Cambricon Technologies
Fabless designer of cloud/edge AI training & inference accelerators (Siyuan/MLU 3xx-6xx cards, IP cores); revenue from accelerator card + system sales to Chinese cloud/internet/state buyers, outsourcing fabrication to SMIC.
Earnings, margins, COGS & capex
Explosive 2025-26 inflection: after years of losses since its 2016 founding, Cambricon posted its first annual profit in FY2025 (revenue CNY 6.497B / +453% YoY, net profit CNY 2.059B / +555% YoY) as Chinese buyers scrambled for domestic AI silicon amid Nvidia's forced high-end China exit. Q1 2026 revenue of CNY 2.885B (~$423M, +160% YoY) already roughly matched all of H1 2025, with net profit up 185% YoY to CNY 1.01B. Gross margin held around 55%. The trajectory is real but hyper-concentrated on a handful of customers and gated by SMIC 7nm capacity and low large-die yields.
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~45¢ is cost of goods and ~0¢ operating expense, leaving ~55¢ of operating profit (~32¢ net).
Revenue trend
Margins
9M 2025 at 55.3%; compressed from ~69% in 2023 as lower-margin card volume scaled
first positive full-year; up sharply from losses
improving on operating leverage
COGS structure
Dominated by SMIC 7nm (N+2) wafer/packaging costs plus scarce HBM memory; large-die yields reported ~20% on the biggest chips, which structurally inflates effective per-good-die cost and caps gross-margin upside despite pricing power.
Capex
Fabless model - no fabs of its own; cash intensity shows up as advance payments / capacity reservations to SMIC and large inventory + WIP builds to hit the 2026 volume target. Explicit capex not disclosed here.
Latest earnings
Absolute growth was huge, but some coverage flagged the print still undershot the loftiest sell-side forecasts even as it beat consensus on profit; specific beat/miss magnitudes not reliably disclosed. Next report: H1 2026 interim, expected ~late Aug 2026
Management targets more than tripling AI-chip output in 2026 (~500K accelerators); no formal revenue guidance disclosed here
- Q1 2026 revenue
- CNY 2.885B (~$423M)
- Q1 2026 net profit
- CNY 1.01B (+185% YoY)
- Gross margin (9M 2025)
- ~55.3%
- Top-5 customer concentration (H1 2025)
- ~94% of revenue; largest ~80%
- Trailing PE
- ~317x
Growth drivers
- Nvidia's forced exit from the China high-end data-center GPU market opening a large domestic void
- Chinese cloud/internet capex surge (LLM training/inference) with a domestic-sourcing mandate
- Ramp of the Siyuan/MLU 590 and forthcoming 690 (690 still in testing; mass production may slip to H2 2026)
- Stated 2026 target of ~500K AI accelerators (vs ~142K in 2025), incl. up to ~300K advanced 590/690 units
- State/industrial-policy tailwind for domestic AI compute self-sufficiency
Bull & bear
Cambricon is the purest listed proxy for China building its own AI-compute stack: Nvidia is largely locked out of the high end, domestic demand is enormous and policy-backed, and Cambricon just proved the model can be profitable at scale with mid-50s% gross margins and triple-digit growth.
- First-ever annual profit in FY2025 (net CNY 2.06B) and Q1 2026 already matching prior-year half-year revenue - the inflection is reported, not projected
- Structural demand: Chinese hyperscalers must source domestically, and Cambricon is the leading independent (non-Huawei) option
- Siyuan 690 targeting H100-class performance would move it up-stack into lucrative training workloads
- Stated path to ~500K accelerators in 2026 (more than triple 2025's ~142K) implies another step-change in revenue if yields/supply cooperate
- Pricing power: scarcity of domestic high-end AI silicon supports gross margins even at low yields
A ~$120B market cap on ~$1.2B of TTM revenue and ~317x trailing earnings prices Cambricon as if domestic-Nvidia dominance is already locked - while it actually depends on one customer, one constrained foundry, and a policy shield that could shift in either direction.
- ~80% of H1 2025 revenue from a single customer and ~94% from five - a demand cliff if ByteDance/hyperscaler capex or vendor mix shifts
- ~20% large-die yields at SMIC and no HBM access structurally cap volume and margins regardless of demand
- Valuation (~317x trailing / still triple-digit forward PE) vs Nvidia's ~30-40x leaves no margin for error
- Huawei Ascend is bigger, vertically integrated, and state-preferred - Cambricon may end up the #2 domestic option
- Roadmap is hostage to export policy: tighter tool/HBM controls throttle it; a Nvidia thaw undercuts pricing; the 690 ramp could slip to H2 2026
What it is worth
Relative multiples + qualitative; no independent DCF given supply/policy opacity.
A customer-concentration, yield/HBM, competitive (Huawei), or policy (Nvidia thaw / tighter controls) shock breaks the growth narrative; a re-rate from ~317x toward global-semis norms implies severe downside.
Growth stays strong but decelerates and stays customer-concentrated and supply-capped; the extreme multiple compresses toward peers even as earnings rise - a flat-to-modest price outcome.
Sustained triple-digit growth into 2027 on the domestic-compute build-out, margins holding around mid-50s%, and customer diversification could let earnings grow into (not de-rate from) the multiple.
At ~CNY 860.5B (~$120B) market cap on ~CNY 8.3B TTM revenue and ~CNY 2.7B TTM net income, Cambricon trades at ~317x trailing earnings, a still triple-digit forward multiple, and ~100x sales - vs Nvidia at roughly 30-40x earnings and AMD lower. The stock prices in years of near-flawless domestic-Nvidia-replacement execution. Justifiable only if 2026-27 volume (the ~500K-unit ambition) and margins compound while the customer base diversifies; any single-customer, yield, or policy shock is not cushioned. US-first framing: this is a China A-share, Entity-Listed name - included for competitive-landscape context, not as a buy/own recommendation. Not financial advice.
SWOT
Strengths
- Leading independent Chinese AI-accelerator IP/design house with a decade of software-stack (Cambricon Neuware) and multi-generation silicon history
- First-mover beneficiary of the domestic-substitution wave; first-ever annual profit achieved in FY2025
- Full-stack: cloud training/inference cards, edge, and licensable IP cores
- Deep-pocketed, sticky customer base of large Chinese cloud/internet buyers with a policy mandate to localize
Weaknesses
- Extreme customer concentration — top 5 ~94% of H1 2025 revenue, single largest ~80% (widely reported as ByteDance); loss of one buyer is existential
- Dependent on SMIC 7nm with reported ~20% large-die yields, capping supply and effective margins
- No access to HBM at scale, TSMC, or leading-edge EDA/tools due to US Entity List (since Dec 2022)
- Valuation (~317x trailing PE) leaves zero room for execution slips
- Software/ecosystem maturity still trails Nvidia CUDA by a wide margin
Opportunities
- Multi-year domestic AI-compute build-out with Nvidia largely excluded from China's high-end market
- Siyuan 690 approaching H100-class performance could expand into training clusters, not just inference - if it clears testing and ramps
- State-backed data-center and 'AI+' industrial programs as a durable demand floor
- Potential to license IP / expand into edge and automotive AI
Threats
- Huawei Ascend - a far larger, vertically integrated, state-favored rival for the same domestic sockets
- Any partial Nvidia re-entry (policy thaw) would immediately re-anchor performance/price expectations
- SMIC capacity is shared across many Chinese chip firms; allocation risk
- Tighter US export controls on tools/HBM/foundry could throttle the roadmap
- Multiple well-funded Chinese rivals (Biren, Moore Threads, Enflame, MetaX) plus Hygon crowding the field
Moats, dependencies & bottlenecks
Moats
Strong (currently) Real but partly policy-conferred; a rules change on either side (US thaw or tighter controls) alters it fast.
Meaningful switching cost vs startups, but far behind CUDA's ecosystem depth.
Rooted in SMIC allocation and Nvidia's absence - both externally controlled, not owned.
Buyers wary of Huawei's competitive overlap prefer a neutral supplier - a genuine but soft advantage.
Dependencies
SSE 688981 / HKEX 0981. Sole viable advanced foundry post-TSMC ban; ~20% large-die yields and shared 7nm capacity gate all volume.
Export-restricted; scarce domestic HBM constrains high-end card output.
~80% of H1 2025 revenue single-buyer concentration; Caixin cited ~200K Siyuan 590 pre-orders.
On the Entity List since Dec 2022; roadmap and inputs are policy-sensitive in both directions.
Access restricted; must lean on domestic/legacy toolchains.
Advantages
- Leading independent Chinese AI-accelerator brand with real shipping product
- First to profitability among Chinese AI-chip pure-plays
- Neutral-supplier appeal vs Huawei for competitively wary buyers
- Full-stack cloud-to-edge portfolio plus licensable IP
- Direct beneficiary of the largest structural demand shift in Chinese semis
Weaknesses
- Existential single-customer / top-5 revenue concentration
- Supply-side gated by SMIC yields and HBM scarcity
- No access to TSMC, leading-edge tools, or global HBM
- Valuation disconnected from current fundamentals
- Second to Huawei in scale and state preference
Bottlenecks
- SMIC 7nm capacity allocation and ~20% large-die yields
- HBM and advanced-packaging supply under export controls
- Leading-edge EDA/tool access
- Single-customer concentration limiting revenue diversification
- Software ecosystem breadth vs CUDA
Top signals & trends
Top signals
Demand ramp is materializing in reported numbers.
Even hyper-growth is testing a priced-for-perfection bar.
Aggressive volume ambition - contingent on yields/HBM and the 690 ramp.
Valuation leaves no room for execution slippage.
Bullish (for domestic demand) · Protective moat while it lasts - but externally controlled.
Trends
Primary demand catalyst for domestic accelerators.
Durable, state-backed demand floor.
Caps roadmap and supply on the input side.
Validates the market but compresses long-run share/pricing.
Expands the total addressable domestic compute market.
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Advanced-node foundry (7nm N+2) - the binding supply constraint post-TSMC ban.
Former foundry, now blocked to Cambricon under US controls - a lost supplier.
HBM source, export-restricted for high-end China AI; HBM scarcity is a hard bottleneck.
HBM/memory source, similarly export-restricted at the high end.
EDA/IP tooling with restricted access; forces reliance on domestic/legacy tools.
EDA/IP tooling with restricted access.
Widely reported single largest buyer (~80% of H1 2025 revenue; Caixin cited ~200K Siyuan 590 pre-orders).
Chinese hyperscaler building domestic AI clusters under localization mandates.
Chinese hyperscaler with large domestic AI-compute demand.
Public-sector 'AI+' and compute-infrastructure programs.
Largest, vertically integrated, state-favored domestic AI-accelerator rival; also SMIC-dependent.
Global AI-GPU leader; largely excluded from China high-end today, but any re-entry re-anchors the market. Context only, not a China-exposure call.
Chinese x86/DCU accelerator maker; overlapping domestic data-center demand.
MI-series AI GPUs; China-restricted at the high end but shapes the global price/performance bar.
Well-funded Chinese GPGPU startup; also Entity-Listed.
Chinese GPU startup targeting AI + graphics.
Tencent-backed Chinese AI-training chip startup.