE, 37B active parameters, and remaining open source; then DeepSeek-R1 was released in January 2025, emphasizing open source, MIT licensing, and reinforcement-learning reasoning capability; by April 2026, DeepSeek-V4 Preview had gone live and open source, focusing on 1M context, Agent capability, and OpenAI/Anthropic API compatibility. From V3’s foundation model to R1’s strengthened reasoning and V4’s long context and Agent task optimization, this is not a linear series of small tweaks to one product, but a signal that the organization can rapidly rebuild its technology stack as the industry’s center of gravity changes.

Its handling of “mistakes” also earns some credit at the technical level. The official R1-0528 update directly listed “reducing hallucinations,” strengthening front-end capability, supporting JSON output, and function calling as improvements, showing that the team is at least willing to convert model defects into targets for the next iteration rather than only talking about performance mythology. Open-source weights, technical reports, and low-price APIs also make it easier for external developers to verify, challenge, and replace it, forcing the organization to keep facing real feedback.

But the issue is that technical mistakes are not the same as operating bad news. DeepSeek’s history is too short: it was founded only in 2023 and has not yet gone through a full funding winter, commercialization failure cycle, major customer churn, valuation markdown, post-regulatory-penalty review, or public-company-style consecutive earnings accountability. The report also repeatedly notes that the company has no publicly audited financials, no 10-K/10-Q, no shareholder letters, and no verifiable revenue structure, cash flow, or capital expenditure. In other words, we can see how it fixes models, but not how it provides capital with true, unpleasant, yet necessary operating information.

So the Q5 score should be “moderately above average but not full marks”: the reinvention gene is strong in the research organization, compute scheduling, model roadmap, and open-source ecosystem; it is weak in business model, governance disclosure, and bad-news transparency. If the core business is disrupted in the future, DeepSeek is unlikely to be stuck in old technology; but whether it can turn reinvented technology into sustainable cash flow and honestly face shareholders and customers in the process still lacks enough public evidence.

Jun 9, 2026
Does management, especially the founder, have a long-term view and deep alignment with the company? Is it willing to sacrifice current profits for the next five to ten years?6/10

Conclusion: Q6 is one of DeepSeek’s strengths, but it is not a full governance pass. Liang Wenfeng’s long-termism, technical obsession, and economic alignment are all strong; the deduction is not about founder motivation, but about the company being unlisted with insufficient disclosure, while minority-shareholder protection and capital-allocation returns cannot yet be verified.

The positive evidence is concentrated. DeepSeek is not a company that first tells a capital-market story and then looks for a technology narrative; it is an AGI project that grew out of High-Flyer’s capital, compute, and research culture. External reporting verification shows that Liang Wenfeng first founded the High-Flyer quant fund and used machine learning to improve trading models, while High-Flyer also owns resources that support DeepSeek AI research. Reuters further verified that High-Flyer announced in 2023 that it would concentrate resources on exploring AGI, and DeepSeek was founded afterward; Liang Wenfeng also said he did not plan to raise money in the short term and that the real bottleneck was high-end chips. This shows his actions look more like a bet on foundation-model capabilities five to ten years out than a pursuit of current-period profit maximization.

Economic alignment is also strong. On the latest financing, Axios, citing Bloomberg, reported that DeepSeek planned to raise about 7.4 billion dollars at a valuation of about 52 billion dollars, with Liang Wenfeng personally planning to invest about 2.85 billion dollars, and that the company emphasized breakthrough AI research over short-term commercialization. If this transaction closes, such a large personal co-investment is a strong positive signal for long-term shareholders: the founder not only controls the direction, but also keeps putting personal wealth on the same long-term technology path.

But “the founder is credible” must be separated from “minority shareholders can rest easy.” DeepSeek is still a private company, with no public-company-level audited statements, shareholder letters, equity-incentive details, capital-expenditure explanations, ROIC/FCF record, and no visible minority-shareholder terms, liquidation preferences, anti-dilution arrangements, or governance-rights structure. Liang Wenfeng’s strong tie to High-Flyer raises long-term alignment on one hand, but also magnifies key-person and concentrated-control risk on the other.

So my judgment is: management long-term view: strong; economic alignment: strong; willingness to sacrifice short-term profit for long-term technology leadership: very likely; but governance transparency, minority-shareholder protection, and capital-allocation returns: unverified. In the Baillie framework, this question should add clear points for DeepSeek, but strong founder long-termism cannot automatically imply that outside investors will share equally high-quality long-term returns.

Jun 9, 2026
If it disappeared tomorrow, how much would customers miss it? Is its growth model sustainable and not dependent on harming society or regulation?4/10

Conclusion: DeepSeek would be clearly missed by developers and cost-sensitive enterprises, but it has not reached the point of being irreplaceable if it disappeared tomorrow; its growth model has sustainable elements, but regulation, data security, international availability, and compute supply are hard constraints.

The customers who would miss it most are two groups: developers who plug models into code, customer service, RAG, and Agent workflows; and budget-sensitive enterprises that still want to use strong models. The reason is straightforward: DeepSeek’s official pricing page shows V4-Flash at 0.14 dollars per million cache miss input tokens and 0.28 dollars for output, with V4-Pro input at 0.435 dollars and output at 0.87 dollars, and the same page provides OpenAI Format and Anthropic Format base URLs, making access easy. Add to this that when DeepSeek-R1 was released, it explicitly said the code and model used the MIT license, could be distilled, and could be commercialized, and its appeal to open-source developers, private-deployment teams, and small businesses is real. If DeepSeek disappeared tomorrow, customers would miss the combination of “cheap, strong, open, and fast-iterating.”

But this kind of missing is not strong lock-in. API customers can usually migrate to Qwen, Doubao, Kimi, OpenAI, Anthropic, or Gemini. The cost is mainly re-evaluation, prompt changes, safety-policy redesign, and cost optimization, not rebuilding the entire business. DeepSeek’s interface compatibility is a customer acquisition advantage and also shows that switching costs are not high. The teams for which migration is truly harder are those with deep customization, local deployment, or fine-tuning around DeepSeek weights; ordinary consumer users and shallow API callers will have much weaker stickiness.

The growth model itself does not necessarily depend on harming society. Low-price models and open weights lower the threshold for AI use and create positive externalities for the developer ecosystem; it is not acquiring users through ad addiction or gray-area finance. But sustainability should be discounted: low prices transfer value to customers but may not settle into high gross margin and free cash flow; open source expands influence while weakening exclusive monetization; enterprise customers will especially care about data, compliance, and availability. DeepSeek’s privacy policy says it collects user inputs and directly processes and stores personal data in China, which is a material obstacle for government, finance, healthcare, and cross-border enterprise procurement. There are already real regulatory cases: Italy’s data-protection authority urgently restricted DeepSeek’s processing of Italian users’ data in January 2025 and launched an investigation, and South Korea also once suspended downloads of the DeepSeek app in local app stores over privacy concerns.

So the Q7 answer is: customers would miss DeepSeek, but more as the regret of losing a high-value-for-money supplier, not the panic of losing the only critical infrastructure. For its growth to be sustainable over the long term, it must prove that low prices are not subsidy-driven price war, that enterprise adoption can turn into stable revenue, that data compliance can clear overseas and highly regulated industry thresholds, and it must also face U.S. advanced semiconductor export controls that restrict China’s access to AI compute. Until these conditions are proved, DeepSeek’s customer value is strong, but irreplaceability and regulation-friendly compounding are not fully established.

Jun 9, 2026
What are the unit economics of this business (gross margin, incremental returns)? Do they improve or worsen as scale grows? Where does the money it earns go?4/10

Conclusion: DeepSeek’s “narrow inference unit economics” look attractive, but “company-level unit economics” have not been proved. The theoretical gross contribution of model calls cannot be treated directly as real gross margin, FCF, or owner earnings; publicly, it lacks the key ledger of audited revenue, gross margin, operating cash flow, capital expenditure, and ROIC.

The optimistic side is that cost efficiency is genuinely strong. DeepSeek’s official pricing page shows V4-Flash at 0.14 dollars per million token cache miss input and 0.28 dollars for output, V4-Pro at 0.435 dollars for input and 0.87 dollars for output, with support for 1M context and relatively high concurrency. This shows it has strong pricing power for API customer acquisition: official V4 API pricing is enough to prove that “cheap” is not hearsay. DeepSeek itself also disclosed a 24-hour stress sample for the V3/R1 online inference system: if all billed at R1 pricing, theoretical daily revenue would be 562,027 dollars, and H800 GPU rental cost would be 87,072 dollars; but the same disclosure explicitly said that actual revenue was significantly lower because Web/App were free, only part of the service was monetized, and night-time discounts applied.

So the real judgment should be split into two layers: the marginal contribution of a single inference may be positive, and technical efficiency is excellent; but company-wide gross margin and incremental returns are unknown. For frontier model companies, costs are not just “how much GPU this call used.” They also include next-generation model training, inference capacity expansion, engineering infrastructure, R&D staff, talent competition, compliance, and enterprise delivery. Reuters-cited reporting also shows that DeepSeek made V4-Pro’s 75% price cut permanent, keeping the price at one quarter of the original; this is good for users, but for shareholders it means the model-layer profit pool may keep being handed to customers: V4-Pro permanent 75% price cut.

Will scale make things better? Technically, possibly; financially, not necessarily. If request volume rises, cache hit rates improve, and compute utilization gets better, unit inference cost will fall; but if scaling also brings more free traffic, lower API prices, more training iterations, and fiercer talent competition, true FCF may become worse. AI foundation models are not traditional SaaS: a larger user base does not necessarily mean an asset-light high-gross-margin flywheel, and may first show up as larger compute and R&D investment.

The money it earns most likely does not go to dividends or buybacks, but continues to be invested in three things: compute and inference capacity, next-generation model training, R&D and top talent. This fits DeepSeek’s technology roadmap and explains why Q8 cannot score highly: it has strong engineering efficiency, but has not publicly proved that “the larger the scale, the more distributable cash for shareholders.” The more appropriate current formulation is: unit economics have potential, but have not been validated by audited financials; during the scaling phase, owner earnings are likely to be depressed first rather than lifted.

Jun 9, 2026
What conditions must all hold for it to rise fivefold in ten years? Are these conditions realistic? What expectations are implied by today’s share price?2/10

Conclusion: for DeepSeek to rise fivefold in ten years, “strong model, low price” is not enough. It must evolve from a low-price model supplier into a global AI platform that can consistently collect high-quality enterprise revenue, while proving positive free cash flow, controllable regulation, and investor exits. Starting from the current valuation of about 52 billion to 59 billion dollars, a fivefold return implies company value of about 260 billion to 295 billion dollars; each condition is possible individually, but together they are demanding as an investment premise.

First clarify “today’s share price”: DeepSeek is still a private company, Crunchbase lists it as Private and founded in 2023, and it has no public share price, public share count, or per-share data. So this is not about a secondary-market share price, but about private financing valuation. According to Reuters’ relay of the latest financing report, DeepSeek planned to raise about 50 billion yuan, with a post-money valuation of about 350 billion to 400 billion yuan, or 52 billion to 59 billion dollars. A fivefold return in ten years would require pushing overall company value to 260 billion to 295 billion dollars.

At least five things need to be true at the same time.

First, enterprise revenue must form at scale. Free apps, open-source models, and developer buzz prove adoption, but they do not automatically prove revenue quality. DeepSeek needs to win long-term contracts in enterprise API, private deployment, Agent workflows, coding/customer service/office automation, and these customers cannot be coming only because it is cheap and ready to leave whenever cheaper Qwen, Doubao, Kimi, OpenAI, or Anthropic options appear.

Second, low-price advantage must turn into pricing power, not evidence of a price war. The official pricing page shows that DeepSeek V4-Pro’s current API price is 0.435 dollars per million input tokens and 0.87 dollars per million output tokens, which is strong and shows high engineering efficiency; but for shareholders, low price itself is not a moat. The real question is whether, as model-layer prices keep falling, DeepSeek can still charge enough gross margin through reliability, ecosystem, data security, deployment convenience, and enterprise integration.

Third, free cash flow must turn positive and be verifiable. The most cautious part of the report is exactly here: DeepSeek has no publicly audited revenue, net profit, operating cash flow, capital expenditure, or shareholder-rights structure. If the next ten years still require continuous investment in training, inference compute, talent, and subsidies without settling into positive FCF, then even a large user base will have difficulty translating a fivefold valuation into shareholder returns.

Fourth, regulation and data compliance must be controllable. DeepSeek’s international enterprise path depends on more than model capability. Reuters’ summary says DeepSeek has faced scrutiny in multiple countries over privacy and security policies, and mentions that under its privacy policy personal data such as user requests and uploaded files is stored on computers in China. If governments, finance, healthcare, and large multinationals restrict use because of data and geopolitical risks, its high-value enterprise market will be weakened.

Fifth, there must be a clear liquidity path. A rising private valuation does not mean investors truly earn fivefold. Minority shareholders also need to examine financing terms, liquidation preferences, dilution, repurchase arrangements, IPO prospects, or secondary transfer windows. Without these, even a much higher paper valuation may remain only paper gains.

So today’s valuation implies very full expectations: the market is already assuming DeepSeek will maintain frontier-model competitiveness, convert low-price traffic into massive enterprise revenue, preserve pricing power under fierce competition, cross regulatory and compute constraints, and eventually provide investors with a realizable exit. My judgment is: this is not impossible, but it requires near-perfect execution; buying at the current valuation leaves a very thin margin of safety.

Jun 9, 2026
Why has the market not recognized all of this yet? Is it because the market does not understand, looks down on it, or cannot look far enough? What will become the “narrative inflection point”?2/10

Conclusion: the market does not misunderstand DeepSeek; it has already priced the “strong technology, low price, open source, domestic frontier model” narrative quite fully. What remains unverified is whether it can turn these advantages into audited revenue, renewable enterprise contracts, positive free cash flow, and financing terms acceptable to minority shareholders. So the expectation gap in Q10 is not “whether DeepSeek will be discovered,” but “whether hype can land in operating numbers.” The narrative inflection point will not be simply another strong model release; it will be the first audited disclosure, major enterprise API customer renewals, FCF turning positive, or a valuation reset from above 52 billion dollars to a range that compensates for these uncertainties.

First, what has the market already seen? The technology narrative is no longer obscure information: DeepSeek’s official API pricing page shows V4-Flash cache miss input at 0.14 dollars per million tokens and output at 0.28 dollars, V4-Pro input at 0.435 dollars and output at 0.87 dollars, with 1M context length; Hugging Face model pages also disclose that V4-Pro has 1.6T total parameters and 49B active parameters, while V4-Flash has 284B total parameters and 13B active parameters, both supporting 1M context. This information makes it easy for the market to understand its “cheap, competitive, open weights/open ecosystem” label. The financing narrative has also been fully priced: Axios, citing Bloomberg, said DeepSeek was raising about 7.4 billion dollars at a valuation of about 52 billion dollars. If an unlisted model company is already raising capital at this scale, it cannot be said that the market still “looks down on it.”

What the market truly hesitates to fully reward is that it “cannot see the books.” DeepSeek once provided a theoretical operating anchor, but this exposed the problem precisely: TechCrunch reported that its theoretical daily revenue was 562,027 dollars and GPU rental cost was 87,072 dollars, while the company also admitted actual revenue was significantly lower. This is not audited revenue, not FCF, and not owner earnings. Free Web/App access, low-price APIs, and open-source models can all bring adoption, but they may also transfer industry profit to customers and ecosystem participants rather than to shareholders. Under the Baillie framework, the issue is not whether there is upside imagination, but that a fivefold return in ten years requires adoption rate, paid conversion, renewal rate, gross margin, capital-expenditure discipline, controllable regulation, and financing dilution friendliness to hold at the same time.

Therefore, the part where the market “cannot look far enough” is mainly enterprise commercialization, not model capability itself. The market still cannot confirm whether enterprises are willing to entrust core workflows to DeepSeek for the long term; whether API customers are merely price-sensitive traffic; whether large customers can accept price increases or at least stable pricing at renewal; whether FCF can turn positive after training, inference, talent, and compute investment expands; and whether data compliance limits adoption by finance, government, and multinational enterprises. Regulation is not an abstract risk either. DeepSeek’s old privacy policy explicitly stated that user information was stored on servers in China and that information would be shared under applicable laws, legal processes, or government requests, which will make some overseas enterprise customers view it as a compliance cost rather than a pure technology choice.

There are four types of narrative inflection point. First, an audit-disclosure inflection point: first publication of audited revenue, gross margin, operating cash flow, capital expenditure, and customer concentration, proving that theoretical revenue is not a slide-deck number. Second, an enterprise paid-adoption inflection point: disclosure of renewals, expansion, and net revenue retention for several high-quality enterprise API/private-deployment customers, rather than only downloads, call volume, or weekly active users. Third, a cash-flow inflection point: operating cash flow and FCF turn positive while prices remain low and model iteration continues, proving cost advantage is not bought with subsidies. Fourth, a valuation inflection point: financing terms become more transparent, or valuation resets to a level that covers regulation, dilution, liquidity, and commercialization uncertainty. Without these inflection points, DeepSeek can still be a good company and a strong technology platform, but it is not yet a proven good investment.

Jun 9, 2026

Ask about this report

Members can ask about this report; once answered it appears under "Reader Q&A" on this page. You can also highlight a passage in the text to ask about it directly.