护城河对比

The FY2025 20-F positions Nebius as 'one of the few global, at scale, multi-tenant clouds purpose built for AI,' but the same filing names its central dependency plainly: 'We currently rely on Nvidia for the GPU chips we use' — a purpose-built neocloud renting a supply the hyperscalers it competes with also control, which is a real but narrow position.

来源：sec.gov

The FY2025 10-K says 'we believe we have established a defensible competitive advantage through our power-first, innovation-driven strategy', resting on a power-native team, an application-agnostic framework for digital infrastructure design, end-to-end greenfield development capabilities, and the use of ASIC compute builds to rapidly and cost-effectively secure and monetize power. What separates this from no moat is scarce power: the filing says demand for energy capacity continues to outpace supply and that 'grid interconnection bottlenecks have further constrained access to power and digital infrastructure development', and Hut 8 reports 330 MW of utility capacity under construction at River Bend, commercialized in December 2025 through a 15-year triple-net lease supporting 245 MW of IT capacity, plus 1,230 MW of utility capacity under development. The same filing says 'The industries in which we operate are highly competitive and continuously evolving.', warns that these factors 'might not provide the competitive advantage we anticipate, or if they do, such competitive advantage might not endure', cites the River Bend lease (valued at $7.0 billion over the base term) as an example of expecting a significant portion of revenue from a limited number of customers, and says the company has not maintained consistent profitability. That supports narrow at most, not wide.

来源：sec.gov

The FY2026 10-K grounds a real but bounded defence. On the defence side: Item 1 states that in the manufacturing services market "there are significant barriers to entry in our existing and target markets, including lengthy sales cycles, the need to demonstrate complex precision optical and electro-mechanical engineering and manufacturing capabilities to a prospective customer and the ability to protect a customer's intellectual property," and that qualification of a program "may take three to six months or longer to complete." On the limiting side, the same filing shows the protection does not reach price: gross profit was 12.0% of revenues in FY2026 against 12.1% in FY2025 and 12.4% in FY2024 even as revenue rose 35.7% to $4.64 billion, four customers each exceeded 10% of revenue (Cisco 19.9%, NVIDIA 16.3%, Nokia 10.7%, Amazon 10.5%), and "reliance on a small number of customers gives those customers substantial purchasing power and leverage in negotiating contracts with us." Sales are made on "individual purchase orders that have short lead times and are subject to revision or cancellation," so the qualified program is sticky while the contract is not.

来源：sec.gov

The moat the filing describes is engineering: designs that 'optimize power and cooling efficiency, lower latency' plus 'a consistent track record of being one of the first-to-deploy the latest generation of NVIDIA GPU chips' — density and time-to-deploy at scale, not a switching lock.

来源：sec.gov

The 10-K says 'we believe success depends on the ability to secure scarce assets like power, data center equipment, and customers', and places part of its claimed edge in 'our ability to use ASIC compute infrastructure development to rapidly and cost-effectively secure and monetize power'. Its examples: Vega was energized less than a year after acquisition for an all-in cost of about $455,000 per megawatt, which the filing calls 'a fraction of traditional data center costs', and Salt Creek was completed for about $250,000 per MW; Vega also has immediate access to some of the lowest locational wholesale power prices in North America. These are low-redundancy ASIC compute builds, which the 10-K says carry relatively low capital intensity by design. The 'fraction of traditional data center costs' comparison is against a different facility tier, and the filing gives no peer cost comparison and no cost figure for its AI campuses.

来源：sec.gov

The 10-K locates the stickiness in qualification rather than in patents or scale. Customers "do not purchase our services until they qualify the services and satisfactorily complete factory audits and vendor evaluations"; qualification "may take three to six months or longer"; production is transferred "copy-exact: the setup of a production process identical to the one used by our customer"; and the filing's own experience of changing a component source is that it "resulted in our customers or their end customers requiring requalification and validation of components, a process that can often be lengthy and has negatively impacted the timing of our revenue" — the same friction a customer would face moving the program elsewhere. The filing rules out an IP-based moat itself: "Historically, patents have not played a significant role in the protection of our proprietary rights," and any process improvement developed for a customer's product is "immediately assigned to that customer." The Thailand cost base is a genuine advantage but the filing warns it is not durable on its own — "Wage increases may impact our ability to sustain our competitive advantage and may reduce our profit margin."

来源：sec.gov

The 20-F's own claim is 'one of the FEW global, at scale' AI clouds — a differentiated challenger to the hyperscalers it names as the competitive field, not a claimed leader of it; the first-to-deploy record is a follower's speed advantage, not category leadership.

来源：sec.gov

The 10-K says 'For example, we signed our first large scale single tenant AI data center lease and commenced construction of the site at our River Bend campus.', offered under the statement that to remain competitive with peers it may need to modify aspects of its business model. River Bend was commercialized in December 2025 through a 15-year triple-net lease supporting 245 MW of IT capacity; Fluidstack is expected to serve as tenant and Google is expected to provide a financial backstop, and initial delivery is targeted for Q2 2027. On speed, the filing says 'Through early engagement with Entergy Louisiana, we assumed key interconnection and development functions typically managed by the utility, such as the development of transmission and distribution infrastructure, materially accelerating originally quoted power delivery timelines.', and that Vega was energized 'less than a year after acquiring it'. It also warns that Hut 8 competes against companies 'that may be more established or have greater financial and other resources and/or expertise'. That describes a newer entrant moving quickly, not an established leader.

来源：sec.gov

The leadership claim in the filing is the company's own qualitative assessment: "We believe we are a leader in manufacturing products for the optical communications market," and, on breadth of process technology, "Based on our experience with customers and our qualitative assessment of our capabilities, we believe we provide a broader array of process technologies to the optics industry than any other manufacturing services provider." The same Item 1 names a crowded field against it — "Benchmark Electronics, Inc., Celestica Inc., InnoLight Technology (Suzhou) Ltd., Jabil Inc., Sanmina Corporation, Venture Corporation Limited and Eoptolink Technology Inc., Ltd., as well as the internal manufacturing capabilities of our customers" — and no independent share ranking is given, so the record supports front-of-field standing but not sole primacy.

来源：sec.gov

The filing frames a 'highly competitive' market with 'frequent introduction of new or improved solutions' and a GPU cost base set by a single supplier (NVIDIA) — a renter of compute competing on efficiency has little list-price control.

来源：sec.gov

The 10-K says ASIC compute, run through American Bitcoin, can only be profitable if hardware and electricity costs are below the price of the Bitcoin mined, which makes it a price-taking business. It also says Traditional Cloud and AI Cloud compete with cloud services providers for customers, and that the company has not maintained consistent profitability. River Bend's contracted triple-net lease economics are not in service yet; initial delivery is targeted for Q2 2027.

来源：sec.gov

Gross profit was 12.0% of revenues in FY2026, 12.1% in FY2025 and 12.4% in FY2024 — flat to down across three years in which revenue grew from $2.88 billion to $4.64 billion — and the FY2026 increase in gross profit is attributed to "sales volume and product mix," not price. Item 1 says so directly: "we expect the prices we charge for our manufactured products to decrease over time (partly as a result of competitive market forces)," with the offset coming from cycle-time, mix, yield and material-cost work rather than from rate. The risk factors add that customer consolidation gives buyers "increased leverage that may result in, among other things, decreases in our average selling prices," and that new competition "could result in price reductions for our services, reduced gross profit margins or loss of market share."

来源：sec.gov

Layer-8 neocloud — a purpose-built AI compute provider reselling NVIDIA silicon at scale.

Power-first developer building AI data center capacity, beginning with River Bend under a 15-year lease (Fluidstack is the expected tenant and Google is expected to provide a financial backstop; initial delivery is targeted for Q2 2027), and runs Bitcoin mining through majority-owned American Bitcoin.

A contract manufacturer sitting between optical components and the network and AI-datacenter OEMs whose brands ship the finished box. FY2026 revenue is 47.9% data center, 33.3% communications infrastructure and 18.8% automotive, industrial and other markets (FY2025: 46.2% / 30.7% / 23.1%), with Cisco, NVIDIA, Nokia and Amazon each above 10% of revenue. Fabrinet also integrates one layer down, designing and fabricating its own customized optics and glass — crystals, ferrules, precision glass tubing — at Fuzhou, China and Mountain Lakes, New Jersey, both for its own assemblies and for the merchant market, while volume manufacturing runs from the Pinehurst and Chonburi campuses in Thailand.

• Nebius AI cloud (GPU compute + managed AI/ML platform) 挑战者 · 中等 来源：sec.gov
• Toloka / Avride / TripleTen (non-cloud segments) 未知 · 低 来源：sec.gov

• AI data center campuses (River Bend, Beacon Point) 未知 · 中等 来源：sec.gov
• ASIC compute (American Bitcoin) 未知 · 低 来源：sec.gov
• Highrise AI 细分市场 · 低 来源：sec.gov
• Hut 8 Canada (traditional cloud and colocation) 细分市场 · 低 来源：sec.gov

• Automotive and medical sensors 细分市场 · 中等 来源：sec.gov
• Communications infrastructure optical products 前三 · 深 来源：sec.gov
• Customized optics and glass 挑战者 · 中等 来源：sec.gov
• Data center optical modules and transceivers 前三 · 深 来源：sec.gov
• Industrial laser manufacturing 细分市场 · 中等 来源：sec.gov

编辑先验，未经回测。

查看它究竟如何投票 →

编辑先验，未经回测。

查看它究竟如何投票 →

编辑先验，未经回测。

查看它究竟如何投票 →