DeepSeek

1.1 Identity and operating model

DeepSeek is described by independent reporting and company-profile coverage as a Chinese AI company and research lab headquartered in Hangzhou. Its product surface combines a free chatbot, developer API access, and open-weight large language models, with R1 positioned as a reasoning model and later model families continuing the open-weight strategy. The strongest evidence distinguishes the model license from the service: the models are described as MIT-licensed or open-weight, while the chatbot and API are commercial delivery channels. That combination gives developers a self-hosting option and gives DeepSeek distribution beyond a single hosted application. The evidence is less conclusive on the legal incorporation date because the assigned sources report both a 2023 incorporation/spin-out narrative and older High-Flyer roots. I therefore treat 2023 as the operating-company milestone and 2015 as the associated founder/parent origin, rather than collapsing them into one founding date.[CO001, CO002, CO003, CO004]

1.2 Leadership, control, and governance

Liang Wenfeng is the only consistently identified senior leader in the assigned evidence. BBC reporting describes him as DeepSeek's founder and links him to High-Flyer, while profile coverage describes the lab as closely associated with the hedge fund he co-founded. CNBC reporting on the 2026 external round also describes investors being told not to poach DeepSeek personnel, which makes retention and founder-led control material diligence considerations. No assigned source establishes a complete board, executive bench, or formal governance chart. That absence is not evidence that governance is deficient, but it limits confidence in succession planning, independent oversight, and key-person resilience. The company therefore appears founder-centered and High-Flyer-linked, with governance depth still requiring primary documentation.[CO005, CO006, CO007, CO008]

1.3 Capital, scale, and disclosure

The refresh evidence indicates a major change from the prior self-funded narrative: CNBC reported that DeepSeek closed a first external round valuing the lab above $50 billion in June 2026, while TechCrunch reported in July that the company was seeking approximately $1.5 billion at about a $71 billion valuation. These are reported or proposed transaction figures, not audited valuation marks or cash available to the company. Other assigned profiles publish materially different funding and valuation numbers, and one profile gives an employee count of 422 without a primary filing. The Information source is retained as a net-new lead for reported revenue, but its fetched text does not expose enough underlying article detail to treat the preview as audited financial evidence. Revenue, ARR, gross margin, customer count, and cash runway therefore remain gaps.[CO009, CO010, CO011, CO012, CO013]

1.4 Milestones and material risk boundaries

DeepSeek's visible trajectory runs from High-Flyer-linked research through the R1 release and subsequent financing interest. BBC reporting links R1's low reported training cost and app adoption to the company's global breakthrough, while technical analysis attributes the cost narrative to mixture-of-experts architecture, compression, and caching. The same evidence shows that technical influence does not remove operating and policy risk: CNBC reports a U.S. ban on DeepSeek-related technology in federal civilian agency systems, citing potential sensitive-data transfers and cybersecurity vulnerabilities. Leadership retention pressure, export-control exposure, and service outages or attacks should consequently be tracked alongside model releases. The company has meaningful milestone evidence, but public disclosure remains insufficient for a complete operating KPI baseline.[CO003, CO014, CO015, CO016]

1.5 Exhibits

2.1 Market boundary and substitutes

DeepSeek's relevant market comprises hosted model inference through chat and APIs plus open-weight deployment on customer-controlled infrastructure. The boundary excludes chips, general cloud capacity, downstream application revenue, and the broader AI development-platform market unless DeepSeek captures that spend directly. Layer3 Labs describes consumer chat, a developer API, and downloadable weights as the visible routes to use, while reporting no formal enterprise sales, professional-services, or support tier at publication. Buyers can substitute closed-model APIs, other open-weight families, or internal deployment. Those alternatives mean that low price and weight availability expand usage but also constrain vendor lock-in and the share of ecosystem value that accrues to the model author.[CM001, CM012, CM013, CM014]

2.2 Sizing and observed adoption

Gartner forecasts worldwide end-user spending on AI models and platforms at $64.252 billion in 2026, up 63.4% from $39.311 billion in 2025. Within that total, foundation generative-AI models are forecast at $23.356 billion and specialized generative-AI models at $4.910 billion. These category figures are not DeepSeek revenue or serviceable market. Direct platform observations provide a narrower adoption lens: Vercel reports that DeepSeek moved from under 1% to 17% of its gateway tokens in May 2026 while remaining near 1% of spend, and OpenRouter reports DeepSeek's token share rose from 9% in January to 18% in June. The evidence establishes usage momentum and price leverage, not monetization or durable share.[CM002, CM003, CM004, CM006, CM007, CM016]

2.3 Buyers, users, and budget ownership

Adoption splits into distinct purchasing paths rather than a single funnel. Application teams can trial a low-cost hosted endpoint; platform teams can route production workloads across several models; infrastructure and security teams can self-host weights when data control outweighs operational simplicity. Gartner says enterprise buyers increasingly evaluate cost, latency, performance, reliability, evaluation, and usage tracking together. Darden's Michael Albert argues that value is likely to accrue to end uses and to firms controlling platforms and data rather than automatically to the foundation model. Consumer distribution is visible—Digital in Asia reports 130 million monthly active DeepSeek app users as of May 2026—but the assigned evidence does not identify the payer, conversion rate, or budget owner behind that audience.[CM005, CM014, CM015]

2.4 Drivers, constraints, and valuation relevance

The strongest adoption driver is economical production usage. OpenRouter reports that V4-Flash represented 70% of DeepSeek's agentic token flow by the end of May 2026, while the State of AI Report cites an estimated 47% quarterly decline since 2023 in the cost of reaching a fixed benchmark score. Capability remains differentiated: the fetched Artificial Analysis comparison places DeepSeek V4.1 Flash tenth in both Finance & Accounting and Strategy & Ops, not at the frontier across every domain. Constraints are commercial and political as well as technical. CNBC reports a US federal civilian-agency ban covering DeepSeek technology and third-party products built with its models. For valuation, broad category growth and routed-token share should therefore be treated as demand signals, not as substitutes for revenue, retention, margin, or a verified serviceable market.[CM008, CM009, CM010, CM011, CM016]

2.5 Exhibits

3.1 Competitive landscape and differentiation

DeepSeek sits in a crowded model market rather than a one-company category. The available evidence describes a direct frontier set of OpenAI, Anthropic, Google, xAI, and Meta, alongside Chinese peers including Moonshot's Kimi, Alibaba's Qwen, Doubao, and other open-weight providers. DeepSeek's differentiated choice is to release weights under MIT licensing while competing on inference economics; a comparison of Grok, Meta AI, and DeepSeek describes DeepSeek as the only one of that set offering self-hostable weights and identifies Meta's embedded distribution and Grok's live X data as different competitive advantages. This is a buyer-choice distinction, not proof that DeepSeek wins every benchmark or workflow. Its durability is mixed. A refreshed technical profile says Moonshot Kimi 2.6 and Alibaba Qwen have narrowed the gap and that the Chinese market is now defined by multiple capable providers. DeepSeek's efficiency and openness are still described as difficult-to-replicate advantages, while Alibaba's cloud-only proprietary releases represent the opposite strategic bet. Internal build and self-hosting are substitutes for hosted API consumption, but they transfer infrastructure, security, and operations responsibility to the buyer.[CP001, CP002, CP003, CP004, CP005]

3.2 Capability, price, distribution, and trust trade-offs

Price-performance is DeepSeek's clearest commercial wedge. The retained comparison reports V4 Pro API pricing of $0.435 per million input tokens and $0.87 per million output tokens, versus Grok at $2 and $6 in the same table; it also reports V4 Flash at $0.14 and $0.28. Another technical comparison estimates that V4 can reduce inference costs by 90% or more versus top US models for high-volume workloads, but explicitly cautions that benchmark parity does not guarantee task parity. These are published comparisons and estimates, not audited realized customer economics. Distribution and compliance pull in the opposite direction. Meta AI benefits from being embedded in WhatsApp, Instagram, Facebook, and Messenger, while Grok offers live X data and coding-workflow integrations. CNBC reports that a US federal action bans DeepSeek-related technology from government agency systems, including third-party products and services built using its models. That is a direct loss of federal procurement access and a reputational signal for risk-sensitive private buyers, although the reporting does not describe it as a blanket private-sector ban.[CP006, CP007, CP008, CP009, CP010]

3.3 Switching costs and moat durability

Switching costs are highest where a buyer has embedded model-specific prompts, evaluation suites, fine-tuning, or self-hosted serving infrastructure. MIT-licensed weights lower licensing lock-in and allow a buyer to move from hosted API usage to controlled infrastructure, but that path is not frictionless: the evidence notes that self-hosting a frontier-class model is operationally substantial. Hosted customers retain exposure to provider availability, rate limits, data-routing rules, and changing model behavior. The main displacement risk is that openness and low price become table stakes. The evidence says Kimi, Qwen, Doubao, and other Chinese providers have caught up, while US frontier vendors retain advantages in mature enterprise support, alignment, and some complex agentic workflows. The October 2026 federal exclusion also shows that model lineage can become a procurement constraint even when weights are open and easily repackaged. DeepSeek therefore has a strong technical wedge but an uncertain enterprise moat: retention depends on workload-level evaluations, deployment control, and trust rather than on benchmark leadership alone.[CP011, CP012, CP013, CP014]

3.4 Competitive diligence gaps

Public evidence is sufficient to map the strategic trade-offs but not to underwrite durable share or customer retention. The assigned sources do not provide comparable audited win rates, net retention, enterprise contract volumes, migration rates, or a consistent independent benchmark across all named vendors. Several comparison pages are vendor-adjacent or analytical summaries, so their pricing and benchmark claims should be re-tested on representative workloads. The federal ban is better evidence of procurement posture than of technical inferiority. A decision-maker should require a workload bake-off covering quality, latency, safety, data residency, total serving cost, and migration effort before treating the price gap as a durable moat.[CP006, CP009, CP010, CP014]

3.5 Exhibits

4.1 Revenue model and public traction

DeepSeek monetizes access to its models through usage-based APIs while also distributing a free chatbot and open-weight models. The API mechanism is explicit: developers and businesses pay for processed tokens, whereas the free surface is a distribution and adoption channel rather than recognized revenue. Public reporting gives a directional traction signal but not an audited revenue bridge. StoneX relays a September 2026 estimate of approximately $500 million of annual recurring revenue and separately cites 475 million yuan, or $70.7 million, of revenue in the first seven months of 2026. Those measures are not interchangeable: ARR is an annualized run-rate and the seven-month figure is reported revenue. Neither source establishes customer concentration, retention, recognized gross margin, or the share of revenue from API versus other channels. The underwriting implication is that usage appears commercially meaningful, but revenue quality remains unverified until management provides accounting definitions, cohort data, and customer-level concentration.[CI001, CI002, CI003, CI004]

4.2 Pricing, cost drivers, and unit economics

The public pricing evidence supports a low-price, high-volume strategy rather than premium software monetization. Wikipedia records a usage-based API and approximately $0.28 per million input tokens and $0.42 per million output tokens as of February 2025; a later Tech Insider analysis reports $0.55 and $2.19 for a different R1 comparison, so the figures should not be treated as one stable tariff without model, date, and billing-schedule reconciliation. The same reporting describes a 671-billion-parameter mixture-of-experts model with 37 billion parameters activated per token and a reported $5.6 million V3 GPU-rental cost, while noting that the training-cost claim is debated. These are efficiency and cost proxies, not DeepSeek's gross margin. Electricity, chips, data-center leases, research compensation, inference utilization, discounts, cloud take rates, and support costs are not publicly quantified. The margin path therefore depends on whether algorithmic efficiency and scale offset aggressive pricing and frontier-model capacity requirements.[CI005, CI006, CI007, CI009]

4.3 Capital intensity and financing dependency

DeepSeek's financing profile has shifted from founder-linked internal funding toward external capital, but the latest reported raise still does not establish capital sufficiency. A Forbes profile states that DeepSeek raised $14.9 billion in October 2026 at a $74.7 billion valuation. That is a funding report, not a cash balance: it does not disclose the amount funded at close, restrictions on proceeds, debt-like obligations, committed infrastructure purchases, or remaining liquidity. StoneX separately relays approximately $1.6 billion of expected AI infrastructure expenditure for 2026 and a 715 million yuan net loss for the first seven months of the year. The periods and accounting measures must remain separate. Net loss is not cash burn, expected infrastructure expenditure is not total operating cost, and funding raised is not cash available. Forward underwriting therefore requires current cash, monthly burn, working-capital needs, debt and capex commitments, and the milestone that would trigger another financing.[CI010, CI011, CI012, CI013, CI015]

4.4 Underwriting verdict and diligence blockers

Revenue quality is promising but not yet underwritable on public evidence. The API is a plausible recurring mechanism, and the open-weight strategy can widen adoption, but free distribution and low per-token prices make usage volume an insufficient substitute for realized revenue, contribution margin, and retention. Capital intensity is the more immediate financial risk: reported infrastructure spending and losses are large relative to the limited public operating detail, and the financing reports conflict on transaction size and valuation. Regulatory restrictions create an additional route to revenue volatility. CNBC reports that a July DHS order made public in October bars federal civilian agencies from procuring DeepSeek products and requires removal of DeepSeek-related technology, citing sensitive-data transfer and cybersecurity concerns. That order does not quantify total commercial revenue exposure, but it is adverse evidence against assuming unrestricted enterprise or public-sector distribution. Underwriting should remain conditional on audited revenue recognition, model-level price cards, gross margin by deployment mode, customer concentration, cash and commitments, and a financing plan tied to next-model capex.[CI015, CI016, CI017, CI018]

4.5 Exhibits

5.1 Product surface and maturity

DeepSeek is simultaneously a research lab, a model family, a free web and mobile assistant, and an API platform. The official transparency page lists DeepSeek Web, Harness, the API Platform, API documentation, downloads, service status, and a sequence of released models; it also records V4.0 as released on April 24, 2026 and V3.2 on December 1, 2025. A Chinese company profile describes the product progression from the first DeepSeek LLM through V2, V3, and R1, alongside app and API access. This is a broad product surface rather than a single SKU, but the public material does not disclose service-level reliability, customer-level deployment counts, or a complete release-management history. [CE001, CE002, CE003]

5.2 Architecture, efficiency, and deployment

The strongest technical evidence is the DeepSeek-V3 technical report, which specifies a 671B-total-parameter mixture-of-experts model with 37B parameters activated per token, 14.8 trillion pre-training tokens, MLA and DeepSeekMoE architectures, auxiliary-loss-free load balancing, multi-token prediction, and 2.788 million H800 GPU hours for the full training run. IBM's technical explainer provides the mechanism: routing only selected experts reduces computation, while the full model still creates a memory bottleneck. NVIDIA's deployment account shows the trade-off in production: the full R1 model can run as a NIM microservice on an eight-H200 system, with reported throughput up to 3,872 tokens per second. This is a credible efficiency story, not a claim that frontier serving is lightweight. [CE004, CE005, CE006, CE007, CE008]

5.3 Trust, security, and implementation risk

Open weights improve inspectability and portability, but they also transfer more control and responsibility to deployers. Stanford HAI faculty describe the open technical paper and weights as valuable for reproducibility while calling DeepSeek opaque on privacy protection, data sourcing, and copyright. The CSIS analysis raises separate concerns about weak safeguards and jailbreak exposure, while CIO reports that the October 2026 DHS order covers federal procurement, removal, and third-party implementations and cites data-sovereignty and security concerns. Those sources do not prove that every deployment is compromised; they do establish that enterprise adoption requires model provenance, data-flow, permission, monitoring, and fallback controls. The public record also lacks a complete independent security-assurance package, incident history, and reproducible end-to-end training pipeline. [CE009, CE010, CE011, CE012, CE013]

5.4 Product and technology verdict

DeepSeek's product advantage is a combination of capable open-weight models, unusually visible research artifacts, and engineering choices that lower active compute relative to total capacity. The architecture is differentiated enough to matter, but not self-sufficient: memory, multi-GPU networking, inference operations, data governance, and safety controls remain part of the buyer's problem. Official materials show continuing releases and a developer-facing surface, while independent technical and policy sources show why deployment maturity should not be inferred from benchmark performance or open licensing. The underwriting conclusion is therefore selective: DeepSeek is technically credible for experimentation and controlled self-hosting, but production use in sensitive workflows depends on verification of data handling, model integrity, security testing, service continuity, and the exact dependency chain behind any third-party product. [CE006, CE009, CE011, CE013]

6.1 Buyer, user, and payer segmentation

DeepSeek serves at least three distinct customer surfaces rather than one homogeneous customer base. Consumers use a free web and mobile chatbot; developers and businesses can call usage-priced APIs; and technical teams can download or self-host open-weight models. The available evidence supports broad user interest and developer accessibility, but it does not disclose a verified split between free users, paid API accounts, enterprise contracts, or self-hosted production deployments. That distinction matters because downloads and model followers measure reach, not recurring revenue or durable usage. The segment with the clearest conversion potential is high-volume technical work where low inference prices, local deployment, or model customization can offset integration and governance costs. The enterprise segment is more constrained: public-sector restrictions and privacy questions can block procurement even where model capability is attractive.[CU001, CU002, CU003, CU004]

6.2 Adoption trajectory and evidence quality

Public usage indicators show a sharp launch-driven adoption event, followed by an incomplete commercial picture. Sensor Tower estimated more than 23 million global mobile downloads in DeepSeek's first 19 days and a more than 700% week-over-week increase in average mobile daily active users during the measured January 2025 window. AppFigures separately estimated more than one million downloads across the App Store and Google Play at an earlier point. These measures are useful for distribution and awareness, but they are not active-user cohorts, paid accounts, retention, or revenue. The later refresh pool contains a June 2026 user-ranking page, yet the fetched evidence does not provide a comparable current denominator or a verified retention series. Investors should therefore treat adoption as high-reach but low-visibility on durability and monetization.[CU005, CU006, CU007, CU008, CU016]

6.3 Named customer proof and deployment quality

The evidence identifies user populations and access channels more reliably than named commercial customers. DeepSeek's public chatbot and API are available to consumers, developers, and businesses, while open-weight models can be downloaded and deployed by third parties. That is evidence of product availability and ecosystem reach, not proof that a named enterprise has deployed the product in production or achieved an outcome. The assigned sources do not supply three independently verified customer case studies with named buyers, production status, measurable outcomes, and contract context. A disciplined underwriting view is to separate public distribution from customer proof: the former is confirmed, while the latter remains a material diligence gap. A customer reference list, anonymized account cohort, and deployment telemetry would materially improve confidence.[CU001, CU002, CU009, CU010]

6.4 Retention, expansion, and concentration risks

No retained source reports NRR, GRR, churn, renewal rates, contract length, cohort retention, or customer concentration. The open-weight model and free-chatbot strategy can expand the top of the funnel while allowing users to switch among models or self-host, so reach should not be converted into retention without account-level evidence. Expansion may occur through API usage, local deployment, or enterprise workflow integration, but the sources do not quantify conversion between those surfaces. Conversely, regulatory and security restrictions can reduce enterprise expansion: CNBC reports a U.S. federal civilian-agency ban that also covers third-party products and services built using DeepSeek technology, while other reporting documents restrictions tied to sensitive-data and privacy concerns. The immediate diligence priority is a segmented cohort bridge from download or registration to active use, paid usage, renewal, and expansion, with customer and geography concentration disclosed.[CU011, CU012, CU013, CU014]

6.5 Exhibits

7.1 Regulatory and legal exposure

DeepSeek faces a widening policy perimeter in which model access, data handling, and downstream deployment are treated as national-security or privacy questions rather than ordinary software procurement. The October 2026 DHS action reportedly requires federal civilian agencies to remove covered DeepSeek products and prohibits both direct acquisition and third-party offerings developed using DeepSeek implementations, extending exposure beyond the branded application (CR001, CR002, CR003). The reported rationale includes possible sensitive-data transfer to China, intelligence value, and embedded vulnerabilities, but the same reporting notes that implementation details and a specific compliance deadline were not yet published (CR002, CR003). This creates a moving compliance surface for enterprise customers and integrators. The risk is high because an exclusion can remove a channel even when model quality and price remain competitive; the monitoring trigger is a final rule, agency implementation guidance, or expansion to commercial critical infrastructure. DeepSeek's own terms say service availability may vary by jurisdiction and that the company may modify, suspend, or terminate services as laws and regulations change (CR006). The diligence path is to obtain the DHS docket, map affected jurisdictions and indirect model dependencies, and test whether customer contracts allocate data residency, indemnity, and continuity obligations.[CR001, CR002, CR003, CR004, CR006]

7.2 Operational, quality, and security risk

The product combines meaningful technical efficiency with unresolved safety and reliability questions. IBM describes DeepSeek-V3 as a 671-billion-parameter mixture-of-experts model with 37 billion active parameters, while also noting that mixture-of-experts systems retain a large memory footprint and are harder to train and fine-tune (CR007, CR008). That architecture can support lower inference cost, but it leaves compute, memory, and operational complexity as material constraints. IBM also reports that DeepSeek-R1 has performed poorly in security and jailbreaking tests, an adverse signal for enterprise deployment even though the article does not provide a standardized scorecard (CR009). The terms promise industry-practice cyber-security measures and stable operation, but expressly provide no warranty that the services will remain available in particular jurisdictions (CR010, CR011). A prior cyberattack disrupted service and led to restricted new-user registration according to the maintained public chronology (CR012). Residual exposure is therefore medium-high: open weights can enable local deployment, but customers must own evaluation, patching, abuse monitoring, fallback routing, and model provenance. Kill criteria should include a material incident without timely disclosure, repeated availability failures, or an evaluation showing unacceptable harmful-output or data-exfiltration rates in the intended workload.[CR007, CR009, CR012]

7.3 Partner, financial, and execution dependencies

DeepSeek's model economics are attractive but do not establish durable margins or sufficient capital. Gartner forecasts worldwide end-user spending on AI models and platforms at $64.252 billion in 2026, while explicitly warning that providers face pressure to demonstrate sustained use and durable margins as usage becomes more cost-driven (CR013, CR014). Vercel's anonymized AI Gateway data shows DeepSeek's token share rising to 17% in May 2026 while its spend share stayed near 1%, and reports a 49% token share but only 4% cost share in coding-agent traffic (CR015, CR016). Those figures indicate usage traction and severe price capture pressure, not revenue or cash generation. A separate report describes a proposed first external round of at least $300 million at a valuation above $10 billion, but the article says the company had not confirmed an official timeline and declined requests for comment (CR017, CR018). CNBC later reported that DeepSeek reportedly closed a first external round valuing it above $50 billion, while stating it could not independently verify the underlying no-poaching report (CR019, CR020). The conflicting financing signals make capital structure, ownership, and available cash a diligence gap. Talent concentration is also material: CNBC reported that the company sought funding after losing key researchers and that a reported investor condition barred poaching staff or encouraging spinouts (CR021, CR022). Monitor disclosed financing documents, retention of named technical leaders, GPU availability and utilization, price changes, and customer-level gross margin; a failed next financing, sustained senior-researcher attrition, or inability to secure compliant compute would break the operating thesis.[CR015, CR016, CR020, CR022]

7.4 Residual exposure and decision triggers

The highest-priority diligence is not another benchmark comparison; it is proof that DeepSeek can convert open-weight adoption into compliant, dependable, and financeable production usage. Stanford's policy brief says commercial strategies for translating adoption into business success are emerging but long-term viability remains uncertain, and it cautions that government support is not guaranteed (CR023, CR024). The appropriate mitigation is a deployment-control package: customer-specific data residency and logging, independent safety evaluation, incident response, model and dependency inventories, multi-provider failover, and evidence of renewal or paid usage by segment. A risk can be downgraded only when the company provides auditable answers, not when a model leaderboard improves. Key thesis-break triggers are a binding expansion of public-sector exclusions, a confirmed material privacy or security incident, a sustained gap between token volume and monetization, inability to retain core researchers, or a capital event that materially changes control or financing terms. Until those tests are passed, the investment implication is to underwrite technical capability separately from commercial durability and to treat any valuation as exposed to policy and execution discounts.[CR023, CR025, CR026, CR027]

7.5 Exhibits

8.1 Investment thesis and anti-thesis

The positive case rests on a differentiated combination of open-weight distribution, efficient model architecture, and a usage-priced API that can convert ecosystem adoption into revenue. Sacra estimates $1.00B of annualized revenue in September 2026 and reports a 50%+ gross margin on model sales by July, with 82.9% API gross margin during the first seven months; these are analyst estimates, not audited company disclosures. The countercase is equally material: IBM notes that mixture-of- experts reduces compute but not memory costs, while CNBC reports a U.S. federal civilian-agency ban that includes third-party products built with DeepSeek models. The result is strategic reach with uncertain direct value capture and a meaningful discount for regulatory, security, and disclosure risk.[CV001, CV002, CV003, CV004]

8.2 Financing context and entry discipline

The financing evidence is inconsistent across the refresh pool and should not be collapsed into one precise valuation. Sacra reports a roughly RMB350.88 billion ($51.82 billion) post-money valuation from a June 2026 external round and says a second round was being sought at roughly $75 billion; an April 2026 report instead described advanced discussions for at least $300 million above $10 billion. The difference may reflect separate financing stages, stale reporting, or source error, so the current price is a blocking diligence question rather than a settled fact. Publicly available evidence also does not disclose liquidation preferences, option dilution, debt, or a cap table sufficient to translate headline valuation into common-equity value. Entry discipline therefore requires verified term sheets and audited revenue definitions before assigning a buy price.[CV005, CV006, CV007, CV008]

8.3 Scenarios and comparable evidence

A scenario framework is more defensible than a point estimate because the evidence supports directional drivers but not a complete forecast. The bull case assumes API demand converts into durable, high-margin usage while open-weight distribution expands enterprise and sovereign deployments. The base case assumes technical leadership remains valuable but direct monetization is constrained by self-hosting, substitutes, and geopolitical limits. The bear case assumes valuation compression as government exclusions, security scrutiny, or capital constraints reduce addressable demand. Public comparables are imperfect: TechNode reports MiniMax at more than $11.5 billion after its Hong Kong debut, with a $512 million net loss in the first three quarters of 2025, while StepFun reporting describes a nearly $2.5 billion financing round and an intended Hong Kong listing. These are market context, not directly comparable operating multiples for DeepSeek.[CV009, CV010, CV011, CV012]

8.4 Decision, exit readiness, and thesis breaks

The appropriate decision is track/research-more, not buy, because valuation support is less reliable than the product signal. A price-sensitive upgrade requires three confirmations: audited revenue separating recognized revenue from annualized run rate, financing documents that establish ownership and preference economics, and evidence that enterprise customers can deploy the models without unacceptable data, export-control, or security exposure. IBM cautions that the widely repeated $5.576 million training figure is only an estimate for a final pre-training run and excludes prior research, data, infrastructure, salaries, and facilities. A thesis break would be a failed financing, sustained loss of access to critical compute, a material security incident, or evidence that usage growth does not translate into paid, repeatable contribution margin. Exit readiness is therefore limited until reporting and governance improve.[CV013, CV014, CV015, CV016]

8.5 Exhibits