StepFun

1.1 Identity, founding, and product/technology architecture

StepFun is the trade name of Shanghai Jieyue Xingchen Intelligent Technology Co., Ltd, incorporated in Shanghai's Xuhui district. Baidu Baike's corporate-registry entry dates founding to April 6, 2023, with Jiang Daxin as legal representative and initial registered capital of RMB 20 million; English-language coverage, including Wikipedia and Reuters, converges on "April 2023" without the more precise street address. Reuters and Wikipedia both name Jiang Daxin -- a former Microsoft global/senior vice president who worked on Bing and Microsoft 365 -- as founder, alongside co-founders Zhu Yibo and Jiao Binxing. The company's technical identity centers on the "Step" family of multimodal foundation models (language, vision, video, audio) and, per MIT Technology Review, an explicit and still-active pursuit of artificial general intelligence that many Chinese peers have deprioritized. StepFun's official platform currently headlines "Step 3.7 Flash," a production-oriented agentic Flash model, while its consumer chat product "跃问" (Yuewen) and a companion app called 冒泡鸭 give it direct-to-user distribution alongside its developer API. Press coverage repeatedly places StepFun among China's "Six AI Tigers" (六小虎) together with Zhipu AI, MiniMax, Moonshot AI, 01.AI, and Baichuan Intelligence, and describes its strategy -- per a StepFun executive quoted by 36Kr -- as closer to "China's xAI plus Tesla" than a pure chat-app play, reflecting its emphasis on embedding models into physical terminals rather than cloud subscriptions alone.[CO001, CO002, CO003, CO004, CO005, CO006]

1.2 Leadership, governance, and key-person dependence

StepFun's public leadership bench is more populated than many peers at this stage, but still leaves real governance gaps. KrASIA's profile names four technical leaders: founder-CEO Jiang Daxin; CTO Zhu Yibo, who previously led AI infrastructure and large-cluster operations at ByteDance; data lead Jiao Binxing, formerly head of Microsoft Bing's core search team; and chief scientist Zhang Xiangyu, a core author of the ResNet architecture. On January 26, 2026, StepFun announced that Yin Qi -- co-founder and former CEO of Megvii (Face++), one of China's earlier "AI Dragons," and concurrent chairman of the Geely-linked smart-driving firm Qianli Technology -- had been appointed chairman, tasked with overall strategy and technical direction. Gasgoo reported that Yin's dual role is explicitly meant to connect StepFun's model R&D to real-world hardware deployment data, forming a "Physical AI" feedback loop rather than a purely advisory chairmanship. The Standard reported, citing an interview with Jiang Daxin, that StepFun's team now exceeds 500 people with roughly 80% in algorithm and technical roles -- a marked jump from the roughly 150-person team Chinese Wikipedia describes from an earlier period. What the public record does not supply is a complete current board roster, committee structure, or cap-table/ownership breakdown, which matters given how much of StepFun's capital base is state-linked and how concentrated strategic influence now sits with Jiang and Yin.[CO013, CO014, CO015, CO016, CO017, CO018]

1.3 Funding history, investors, and valuation trajectory

StepFun's financing history shows a company that started slowly but has since compounded quickly. KrASIA and BigGo Finance both describe an angel/first round that reached unicorn valuation on the strength of founder networks, backed by Sequoia China (HongShan), Qiming Venture Partners, and IDG Capital; a 2024 Series A added US-dollar investors 5Y Capital and Shunwei Capital alongside Legend Capital; and a Series B, launched mid-2024 and closed around December 2024, brought in Shanghai State-owned Capital Investment and affiliated funds, the Hong Kong Investment Corporation (HKIC), and Tencent. On January 26, 2026, TechNode, KrASIA, Yicai, and C114 all reported that StepFun closed a Series B+ round exceeding RMB 5 billion (roughly $718-733 million) -- called the largest single financing in China's large-model sector over the prior 12 months -- with new industrial and state-linked investors (Shanghai SDIC/S&T Capital InnoFund, China Life Private Equity, Pudong Venture Capital, Xuhui Capital, Wuxi Liangxi Fund, Xiamen ITG Group, Huaqin Technology) alongside follow-on money from Tencent, Qiming, and 5Y Capital. By April 2026, StepFun had converted to a joint-stock company and dismantled its Cayman Islands "red-chip" structure, and by May 2026 KrASIA and CnTechPost reported it was nearing an almost $2.5 billion pre-IPO round -- with HKIC's only disclosed generative-AI investment to date, plus supply-chain investors Huaqin, Longcheer, OmniVision, and ZTE. Reuters cited Caijing reporting that the pre-IPO round targeted up to a $6 billion valuation ahead of a Hong Kong filing near a $10 billion anchor valuation; FAQ.com.tw reported StepFun filed for that Hong Kong IPO on June 9, 2026, targeting $10-12 billion and roughly a $500 million raise.[CO023, CO024, CO025, CO026, CO027, CO028]

1.4 Commercialization milestones -- models, partnerships, and scale

StepFun's product chronology moved fast from the start: Chinese Wikipedia and KrASIA both place the roughly 100-billion-parameter Step-1 launch in August 2023, months after founding, followed by the multimodal Step-1V in November 2023 and a trillion-parameter Step-2 Mixture-of-Experts preview in March 2024. At WAIC 2024 in July, StepFun officially launched Step-2, the Step-1.5V multimodal model, and the Step-1X image generator. February 2025 brought open-sourced Step-Video-T2V and Step-Audio, developed jointly with Geely; by WAIC 2025 that automotive relationship had matured into a joint showcase of the Galaxy M9 -- billed as the world's first vehicle with a "human-like AI agent" -- powered by StepFun's end-to-end voice model and a jointly built Agent OS smart-cockpit system, per BusinessWire. StepFun open-sourced its flagship Step 3 (321 billion total, 38 billion active parameters, Apache 2.0) on July 31, 2025, with a technical paper describing Multi-Matrix Factorization Attention and Attention-FFN Disaggregation designed to cut decoding costs versus DeepSeek-V3 and Qwen3 MoE 235B; February 2026 followed with the open-source Step 3.5 Flash (196B/11B active, 256k context). Reuters reported Step 3.5 Flash has since ranked among the three most-used models on the OpenClaw agent platform. On the deployment side, FAQ.com.tw and BigGo Finance both report StepFun models powering AI features on more than 40 million smartphones (OPPO, Honor, ZTE) and targeting more than one million Geely vehicles, serving an estimated 20 million daily users.[CO039, CO040, CO041, CO044, CO045, CO046]

1.5 Adverse signals -- regulatory restructuring, disclosure gaps, and IPO execution risk

The clearest adverse thread in StepFun's chapter-one record is regulatory rather than reputational. Reuters reported on April 13, 2026 that StepFun was unwinding its offshore "red-chip" incorporation structure after China's securities regulator instructed some red-chip companies to re-domicile before pursuing overseas listings, and that StepFun -- heavily backed by state capital -- judged an onshore structure more appropriate. BigGo Finance's analysis, and Reuters' own reporting that roughly a fifth of a recent cohort of approved Hong Kong listings involved offshore holdings, frame this as a sector-wide compliance wave that can delay listings or, in some cases, cause companies to abandon IPO plans given the cost of changing legal structure. Layered on top of that timing risk is a valuation-support question: BigGo Finance's compiled figures put StepFun's 2025 revenue at roughly CNY 500 million (about $73.3 million) against a Hong Kong IPO filed at a $10-12 billion target valuation -- a gap this chapter treats as an explicit diligence flag rather than a resolved fact, since the revenue figures come from secondary market reporting rather than an audited filing. Finally, the same opacity that limits leadership visibility (section 2) extends to ownership: reviewed public sources do not disclose a full board roster or cap table, which is notable for a company converting to a joint-stock structure and courting public-market capital within months of that conversion.[CO022, CO030, CO031, CO032, CO033, CO042]

1.6 Exhibits

2.1 Market boundary, included spend, and status-quo substitutes

StepFun should be read as a company competing inside China's foundation-model application and embedded-device AI layer, not as a stand-in for China's entire AI or semiconductor economy. IDC's China AI agent market-glance draws exactly this line, separating model-development, application-development, and agent-development platforms from the underlying compute, storage, network, and semiconductor infrastructure stack beneath them. Inside that boundary, StepFun's own platform.stepfun.com documentation shows the company monetizes per-token API metering across chat, vision, voice, and agentic 'step-3.x' models rather than flat software licensing, and its Step-Audio voice model already ships as embedded cockpit software inside Geely's production Galaxy M9 -- putting automotive smart-cockpit licensing, not just consumer chat, squarely inside the addressable spend. A fourth, easily overlooked segment is compliance: as of April 30, 2026 China had 868 nationally CAC-registered generative-AI services and 530 locally registered applications, and every one of StepFun's public-facing surfaces must clear the same registration, security-assessment, and content-labeling gates that make compliance a recurring cost of doing business rather than a legal footnote. Excluded from this boundary is the much larger infrastructure layer -- Forrester puts China's 2026 AI-infrastructure spending above $70 billion -- which StepFun consumes as a cost input rather than monetizes directly. Status-quo substitutes for StepFun's surfaces range from native phone or car software without embedded generative AI, to larger hyperscaler-distributed alternatives (Doubao, Qwen) that OEM partners could choose instead.[CM001, CM002, CM003, CM004, CM007]

2.2 Evidence-constrained sizing lenses for a market without one clean number

No single public figure cleanly sizes 'StepFun's market,' so this chapter preserves several source-backed lenses rather than manufacturing one TAM. The narrowest applications lens, IMARC's China generative-AI (AIGC) market estimate, puts 2025 value at roughly $5.16 billion, growing to $19.56 billion by 2034 at a 15.96% CAGR -- a lens that likely understates StepFun's true addressable pool because it excludes embedded-device licensing. A much larger, adjacent lens is infrastructure: Forrester expects China's AI-infrastructure spending to exceed $70 billion in 2026, and Gartner puts worldwide AI spending at $2.59 trillion (+47% year-over-year), figures that sit upstream of the layer StepFun actually monetizes. A usage lens comes from China's public-cloud model-as-a-service consumption, which reached 1,944 trillion tokens in 2025 (16x year-over-year), with ByteDance's Volcano Engine alone holding a 49.5% share -- evidence that usage is scaling far faster than any settled revenue pool. A regulatory-scale lens shows 868 nationally registered generative-AI services as of April 30, 2026, growing by 72 in just two months, while a global-share lens (Nikkei, via TrendForce) found Chinese open-source models jumped from about 1% to 15% of global usage share in a single year. Usefully, IDC's own China foundation-model tracker (CHC5530125) independently lists StepFun's 'Step Reasoner mini' alongside Qwen3-Max, DeepSeek, Kimi K2, and GLM-4.6 as one of the frontier models it follows, and separately sets national targets of 70% smart-device/agent penetration by 2027 and 90% by 2030, with 80% of China's top-1,000 enterprises expected to prioritize AI sovereignty by 2027. None of these lenses can be added together into a single number; taken together they show a fast-growing, usage-rich, but still commercially unsettled market.[CM008, CM009, CM010, CM011, CM012, CM013]

2.3 Device-embedded AI: the phone, PC, and cockpit surfaces StepFun is actually chasing

StepFun's own framing is explicit: rather than compete purely on cloud chat traffic, it embeds models inside smartphones, vehicles, and industrial hardware. Public reporting credits StepFun's flagship models with powering AI features in more than 40 million smartphones and targeting deployment in more than one million vehicles through its Geely partnership -- numbers that only matter if the underlying device categories are themselves growing. On phones, IDC (corroborated by a second Chinese outlet) forecasts 2026 China shipments will fall 2.2% to about 278 million units, even as next-generation AI-phone shipments rise to 147 million units, 53% of the total and the first time AI phones cross half the market, under an 'edge-cloud synergy' model. On PCs, IDC projects a roughly flat 42.22 million-unit China market in 2026 but 146.5% year-over-year growth in GenAI PCs, reaching an estimated 36.5% of the category by 2029. On automotive cockpits, 中商产业研究院's industry forecast puts China's smart-cockpit solutions market at RMB 182.8 billion in 2026, up from RMB 154.5 billion in 2025, with voice-interaction penetration nearly doubling from 26% to 48% in a year. Geely's Galaxy M9, unveiled at WAIC 2025, is cited by both companies as the industry's first mass-produced vehicle with an end-to-end voice-LLM-powered 'Agent OS' cockpit built on StepFun's Step-Audio model. Each of these device categories is expanding, but StepFun is one of several embedded-model suppliers competing for OEM design wins, and only one relationship -- Geely -- is publicly confirmed at scale, a concentration that independent analysis flags as a potential ceiling on how independent StepFun's addressable market really is.[CM017, CM018, CM019, CM020, CM021, CM022]

2.4 Buyer, user, payer, and adoption path across four distinct surfaces

StepFun's buyer map is not one funnel but at least four parallel ones. Individual consumers self-pay (or use free tiers) for its consumer chat and voice apps. Developers and agent builders meter usage directly against platform.stepfun.com, and StepFun's March 2026 Step Plan subscription -- four tiers from a standard RMB 49 to RMB 699 per month, with a standing developer-community discount -- is purpose-built for OpenClaw-agent and AI-coding workflows specifically; independent, usage-based evidence backs this segment's traction, with OpenRouter's real-call leaderboard placing Step 3.5 Flash (free) second globally in the week of March 9-15, 2026 at 1.34 trillion tokens, up 79% week-on-week. Smartphone and automotive OEMs are a third, structurally different buyer: their adoption trigger is a flagship product cycle or joint showcase -- the Galaxy M9's WAIC 2025 unveiling, for instance -- rather than a self-serve trial, so this channel depends on a small number of high-value relationships rather than broad conversion. A fourth buyer, the enterprise or regulated deployer, must clear CAC registration and content-labeling requirements before go-live, and IDC's own China AI agent market-glance shows this buyer segmented across many functions (finance, HR, customer service, supply chain) and industries (FSI, manufacturing, automobile, retail), meaning no single enterprise vertical dominates the opportunity. Crowding this same developer segment, IDC also tracks at least a dozen competing China AI-coding-IDE vendors -- Alibaba Qoder, ByteDance Trae, Baidu Comate, Tencent Codebuddy, and others -- even before StepFun's own share is counted, while IDC forecasts 35% of China's professional developers will adopt AI-native coding platforms by 2027 even as top vendors' 2025 coding revenue stayed below RMB 100 million despite triple-digit usage growth.[CM024, CM025, CM026, CM027, CM028, CM038]

2.5 Growth drivers, price competition, compute constraints, and regulation

Three structural drivers push in StepFun's favor. China's May 2026 CAC/NDRC/MIIT 'Implementation Opinions' is the country's first dedicated national policy framework for AI agents, defining them formally and naming 19 priority deployment scenarios across research, manufacturing, transport, finance, healthcare, and government. IDC's 'AI+' targets -- 70% smart-device/agent penetration by 2027, 90% by 2030, and 80% of China's top-1,000 enterprises prioritizing AI sovereignty by 2027 -- create durable demand for domestic, embeddable models. And China's device-replacement cycles (AI phones, GenAI PCs, smart cockpits) are all expanding the physical surface area for embedded AI. But the same forces double as constraints. China's 2023 generative-AI measures and its 2025 AI-content labeling rules (effective September 1, 2025) impose recurring content-security, data, and explicit/implicit labeling obligations on every deployed surface. Compute access tightened further on May 31, 2026, when the U.S. Commerce Department closed a loophole that had let China-headquartered firms buy advanced Nvidia/AMD chips through overseas subsidiaries without a license -- extending export controls by ownership rather than delivery location. Pricing is compressing just as fast: Chinese labs cut API prices at least six times in H1 2026, with DeepSeek, Xiaomi MiMo, and Kimi making cuts permanent, pushing per-token prices 5-34x below Western frontier APIs; StepFun's own published step-3.5-flash and step-3.7-flash rates sit in the same competitive band rather than at a premium. Trust is also under strain: Chinese authorities have issued high-level warnings about Trojan-laced counterfeit skill packages on the OpenClaw agent platform -- the same ecosystem where Step 3.5 Flash ranks near the top of usage charts. Finally, distribution is asymmetric: ByteDance's Doubao alone reaches over 500 million phone terminals and more than 7 million vehicles, roughly an order of magnitude beyond StepFun's disclosed footprint, while global automakers increasingly co-develop China-specific cockpit AI with local partners (Huawei HarmonyOS, XPeng, Momenta) rather than a single vendor -- and China's foundation-model layer saw financing roughly halve in 2025 to about 22 deals worth RMB 9.4 billion, tightening the capital available to independent players competing against these larger platforms.[CM029, CM030, CM031, CM032, CM033, CM034]

2.6 Exhibits

3.1 Landscape: peer startups, big-tech-embedded models, and adjacent entrants

StepFun's real competitive set is broader than the "Six AI Tigers" framing chapter 1 uses to describe its founding cohort. By Q2 2026 the Chinese foundation-model field has consolidated to roughly ten providers with meaningful public usage share -- Xiaomi, Alibaba (Qwen), Zhipu AI (rebranded internationally as Z.ai), DeepSeek, Moonshot AI, MiniMax, StepFun, ByteDance, Baidu, and Tencent -- with a longer tail of vertical or niche players (Baichuan Intelligence, Yi, Xunfei, KwaiKAT) behind them. Inside that set, StepFun sits alongside Zhipu AI, Moonshot AI, MiniMax, and Baichuan Intelligence as a venture-funded, pure-play foundation-model startup, distinct from big-tech-embedded programs -- Alibaba's Qwen, Baidu's ERNIE, and ByteDance's Doubao/Seed -- that ride an existing cloud, search, or consumer-app distribution stack, and distinct again from DeepSeek, which is investor-independent (High-Flyer-funded) until its first external round in mid-2026. Two adjacent classes matter as much as the named peer set. First, Xiaomi's MiMo model family is not one of StepFun's traditionally cited peers, yet it alone commands roughly 21% of weekly Chinese-provider OpenRouter token volume in Q2 2026 -- more than any single pure-play startup -- showing that a device incumbent can out-distribute every funded AI lab purely through smartphone and developer-tooling bundling, the same channel StepFun is trying to use as its own wedge. Second, internal build or reliance on blocked or export-controlled overseas frontier APIs remains a status-quo substitute for some regulated Chinese buyers, but data-localization rules and chip export controls push most domestic demand toward the ten-provider Chinese set profiled in this chapter.[CP001, CP002, CP003, CP004, CP041]

3.2 Competitor profiles: scale, funding, and strategic direction

Zhipu AI, rebranded internationally as Z.ai, became the first listed large language model company via a January 2026 Hong Kong IPO; its FY2025 annual report showed revenue of roughly CNY 724 million (about $105 million, up 131.9% year over year) against a net loss of about CNY 4.72 billion, with its market capitalization trading above HK$400 billion by mid-2026. Its GLM-5 flagship is a 744-billion-parameter MoE model released under an MIT license and trained end-to-end on Huawei Ascend silicon, positioning Zhipu as the export-control-resilient choice for state-linked and enterprise buyers. Moonshot AI, maker of the Kimi model family, raised about $2 billion in May 2026 at a $20 billion valuation -- up from $4.3 billion at the end of 2025 -- led by Meituan's Long-Z Investments; its Kimi K2.6/K2.5 models compete on long-context, agent-swarm coordination, and are embedded in Cursor Composer 2, a Western developer tool. MiniMax listed in Hong Kong the same January 2026 window, raising about $619 million at a roughly $6.5 billion valuation with GIC, Baillie Gifford, and the Abu Dhabi Investment Authority as cornerstone investors, and ships a self-evolving agentic M-series (M2-M2.7). Baichuan Intelligence took a different path, pivoting from a broader platform strategy to vertical healthcare AI -- disbanding its finance and education teams to focus on models like Baichuan-M2, an open-source medical LLM that scored above OpenAI's gpt-oss-120b on the HealthBench benchmark -- while retaining backing from Alibaba, Tencent, Xiaomi, and multiple state-backed investment funds. DeepSeek, a High-Flyer hedge-fund spinout, operated with zero external investors until a June 2026 round that raised over CNY 50 billion (about $7.4 billion) at a valuation above CNY 330 billion, structured so most outside investors get no voting rights while the National AI Industry Investment Fund invested directly with full voting rights. Alibaba's Qwen, Baidu's ERNIE, and ByteDance's Doubao/Seed round out the big-tech-embedded tier: Qwen3.6-Plus is wired into Alibaba's Wukong enterprise platform and consumer Qwen App; ERNIE 5.1 (not open-sourced, unlike ERNIE 4.5) claims 4th globally on the Arena Search leaderboard at about 6% of peer training cost; and Doubao/Seed 2.0 is ByteDance's consumer-first flagship with no separately disclosed funding round of its own.[CP005, CP006, CP007, CP009, CP010, CP013]

3.3 Comparative dimensions: capability, pricing, distribution, and trust posture

On raw capability, independent July 2026 benchmark tracking places Qwen3.7 Max and GLM-5.2 at the top of the Chinese-model leaderboard (score 90 each), DeepSeek V4 Pro Max close behind at 87, and Kimi K2.6 at 81; StepFun's models, while a real "surprise performer" via Step 3.5 Flash's #3 rank on OpenRouter's free-model leaderboard, are not listed among the top five nationally ranked Chinese models on this leaderboard. On pricing, StepFun's own step-3.7-flash (CNY 0.27 cache-hit / 1.35 standard input / 8.1 output per million tokens) and step-3.5-flash (0.14/0.7/2.1) sit inside the same broad CNY-denominated band as DeepSeek, Qwen, and Kimi's tiers, but the 2026 API price war has been driven almost entirely by other labs: DeepSeek, Xiaomi MiMo, and Moonshot Kimi each declared 2026 price cuts permanent, with DeepSeek V4-Pro now at $0.87 output (about 34x below GPT-5.5) and Kimi K2.6 holding the lowest cache-hit floor in the market at $0.07. Distribution power diverges even more sharply than pricing. ByteDance's Doubao is China's most-used consumer AI app at 336-345 million monthly active users; Alibaba's Qianwen and Quark apps combine for roughly 380 million; Baidu's ERNIE-powered assistant (Wenxiaoyan) has about 200 million monthly and over 300 million registered users; and even DeepSeek's own consumer app ranks #4 nationally at roughly 139 million MAU. Kimi's consumer app, by contrast, ranks 8th at about 25 million MAU, and StepFun's Yuewen does not appear in China's top-50 AI apps by MAU at all. On trust and hardware posture, GLM-5 (Huawei Ascend) and Baidu ERNIE (Baidu Kunlun) lead on domestic-hardware independence relevant to export-control-sensitive buyers, while StepFun trains on NVIDIA Hopper -- a posture shared with most of the field but not a differentiator. Because most of these APIs are OpenAI-compatible and several ship permissively licensed open weights, enterprise and developer buyers can multi-home or switch providers with comparatively low technical friction, which caps how much lock-in any single vendor, including StepFun, can extract from pricing or model quality alone.[CP008, CP011, CP012, CP015, CP021, CP022]

3.4 Moat durability, commoditization risk, and adverse competitor evidence

StepFun's most durable-looking claim is not benchmark leadership or consumer reach -- it is embedded multimodal device distribution (smartphone pre-installs and its Geely automotive relationship, detailed in chapter 2) combined with an explicit, still-active pursuit of AGI-oriented model scope that several peers have deprioritized. That wedge is real but conditional: Xiaomi's MiMo family alone already out-distributes StepFun by volume through the same device channel, and Geely's own April 2026 "Full-Domain AI 2.0" materials describe a multi-vendor, increasingly in-house cockpit and ADAS strategy rather than exclusive dependence on any single outside model supplier -- a genuine displacement risk for a concentrated OEM relationship. The adverse evidence is broader than any one competitor. First, capital access: Zhipu and MiniMax are already listed, Moonshot raised at $20 billion and DeepSeek at over $50 billion (the latter with a direct, voting investment from the National AI Industry Investment Fund), all materially deeper and more transparent capital bases than StepFun's reported near-$2.5 billion pre-IPO round. Second, consumer brand: StepFun's Yuewen app trails every other named competitor in this chapter on disclosed MAU, and the gap versus Doubao (336-345 million) or even DeepSeek's own consumer app (about 139 million) is an order of magnitude or more. Third, benchmark opacity: independent commentary on Chinese LLM leaderboards flags test-set leak and gaming concerns (disputed C-Eval results, self-reported scores without disclosed methodology), which should discount confidence in any single provider's self-reported ranking, StepFun's included, until independently reproduced. Fourth, commoditization: the 2026 API price war was started and sustained by DeepSeek, Xiaomi, and Moonshot, not StepFun, meaning StepFun must defend margin inside a pricing floor set by better-capitalized rivals rather than one it controls. None of this erases StepFun's device wedge, but it means that wedge -- not benchmark rank, app scale, or balance-sheet depth -- is the only place this chapter finds durable relative advantage.[CP037, CP040, CP041, CP042, CP043, CP045]

3.5 Exhibits

4.1 Revenue model and pricing: official rate cards are visible, realized economics are not

StepFun's official open platform publishes an unusually granular rate card across text, vision, and audio models, with per-1M-token pricing for step-3.7-flash, step-3.5-flash, and step-1o-turbo-vision, plus separate usage pricing for realtime speech, text-to-speech, internet search, and file storage. That list-price transparency is real, but it sits alongside an explicit strategy statement: in July 2025 a StepFun vice president said the company deliberately avoids both highly customized project delivery, which loses to open-source substitution, and pure token-priced APIs, which large platforms can commoditize by bundling model access with cloud sales. Instead, StepFun frames its commercialization around productized to-B and to-C revenue with device partners across phones, cars, IoT, and embodied robotics, plus vertical partnerships such as a dedicated finance joint venture and cooperation with Guotai Haitong. Independent reporting adds that enterprise services reportedly cover multiple Fortune 500 clients and that AgentAPI call volumes have grown significantly, but neither claim is independently corroborated with company-disclosed figures, and a chip-adaptation licensing channel with domestic chip vendors adds a further, still-unquantified revenue surface. The result is a company with more visible pricing mechanics than most private AI peers, but no public bridge from any of these surfaces to realized, segment-level revenue.[CI001, CI002, CI003, CI004, CI005, CI006]

4.2 GTM motion and device partnerships: distribution is real, sales efficiency proxies are not

StepFun's go-to-market motion looks device-led rather than purely developer-led. ZTE's Nubia Z80 Ultra was reported as the first smartphone to natively embed StepFun's OpenClaw AI agent inside the operating system rather than as a separate app, and Tencent Cloud signed a strategic cooperation agreement with StepFun to build an 'AI Cockpit Agent' linking Tencent's music, video, maps, payment, and mobility services for in-vehicle use, deepening a relationship that started with Tencent's Series B investment. Reporting also describes StepFun's recent fundraising rounds as concentrated with consumer-electronics supply-chain investors -- Huaqin, Longcheer, OmniVision, and ZTE -- rather than purely financial backers, and says StepFun's smartphone Agent partnerships covered more than half of China's leading domestic phone brands as of mid-2025. That concentration cuts both ways: it gives StepFun real, named distribution partners, but it also ties near-term revenue growth to the shipment cycles and negotiated terms of a small number of manufacturers. None of the sources reviewed here disclose a customer-acquisition cost, payback period, net revenue retention figure, or contract-level pricing for any device or enterprise deal, so this chapter cannot verify whether the distribution is efficient or merely visible.[CI008, CI009, CI010, CI011, CI012, CI013]

4.3 Reported revenue, funding, and capital access: real signals, a missed guidance target, and no audited baseline

Company Overview already covers StepFun's round-by-round funding chronology; this chapter's question is what recent reporting implies about present-day capital access and revenue quality. Caijing reported StepFun's 2025 revenue at roughly RMB500 million (about $73.3 million) with a 2026 projection near RMB1.2 billion (about $175.9 million). That is a meaningfully lower base than the roughly RMB1 billion full-year 2025 revenue target StepFun's own leadership stated in July 2025, when several hundred million RMB of contracted revenue had reportedly already been booked in H1 -- implying the company fell short of its own mid-year guidance. On the capital-access side, a February 2026 report described a two-tranche Pre-IPO round (roughly $4 billion and $5-6 billion pre-money) and a planned Hong Kong filing by mid-2026 with cornerstone pricing near $10 billion, consistent with the $10-12 billion range and June 9, 2026 filing date already established in Company Overview; internal documents cited separately point to a possible mid-2027 share lock-up. Independently reviewed registered-capital filings show StepFun raised its onshore registered capital twice in 2026 and registered a Hong Kong entity in May 2026, both consistent with active listing preparation. Set against already-listed peers, Zhipu AI reported audited 2025 revenue of RMB724 million at a 41% gross margin (still with a RMB4.72 billion net loss), and MiniMax reported roughly RMB543 million of 2025 revenue -- both larger and more disclosed than StepFun's still-unaudited figures.[CI015, CI016, CI017, CI018, CI019, CI020]

4.4 Cost structure and capital intensity: a domestic-chip alliance does not remove the compute bill

StepFun's public cost story centers on compute rather than customer-facing infrastructure. In July 2025 it co-founded the Model-Chip Ecosystem Innovation Alliance with roughly ten domestic chip and AI-infrastructure vendors -- including Huawei Ascend, MetaX, Biren, Enflame, Iluvatar CoreX, Moore Threads, and Cambricon -- explicitly to reduce dependence on foreign GPUs, and StepFun says its Step 3 model reaches up to roughly 3x DeepSeek-R1's inference throughput on domestic chips and a 70%-plus throughput gain over DeepSeek-R1 on Nvidia Hopper hardware. By June 2026, however, industry commentary summarized by Baidu Baike said the alliance's progress remained concentrated in hardware adaptation and basic integration testing, with limited progress toward large-scale deployment or the kind of broad developer ecosystem that Nvidia-based platforms already have. That matters because frontier inference stays expensive regardless of chip origin: independent analyst data put global hyperscaler AI capex above $130 billion in a single quarter, projected combined cloud-provider capex above $520 billion for 2026, single H100 GPUs at $25,000-$40,000 and B200s at $30,000-$70,000, and GPU-collateralized asset-backed financing as an increasingly common way to fund 100-10,000-GPU deployments without diluting equity. Set against a Chinese large-model market 36Kr estimates at under RMB30 billion in 2024 and growing toward RMB70 billion-plus by 2026, StepFun's compute ambitions look proportionate to the opportunity, but nothing in the public record quantifies StepFun's own compute spend, chip mix, or resulting unit-cost savings.[CI026, CI027, CI028, CI029, CI030, CI031]

4.5 Financial verdict and diligence blockers: real commercial traction, no audited baseline, and an unreconciled guidance miss

StepFun's positive financial case is visible: official pricing spans five-plus monetization surfaces, named device and enterprise partners are real, and 2026 registered-capital and Hong Kong entity filings confirm active, well-financed listing preparation. The negative case is just as concrete. No source reviewed here discloses StepFun's cash on hand, monthly burn, runway, gross margin, or unit economics by product line, so capital adequacy and margin quality cannot be independently underwritten. AsiaICT's analysis frames the starkest version of the concern: a roughly $10 billion cornerstone valuation target against a low-hundreds-of-millions revenue base implies a revenue multiple in the tens of times, and StepFun had not filed a formal HKEX Application Proof as of May 20, 2026 despite a widely reported June 2026 filing plan, adding execution-timing risk on top of the CSRC red-chip unwind already covered in Company Overview. The roughly RMB1 billion mid-2025 revenue guidance versus the roughly RMB500 million reported 2025 outturn remains unreconciled in every source reviewed, and the 30%-60% device-partner revenue-share norm cited in adverse reporting is not quantified for StepFun specifically. Until an audited statement or filed prospectus closes these gaps, this chapter's conclusion is cautious: StepFun looks like a real, fast-scaling commercial business, but public evidence still supports a research posture rather than a clean financial underwrite.[CI034, CI035, CI036, CI037, CI038, CI039]

4.6 Exhibits

5.1 Product line: from Step 3.7 Flash down to open-sourced audio, video, and image tooling

StepFun's commercial surface is a single API platform (platform.stepfun.com in China, platform.stepfun.ai globally) that meters a family of models rather than one flagship product. The current flagship, Step 3.7 Flash, is a 198B-parameter sparse Mixture-of-Experts vision-language model (196B language backbone plus a 1.8B vision encoder) that activates roughly 11B parameters per token, supports a 256K context window, and is priced at RMB1.35/RMB0.27 (cache-miss/cache-hit) input and RMB8.1 output per million tokens. Its predecessor, Step 3.5 Flash (196B/11B active, 262K context), remains served and ranked #3 on OpenRouter's free-model leaderboard shortly after its February 2026 release. Below the Flash tier sits Step-2, StepFun's original trillion-parameter MoE model launched at WAIC 2024, and the Step-1V/Step-1.5V multimodal line that StepFun has said powers its more-than-40-million embedded smartphone AI features. A separate creative-tooling line covers text-to-image and image editing (step-2x-large, step-image-edit-2, billed per generated image) and an extensive audio family: Step-Audio 2, an open-source end-to-end speech-conversation LLM with retrieval-augmented generation and tool-calling, plus realtime, TTS, voice-cloning, and ASR product tiers each billed on their own metering unit, and installable via pip as well as GitHub. StepFun also open-sourced Step-Video-T2V, a 30B-parameter text-to-video model, and maintains developer-facing tooling beyond the models themselves, including the STEP-GUI GUI-agent framework (repository gelab-zero) and Step-Realtime-CLI/Console voice-agent tooling. On the consumer side, none of these models carry a separate consumer brand; they sit behind the 阶跃AI (跃问 / Yuewen) chat app and embedded device assistants, meaning StepFun's consumer distribution is entirely a reasoning-backend relationship rather than a stand-alone product.[CE001, CE002, CE004, CE005, CE006, CE007]

5.2 Architecture: sparse MoE, attention disaggregation, and the domestic-chip wedge

Step 3.7 Flash's cost efficiency rests on fine-grained MoE routing: independent technical analysis of the release describes 42 of 45 language layers as routed MoE layers, each holding 288 experts, with a learned router selecting only the top 8 per token -- roughly 5.6% of weights computing on any given token. A separate, larger model in the family, Step-3 (321B total / 38B active parameters), pushes further on hardware-aware efficiency: its co-design paper introduces Multi-Matrix Factorization Attention (MFA) to shrink KV-cache size and Attention-FFN Disaggregation (AFD) to split attention and FFN compute into specialized subsystems, claiming a decoding throughput of 4,039 tokens/second/GPU on Hopper hardware versus 2,324 for DeepSeek-V3 in the same test setup -- a StepFun-run, StepFun-favorable comparison rather than an independent benchmark. StepFun has also tried to reduce compute-supply risk directly: in July 2025 it co-founded the Model-Chip Ecosystem Innovation Alliance with roughly ten domestic chip and platform vendors (Huawei Ascend, MetaX, Biren, Enflame, Iluvatar CoreX, Moore Threads, Cambricon among them) and adapted Step 3's MFA architecture to run on domestic accelerators including MetaX GPUs. As of June 2026, however, industry commentary characterized the alliance's progress as still concentrated in hardware adaptation and basic integration testing rather than large-scale production deployment. On the product-efficiency side, Step 3.7 Flash's Advisor Mode -- routing most of an agent trajectory through the cheap executor and escalating to a larger model only at planning or failure-recovery inflection points -- is StepFun's own implementation of an escalation pattern it attributes to Anthropic, and its reported 'emergent' cross-tool combination behavior (independently calling Python crop/zoom utilities inside visual-search loops) is a vendor observation StepFun itself has not put through independent reproduction.[CE002, CE013, CE014, CE015, CE044, CE042]

5.3 Customer workflows, deployment footprint, and release roadmap

StepFun's most concrete efficiency pitch is agentic coding economics: Step 3.7 Flash's Advisor Mode is claimed to run an agent task for about $0.19 versus roughly $1.76 for a larger advisor model, a claim StepFun itself frames as an internal scaffold benchmark rather than an independently audited comparison. Beyond coding, the model line covers enterprise document/screenshot parsing (native multimodal perception), web/visual search-augmented question answering, end-to-end conversational voice agents via Step-Audio 2, and GUI-driven desktop automation via the STEP-GUI framework. Deployment is genuinely flexible on paper: Step 3.7 Flash ships for cloud, data-center, and local use, is integrated with vLLM, SGLang, Hugging Face Transformers, llama.cpp, the NVIDIA NeMo ecosystem, and NVIDIA NIM, and is available at release on StepFun's own platform, OpenRouter, and NVIDIA NIM (with DeepInfra, Fireworks AI, and Modal announced as forthcoming). Standard API access is rate-limited on a five-tier ladder (V0 through V5) gated by cumulative account recharge rather than a fixed developer plan, separate from a newer subscription 'Step Plan.' Genuinely local deployment, however, has a real hardware floor: even the smallest published quantization needs at least 128GB of unified memory, which rules out ordinary consumer laptops and confines local runs to workstation-class hardware such as DGX Spark, Mac Studio M4 Max, or AMD Ryzen AI Max+ 395 systems. StepFun's release cadence across 2025-2026 -- Step-Audio 2 and Step-3's technical papers in July 2025, open-sourced Step-Audio 2 mini variants and a vLLM backend that August-September, Step 3.5 Flash in February 2026, and Step 3.7 Flash in May 2026 -- shows a major model-family update roughly every one to three months, with Step-Audio 2 also distributed as a standard pip-installable PyPI package alongside its GitHub source.[CE003, CE016, CE029, CE030, CE033, CE036]

5.4 Benchmark credibility and competitive positioning

StepFun's own release material is unusually explicit that some of its headline comparisons are vendor-stated rather than independently reproduced: its Step 3.7 Flash benchmark chart notes that a rival's Terminal-Bench 2.1 score came from StepFun's own internal testing while other vendors' figures are their own reported results on a different benchmark version (Terminal-Bench 2.0), and it leads with the higher, contamination-flagged 76.5% SWE-Bench Verified score even though the more contamination-resistant SWE-Bench Pro figure (56.3%, good for 15th of 35 tracked models) is the cleaner signal per independent analysis. Independent hands-on testing on an NVIDIA DGX Spark did report a genuine positive: a 100% tool-call success rate across a multi-step evaluation suite, consistent with (though not proof of) the model's first-place 67.1 ClawEval-1.1 score. But a broader, vendor-neutral aggregator (Crafiq) ranks Step 3.7 Flash a comparatively modest #108 overall on intelligence and #89 on coding -- a useful corrective to the narrower agentic-benchmark framing StepFun's own material emphasizes. StepFun also reports strong search-augmented scores (79.2 SimpleVQA-with-tool, 75.82% BrowseComp, 92.82% DeepSearchQA) and enterprise-occupation coverage (45.8% GDPval across 44 occupations, 98%+ pass rate on Tau2-bench Telecom tiers), and narrowed cross-harness performance variance from 43-73% (Step 3.5 Flash) to 64.5-71.5% on its own Step-SWE-Bench -- all still self-measured. Independent review coverage of Step 3.5 Flash's February 2026 debut, by contrast, generated real Hacker News and local-inference community buzz, and StepFun's GitHub organization backs that up with roughly 1,700 followers and 32 public repositories, including a 2,100+-star Step-3.5-Flash repo and a rapidly-growing Step-3.7-Flash repo.[CE017, CE018, CE019, CE020, CE021, CE022]

5.5 Trust, safety, security, and compliance

StepFun's open-weight releases (Step 3.7 Flash, Step 3.5 Flash, Step-Audio 2, Step-Video-T2V, Step1X-Edit) are consistently distributed under the permissive Apache 2.0 license, which is a genuine trust and transparency positive versus fully closed competitors. Set against that, several trust gaps remain undocumented in the public record reviewed for this chapter. China's national AI-Generated Synthetic Content Labeling Measures, effective from their March 2025 issuance, require visible and technical (watermark/metadata) labeling of AI-generated audio, image, and video content -- an obligation that squarely covers Step-Audio's voice-cloning/TTS outputs and the step-2x-large/step-image-edit-2 image models, yet no StepFun-specific labeling or watermarking implementation is described in the sources reviewed. Separately, Chinese authorities (CNCERT/CC, MIIT, and the National Computer Virus Emergency Response Center) have warned about Trojan-laced counterfeit skill packages on the OpenClaw agent platform -- the same ecosystem StepFun's own release material lists as a supported, highly-ranked integration for both Step 3.5 and Step 3.7 Flash -- without any disclosed StepFun-specific skill-verification or sandboxing control. No independent SOC 2, ISO 27001, or comparable security certification is disclosed for the platform.stepfun.com API service, and the only available reliability signal is a low-reputation, crowd-sourced uptime monitor (operating normally as of July 1, 2026, 26 problem reports in 30 days) rather than a StepFun-published SLA or incident postmortem. Finally, Step-Audio 2's claimed state-of-the-art benchmark performance versus other open-source and commercial audio models is a comparison authored entirely by the StepFun team, not an independent evaluator.[CE024, CE025, CE026, CE027, CE028, CE040]

5.6 Exhibits

6.1 Consumer app and developer ecosystem: broad reach, weak independent measurement

StepFun's most direct customer relationships sit at the two ends of its distribution stack: the free-to-use Yuewen (跃问) consumer app and the open-weight developer ecosystem built around Hugging Face and GitHub. Neither shows strong independently-measured traction on its own. Yuewen's Google Play listing carries only 58 reviews at a 4.2 rating as of its May 2026 update, and multiple named reviewers complain that the app blocks Google, email, or social login and only accepts a phone number -- friction serious enough that one reviewer reports being unable to chat or copy text at all without providing one. Independently, AICPB's April 2026 China AI app ranking lists more than two dozen apps by monthly active users, from ByteDance's Doubao at 336 million down to under a million for the smallest ranked entry, and Yuewen does not appear anywhere on that list. The developer side looks healthier by comparison: the stepfun-ai Hugging Face organization shows individual open-weight model variants with download counts in the hundreds of thousands, and the stepfun-ai GitHub organization maintains active repositories, including a vLLM fork, with ongoing commit activity through June 2026. Step 3.7 Flash was also one of five major new models OpenRouter listed in a ten-day window ending June 4, 2026, alongside frontier releases from Anthropic and others, indicating the model remains commercially current on third-party marketplaces. None of this developer signal, however, distinguishes paying API accounts from free experimentation, and no source reviewed for this chapter discloses a developer-to-paid conversion rate.[CU001, CU002, CU020, CU003, CU004, CU033]

6.2 Device, automotive, and enterprise customer proof: real named deployments, thin per-account detail

StepFun's strongest customer evidence is device-led. Independent Chinese tech press directly ties named, shipped features to StepFun: 科创板日报 reported that OPPO's Find X8 Ultra flagship's 'one-tap memory capture' feature runs on StepFun's multimodal technology, Honor gave StepFun a 'Best Collaboration Award' for powering the Magic8 series' YOYO assistant memory feature, and ZTE co-launched an elder-care voice assistant with StepFun in July 2025 while its Nubia Z80 Ultra phone became the first device to natively embed the OpenClaw agent StepFun's release material lists as a supported integration. Sina Technology's coverage of StepFun's late-2025 GELab-Zero open-source launch puts a number on this reach: over 60% of China's domestic phone brands. On the automotive side, Geely Auto Group and Qianli Technology jointly showcased the AgentOS smart-cockpit voice agent with StepFun at WAIC 2025, and the lead vehicle, the Geely Galaxy M9, sold roughly 40,000 units in its first three months, with StepFun targeting over one million vehicles for 2026. Tencent Cloud has separately announced a strategic partnership to build AI cockpit agents on StepFun's models, and Huanxi Media Group signed a Hong-Kong-filed framework agreement -- confirmed by both the HKEXnews regulatory filing and law firm Wilson Sonsini -- to form a joint venture applying StepFun's models to film, TV, and game content. What is consistently missing across every one of these relationships is per-account commercial detail: no disclosed unit royalty, contract length, or revenue-share figure exists for any single OEM, automotive, or enterprise partner, so this evidence proves feature-level production use, not verified commercial scale.[CU024, CU025, CU027, CU012, CU026, CU007]

6.3 Retention and durability: no disclosed metric in any customer segment

No source reviewed for this chapter discloses net revenue retention, gross revenue retention, renewal rate, or churn for any StepFun customer segment -- device, automotive, developer, or enterprise. The closest available proxy is a StepFun-sourced claim that terminal Agent API call volume grew nearly 170% quarter-on-quarter for three consecutive quarters through late 2025, but this figure has no disclosed absolute base, is not corroborated by an independent analytics provider, and measures call volume rather than customer retention. On the consumer side, the Yuewen app's thin, mixed Google Play review base (58 reviews, 4.2 rating, with recurring login-friction complaints) is the only publicly visible satisfaction signal, and it is too small a sample to generalize about retention. Contract-level detail is equally absent: no source specifies a contract length, renewal date, or minimum-volume commitment for any named OEM, automotive, or enterprise relationship, which means it is impossible to tell whether StepFun's device and vehicle partnerships are durable multi-year commitments or agreements that could lapse or be renegotiated on short notice. This chapter treats retention, paying-customer count, and contract-term visibility as explicit, material evidence gaps rather than estimating placeholder figures.[CU030, CU034, CU035, CU036, CU001, CU002]

6.4 Expansion and concentration risk: a wide device footprint riding on a narrow partner set

StepFun's expansion story reads as broad -- phones, cars, cloud, media, robotics -- but the underlying distribution depends on a small number of relationships. Three phone OEM groups (OPPO, Honor, ZTE) carry the entire 42-million-device, 60%-brand-share headline, and Qianli Technology, StepFun's automotive cockpit vehicle, is itself financially exposed: its own IPO prospectus discloses that its top five customers took 41.5%-48.1% of revenue between 2022 and mid-2025, with Geely Group alone contributing 30.8%-39.7%, while Qianli posted net losses in three of the last four disclosed periods. A June 2026 Sina Finance report adds that Qianli's controlling group raised 2026 related-party transaction limits over 80% year-on-year and approved a roughly 2x-premium acquisition of a related, insolvent radar-sensing company without a performance-compensation clause -- governance signals that should concern anyone relying on Qianli as StepFun's primary automotive distribution channel. A separate, independent business profile notes that StepFun's own chairman also controls Qianli, meaning the two companies sit inside one Geely-anchored group rather than being arm's-length counterparties. Tencent compounds this pattern: it is simultaneously a multi-round StepFun shareholder and the named commercial partner for AI cockpit agents, a structure that makes it harder to confirm the automotive partnership was priced at arm's length. Enterprise-vertical diversification (Huanxi Media, Shanghai United Media Group, Zhiyuan Robotics) remains at the framework-agreement or early-partnership stage, without a completed joint venture or disclosed revenue, so it is not yet a meaningful counterweight to the device/automotive concentration.[CU037, CU038, CU015, CU016, CU017, CU018]

6.5 Distribution channel dependency and diligence verdict

Ranking StepFun's customer channels by how independently their evidence can be verified produces a clear ordering. Regulatory filings and third-party platform telemetry sit at the top: the HKEXnews filing and Wilson Sonsini's legal confirmation independently corroborate the Huanxi Media joint venture, and Hugging Face/GitHub download and commit data independently corroborate developer-ecosystem activity. Device and automotive evidence sits in the middle: named feature press coverage (OPPO, Honor, ZTE) and joint-showcase or press-release material (Geely/Qianli, Tencent) are independently reported but ultimately trace back to company or partner disclosure of a specific feature or event, not to audited usage or revenue data. The weakest evidence is the consumer app: Yuewen's absence from AICPB's independent MAU ranking and its thin, complaint-heavy Google Play review base both point the same direction -- StepFun has not converted its device-level reach into an independently measurable consumer audience of its own. Overall, this chapter finds genuine, verifiable customer proof at the device and automotive feature level, but no disclosed paying-customer count, retention metric, or contract term anywhere in the portfolio, and a Qianli/Geely/Tencent ownership and investment structure that concentrates the automotive channel inside a small, interlinked group rather than a diversified customer base.[CU040, CU041, CU042, CU020, CU001]

6.6 Exhibits

7.1 Regulatory and listing risk: a red-chip unwind and a stacked 2026 compliance calendar

StepFun's most immediate structural risk is regulatory rather than technical. Reuters reported on April 13, 2026 that StepFun is unwinding its offshore Cayman Islands 'red-chip' incorporation structure to pave the way for a planned Hong Kong IPO, after China's securities regulator instructed some offshore-registered companies to dismantle such structures over shareholder-transparency and compliance-risk concerns. Two sources told Reuters the decision reflected StepFun being 'heavily backed by state capital,' including Shanghai municipal and district government investment vehicles, and by April 2026 the company had completed conversion into a joint-stock company, with Qianli Technology chairman Yin Qi formally registered as StepFun's chairman. Under the Trial Measures that have governed overseas listings since March 2023, indirect listings via red-chip structures require CSRC filing within three working days of an overseas listing application and face a 20-working-day review window; legal commentary notes that unwinding an existing red-chip structure can be legally complex and cost-prohibitive, with some companies delaying or abandoning IPO plans altogether. Layered on top of that listing-structure risk is a stacked compliance calendar: China's amended Cybersecurity Law took effect January 1, 2026 with a materially heavier tiered penalty regime (fines up to RMB 10 million or more for critical-infrastructure violations) and explicit AI-governance language, while the AI-generated-content Labeling Measures and mandatory GB 45438-2025 standard have required dual explicit/implicit labeling since September 1, 2025. On the opportunity side, HKEX's March 2026 consultation proposes lowering weighted-voting-rights listing thresholds, which could ease (not tighten) StepFun's path to a founder-friendly listing structure if adopted before its filing window closes.[CR001, CR002, CR003, CR004, CR005, CR006]

7.2 Compute and supply-chain risk: export-control tightening meets a still-trailing domestic chip stack

StepFun's second structural risk sits upstream, in the compute it needs to keep training frontier-competitive models. On May 31, 2026 the U.S. Bureau of Industry and Security closed a loophole that had let China-headquartered firms buy advanced Nvidia/AMD chips through overseas subsidiaries in Singapore, Malaysia, and Taiwan without an export license, resting on an 'ultimate parent company' test that has technically been in force since November 2023 and now creates retroactive compliance exposure for any such transactions. Tech Times reported that Trump administration officials were internally divided in June 2026 over whether this had been an inadvertent policy gap following a 2025 decision to stop enforcing the Biden-era AI Diffusion Rule. The same month, BIS added dozens more Chinese entities to the Entity List for advanced-AI and high-performance-chip development tied to China's military-industrial complex, even as CNBC reported the U.S. was holding off formally blacklisting DeepSeek, CXMT, and 100-plus other flagged firms to avoid escalating tensions with Beijing -- and the Pentagon's June 2026 '1260H list' update added Alibaba, Baidu, and BYD. No source reviewed names StepFun directly on either list as of this chapter's run date, but the rapid expansion of both rosters signals adjacency risk for any state-backed Chinese AI unicorn. StepFun's own hedge is the Model-Chip Ecosystem Innovation Alliance it co-founded in July 2025 with roughly ten domestic chip vendors including Huawei Ascend, MetaX, Biren, Enflame, Iluvatar CoreX, and Moore Threads; Huawei's newest Ascend 950PR delivers roughly 2.8x the inference throughput of Nvidia's export-approved H20 but draws about 200W more power, and independent analysis finds the domestic stack still trails Nvidia's Blackwell-class chips on per-chip efficiency and CUDA-ecosystem maturity, compensating with cluster-scale aggregation at nearly 4x the power draw. Beneath both risks sits the sheer capital intensity of frontier AI: a 100-to-10,000-GPU cluster requires $5 million to over $500 million upfront in 2026, with hardware itself representing only about 35% of five-year total cost of ownership once networking, power, cooling, and staffing are included -- a capital bar StepFun must clear against far larger competitors.[CR016, CR017, CR018, CR019, CR020, CR021]

7.3 Competitive and pricing risk: a consolidating field, an active price war, and a widening distribution gap

StepFun competes inside a China foundation-model field that KrASIA and JPMorgan both describe as consolidating fast. KrASIA groups StepFun with 01.AI, Baichuan AI, MiniMax, Moonshot AI, and Zhipu as China's 'AI tigers,' noting that Zhipu and MiniMax had already listed in Hong Kong by early 2026 while the rest, including StepFun, remain private and under funding pressure. JPMorgan's February 2026 sector note argues the number of viable Chinese model developers has shrunk from more than 200 to fewer than ten, with future winners determined by deployment scale and monetization efficiency rather than model-release cadence -- and the bank initiated public analyst coverage on Zhipu and MiniMax with 'Overweight' ratings, but not on StepFun, consistent with StepFun's comparatively thin public disclosure. The revenue gap is material: Zhipu's 2025 revenue reached RMB 724.3 million (up from RMB 312.4 million in 2024), well above StepFun's previously reported roughly RMB 500 million, while ByteDance alone plans to spend more than RMB 160 billion on AI procurement in 2026 and Tencent's 2025 capital expenditure reached RMB 79.2 billion -- an order of magnitude beyond the infrastructure scale StepFun has disclosed. This consolidation is compounding an active price war: Chinese labs cut API prices repeatedly through 2026, compressing per-token revenue just as compute costs remain capital-intensive, squeezing StepFun's unit economics from both the demand and supply side simultaneously, with no evidence in the sources reviewed here that StepFun has achieved differentiated pricing power versus larger, better-capitalized rivals.[CR028, CR029, CR030, CR031, CR032]

7.4 Customer concentration and channel risk: a device footprint riding on a financially fragile automotive partner

StepFun's clearest device-adjacent proof points -- smartphone embedding with OPPO, Honor, and ZTE, and automotive cockpit AI with Geely -- are also its narrowest channel dependencies. On the smartphone side, three OEM groups reportedly carry StepFun's entire roughly 42-million-device distribution headline; OPPO's own public compliance statement commits only to general legal and data-protection compliance across its 60-plus countries, with no StepFun-specific service-level terms disclosed. On the automotive side, the channel runs entirely through Qianli Technology, the Geely-linked automotive-AI vehicle that is also chaired by StepFun's own chairman, Yin Qi. Qianli's 2025 annual report shows sales to Geely-affiliated companies at 29.26% of total revenue (ranging 30.8%-39.7% across 2022-2025), and its shareholders approved 2025-2026 related-party-transaction limits with Geely entities on December 31, 2025 with over 99% approval; a June 2025 cninfo filing shows Qianli, Geely, Megvii-linked Yin Qi entities, and Chongqing state funds jointly structuring an intelligent-driving joint venture with RMB 1.5 billion in committed capital, deepening rather than diversifying the relationship. Critically, Qianli's own second-quarter-2026 prospectus update shows its core, non-recurring-excluded net profit has been negative for three straight years despite 42.1% revenue growth in 2025, with its automotive-business gross margin at just 2.8% and the parent company's cumulative unrecovered losses near RMB 948 million -- meaning the channel StepFun's automotive proof runs through is itself under real financial strain. StepFun's parallel hedge against smartphone concentration is its Model-Chip Ecosystem Innovation Alliance and its state-linked capital base (Shanghai municipal/district funds and the Hong Kong Investment Corporation), but none of these reduce the single-point dependency on Qianli for automotive revenue.[CR033, CR034, CR035, CR036, CR037, CR038]

7.5 Governance and disclosure risk: an interlocking directorate and no audited baseline

StepFun's governance profile carries two compounding gaps. First, its chairman, Yin Qi, concurrently chairs Qianli Technology -- the same Geely-linked entity that is StepFun's primary, and only proven, automotive distribution counterparty -- meaning the two companies' most important commercial relationship is negotiated across an interlocking directorate rather than at arm's length. No board minutes, related-party-transaction approval process, or independent-director attestation covering StepFun-Qianli dealings was found among the sources reviewed for this chapter. Second, no audited 2025 financial statements or filed HKEX prospectus for StepFun were retrievable as of this chapter's run date, leaving its reported roughly RMB 500 million 2025 revenue and targeted $10-12 billion valuation without an independently verified baseline -- a gap that matters given the scale of the implied revenue multiple. StepFun's investor base compounds the transparency question: Reuters reports state-linked capital including Shanghai municipal and district investment vehicles, and the 2026 pre-IPO round separately drew participation from the Hong Kong Investment Corporation, a fund wholly owned by the Hong Kong government whose own disclosure describes a dual mandate of financial return and enhancing Hong Kong's competitiveness. That dual mandate is not inherently adverse, but it means StepFun's ownership and governance decisions -- including the IPO domicile choice itself -- are shaped in part by public-policy objectives that a purely private-capital company would not carry, and the full cap table across StepFun's angel-through-pre-IPO rounds remains only partially disclosed via press reporting rather than a primary filing.[CR041, CR042, CR043, CR044]

7.6 Technical and product risk: benchmark opacity, unverified security posture, and a growing labeling burden

StepFun's technical risk sits less in raw model quality than in what cannot be independently verified about it. Industry-wide analysis of 2.8 million LMArena model-comparison records found selective submissions had inflated leaderboard scores by up to 100 points through cherry-picking, with major labs submitting only their best-performing private-test variants rather than shipped production models; this pattern is not StepFun-specific, but no independent, StepFun-authorized reproduction of its own flagship agentic-benchmark claims was found among the sources reviewed here, leaving the same opacity risk unresolved for StepFun specifically. Separately, no independent security certification, bug-bounty program, or public incident-response disclosure for StepFun's platform, Step-Audio, or Step-Video was located, and no public litigation, data-breach disclosure, or class-action record naming StepFun was found either -- an absence that should be read as an evidence gap rather than a confirmed clean record. Compliance-wise, China's dual-track explicit/implicit AI-content-labeling regime has been mandatory since September 2025, and major platforms including Tencent, Douyin, RedNote, Weibo, and DeepSeek have already published compliance statements, while no equivalent public statement from StepFun was found for its Yuewen consumer app or API outputs. The compliance surface is still expanding: the Cyberspace Administration circulated a draft in April 2026 proposing additional mandatory ethics review, emergency procedures, and content-governance obligations specifically for AI services that simulate human interaction, expected to take effect around July 2026, which would add further recurring engineering and audit burden on top of the labeling regime StepFun's consumer surfaces must already clear.[CR045, CR046, CR047, CR048, CR049, CR050]

7.7 Synthesis: mitigation maturity, monitorable triggers, and where the thesis breaks

Read together, StepFun's highest-severity risks cluster in three places: listing-execution risk from the CSRC red-chip unwind and stacked 2026 compliance calendar; compute-supply risk from tightening US export-control enforcement against a domestic-chip alternative that still trails Nvidia on efficiency and software maturity; and a governance/channel risk compounded by an interlocking directorate between StepFun's chairman and its sole proven automotive-channel counterparty, Qianli Technology, which is itself under real financial strain. None of these risks are individually fatal -- red-chip unwinds have completed successfully elsewhere, domestic chips are improving generation over generation, and interlocking directorates are common in China's state-linked technology sector -- but their mitigation maturity is uniformly low: almost none of the risk items in this chapter's registers have a StepFun-disclosed formal mitigation beyond the underlying relationship or restructuring itself. The clearest thesis-break triggers are therefore observable from outside the company: CSRC filing closure or an HKEX prospectus withdrawal/delay past year-end 2026; Entity List or Pentagon 1260H designation of StepFun, Qianli, or a named chip supplier; a documented CAC enforcement action against StepFun specifically; or a related-party-transaction disclosure showing non-arm's-length terms between StepFun and Qianli. Each of these is a public, monitorable event rather than a private one, which means diligence can track StepFun's risk trajectory between report refreshes without relying on company disclosure alone.[CR001, CR009, CR016, CR028, CR030, CR033]

7.8 Exhibits

8.1 Recommendation: a real device wedge priced at a level public disclosure cannot yet underwrite

StepFun enters this diligence with a genuinely differentiated distribution story: a completed share reform, a dismantled red-chip structure, and a nearly $2.5 billion pre-IPO round that drew the Hong Kong Investment Corporation alongside smartphone-ODM and component investors Huaqin, Longcheer, OmniVision, and ZTE. That investor mix is not generic financial capital; it is the same supply chain StepFun needs to keep its models embedded on device. Founder Jiang Daxin, a former Microsoft vice president, has built a roughly 500-person, 80%-technical team since April 2023, and the company has drawn outside recognition as one of the Chinese AI labs worth watching beyond DeepSeek. None of that resolves the pricing question this chapter must answer. StepFun's Hong Kong IPO targets a $10-12 billion valuation against a reported $73.3 million of 2025 revenue and a market projection of $175.9 million for 2026 -- implied multiples of roughly 137-164x trailing and 57-68x forward revenue. Independent commentary explicitly ties the $12 billion figure to 'national champion' sentiment and state-backed capital rather than to disclosed fundamentals, and no source reviewed here supplies an audited financial statement, cash position, or filed prospectus. Given a real but unverified thesis, a stretched-but-not-extreme multiple, and a thin evidence base, the appropriate call is research-more, not buy: track the filing and first two post-listing quarters, and do not commit capital above the low end of the range absent audited numbers.[CV001, CV002, CV003, CV005, CV036, CV038]

8.2 Price context: stretched versus public benchmarks, mid-pack versus its own Chinese AI-tiger peers

The most useful lens here is not one multiple but a stack of them across a genuine peer set. Zhipu AI, the first Chinese large-model company to list in Hong Kong (January 8, 2026), reported FY2025 revenue of RMB724 million against a net loss of RMB4.72 billion; on a mid-2026 market value of roughly HK$495.8 billion, that implies a static price-to-sales ratio near 576x -- an outlier even against the 20-40x ceiling independent analysis treats as extreme for top-tier global SaaS. MiniMax, listed one day later, reported $79 million of FY2025 revenue at a 25.4% gross margin and now trades near a $16.8 billion market cap, more than two-thirds below its 52-week high, implying roughly 212x revenue. Moonshot AI's private valuation rose from $4.3 billion to $20 billion in five months on the back of $200 million of annualized recurring revenue, near 100x ARR; DeepSeek's first external round priced it above $50 billion with no disclosed revenue metric at all; Baichuan Intelligence, pursuing a narrower healthcare-AI pivot, sits near $2.7-2.8 billion, the low end of this comparable set. Against that backdrop, StepFun's implied 57-68x forward multiple is stretched relative to Palantir's ~53.5x public-market benchmark, but it is not the richest mark in its own comp set -- Zhipu and MiniMax both trade far higher on a revenue-multiple basis, propped up in part by thin free float, Hang Seng Tech Index inclusion flows, and cornerstone lock-up mechanics rather than by audited margin or cash-generation evidence. StepFun's own pricing therefore looks defensible only if the 2026 revenue projection holds and the deal prices toward $10 billion rather than $12 billion.[CV010, CV011, CV012, CV013, CV020, CV021]

8.3 Scenario logic: index-flow-driven upside is real, but so is a fast reset if execution slips

The bull case is straightforward: if StepFun's device and automotive footprint (40M+ smartphones, 1M+ target Geely vehicles) converts into monetizable API and licensing revenue at or above the $175.9 million 2026 projection, and the IPO prices near the top of its range with the same index-inclusion and cornerstone-lock-up support that has cushioned Zhipu's share price, StepFun could plausibly compound toward the $12-20 billion range over the following year -- approaching Moonshot's current mark. That outcome requires flawless execution: no repeat of the mid-2025 revenue-guidance miss already flagged in Chapter 4, continued device-partner cooperation, and no adverse turn in the CSRC-driven listing review. The base and bear cases are easier to defend from public evidence. Both Zhipu and MiniMax have swung more than 40-70% from their post-listing highs within six months of trading, so even a clean StepFun listing near $10-11 billion offers no guarantee of price stability -- Chinese AI-lab valuations in 2026 are as much a function of index flows, cornerstone lock-ups, and scarcity narratives as of underlying economics. The bear case adds a further downside path: a delayed or downsized listing tied to the ongoing red-chip and CSRC review, a repeat of 2025's guidance miss, or a punitive device-partner revenue share, any of which could reset the mark back toward the $4-6 billion pre-money range earlier 2026 reporting had already disclosed for StepFun.[CV026, CV027, CV036, CV040, CV041, CV042]

8.4 Exit readiness and final diligence: the asks that would change this call

StepFun has cleared the structural prerequisites for a listing -- share reform, red-chip unwind, and a cornerstone lineup that already includes state-linked and industrial capital of the kind that supported Zhipu's post-listing resilience. That is a necessary but not sufficient condition for a constructive valuation call. The evidence gap is specific and closeable: an audited prospectus with cash, burn, gross margin, and segment revenue; a cap table showing what HKIC, Shanghai state capital, and industrial investors actually own after the round; the final IPO pricing and raise size; and StepFun-specific device and OEM revenue-share terms rather than the industry-wide 30-60% norm cited elsewhere in this diligence. Each of these closes a distinct thesis-break trigger. A listing priced at or above $12 billion without new disclosure pushes the forward multiple past even Zhipu's stretched, index-inflated benchmark once those flow effects are stripped out. A post-listing swing on the scale Zhipu and MiniMax both experienced would confirm that this comp set's liquidity-driven repricing pattern applies to StepFun too, meaning any entry price at listing should be treated as provisional. And a further round of API price cuts across Zhipu, Kimi, and MiniMax -- already documented through mid-2026 -- signals a margin ceiling that applies to StepFun's own token-priced business regardless of how the IPO prices.[CV002, CV038, CV039, CV044, CV047, CV048]

8.5 Exhibits