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

宽阔：CUDA 软件平台及其开发者既有用户基础在整个 AI 训练技术栈上形成高昂的转换成本，由累计 $76.7B 研发投入护持，并被 FY2026 约 71% 的毛利率以及同比增长 68% 的数据中心营收（约占总额 90%）所强化——按 FY2026 10-K，这些是持久且难以复制的优势。

来源：sec.gov

The FY2025 20-F Competition section cuts both ways: UMC calls itself "one of the world's largest independent semiconductor foundries" with production capacity "currently among the largest of all semiconductor foundries in the world" and says it believes it competes favorably with competitors on each principal element of foundry competition, yet in the same section concedes that "[s]ome of our competitors have substantially greater production, financial, research and development and marketing resources than we have" and that "several new dedicated foundries have commenced operations and compete directly with us," whose entry is "likely to initiate a trend of competitive pricing and create potential overcapacity in legacy technology" that "may erode our profit margins." A real but bounded advantage, not a durable wide one.

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

持久优势建立在专有 IP 与软件之上：可在所有 NVIDIA GPU 上运行的全栈 CUDA 开发平台，加上数百个专有领域库/SDK/API，以及累计 $76.7B 研发投入所带来的 "inventions that are essential to modern computing"（NVIDIA 于 1999 年发明了 GPU）。按 FY2026 10-K 的业务章节，这一优势还被开发者生态系统的网络效应与 CUDA 转换成本进一步强化。

来源：sec.gov

The 20-F states plainly that "Our success depends on our ability to obtain patents, licenses and other intellectual property (IP) rights covering our production processes and activities," and reports more than 16,700 patents worldwide as of December 31, 2025, of which more than 7,490 are in the U.S., 623 domestic and foreign patents granted in 2025 alone, NT$17,725 million of R&D (7.5% of operating revenues) and 1,591 R&D professionals. The Competition section names where UMC believes it wins: it competes favorably "particularly [on] our technical competence and research and development capabilities." The proprietary specialty-process library, not scale or a network, is the cited source of the advantage.

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

FY2026 Data Center revenue was $193.7B (~90% of the $215.9B total), up 68% YoY on the Blackwell ramp, and NVIDIA describes itself as "a data center scale AI infrastructure company reshaping all industries." The 10-K frames named rivals as parties who "provide or intend to provide" GPUs/accelerators — incumbent-leader positioning.

来源：sec.gov

UMC describes a deliberate follow strategy in the 20-F: it intends "to avoid investments in technologies that do not present commercial potential for volume production" and believes "it is beneficial to defer investment in premature equipment needed to claim the earliest advanced technology and instead to purchase more advanced and less expensive versions of such equipment from vendors who design this equipment based on pre-production lessons learned from the earliest technology." The node ladder bears this out: the most advanced process in production in 2025 is Fab 12A's 0.014 micron, wafers at 14 nanometers and under rounded to 0.0% of wafer sales in each of 2023, 2024 and 2025, and the next rung — UMC's own 12nm FinFET process, which Intel replicated in 2025 "based on our electrical design rules and UMC-verified process" for manufacture at its Arizona fab — is only "expected to begin in 2027."

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

FY2026 gross margin was 71.1% (75.0% in FY2025); per the 10-K MD&A the ~3.9pt decline reflects the Hopper HGX→Blackwell full-system mix shift and a one-time $4.5B H20 excess-inventory/purchase-obligation charge, not competitive price erosion. A low-70s% hardware gross margin evidences strong pricing power.

来源：sec.gov

The 20-F claims no premium — "our current level of pricing is comparable to that of other leading foundries in each respective geometry" — and reports that 2025 average selling price declined 5.4%, driving gross margin down from 32.6% in 2024 to 29.0% in 2025 even as 12-inch-equivalent wafer shipments rose from 3,446 thousand to 3,870 thousand and average capacity utilization improved from 68.7% to 75.2%. It also notes that "[p]rices for wafers of a given level of technology generally decline over the processing technology life cycle."

来源：sec.gov

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

Upstream compute-platform supplier: NVIDIA sells full-stack data-center systems (GPU + Arm CPU + DPU + NVLink/InfiniBand networking + CUDA software) to cloud providers and enterprises, and relies on third-party foundry/assembly-test-packaging partners (e.g., SPIL, Amkor, Wistron, Fabrinet). Per the FY2026 10-K, several of its largest customers (hyperscalers such as Amazon, Alphabet, Microsoft) are simultaneously customers and emerging competitors developing internal accelerated-computing silicon.

Mature- and specialty-node wafer supplier: 2025 wafer sales were 41.4% communication, 30.6% consumer and 11.6% computer, with the AI tie indirect — 2.5D silicon interposers with deep-trench capacitors shipped to major customers for HPC, cloud computing and LLM applications.

• 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

• CUDA accelerated-computing software platform (CUDA-X, NVIDIA AI Enterprise) 领先 · 深 来源：sec.gov
• Data Center AI accelerators (GPU — Blackwell/Hopper) 领先 · 深 来源：sec.gov
• Data-center networking / interconnect (NVLink, InfiniBand, Spectrum Ethernet, DPUs) 前三 · 中等 来源：sec.gov

• 12-inch 28nm/22nm logic and specialty foundry 挑战者 · 中等 来源：sec.gov
• 12-inch silicon photonics 未知 · 中等 来源：sec.gov
• 2.5D silicon interposer and 3D hybrid bonding 未知 · 中等 来源：sec.gov
• 8-inch and legacy-node foundry (0.11 micron and above) 挑战者 · 中等 来源：sec.gov
• Embedded non-volatile memory platforms (eFlash, RRAM, MRAM) 未知 · 深 来源：sec.gov
• Specialty mixed-signal platforms (eHV, BCD, BSI-CSI, RF-SOI) 未知 · 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 →