Evoken

1.1 Identity, structure, and business model

Liblib should be treated in this report as the operating AI creative platform brand inside Evoken, the corporate group that had been known as Qidian Xingyu / Yanyu Technology before adopting the Evoken name in June 2026. The hard identity facts are Beijing headquarters, May 2023 operating-company formation, and a flagship LiblibAI product that began as an AI image-generation and model-sharing community. The business model is not a generic chatbot wrapper: it combines a creator community, model marketplace, cloud generation, LoRA training, and paid professional workflows. That structure matters because later chapters can assess market, product, and valuation around a coherent chain: community models and assets feed generation, video production, and design-agent use cases. The official site remains useful only as a company claim because it is JavaScript-only, so the durable identity base here relies on third-party news, registry-like Baike pages, and the CAC filing ecosystem.[CO001, CO002, CO003, CO004, CO005, CO006]

1.2 Founder, leadership, and governance signals

The founder story is unusually central to the investment case. Chen Mian is repeatedly framed as a product and commercialization founder, with prior responsibility for ByteDance's Jianying / CapCut global commercialization and a public narrative around speed, PMF discipline, and professional creator workflows. That background fits LiblibAI's motion: professional image and video creators are asked to pay for production infrastructure, not simply try a novelty model. The concentration risk is also real. Public materials point to Chen as controlling shareholder at roughly 73.95%, while Zhang Zijie appears in early legal-representative and core product-community disclosures and Yang Nan became legal representative in November 2025. The disclosed team bench is impressive but still high level: elite Chinese and overseas universities plus Tencent, Alibaba, and ByteDance backgrounds. Board composition, protective provisions, and post-Series B+ governance are not public, so the report should preserve founder dependence rather than infer institutionalized governance.[CO007, CO008, CO009, CO010, CO011, CO012]

1.3 Funding chronology and investor base

Evoken's financing history is the core timeline of record for the whole report. The company appears to have moved from a US$3.5 million angel round at about a US$15 million valuation in 2023 to multiple 2024 rounds that validated the AI-image community, then to several-hundred-million-RMB financing in early 2025. The October 2025 US$130 million Series B created a public application-layer financing benchmark in China and brought HSG / Sequoia China, CMC Capital, Ant, Lenovo Capital, Shunwei, Source Code, Mingshi, Yingce, and INCE into the story. The June 2026 Series B+ is the step-change: roughly US$300 million at more than US$2 billion post-money, led by Granite Asia, Tencent, and Shunwei, with HT Investment, Times Capital, Gaorong, Ant, and HSG / Sequoia China also cited. The investor map therefore mixes financial venture capital, strategic technology capital, and repeat insiders, while the stated use of proceeds keeps the focus on R&D, global expansion, and product portfolio build-out.[CO013, CO014, CO015, CO016, CO017, CO018]

1.4 Scale, product portfolio, and adverse flags

The headline operating metrics are strong enough to justify unicorn-level attention but must be read with caveats. The canonical current metrics are more than 30 million cumulative LiblibAI users, more than 500,000 original models, more than 5 million images generated per day, and ARR above US$300 million as of May 2026. Earlier markers included roughly 25 million users, 4 million MAU, and RMB 206 million of 2024 revenue, which helps reconcile why the 2026 ARR step-up is a major acceleration rather than a trivial restatement. The product portfolio now spans LiblibAI for community/model assets, LibTV for video production, Xingliu for domestic design-agent workflows, and Lovart as an overseas design-agent experiment. Two diligence negatives belong in chapter 1, not buried later: CCTV named LiblibAI in a porn-generation loophole exposé, and 2025 reporting said the company had not yet turned profitable. Both risks are compatible with fast growth; neither should be ignored. materially. This matters.[CO022, CO028, CO029, CO030, CO031, CO032]

1.5 Exhibits

2.1 Market boundary and sizing range

Liblib should not be valued against every yuan of entertainment or design spend. The investable boundary starts with China AIGC software and services that automate content creation, then narrows to image generation platforms, video and animation generators, model communities, and workflow tooling that creators or studios can buy. Public analyst estimates are wide: IMARC puts the 2025 China AIGC market at US$5.16 billion with a 15.96% CAGR to 2034, while the Grand View lower lens used in the source pool supports a roughly US$2.36 billion current-market anchor. That range is more useful than a single TAM because Liblib monetizes workflow adoption, not all generative-AI infrastructure, and because the company also sits next to a much larger content-spend pool in micro-drama production. A practical diligence model should therefore reconcile two routes: top-down analyst market forecasts and bottom-up wallet-share from studios, agencies, and enterprise creative departments. The first route proves that China AIGC is already large; the second determines how much budget is addressable by Liblib rather than by cloud infrastructure, foundation-model subscriptions, or internal creative labor.[CM001, CM002, CM003, CM004, CM005, CM006]

2.2 Micro-drama demand as the core B-side pull

The clearest B-side driver is China’s micro-drama economy. Multiple sources put the 2026 micro-drama market above RMB120 billion, or about US$16.5 billion, and describe the format as phone-native, high-volume, algorithmically distributed entertainment. AI matters because it compresses both cost and time: reports cite one-tenth live-action cost, one-fifth traditional shoot cost, and production cycles falling from three months to one. This changes buyer behavior. A studio can run many more pilots, outsource fewer low-value assets, and redirect budget toward prompts, model workflows, reusable characters, localization, and post-production. For Liblib, the implication is not that the whole micro-drama market is revenue; rather, that a large recurring production budget now has a reason to buy image, model, and video workflow capacity. This distinction matters because short-drama economics reward repeatable throughput more than one-off image quality. Buyers with daily episode pipelines value reusable characters, consistent style, fast localization, and low-friction collaboration, which are closer to workflow software than to consumer entertainment spend.[CM010, CM011, CM012, CM013, CM014, CM015]

2.3 Segments, buyers, and adoption path

Adoption splits across professional creators, studio buyers, enterprise marketing teams, and platform ecosystems. The end user may be a designer, prompt artist, producer, editor, or operations team, but the payer is typically a studio, agency, brand, game company, e-commerce merchant, or enterprise department that already budgets for faster content output. IMARC’s market segmentation confirms that image generation platforms and video generators are recognized AIGC solution classes, and 6Wresearch shows that Chinese image generation is shaped by Baidu, Alibaba, Tencent, ByteDance, SenseTime, Meitu, and other large ecosystems. Liblib’s market positioning therefore needs two filters: can it own creator workflow and asset reuse despite giant-model alternatives, and can it convert the creator community into paid teams in short-drama, advertising, design, and e-commerce workflows? The buyer journey also implies multi-homing at the model layer but higher stickiness at the asset layer: teams may test several image or video models while retaining the platform that stores approved templates, production histories, and community-trained models.[CM005, CM006, CM007, CM008, CM009, CM023]

2.4 Growth drivers, constraints, and market-shaping regulation

The growth case is strong but not frictionless. Government AI-plus policy, enterprise digitization, automated content needs, and local production subsidies all expand demand. At the same time, regulation is not an afterthought; it shapes which models can launch, how generated content is labeled, and which micro-dramas can be distributed. CAC filing records and Liblib’s own filing history show that compliance is a market-access prerequisite. The adverse case is also market-level, not just company-level: cheap generation can flood platforms with undifferentiated content, creating price competition and lower willingness to pay for generic outputs. For Liblib, the durable opportunity is to be a workflow and asset marketplace that saves money while raising output quality; the risk is being pushed into commodity tooling if buyers can multi-home across large-model ecosystems. The diligence priority is to separate demand creation from value capture. A rising tide in AI content can lift usage metrics, yet only retention, paid conversion, compliance resilience, and differentiated workflows will show whether Liblib captures durable margin.[CM029, CM030, CM031, CM032, CM033, CM034]

2.5 Exhibits

3.1 Landscape: direct model communities, global substitutes, and Chinese platform entrants

Liblib’s competitor set is not a simple list of image generators. The closest direct peers are model-sharing and creation communities that can host or train models, expose LoRA-style customization, and keep creators inside a marketplace or community loop. Tensor.Art, SeaArt AI, Wujie AI, and AituBo are source-backed in this chapter; 6pen and PixAI remain named in the requested scope but not fully profiled because the retained source texts do not contain verifiable current profiles. Global substitutes bracket the same job from two directions: Civitai emphasizes open model discovery and community control, while Midjourney emphasizes proprietary model quality and fast output. The China-platform threat sits above both groups, because Baidu, Alibaba, Tencent, and ByteDance can bundle image and video generation into e-commerce, gaming, social, and creator distribution where Liblib must earn usage rather than inherit it.[CP008, CP009, CP010, CP011, CP012, CP015]

3.2 Capability and scale comparison

Liblib’s scale metrics make it more than a hobbyist image tool: reported mid-2026 figures include more than 30 million cumulative users, more than 500,000 original models, more than 5 million images per day, and ARR above US$300 million. That scale is the basis for its professional studio narrative, but it does not create capability exclusivity. SeaArt has text, image, and video generation, LoRA training, face swap, AI characters, ComfyUI, and upscaling. Tensor.Art has model hosting, online training, ComfyUI workflows, and API services. Wujie adds a domestic copyright-registration path; AituBo attacks the beginner and low-budget segment. The strongest conclusion is that Liblib’s advantage is a China-focused density of community, models, and professional demand, not possession of unique feature checkboxes. The traffic benchmark cited in the brief needs a fresh independent traffic-panel pull before it is used quantitatively. This is also why pricing and plan mechanics matter: even where Liblib has stronger local supply, a low-cost rival can win exploratory usage and then upsell once a creator has built habits around its interface.[CP002, CP003, CP004, CP018, CP020, CP023]

3.3 Moat durability, switching costs, and multi-homing

Liblib’s moat is strongest where Chinese-language community, model supply, workflow know-how, and professional B-side demand reinforce one another. A designer or short-drama studio that already depends on Liblib-hosted models, LoRA training, workflow templates, and Chinese moderation/compliance infrastructure faces a higher switching cost than a casual prompt user. Yet the evidence also supports a multi-homing thesis. Reviews describe Civitai as the control and model-ownership option, Midjourney as the speed and quality option, and SeaArt/Tensor/Wujie/AituBo as overlapping direct tools. Those alternatives reduce Liblib’s ability to monetize every use case, especially when teams can mix one platform for exploration, another for polished outputs, and local Stable Diffusion pipelines for sensitive assets. The practical diligence issue is not whether Liblib has a moat; it is whether the moat is deep enough to protect take-rate and margins when feature parity arrives quickly. Multi-homing should therefore be treated as the default user behavior until Liblib proves exclusive assets, superior Chinese creator liquidity, or enterprise workflow data that competitors cannot easily copy.[CP014, CP015, CP016, CP018, CP023, CP026]

3.4 Adverse evidence and open diligence items

The adverse evidence is material. Skywork’s SeaArt review highlights content-safety loopholes around explicit material, and Baidu Baike records a separate CCTV-related content-safety issue for Liblib itself. Those facts do not prove SeaArt is stronger than Liblib, but they show that platforms competing for open community creation inherit moderation, licensing, and brand-safety burdens. Cybernews adds a commercial-risk angle by flagging SeaArt’s pricing and licensing clarity; Recatools flags Wujie’s dense Chinese-only workflow; and the Civitai/Midjourney comparisons show that professional users may rationally maintain multiple tools. The open-source and third-party model ecosystem also means workflows can be copied faster than communities can. Before underwriting a US$2 billion-plus valuation premium, diligence should refresh traffic metrics, directly test creator migration friction, collect enterprise contract retention data, and obtain current evidence for under-sourced rivals such as 6pen and PixAI.[CP007, CP019, CP021, CP027, CP035, CP036]

3.5 Exhibits

4.1 Funding runway is abundant, but capital has become part of the moat

Financially, Evoken is no longer an undercapitalized AI-art community; it is a late-stage application-layer company being funded as a category leader. The local funding facts needed for this chapter are straightforward: a small 2023 angel round, a cluster of 2024 and early-2025 rounds, a US$130 million Series B in October 2025, and a June 2026 Series B+ of roughly US$300 million at more than US$2 billion post-money. The important underwriting point is not simply the chronology, but what it says about capital adequacy. Management says the B+ proceeds will fund R&D, global expansion, AI creative product capability, and portfolio build-out. That gives Evoken a near-term survival buffer in a market where compute, model-API resale, creator subsidies, and paid traffic can quickly overwhelm self-generated cash. The cap table also matters: Granite Asia, Tencent, Shunwei, HSG, Gaorong, Ant, CMC, Source Code, Mingshi, Yingce, INCE, and Lenovo-related capital create unusually broad strategic and financial access, while Chen Mian's reported controlling stake keeps founder incentives concentrated.[CI001, CI002, CI003, CI004, CI005, CI006]

4.2 ARR scale is exceptional; recognized revenue and revenue mix remain thinner disclosures

The strongest financial fact is the step-change from 2024 revenue to 2026 ARR. Tencent News reports 2024 full-year revenue of RMB206 million, while multiple 2026 reports cite US$300 million of ARR as of May 2026 and more than 30 times, or more than 3,000%, year-over-year growth. The public story is therefore less about whether demand exists and more about how durable the mix is. Revenue appears to come from four linked streams: C-side creator subscriptions and credits, professional creator compute consumption, B-side API or customized enterprise services, and LibTV's production-workflow monetization. LibTV is particularly important because it moved the company from image-generation hobby and design workflows into short-drama, film, advertising, and brand-video budgets. Reported first-month single-day revenue above US$1 million, May revenue more than 13 times the launch-month level, and nearly 1,000 served teams suggest real willingness to pay. But revenue recognition by stream, retention, refunds, credit breakage, and enterprise contract terms remain private.[CI011, CI012, CI013, CI014, CI015, CI016]

4.3 The valuation is modest on ARR, but the margin question is the fulcrum

On headline math, Evoken does not look expensive relative to global AI creative peers: US$2 billion on US$300 million ARR is roughly 6.7 times ARR. That is below the Canva reference range of about 10.5 times to 16 times ARR and below Sacra's Gamma example at roughly 20.6 times. Midjourney's reported US$500 million revenue and US$10 billion valuation imply a much richer creative-AI benchmark. The discount is rational because Liblib is not yet proven to own the gross-margin stack. China Biz Insider's adverse read is central: Evoken does not train its own foundation models, and each generation job depends on third-party model providers, so the company's economics are a spread between retail pricing and wholesale model-API or compute costs. The best upside case is that community assets, workflows, and B-side production integration create enough lock-in to negotiate volume discounts and preserve spread. The downside case is that first-party model vendors cut prices, bundle the workflow, or force Liblib to subsidize usage to defend share.[CI026, CI027, CI028, CI029, CI030, CI031]

4.4 Financial verdict: real PMF, unproven contribution margin

The financial verdict is constructive but conditional. Liblib has the rare application-layer combination of public ARR scale, recent growth, multi-product expansion, and a blue-chip investor syndicate; these are not vanity signals if the professional-content budget evidence holds. The B+ round should cover near-term product and global expansion needs, and the comparatively low ARR multiple leaves room for upside if Evoken proves that it is becoming a creator operating system rather than a thin model reseller. The unresolved diligence items are also large enough to determine the investment case. Buyers need actual gross margin by product, wholesale model-API price schedules, cloud commitments, cohort retention, enterprise revenue share, credit liability, monthly burn, post-money cash, and runway. Regulatory continuity belongs in the financial chapter because CAC filing status and content-compliance failures can change availability, payment conversion, and enterprise procurement. Until those private inputs are verified, the chapter should underwrite revenue quality as strong, margin quality as unproven, and financing dependency as improved but still strategically important.[CI022, CI023, CI024, CI025, CI043, CI044]

4.5 Exhibits

5.1 Portfolio has expanded from image community to creator operating system

Evoken should be evaluated as a multi-product creative workflow company, not only as an AI image website. The flagship LiblibAI surface remains the core traffic and asset layer: public reports cite more than 30 million cumulative users, more than 500,000 original models, and more than 5 million images generated per day. LiblibAI 2.0 then reframes that community as a professional creation studio by adding integrated image and video generation, top video models, effects templates, and a broader open-plus-closed model menu. The adjacent products extend that same asset base into new budgets: LibTV targets video and short-drama teams, Xingliu targets domestic design-agent workflows, and Lovart carries the design-agent thesis internationally. The portfolio logic is therefore sequential: creator community creates assets, studio workflows raise frequency, and agents/video products monetize professional production use cases. This matters for diligence because the product promise is only durable if these workflow surfaces convert into repeated paid jobs, not one-time experiments subsidized by promotional compute or temporarily cheap upstream model access.[CE001, CE002, CE003, CE004, CE005, CE006]

5.2 The technical center is workflow depth around diffusion, LoRA, and ComfyUI-style graphs

The strongest public technical evidence points to an application stack assembled around established diffusion models and workflow orchestration rather than a proprietary foundation model. At the model layer, retained technical sources cover Stable Diffusion, SDXL, Stable Diffusion 3.5, FLUX.1, ControlNet, image-to-image, and LoRA. At the workflow layer, ComfyUI matters because it turns generation into an explicit node graph: checkpoint load, prompt encoding, sampling, VAE decode, post-processing, and custom nodes can be swapped, versioned, and submitted through an API. That architecture maps well to LiblibAI’s public positioning as a model marketplace and creation studio; creators can use specialized Chinese-style model assets, combine base checkpoints with LoRA adapters, and package repeatable workflows. The evidence supports workflow sophistication, but it does not prove ownership of the underlying base models or differentiated inference infrastructure. Those details define practical technical leverage.[CE012, CE014, CE015, CE016, CE018, CE019]

5.3 Developer surface exists, but integration evidence is thinner than product traction

Developer-signal sources indicate that LiblibAI can be treated as a programmatic generation surface, but public evidence is not yet equivalent to a mature enterprise API program with transparent pricing, SLAs, SDKs, and security documentation. GPTProto describes a LiblibAI API ecosystem that exposes text-to-image, image-to-image, ControlNet, ADetailer, LoRA training, ComfyUI online, and complex node workflows. Runflow’s ComfyUI guide shows the pattern such a backend usually follows: clients submit workflow JSON to /prompt, watch execution through WebSocket or history endpoints, upload inputs, and retrieve generated files. This is useful for studios that need repeatable production pipelines, but the diligence gap is real. Public sources do not show Evoken’s API rate limits, customer authentication model, data-retention policy, uptime history, or model-provider pass-through terms; those private controls determine whether developer access is a scalable product line or merely an advanced user convenience.[CE016, CE017, CE021, CE022, CE023, CE043]

5.4 Differentiation is localized workflow density; the risk is third-party model dependence

The constructive technology thesis is that Liblib has accumulated a localized Chinese creator ecosystem: creators upload, train, and share LoRA models and workflows, while ordinary users reuse those resources without owning GPUs or mastering local tooling. That can be more defensible than a single prompt interface because community assets, style libraries, workflow templates, and production habits create switching costs. The adverse thesis is equally important. China Biz Insider reports that Evoken does not develop proprietary foundation models and that LibTV uses third-party model APIs, including Seedance 2.0 through Volcano Engine. If model owners lower direct prices, restrict access, or bundle first-party workflow applications, Evoken’s gross margin and differentiation could compress quickly. Compliance is another product risk: CAC filing context and the reported CCTV content-safety incident show that moderation and labeling are not afterthoughts. The underwriting conclusion is a workflow-application moat with strong traction, but one that requires private diligence on model contracts, safety remediation, API controls, and contribution margin by product line.[CE013, CE032, CE033, CE034, CE035, CE036]

5.5 Exhibits

6.1 Customer base: a mass creator community feeding B-side workflows

Liblib's customer base should be underwritten as a dual-market system rather than a pure SaaS account list. On the C side, public reports cite more than 30 million cumulative users by mid-2026, 25 million total users and about 4 million MAU in late 2025, more than 20 million AI creators, more than 500,000 original models, and more than 5 million images generated per day. That creates a large supply side of prompt users, LoRA trainers, designers, illustrators, self-media creators and professional visual workers. On the B side, the verified public proof is narrower but strategically important: Jiemian says LiblibAI provides professional AI-image scenario solutions to Kingsoft Office/WPS, Wondershare, G-bits Games, Tmall Campus and Tsinghua University. The strongest segment lens is therefore C-side creators for liquidity and data, professional creators for subscriptions and credits, and B-side studios or enterprises for API, customized services and higher-value workflow spend.[CU001, CU002, CU003, CU004, CU005, CU006]

6.2 Named logos are real, but LibTV and vertical budgets are stronger adoption signals

The named customer list matters because it satisfies a customer-proof standard, but it does not by itself prove recurring production deployment or outcome specificity. The more compelling B-side evidence comes from vertical budgets around games, short drama, film and advertising. LibTV reportedly moved from launch to more than US$1 million of single-day revenue in its first month, May revenue more than 13 times first-month revenue, more than 300 short-film and film companies integrated within a month, and nearly 1,000 agencies, production houses and brand clients served. Gaming adds a second large budget pool: miHoYo's RMB100 billion AI commitment, NetEase and Tencent AI production pipelines, and broad generative-AI adoption by Chinese game developers all support the thesis that art, asset, NPC and UGC workflows are becoming procurement priorities. These facts imply real demand, while leaving contract size, production status and logo-by-logo outcomes private.[CU009, CU010, CU011, CU012, CU013, CU015]

6.3 Go-to-market: freemium activation into credits, subscriptions and enterprise/API spend

The observable go-to-market motion starts with low-friction community creation rather than top-down enterprise sales. Free daily generations, creator incentives and model browsing bring users into the ecosystem; high-volume usage, priority GPU access, training and professional features convert into subscriptions or credits; API and customized enterprise needs move B-side customers into higher-value contracts. This explains why the same company can show both creator-community metrics and US$300 million ARR: the creator base supplies content, workflows, models and distribution, while professional producers and enterprises supply willingness to pay. Expansion depends on cross-product loops. A creator or studio can begin with image models on LiblibAI, move into video workflows on LibTV, and later buy API or custom services. The diligence question is how much of that loop is recurring and defensible versus subsidy-driven, especially because API setup friction and third-party model alternatives can encourage multi-homing.[CU019, CU028, CU029, CU030, CU031, CU032]

6.4 Durability remains unproven: retention, concentration and reputational risk are the gaps

The customer chapter's constructive conclusion is that demand exists; its caution is that public retention proof does not. No retained source discloses NRR, GRR, churn, renewal rates, contract length, top-customer share, enterprise pipeline conversion or satisfaction. Daily image generation, creator incentives, model supply and LibTV team adoption are useful repeat-usage proxies, but they are not cohort retention. Demand-side risk is also not theoretical. Tech Times reports an industrial-scale AI short-drama boom with 470 AI-produced titles per day, a content glut, real worker displacement, canceled writing projects and unauthorized likeness concerns. That creates two possible demand paths: more studios buy AI tools to compete on volume, or brands and institutions slow adoption because AI content becomes reputationally sensitive. Investment diligence should therefore obtain cohort retention by segment, enterprise revenue concentration, contract renewal data, API/private-deployment terms, and brand-safety procurement objections before treating Liblib's customer base as durable. The practical test is whether usage survives after free compute, creator rewards and launch publicity normalize. Without that evidence, ARR should be treated as strong demand proof but not yet as proof of low-churn customer durability.[CU024, CU025, CU026, CU027, CU037, CU038]

6.5 Exhibits

7.1 China's AIGC regime is now an operating system, not a launch checklist

Liblib's first risk layer is regulatory continuity. China now regulates public-facing generative AI through a stack of instruments rather than a single statute: the January 2023 Deep Synthesis Provisions, the August 2023 Interim Measures for Generative AI Services, filing and registration registries administered through CAC and local cyberspace offices, and the September 2025 AIGC labeling Measures plus GB 45438-2025. For Liblib, this stack maps directly to image, video, LoRA training, downloads, user uploads, and distribution workflows. The favorable fact is that 36Kr reports Liblib cleared the deep-synthesis algorithm filing in February 2024 and became the first AI-community platform to complete generative-AI-service filing in March 2024. The adverse underwriting point is that filings are necessary but not sufficient: labeling files, metadata, user declarations, incident response, minors protection, and periodic registry visibility must keep pace with product expansion into LibTV, Xingliu, Lovart, and cross-border workflows.[CR001, CR002, CR003, CR004, CR005, CR006]

7.2 The biggest legal tail risk is not a fine; it is interruption of creator workflows

The hardest risk to underwrite is the combination of user-generated models, generated outputs, and content moderation. CCTV's 2026 exposure that LiblibAI could be prompted around filters to generate pornographic or inappropriate content is directly relevant because the Chinese AIGC regime is enforcement-oriented and increasingly labeling-centered. The company reportedly undertook technical rectification, but investors need evidence that fixes are durable under adversarial prompts and not merely reactive. Copyright risk is similarly workflow-specific. Ultraman v Acgnai did not make every training act illegal; the Hangzhou courts were comparatively tolerant of input-stage training and focused liability on stable infringing outputs, promoted LoRA models, platform profit, obviousness, and failure to take preventive measures. That is precisely why Liblib's model-sharing and LoRA marketplace needs robust IP review, complaint channels, model blocking, labeling, and takedown operations. The damages number was only RMB 30,000, but injunctions, delisting, or enterprise procurement blocks would be much more material.[CR012, CR013, CR014, CR015, CR016, CR017]

7.3 Aggregator economics are investable only if workflow lock-in outruns model commoditization

Evoken's business risk is not demand: public sources cite more than 30 million users, 500,000 original models, and US$300 million ARR by May 2026. The issue is whether those users and workflows create enough margin control. China Biz Insider's adverse framing is central: Evoken does not own proprietary foundation models and depends on third-party APIs, model providers, and cloud or compute economics. That exposes the company to repricing, degraded access, first-party bundling, and direct substitution if model vendors enter the workflow layer. LibTV's growth after Jimeng pricing changes shows Evoken can arbitrage user pain quickly, but it also makes queue times and API cost schedules thesis variables. Not-yet-profitable status and a large B+ round imply that capital is still part of the competitive apparatus. The core diligence ask is product-level gross margin by model family, not just ARR, because a model reseller can grow revenue while destroying contribution margin.[CR026, CR027, CR029, CR030, CR031, CR032]

7.4 Governance diligence should focus on concentration, controls, and measurable tripwires

People and governance risk is moderate but real. Chen Mian's prior CapCut commercialization experience is an asset because the company depends on rapid product-market-fit judgment, but the same evidence makes the thesis founder-centered. The source pack's reported 73.95% founder stake and later legal-representative change should not be treated as misconduct; they should be diligence prompts for board composition, reserved matters, related-party controls, information rights, and succession planning. The practical investment process should convert every major risk into a trigger: repeat CAC or app-store action, unresolved AIGC labeling deficiency, another CCTV-grade moderation failure, unresolved IP notice backlog, injunction against LoRA workflows, loss or repricing of critical model APIs, or evidence that product-level gross margin remains structurally negative after scale. Liblib is attractive only if compliance, IP operations, and model-procurement discipline become as repeatable as product launches. The evidence also leaves several quantitative questions open: current compliance staffing, notice volumes, red-team failure rates, and API-provider concentration are not public, so each should be a closing deliverable rather than a post-investment improvement plan.[CR038, CR039, CR040, CR041, CR043, CR044]

7.5 Exhibits

8.1 Headline valuation: a large round but a modest ARR multiple

Evoken’s June 2026 financing looks optically large but not obviously inflated when normalized by the public ARR evidence. Multiple independent reports converge on a roughly US$300 million Series B+ round, a post-money valuation above US$2 billion, and May 2026 ARR of about US$300 million. That places the headline round at approximately 6.7x ARR, before any adjustment for cash on the balance sheet or preferred terms that are not public. For a two-year-old AI application group with more than 30 million users, 500,000 original models, and reported professional-content budgets behind the revenue, the price is surprisingly disciplined. The underwriting question is therefore not whether investors paid a classic AI-bubble multiple; it is whether the reported ARR is recurring, retained, and gross-margin-positive after third-party model and compute costs. That framing also makes the valuation easier to compare with conventional SaaS: if the greater-than-US$2 billion mark is treated as a lower-bound enterprise value, every additional dollar above US$2 billion only moves the multiple modestly unless the true post-money is far higher than disclosed.[CV001, CV002, CV003, CV004, CV005, CV008]

8.2 Benchmarks make 6.7x look cheap, but not every AI dollar deserves a premium

The external benchmark set supports a fair-to-constructive valuation stance. 2026 SaaS sources put public cloud averages near the mid-single digits, conventional SaaS around 4x to 7x ARR, and higher-growth AI-native or vertical AI software much higher. Against that backdrop, Liblib’s 6.7x ARR multiple is below the AI-native top quartile and closer to reset-era public SaaS. Creative-AI comparables sharpen the point: Canva’s US$4 billion ARR supports a US$42 billion to US$65 billion valuation range, Midjourney appears near a 20x revenue multiple, and Sacra’s Gamma datapoint is above 20x ARR. The discount is logical, not punitive. Liblib has unusual growth, but public evidence has not yet proven proprietary model economics, enterprise retention, or a durable workflow moat comparable with the best global AI creative platforms.[CV015, CV016, CV017, CV018, CV019, CV020]

8.3 Scenario analysis: valuation is justified if workflow lock-in offsets aggregator risk

A scenario view points to a conditional “fair” price. The bear case assumes ARR stalls near the disclosed US$300 million level, retail usage remains subsidy-sensitive, and first-party model vendors compress the resale spread; at 4x ARR the implied value is roughly US$1.2 billion. The base case assumes forward ARR rises to about US$420 million while investors apply only a 5x multiple because gross margin, NRR, and API-cost pass-through are still undisclosed; that lands near US$2.1 billion and supports the current round. The bull case requires evidence that LibTV, LiblibAI, Xingliu, and Lovart are becoming a creator operating system with workflow lock-in; if forward ARR reaches US$500 million and the market pays 8x, the company could support about US$4 billion. The asymmetry is attractive only if private diligence validates revenue quality.[CV028, CV029, CV030, CV031, CV032, CV033]

8.4 Final judgment: justified, but with China AI froth and compliance haircuts

China’s 2026 unicorn context adds both support and caution. Nation Press reports 67 new Chinese unicorns in H1 2026, and Hurun’s global index shows AI as the main engine of private-market value creation. That means Liblib is not an isolated financing anomaly; it sits inside a broad capital reallocation toward AI applications, robotics, infrastructure, and workflow automation. The adverse counterweight is equally important: China Biz Insider’s embodied-AI reality check warns that many AI unicorns with short runways may face consolidation, down-rounds, or failure in 2027–2028. Liblib is better positioned than many concept-stage peers because it has reported ARR and a blue-chip syndicate, but regulatory/content-safety exposure and the lack of public margin metrics deserve explicit valuation haircuts. On the public record, the greater-than-US$2 billion valuation is justified, but only as a fair, evidence-sensitive entry price rather than a clear bargain. Therefore the correct IC posture is not to reject the round as AI exuberance, but to demand proof that revenue quality is stronger than the discount implies: audited ARR bridges, customer cohorts, contribution margin, and compliance controls would move the stance from fair to attractive.[CV036, CV037, CV038, CV039, CV040, CV041]

8.5 Exhibits