Shengshu Technology

1.1 Identity, Headquarters, and Product Scope

Shengshu Technology is a Beijing-based generative AI company incorporated on March 6, 2023, with its registered address in Haidian District at Dongsheng Building on Zhongguancun East Road. Multiple sources tie the company to Tsinghua University research roots and to the earlier U-ViT architecture work that predated Vidu commercialization. The company's core product is Vidu, a multimodal video generation platform spanning text-to-video, image-to-video, and reference-to-video workflows, later extended into API, agent, and real-time avatar products. Public company materials position Shengshu as selling both MaaS and SaaS, while third-party profiles indicate a mix of subscription revenue and corporate usage. The identity question matters because later chapters depend on a clean baseline: this is not just a consumer toy app, but a research-heavy Chinese video-model startup with a commercial platform, a Beijing legal entity, and a product family already pushed into global distribution.[CO001, CO002, CO003, CO006, CO007, CO008]

1.2 Founders, Leadership, and Organizational Setup

Leadership disclosure is meaningful but not perfectly clean. Public English-language reporting in late 2024 described Jiayu Tang as Shengshu's co-founder and CEO, while 2026 CNBC coverage quoted Zhu Jun as founder and company statements describe him as founder and chief scientist. Dealroom reconciles the picture best: Tang is presented as co-founder and former CEO, Bao Fan as CTO, Zhu Jun as chief scientist with deep Tsinghua affiliation, and Luo Yihang as the executive brought in from ByteDance's Volcano Engine to lead research, product, and commercialization as CEO while Tang shifted to president. That leadership transition suggests Shengshu moved from founder-led technical incubation into a more scaled operating structure during 2025-2026. What remains missing is public board composition, formal governance structure, or disclosed control rights, which leaves key-person concentration and governance diligence open.[CO003, CO004, CO005, CO027, CO036, CO041]

1.3 Funding, Capitalization, and Stakeholders

Shengshu has raised capital in frequent steps since 2023, but public disclosure quality is uneven. The cleanest hard disclosures are the February 2026 Series A+ of more than RMB 600 million and the April 2026 RMB 2 billion Series B led by Alibaba Cloud, while earlier financing history is reconstructed from Baiduwiki and Dealroom. Those sources point to an angel round near RMB 100 million in June 2023, an angel-plus round in August 2023, several-hundred-million-yuan rounds in March and June 2024, and a several-hundred-million-yuan Series A in September 2025. Dealroom labels Shengshu a unicorn and gives a wide $1-2.5 billion valuation band, but CNBC explicitly notes that the company declined to disclose valuation with the April 2026 round. The practical diligence conclusion is that Shengshu is clearly venture-scale and strategically important enough to attract Alibaba Cloud, Qiming, Baidu-linked investors, the Beijing AI fund, and multiple industrial partners, yet investors still face material opacity on exact cumulative proceeds, dilution, liquidation preferences, and present ownership control.[CO015, CO016, CO017, CO018, CO019, CO020]

1.4 Milestones, Commercialization, and Recognition

Shengshu's most notable strength is product cadence. After releasing Vidu in April 2024 and launching globally in July 2024, the company pushed Vidu 1.5 in November 2024, an enterprise API in February 2025, Vidu 2.0 in January 2025, reference-heavy Q-series upgrades through 2025-2026, TurboDiffusion in December 2025, the one-click Vidu Agent in December 2025, and the real-time Vidu S1 model in July 2026. Company statements tie that roadmap to accelerating commercialization: Vidu is claimed to operate in more than 200 countries and regions, more than 40 million creators and over 10,000 developers or enterprise customers are claimed by mid-2026, and named commercial users span ByteDance, Samsung, TAL, Alipay, JD.com, Amazon, L'Oréal, Tencent Animation, iQIYI, and Mango TV. External validation exists but is mixed. Shengshu's World Economic Forum Technology Pioneer selection and repeated benchmark citations support category relevance, while current Artificial Analysis snapshots place Vidu Q3 Pro in the upper tier but not at the very top globally. The commercialization story is therefore credible, but still anchored heavily in company claims rather than audited usage or revenue disclosures.[CO008, CO009, CO010, CO011, CO012, CO013]

1.5 Adverse Signals and Unresolved Disclosure Gaps

The main company-overview risks are not existential red flags but disclosure and execution caveats. Notebookcheck's hands-on test of Vidu in September 2025 concluded that the product could create striking visuals but remained too glitch-prone and inconsistent for dependable professional work, underscoring a persistent gap between benchmark marketing and production reliability. Public benchmark positioning also appears time-sensitive: company PRs framed Vidu Q3 as No.1 in China and No.2 globally in early 2026, but August 2026 Artificial Analysis snapshots show lower rankings for both text-to-video and image-to-video. Governance disclosure is thinner still: public sources reviewed do not identify the board, independent oversight, ownership rights, or audited financial statements, and even basic current metrics such as headcount and exact lifetime funding remain only partially observable. The jobs portal does suggest recruiting across Beijing, Shanghai, Shenzhen, and San Francisco, but that should not be mistaken for fully documented operating headquarters beyond Beijing. Investors should treat Shengshu as a fast-moving but still opaque private company whose strongest facts are about technical shipping velocity rather than institutional transparency.[CO020, CO029, CO033, CO034, CO035, CO036]

1.6 Exhibits

2.1 Market Boundary and Scope

Shengshu should be analyzed against the narrow AI video generator market, not against all video software. The narrow category, as defined by The Business Research Company and Research and Markets, consists of software and services that generate or transform video from prompts, images, slides, or related structured inputs using AI. That category already spans text-to-video, image-to-video, editing automation, and delivery models ranging from self-serve subscriptions to API and enterprise services. Research and Markets' adjacent "generative AI in video creation" lens is slightly narrower in one respect and broader in another: it emphasizes creation workflows and deployment types such as cloud versus on-premise, but it still covers end users from enterprises to individual creators and media companies. Official market surfaces confirm that the commercial category has moved beyond a single demo model. Runway sells creative SaaS, developer tooling, and robotics/simulation surfaces; Kling exposes API, 4K generation, and mobile distribution; Pika emphasizes agents and workflow automation; PixVerse markets CLI, agent, marketing-hub, and API workflows; and Jimeng optimizes for Chinese-language prompting, community remixing, and frame-control features. The practical inclusion rule for Shengshu is therefore: count spend on AI-native video generation, campaign creation, creator tooling, and developer/enterprise video APIs; exclude legacy NLE suites, surveillance analytics, CDN/streaming infrastructure, and generic social-media ad spend that never touches a video model.[CM001, CM002, CM003, CM004, CM020, CM021]

2.2 Market Sizing and Geographic Shape

Public market estimates cluster in a relatively narrow band only if the analyst uses a similar boundary. The Business Research Company sizes the AI video generator market at $0.85 billion in 2025, $1.04 billion in 2026, and $2.07 billion in 2030, implying 18.9% CAGR from 2026 to 2030. Its adjacent generative-AI-in-video-creation lens is smaller, at $0.39 billion in 2025, $0.47 billion in 2026, and $0.98 billion in 2030, with 20.4% CAGR. Fortune Business Insights lands between those lenses on the current base year, at $716.8 million in 2025 and $847 million in 2026, forecasting $3.35 billion by 2034 at 18.8% CAGR. The spread is meaningful but explainable: some publishers count only pure generator software, others include more creation services, and some regional splits move materially depending on whether analytics, enterprise workflow services, or consumer apps sit inside the boundary. Segment data, however, are directionally useful. Fortune reports text-to-video as 46.25% of the 2026 market, marketing and advertising as the largest application at 33.88%, social media as the fastest-growing application at 23.5% CAGR, and large enterprises as the largest customer class at 50.86% share while SMEs grow fastest at 21.1% CAGR. Geography is less consistent: TBRC calls Asia-Pacific the largest region for AI video generator tools in 2025, while Fortune gives North America a 41.0% share in 2025 and Asia-Pacific 20.9%, with China at $49 million in 2026. For underwriting Shengshu, that disagreement is a warning that top-down China TAM should be treated as directional rather than precise.[CM005, CM006, CM007, CM008, CM009, CM010]

2.3 Buyer, User, and Payer Architecture

The market is not one monolithic creator pool. It breaks into at least six commercially distinct segments with different budget owners and adoption triggers. First, marketing teams and agencies buy speed, iteration volume, and cost compression for campaign creatives; they are the most important near-term segment because analyst data say marketing and advertising is already the largest application. Second, retail and e-commerce operators want dynamic product showcases, localized short video, and catalog-scale asset production. Third, media, entertainment, and studio users care about higher-end controllability, consistency, and production workflows. Fourth, individual creators and prosumers enter through freemium or credit tiers and are highly sensitive to UX, trend tools, and mobile distribution. Fifth, developers and workflow builders buy APIs, CLI tooling, and orchestration primitives rather than consumer-facing subscriptions. Sixth, enterprise innovation, education, and internal-content teams use AI video for training, product explanation, and internal communication. Official competitor surfaces make this segmentation visible without guessing: Runway separates creative, dev, and robotics lines; Kling bundles consumer creation with API and mobile; Pika pushes agents plus apps; PixVerse exposes CLI, agent, marketing hub, and API; Jimeng optimizes for Chinese prompts and community remix loops. For Shengshu, this means Vidu competes in both B2C and B2B corridors, but the economically important buyers are probably enterprise marketers, developers, and media teams rather than casual free creators.[CM015, CM016, CM017, CM018, CM019, CM020]

2.4 Demand Drivers and Workflow Shifts

Demand is being pulled by both media consumption volume and workflow maturation. Pew finds YouTube used by 83% of U.S. adults, Facebook by 68%, Instagram by 47%, and TikTok by 33%; YouTube's own 2024 U.S. impact report says the platform's creator ecosystem contributed $55 billion to U.S. GDP and that YouTube paid more than $70 billion to creators, artists, and media companies between 2021 and 2023. This is the macro context for why AI video tools do not need to replace cinema first; they only need to lower the cost of constant short-form, campaign, education, and product-video production. The second demand driver is capability maturity. a16z wrote in March 2025 that the prior six months delivered major progress in video quality and controllability, that Chinese models Hailuo and Kling had already surpassed Sora in monthly web visits by January 2025, and that provider differentiation was emerging around prompt adherence, lip sync, and camera control. Official product pages reinforce that observation: vendors are shipping workflow features such as reference control, agentic editing, voice/lip-sync, ready-made ad templates, CLI batch execution, and API-first orchestration. That evolution shifts the category from novelty toward workflow software. Constraints still matter. OpenAI's shutdown of Sora's consumer product and pending API sunset in 2026 show that platform availability and commercialization paths can change quickly, and that market leadership is not permanent even for frontier labs.[CM023, CM024, CM025, CM026, CM027, CM028]

2.5 Regulation, Compute, and Cross-Border Access

Regulation is a first-order market variable for Shengshu because its home market is China and its product outputs public-facing synthetic media. China's Interim Measures for the Administration of Generative AI Services took effect on August 15, 2023 and explicitly apply to public services generating text, images, audio, video, and related content. The reviewed regulatory sources are unusually clear about operational obligations: providers must address lawful training data and IP provenance, personal-information protection, content safety, complaint handling, transparency, and where relevant filing or security-assessment requirements. China Law Translate's rendering of the measures makes the video-specific obligation explicit in Article 12: providers shall label generated images and video according to the deep-synthesis rules; Article 14 further requires removal of illegal content and action against abusive users. Those obligations hardened in March 2025 when China published labeling rules that take effect September 1, 2025 and require visible marks plus technical identifiers, while forbidding deletion or concealment of those labels. Cross-border supply adds another constraint. BIS guidance published in May 2026 reaffirmed that a U.S. export license remains required for covered advanced-computing items destined for China- or D:5-headquartered entities, even when those entities receive the items outside China. For a China-based video-model company, that is a direct input-cost and hardware-access risk, not an abstract geopolitics footnote.[CM029, CM030, CM031, CM032, CM033, CM034]

2.6 Shengshu-Relevant SAM and Unresolved Sizing Gaps

The most defensible Shengshu sizing lens is not an undifferentiated multi-billion-dollar global "AI video" TAM. A cleaner near-term lens starts with Fortune's $847 million 2026 AI video generator market, then isolates the 46.25% text-to-video slice to reach an estimated $392 million global text-to-video revenue pool. Applying Fortune's 20.9% Asia-Pacific share implies an approximately $82 million APAC-attributed text-to-video slice, while the same report separately references China at $49 million in 2026. A different but complementary lens isolates the 33.88% marketing-and-advertising slice, yielding roughly $287 million of 2026 spend tied to campaign-production workflows. Those calculations do not produce Shengshu's actual SOM; they simply show that the company's most immediately monetizable arena is probably a few hundred million dollars globally and materially smaller within the observable China/APAC slice. That still may be large enough to support a unicorn if market share, enterprise ACV, and global expansion are strong, but public data do not reveal enough about China-specific enterprise spend, free to paid conversion, API revenue mix, or non-China buyer willingness to adopt Chinese video models. Investors should use TAM work here as triangulation, not as a substitute for customer and revenue diligence.[CM036, CM037, CM038, CM039, CM040]

2.7 Exhibits

3.1 Landscape and Competitive Classes

Shengshu sits inside a crowded but still immature AI video landscape where direct, adjacent, and substitute competitors overlap. The direct peer class includes platforms that already ship text-to-video or image-to-video, creator-facing interfaces, and at least one of API, enterprise, or production workflow features. By that standard, the closest rivals are Runway, Kling, Hailuo/MiniMax, Pika, Luma, Jimeng, PixVerse, and Wan. Third-party evidence supports that grouping. Artificial Analysis compares Hailuo, Kling, Sora, Vidu, and Wan within the same benchmark family, and a16z's March 2025 consumer ranking called out Hailuo, Kling, and Sora as newly relevant web products while putting Runway on the Brink List. The market also contains important substitutes rather than only direct peers: open-model or research alternatives such as Stable Video Diffusion and Wan, internal build-outs by large AI labs or clouds, and the status quo of manual video production with editors, agencies, or in-house creative teams. For Shengshu, the strategic question is therefore not "who else can generate a clip?" but "which rival owns the buyer's preferred combination of quality, workflow depth, price clarity, localization, and distribution?"[CP001, CP002, CP003, CP004, CP022, CP023]

3.2 Direct Peer Profiles and War-Chest Tiers

The most important direct peers do not all compete on equal footing. Runway is the best-capitalized specialist in the set reviewed here, with a reported $315 million Series E at a $5.3 billion valuation in February 2026, a 60-million-plus creative user base claim, and an expanding world-model narrative that extends beyond media into robotics. MiniMax, the company behind Hailuo, appears even broader as a multimodal platform: Sacra describes it as a $4 billion-valued company with roughly $1.15 billion total funding and strong enterprise API traction, of which Hailuo is only one surface. Luma also looks well-capitalized and increasingly enterprise-oriented, with Owler reporting $1.1 billion total funding and a $900 million November 2025 round, while official pages present Luma as a professional creative-agent workflow rather than a novelty toy. Pika is much smaller financially, with Sacra reporting $135 million raised and a $470 million valuation, but it compensates with accessible packaging and Adobe Firefly distribution. Kling, Jimeng, and Wan disclose less investor detail in the sources reviewed, but their product posture suggests strong China ecosystem adjacency. Shengshu therefore competes against three different capital tiers: multibillion-dollar specialists, broad Chinese multimodal platforms, and lighter consumer-creator challengers.[CP005, CP006, CP007, CP008, CP009, CP010]

3.3 Capability and Packaging Comparison

Capability breadth is converging, but packaging remains divergent. Runway markets creative, developer, and robotics surfaces; Luma sells agents, team workspaces, and APIs; PixVerse exposes API, CLI, canvas, and marketing hubs; Pika emphasizes accessible effects, agentic content creation, and mobile-social remix; Kling combines video, image, sound, effects, and native 4K; Jimeng emphasizes Chinese prompt fluency and first-last-frame control; and Hailuo's associated MiniMax stack combines a consumer app with enterprise APIs. In practice, this means buyers are choosing between product philosophies, not only model outputs. Professional teams may favor Runway or Luma because workflow continuity, export formats, shared credits, or API controls matter more than novelty. Consumer creators may prefer Pika, Jimeng, or Hailuo because the interface and viral loop are easier. API-first builders may prefer Vidu, PixVerse, MiniMax, or Luma. Artificial Analysis and official documentation also imply that the frontier now spans quality, speed, and price simultaneously, which compresses any moat based solely on claiming "we have video AI." Shengshu's challenge is that most core features now exist somewhere else; its opportunity is to win specific user cohorts with China-native product fit and high-quality output.[CP011, CP012, CP013, CP014, CP016, CP017]

3.4 Distribution, Switching Costs, and Multi-Homing

Distribution power may matter as much as model quality. Pika's Adobe Firefly placement gives it access to a large professional creative ecosystem without building classic enterprise sales from scratch. Runway has both its own large creator base and external enterprise partnerships, while TechCrunch reports it is also expanding compute capacity and use cases into gaming and robotics. Luma's team workspaces, shared context, API, and enterprise plans push it toward a workflow-system position rather than one-off clip generation. PixVerse similarly bundles canvas, CLI, marketing hub, and API surfaces. By contrast, many creator-facing products remain easy to multi-home because their plans are credit-based, user interfaces are prompt-centric, and content can be regenerated elsewhere with modest switching effort. Stronger lock-in appears when a platform owns team collaboration, pipeline integration, brand assets, moderation controls, or enterprise commitments. That means Shengshu's defensibility will probably come less from casual creator subscriptions than from deeper workflow attachment via API, enterprise deployments, and any localized compliance or distribution advantages it can sustain.[CP020, CP021, CP024, CP025, CP028, CP029]

3.5 Moat Durability and Adverse Competitive Evidence

The adverse evidence is straightforward: Shengshu does not own a clean feature monopoly. Runway and Luma are pushing deeper into professional and enterprise workflows; Kling, Hailuo, and Jimeng bring culturally local product fit and strong consumer momentum; PixVerse is pushing workflow orchestration and API surfaces; and open-model alternatives such as Stable Video Diffusion and Wan reduce the cost of experimenting outside any managed SaaS platform. Even OpenAI's Sora, while no longer an active consumer product, helped reset buyer expectations on what a frontier video model should do. Public disclosures are also uneven. Official sites are usually strongest on feature marketing and weakest on realized enterprise pricing, net retention, or audited usage, which means some rows in any serious comparison remain unknown or secondary-sourced. The competitive verdict is therefore mixed: Shengshu has a plausible right to win in China-native video quality and commercial speed, but its moat is not secure unless it can combine model quality with durable distribution, enterprise workflow stickiness, and compliance execution better than rivals with larger ecosystems or deeper capital bases.[CP022, CP023, CP027, CP031, CP032, CP033]

3.6 Exhibits

4.1 Revenue Surfaces and Monetization Lanes

Public evidence supports a multi-lane monetization model rather than a single subscription product. Shengshu's own materials describe the company as operating both MaaS and SaaS, while the product ecosystem cited in the Series A+ release spans Vidu MaaS, Vidu SaaS, Vidu App, and Vidu Agent. The API launch release adds the most concrete pricing evidence: Vidu API was opened with no application required, entry access starting at $10, and base pricing of $0.05 per credit, with a four-second video consuming 4 to 40 credits depending on feature and aspect ratio. This is a classic usage-based developer funnel. On top of that, the Agent launch reframes video generation as a workflow product for ads, TVCs, e-commerce, and short-form media, which points to higher-value commercial budgets than pure prompt entertainment. Dealroom and company materials together imply that creator subscriptions and credits are not the whole story; corporate clients and platform-level services appear meaningful. The main revenue-model debate is therefore not whether Shengshu monetizes, but how much of revenue is low-ARPU self-serve creator spend versus higher-quality API and enterprise revenue.[CI001, CI002, CI003, CI004, CI005, CI006]

4.2 Pricing Architecture and Go-to-Market Proxies

Vidu's public pricing architecture is a blend of consumer and developer motions. The English pricing page confirms a creator-plan surface and numerous premium tools, but the accessible page does not expose a clean price ladder in readable text; by contrast, the API launch release gives direct per-credit and minimum-spend detail. That asymmetry itself is a useful signal: list pricing is public enough to support self-serve funnel entry, but realized economics likely depend on feature mix, enterprise arrangements, and off-page terms. The API launch also says Shengshu staffs a dedicated B2B service team, indicating an explicit enterprise motion rather than a pure PLG creator business. Customer examples in the Series A+ release—Pollo AI, PhotoGrid, OpenArt, Hubx, Fal.ai, Eachlabs, Freepik, and GensPark—suggest GTM can extend through developer platforms and partner applications, not only direct end-customer sales. The financial implication is that Shengshu may have better distribution leverage than a pure standalone app, but also potentially higher support and integration cost per serious account. Immediate API access broadens the top of funnel; whether it produces efficient payback depends on conversion, usage depth, and support burden, none of which are publicly disclosed.[CI003, CI004, CI012, CI013, CI014, CI015]

4.3 Traction Signals and Revenue Quality

Shengshu's disclosed traction is encouraging but heavily self-reported. The strongest commercial narrative comes from the Series A+ and Global Creativity Week releases: the company says 2025 users and revenue both grew more than 10x, that Vidu has reached more than 200 countries and regions, that it serves more than 40 million creators and over 10,000 developers and enterprise customers, that more than 500 million videos have been generated, and that commercial projects account for over 70% of output. If accurate, that commercial-project share is especially meaningful because it implies Vidu is used for monetizable work rather than only experimentation. The problem is not that these claims are implausible; it is that they are not audited. There is still no disclosed ARR, MRR, revenue mix by product, retention, enterprise concentration, or cohort conversion. Notebookcheck's hands-on review is a useful counterweight here: if production reliability still lags benchmark marketing, some top-of-funnel creator activity may not translate cleanly into durable paid usage. Revenue quality should therefore be treated as promising but unproven—better than a pure consumer toy if partner and enterprise claims hold, but still short of diligence grade.[CI008, CI009, CI010, CI028, CI029, CI030]

4.4 Cost Structure, Unit Economics, and Margin Pressure

The category economics around Shengshu are clear even if Shengshu's own gross margin is undisclosed. Frontier AI video is compute intensive, and both public cloud and peer pricing evidence show why. Google's official agent pricing illustrates the broader point that multimodal model usage is metered and optimized through spend-based commitments, cached tokens, and differentiated model rates rather than flat-cost infrastructure. Luma, Runway, Pika, MiniMax, and Vidu all use credits, usage-based charges, or both—because longer, higher-resolution, and more controllable outputs cost meaningfully more to deliver. MiniMax's docs and Sacra profile further suggest that a serious multimodal vendor carries substantial model-training and inference overhead, while Pika's Sacra profile explicitly warns about "unsustainable compute economics" as user demands scale upward. For Shengshu, this means revenue growth alone is insufficient. The key financial question is whether enterprise mix, workflow automation, and partner channels can lift realized revenue per unit of compute faster than pricing pressure and model-quality arms races compress it. Without gross margin, retry rate, GPU sourcing, or cost-per-generated-second data, the unit economics remain a structured unknown.[CI017, CI018, CI019, CI020, CI021, CI023]

4.5 Capital Adequacy and Financing Dependency

Capital adequacy is the strongest part of Shengshu's public financial story. Official and tier-one reporting show a >RMB 600 million Series A+ in February 2026 followed by a RMB 2 billion Series B in April 2026, a cadence that suggests the company raised a very large amount of fresh capital in a short window. CNBC also notes that the Series B proceeds were intended to support a "general world model," which implies that a meaningful share of funding is earmarked for frontier model R&D and compute, not just go-to-market scaling. That is important because world-model ambitions can consume capital well ahead of revenue. The positive interpretation is that Shengshu appears unlikely to be under immediate financing stress. The negative interpretation is that recent funding size may conceal equally heavy cost ambition, particularly if the company is trying to compete simultaneously on model quality, real-time systems, international expansion, and enterprise deployment. Public sources do not disclose cash on hand, monthly burn, debt, or runway. So while Shengshu looks well funded, it cannot yet be called capital efficient. The right judgment is "buffered but opaque."[CI025, CI026, CI027, CI035]

4.6 Financial Verdict and Diligence Blockers

Shengshu's financial posture is attractive enough to merit deeper diligence but too incomplete for conviction. Publicly, the business looks like a fast-scaling AI video company with real monetization surfaces, strong recent fundraising, global reach claims, and a potential mix shift toward higher-quality commercial usage. Those are the right ingredients for a venture-backed winner in generative media. But the missing variables remain the ones that determine whether this is a sustainable software business or an expensive compute story with thin margins: gross margin by product line, creator-to-paid conversion, enterprise ACV, partner revenue share, customer concentration, support cost, cash burn, and next-round trigger. Compared with the reporting discipline visible in public-company filing systems, Shengshu's disclosure remains sparse. The working verdict is that revenue quality is plausible, capital intensity is definitely high, and capital adequacy is currently strong; however, underwriting should stay conditional until management provides an actual financial model and cohort evidence.[CI027, CI028, CI029, CI035, CI036, CI037]

4.7 Exhibits

5.1 Product Definition and Module Map

In customer workflow terms, Vidu is not just an AI model; it is a creative operating stack. The public surface now covers classic prompt-based generation modes such as text-to-video and image-to-video, reference-heavy generation for character or object continuity, native-audio storytelling, workflow automation for ads and commerce, API access for developers, and real-time interactive avatar generation via Vidu S1. The Q3 page is especially helpful because it frames the product around end use rather than lab metrics: finished clips with dialogue, sound effects, and music generated together; longer 16-second single-pass outputs; camera-language control; multilingual output; and readiness for comic drama, short series, and narrative ads. The S1 / stream surface expands that definition again, positioning Vidu as a live interaction layer for voice-controlled digital characters rather than only batch clip rendering. That breadth matters because Shengshu appears to be solving several adjacent jobs-to-be-done: creative ideation, production acceleration, localization, ad-asset generation, interactive avatars, and developer embedding. The module map therefore looks more like a platform roadmap than a single-model feature list.[CE001, CE002, CE003, CE004, CE005, CE017]

5.2 Core Model Architecture and Engineering Stack

Shengshu's technology stack has unusually visible technical roots. The original Vidu paper describes the model as a diffusion system using U-ViT as its backbone, with explicit claims of 1080p generation and up to 16-second single generation, plus strong coherence and controllable-video experiments such as canny-to-video, video prediction, and subject-driven generation. Later company releases suggest that the commercialization layer has been built by adding controllability, consistency, and inference acceleration rather than by replacing that core architecture. Vidu 1.5 added multiple-entity consistency, multi-angle consistency, and advanced camera control. Vidu 2.0 then emphasized a full-stack inference accelerator, lower latency, lower cost, and template abstractions. By late 2025, TurboDiffusion and SageAttention pushed the engineering story further into inference infrastructure, with open-source materials detailing sparse attention, quantization, sampling-step distillation, and example deployment patterns. The key takeaway is that Vidu's differentiation is not one algorithmic trick. It is the compounding of multimodal video generation, reference consistency, control, audio, and systems engineering.[CE006, CE007, CE008, CE009, CE010, CE011]

5.3 Deployment, Integration, and Operational Workflow

Deployment evidence suggests that Shengshu is trying to behave like a production platform, not just a research lab. The API platform, CLI tooling, GitHub repositories, and help-center content show multiple operational layers: developers can access the API surface, automate uploads and tasks, choose model versions and durations, manage lip sync and text-to-speech jobs, and retrieve outputs programmatically. The MaaS API update adds another important integration signal: Model Context Protocol support lets tools such as Claude and Cursor choose video-generation modes through conversational workflows instead of manual API orchestration. That is strategically significant because it makes Vidu easier to embed inside agentic and workflow tools rather than only standalone web UI usage. Support and onboarding infrastructure are still lighter than large enterprise software norms, but the existence of CLI, skill, docs, and customer-support surfaces means Shengshu already ships an ecosystem around the core model. This section of the stack looks commercially useful today, though it still appears more startup-native than enterprise-hardened.[CE005, CE012, CE013, CE014, CE015, CE024]

5.4 Differentiation, Benchmarks, and Maturity Trajectory

Vidu's product differentiation is strongest where consistency, speed, and workflow packaging intersect. Public releases repeatedly emphasize subject or entity consistency, camera control, long single-pass generation, native audio, and reference-based production rather than generic prompt-to-video novelty. Artificial Analysis leaderboards provide third-party evidence that Vidu remains an upper-tier global contender, even if not unambiguously the global best. The GitHub and PR materials around TurboDiffusion strengthen the case that Shengshu is attacking one of the category's core bottlenecks—latency—at a systems level. Meanwhile, the roadmap from subject consistency in 2024, to Vidu 2.0 speed and templates in early 2025, to audio-capable Q3, MaaS lip-sync/MCP, and real-time S1 in 2026 shows a coherent move from offline clip generation toward a richer production and interaction platform. The maturity caveat is that the most impressive claims still lean heavily on company-controlled materials, while benchmark positions and public hands-on reviews suggest performance leadership is real but not settled.[CE008, CE009, CE010, CE015, CE019, CE025]

5.5 Trust, Safety, Privacy, and Quality Controls

Shengshu has some visible trust and control surfaces, but they are not yet the kind of public assurance package that reduces enterprise adoption friction on their own. The help center explicitly includes content moderation, credits, subscriptions, and blocked-account recovery, implying that moderation, fraud, and account-enforcement workflows do exist. The Vidu S1 page separately warns that the feature involves personal information processing, which matters because the product supports image uploads, voice selection, and voice cloning for digital-human interactions. However, reviewed public materials did not surface a public status page, public incident history, SOC 2 or ISO certifications, detailed model cards, red-team reports, or quantified safety-performance metrics. Notebookcheck's test is therefore relevant not only as a product review but as a quality-control warning: if prompt obedience, physical consistency, and reference fidelity still fail materially in live usage, customer trust and repeat adoption may lag benchmark marketing. The trust story is directionally present but substantively incomplete.[CE020, CE021, CE022, CE023, CE033, CE034]

5.6 Product-Tech Verdict and Blockers

Shengshu clears the most important product-tech bar for diligence: there is a real and fast-moving product platform here, not merely a frontier-model funding narrative. The evidence base supports a technically serious stack with a visible research core, repeat product releases, production-facing APIs and tooling, and concrete attempts to solve latency, consistency, and workflow packaging. That is a materially stronger posture than a startup whose story rests only on a demo website. The blockers sit one layer higher. Buyers and investors still lack confidence on deployment reliability, safety governance, enterprise controls, and how much of the technical edge is durable versus rapidly copied by better-capitalized peers. Product conviction should therefore be high on breadth and cadence, moderate on defensible technical lead, and only moderate-to-low on trust maturity until Shengshu exposes more production-grade control evidence.[CE019, CE020, CE023, CE025, CE032, CE035]

5.7 Exhibits

6.1 Segment Map by Buyer, User, and Payer

Shengshu's customer base is best understood as layered rather than monolithic. At the volume end are creators and self-serve users using Vidu for social content, short videos, and experimentation. At the higher-value end are developers, enterprise customers, marketers, and partner platforms embedding Vidu into their own workflows or end products. The company repeatedly markets to advertising, film, animation, e-commerce, mobile ads, cultural tourism, and education, which suggests Vidu is solving different jobs for different payers: creators buy access and credits, developers buy API consumption, and enterprise or commerce teams buy campaign throughput or workflow acceleration. This distinction matters because user count alone is not the same as revenue quality. Public sources imply that the largest segment by usage is creators, while the strategically most important segments are platform partners and B2B customers that can generate recurring usage at higher average spend. The segment map is therefore broad, global, and commercially interesting, but not yet disclosed finely enough to show where the business is truly anchored.[CU001, CU002, CU003, CU018, CU024, CU027]

6.2 Adoption Trajectory and Public Usage Scale

The adoption trajectory looks fast by any public standard, albeit still company-controlled. Shengshu says Vidu reached 1 million users within its first month, surpassed 10 million within three months, generated more than 100 million videos by month four, exceeded 100 million reference-to-video generations by month eight, and later surpassed 500 million total generated videos while serving more than 40 million creators and over 10,000 developers and enterprise customers. These are large numbers and suggest that Vidu is not a niche prototype. More importantly, the company says more than 70% of output now comes from commercial projects, which—if accurate—indicates meaningful usage beyond hobbyist experimentation. Still, the public record omits the denominators that matter for customer quality: paid share of those users, account activity frequency, active customer definition, enterprise deployment depth, and the portion of usage driven by partner platforms versus first-party Vidu properties. The adoption story is therefore best viewed as impressive scale with incomplete transparency.[CU003, CU004, CU005, CU006, CU016, CU026]

6.3 Named Customer Proof and Reference Quality

Named customer proof exists, but the quality of that proof varies meaningfully. The strongest evidence is not a Fortune 500 logo wall; it is partner or platform pages that publicly package Vidu as an available model or embedded capability. OpenArt has a dedicated Vidu video-generator page. each::labs publicly offers a Vidu model family with API access, positioning Vidu inside a unified developer platform. Shengshu's own releases say PhotoGrid embedded Vidu capabilities into its offering, and direct HTML inspection of PhotoGrid's AI-video page surfaced a Vidu Q3 tile even though the readability extract did not preserve it cleanly. By contrast, named examples such as Pollo AI, Odin, and long-form production teams are strategically interesting but much less independently verifiable from reviewed public sources. Lenovo is a useful separate category: it is less customer proof than channel or ecosystem proof, showing Shengshu can gain distribution through hardware and PC workflows. Overall, customer proof is real, but it is strongest in creator or developer platforms and weaker in fully documented enterprise case studies.[CU007, CU008, CU009, CU010, CU011, CU012]

6.4 Retention, Durability, and Repeat-Usage Gaps

Durability is the biggest unresolved customer question. Public sources provide plenty of scale signals and some named partner evidence, but almost no direct retention evidence. There is no disclosed NRR, GRR, logo churn, contract length, renewal rate, cohort curve, or satisfaction metric. The best public repeat-usage proxy is commercial mix: more than 70% of generated output is said to come from commercial projects, which would imply repeat usefulness if it is measured consistently. Support and billing surfaces also suggest the company has the operational scaffolding for repeat customers, not just one-time novelty traffic. But reliability concerns from Notebookcheck's hands-on review cut the other way: if reference fidelity, prompt obedience, and scene consistency still fail too often, customer usage may be broad but shallow. The right diligence posture is to treat retention as a major open question with only rough estimated proxies—not as a solved positive.[CU015, CU016, CU017, CU021, CU025, CU032]

6.5 Expansion Loops and Concentration Risk

Shengshu appears to have several land-and-expand paths, but the company has not disclosed enough to judge their economics. A creator may start in the self-serve Vidu interface, graduate to paid credits or premium features, then move into more structured workflows. A developer or small platform can start with the API, deepen usage through Vidu's broader model family, and eventually deploy at scale in a marketplace or app. Brand and commerce teams can be pulled in through Vidu Agent, lip-sync, or ready-made product-video workflows. These loops are strategically attractive because they allow the company to expand from consumer or creator usage into higher-value B2B contexts. The main counterweight is concentration opacity. Public sources do not disclose top-customer share, enterprise ACV, partner-revenue concentration, or whether key ecosystem channels can switch among competing video models with limited friction. The business likely has lower geographic concentration than many startups, but partner-platform dependence could still become a meaningful commercial risk.[CU018, CU019, CU020, CU021, CU022, CU023]

6.6 Customer Verdict and Diligence Priorities

The customer verdict is favorable on breadth and channel creativity, but incomplete on durability and quality of spend. Shengshu likely has a genuinely large creator top of funnel and increasingly credible partner-platform distribution, which are the two most important public positives. The best proof of real adoption is that outside platforms like OpenArt and each::labs publicly expose Vidu to their own users, while Shengshu's own releases point to broader adoption by creative platforms and commerce use cases. The weakest areas are classic diligence variables: paid conversion, renewal, contract depth, ACV, customer concentration, and independent ROI evidence. Investors should therefore underwrite customer reach and ecosystem relevance with moderate confidence, but keep revenue durability and reference quality as open diligence gates.[CU020, CU021, CU026, CU032, CU035]

6.7 Exhibits

7.1 Top Risk Picture and Why It Matters

Shengshu's risk stack is highly coupled. A regulatory event can become a customer problem; a compute shortage can become a product-quality and margin problem; a benchmark slip can become a financing problem. The most important insight is therefore not any one isolated risk but the transmission path between them. China-specific AI rules make compliance a product requirement, not merely a legal afterthought. U.S. export controls make infrastructure access a strategic variable, not just a procurement detail. Heavy competition means the company cannot simply slow down to be safer, because model and workflow leadership are part of the commercial thesis. That creates a classic frontier-AI tension: Shengshu must move fast enough to remain relevant while building enough moderation, labeling, privacy, support, and enterprise-control maturity to avoid accidents or enforcement. For investors, the practical takeaway is that the downside cases are correlated; when they hit, they are likely to hit more than one operating dimension at a time.[CR001, CR003, CR006, CR014, CR020, CR031]

7.2 Regulatory and Legal Risk

The Chinese regulatory burden on Shengshu is substantial and ongoing rather than hypothetical. The 2023 Interim Measures for Generative AI Services apply directly to public-facing video generation services and require lawful data sources, prohibited-content controls, privacy protection, complaint handling, and safe stable service delivery. The 2025 labeling measures go further by requiring explicit labels on generated video and implicit labels in metadata, plus checks by internet application distribution platforms when apps provide generative-AI services. Those labeling requirements rest on earlier Deep Synthesis and Algorithmic Recommendation rules, which also inform filing and safety assessment expectations. In practice, Shengshu faces three related legal exposures: first, content compliance and moderation failures; second, data provenance, privacy, or IP challenges if training or inference inputs are mishandled; third, process risk from incomplete algorithm-filing or labeling documentation. The regulatory environment does not prohibit growth, and Chinese rules explicitly talk about encouraging innovation, but it makes compliance a permanent operating cost with service-suspension risk if the company gets it wrong.[CR001, CR002, CR003, CR004, CR005, CR023]

7.3 Operational, Quality, Privacy, and Security Risk

Shengshu's product itself creates meaningful operational and trust risk. AI video generation is unusually exposed to quality failures because users can see prompt drift, physics errors, broken continuity, or weak lip sync instantly. Notebookcheck's hands-on review shows that this remains more than a theoretical issue. Vidu's own product direction also expands the risk surface: native audio, lip sync, voice cloning, live avatar interaction, and template-based automation all create new moderation, privacy, abuse, and reputation burdens. Public sources show that Vidu has a content-moderation surface and account-enforcement processes, which is better than having no visible controls, but they do not show the depth of model-evaluation, enterprise security assurance, or incident transparency that larger buyers often expect. The result is an operational profile where every new feature can improve commercial value while simultaneously raising the cost and difficulty of safe deployment.[CR009, CR010, CR011, CR012, CR013, CR026]

7.4 Partner, Dependency, Customer, and Financial Risk

The dependency layer is where geopolitical, financial, and customer risks converge. Shengshu needs advanced compute, frontier engineering talent, and continued product acceleration to compete with well-funded rivals such as Runway, Kling, Hailuo, and others. U.S. export-control enforcement makes GPU access and related infrastructure more fragile, while Shengshu's world-model ambitions likely keep compute demand elevated. At the same time, publicly visible customer proof is strongest in partner platforms such as OpenArt and each::labs, which is commercially useful but introduces platform dependence and potential switching risk. Financial opacity compounds the problem: the company has raised large rounds, but public sources do not disclose burn, gross margin, or customer concentration. That means an investor can see the likely pressure points—compute supply, partner-platform concentration, pricing compression, and margin dilution—without being able to precisely quantify them. This is a classic case where capital buffer mitigates near-term survival risk but does not eliminate model-risk or execution-risk intensity.[CR006, CR007, CR008, CR014, CR015, CR016]

7.5 People, Execution, and Kill Criteria

The final risk layer is organizational execution. Shengshu's product cadence suggests impressive engineering speed, but it also raises the bar for management systems, compliance operations, enterprise support, and talent retention. The company appears to have evolved from founder-led product momentum under Jiayu Tang into a broader operating phase where Yihang Luo is the public CEO voice; that shift is not automatically negative, but it does increase the need to understand governance continuity and decision rights. The company also needs to keep attracting specialized talent in model research, systems acceleration, safety or moderation operations, and enterprise delivery while rivals with deep capital pools compete for the same people. These are monitorable risks, which means they should feed directly into kill criteria rather than remain vague concerns. For this chapter, the core thesis-break events are export-control shock, regulatory enforcement, benchmark or product-quality slippage, failed customer durability, and evidence that frontier ambition is outrunning operational discipline.[CR021, CR022, CR030, CR041, CR042]

7.6 Exhibits

8.1 Recommendation and price discipline

The core valuation question is not whether Shengshu is interesting; it is whether the current public evidence supports paying an exact late-stage price. On that narrower question, the answer is still no. CNBC confirmed Shengshu's April 2026 Alibaba-led financing but also said the company declined to disclose valuation, while Dealroom only provides a broad public unicorn band of roughly $1 billion to $2.5 billion. At the same time, Shengshu has meaningful positives that stop this from being a simple pass: the company disclosed more than RMB 600 million of Series A+ financing in February 2026, roughly RMB 2 billion of Series B financing in April 2026, a 10x increase in users and revenue during 2025, and a product surface that already spans creator tools, Vidu API, and the ad-oriented Vidu Agent. That combination makes Shengshu watchlist-worthy but still price-sensitive. If investors are underwriting an internal mark around $1.5 billion to $1.7 billion, they are implicitly assuming that the company has already converted product quality and commercial usage into a substantial revenue base. Without audited revenue, gross margin, cap-table, or preference disclosures, there is no evidence-backed way to confirm that assumption. The recommendation is therefore Track rather than buy: keep the company active in coverage, but do not commit capital at an opaque mark unless management opens the revenue bridge, gross-margin profile, customer concentration, and dilution stack. The practical rule is simple: the better the price and the cleaner the financial package, the more Shengshu shifts from compelling story to underwritable asset.[CV001, CV002, CV003, CV004, CV005, CV006]

8.2 Comparable band and supportable range

The cleanest way to avoid false precision is to triangulate Shengshu across three reference sets: direct AI-video comparables, broader AI/software public multiples, and Shengshu's own disclosed commercial signals. On direct private comps, Runway sits at the top end with a February 2026 $5.3 billion valuation and roughly $860 million raised, Luma reached $4 billion with a $900 million Series C, MiniMax reached a later $4 billion growth mark with about $1.1 billion raised, while Pika appears much smaller at about $470 million valuation and $7.6 million of 2024 revenue. Shengshu belongs above Pika in visible scale and strategic capital, but it still lacks the disclosed revenue depth and investor-grade financial transparency seen even in the secondary reporting around Runway or MiniMax. Public market anchors are a useful reality check. Multiples.vc's August 2026 software data show public AI and design-engineering software trading around 4.0x to 4.2x next-twelve-month revenue, while the media-and-entertainment software overview reminds investors that revenue quality varies dramatically between high-margin creative subscriptions and much lower-margin cloud-rendering or infrastructure businesses. Shengshu is not a public SaaS company and should command a private-growth premium if its commercial usage converts well. But that premium cannot be infinite. A $1.5 billion equity value implies about $150 million of sustainable revenue at 10x, about $125 million at 12x, or roughly $107 million at 14x. A $2.0 billion mark implies even more. Those thresholds are plausible for a fast-scaling AI-video leader, yet public evidence does not confirm Shengshu has already reached them. The right conclusion is not that the company is overvalued with certainty, but that the valuation case currently rests on unverified revenue assumptions.[CV008, CV009, CV013, CV014, CV015, CV016]

8.3 Scenario ranges and downside transmission

Scenario work matters because Shengshu's best public metrics are activity signals, not audited revenue metrics. In the bull case, Shengshu's claimed >70% commercial-project mix proves to be a real revenue-quality signal, the developer and enterprise customer base converts into durable API and B2B contracts, product leadership stays near the top of AI-video benchmarks, and management can show gross margins moving toward software-like levels despite compute intensity. Under that path, a roughly $1.8 billion to $3.0 billion band becomes defensible and can support attractive upside from a lower-end entry. The base case is more conservative. It assumes Shengshu is real and growing, but that revenue quality is mixed across self-serve creators, platform partners, and enterprise accounts; compute and compliance costs remain heavy; and the private market applies a premium to public AI/software multiples without treating Shengshu like a fully disclosed scarcity asset. On that path, about $0.9 billion to $1.6 billion looks more supportable. The bear case is not a collapse story; it is a multiple-compression and opacity story. If commercial usage does not translate into durable ARR, if export controls tighten, if regulatory friction slows launches, or if competitive pricing forces thinner margins, a roughly $0.5 billion to $0.9 billion band becomes plausible. That is why the downside transmission runs mainly through missing economics rather than through product irrelevance alone.[CV005, CV008, CV009, CV010, CV017, CV018]

8.4 Exit readiness, diligence asks, and final view

Shengshu is not exit-ready from a diligence standpoint even though it may be exit-capable from a strategic-interest standpoint. The company clearly knows how to ship product, attract capital, and market commercial use cases. What is missing is the package an investment committee needs to underwrite a specific price: ARR by product line, net retention, enterprise concentration, gross margin, compute commitments, cash burn, board structure, and the preference stack created by rapid 2025-2026 financing. Filing infrastructure such as SEC EDGAR and HKEXnews shows the standard public investors eventually demand; Shengshu is still far short of that level of disclosure. The final view is therefore straightforward. Shengshu is one of the more credible Chinese AI-video companies and deserves continued coverage because the product, capital, and commercialization story are all real enough to matter. But the public evidence today supports interest, not conviction. Upgrade the name only if management provides audited or filing-grade economics, or if entry pricing falls low enough to create margin of safety despite the opacity. Until then, Track is the disciplined answer: strong company, incomplete valuation proof.[CV011, CV012, CV026, CV034, CV035, CV036]

8.5 Exhibits