Cambricon Technologies

Phase A — Understand the business

Lens 1 · Company Overview

Cambricon designs AI accelerator chips — it is a fabless chip designer, not a foundry and not a model builder. Founded 2016 as a spin-out of the Institute of Computing Technology, Chinese Academy of Sciences (CAS-ICT); IPO'd on the STAR Market in 2020. The product family is the Siyuan / MLU (Machine Learning Unit) line — cloud accelerator cards (MLU290, MLU370, MLU590 / "Siyuan 590", next-gen MLU690 / "Siyuan 690"), edge modules (Siyuan 220), and terminal IP — all bound to the proprietary Cambricon Neuware software stack, the developer interface to the hardware (the company's CUDA-analog).

How it makes money today: almost entirely cloud AI accelerator cards sold to Chinese internet/data-center buyers. In FY2025 the cloud product line was RMB 6.477B of RMB 6.497B total revenue — i.e. >99% of revenue is cloud silicon; the edge line shrank 48% YoY and IP/software is a rounding line. This is a company that, in one year, went from a diversified-but-tiny edge/IP shop to a single-product cloud-inference-card company.

Customers / concentration (the defining fact): the top-5 clients were ~94% of H1 2025 revenue, with the single largest customer ~80%. Chinese press consistently identifies ByteDance as that dominant buyer — reportedly pre-ordering ~200,000 Siyuan 590 chips, "more than half of Cambricon's orders". Alibaba is the named prospective second leg.

Contract structure: hardware sell-through (not recurring/take-or-pay). Demand is pull-driven — ByteDance's single pre-order alone reportedly exceeds Cambricon's entire projected 2025 shipment capacity, so the constraint is supply, not demand. That inverts the normal chip-cycle risk: the near-term risk is a production/yield miss, not a demand air-pocket.

Competitors: NVIDIA (the incumbent being displaced by export controls), Huawei Ascend (the domestic 800-lb gorilla), and the "little dragons" — Moore Threads, MetaX, Biren, Enflame, Iluvatar.

Lens 2 · Supply Chain

Map: upstream inputs → Cambricon (fabless design) → cloud/data-center buyer. Named stakeholders and chokepoints:

• EDA / IP (upstream, sanctioned-off): Cambricon has been on the US Entity List since Dec 2022 — it cannot legally buy US-origin EDA (Synopsys/Cadence), design IP, or manufacturing tech. This is a permanent structural input constraint, not a passing headwind.
• Foundry — THE chokepoint: the Siyuan 590/690 are fabbed at SMIC's N+2 (7nm-class) DUV process. SMIC is itself Entity-Listed and EUV-denied, so it multi-patterns 7nm on DUV. Yields are ~20% on Cambricon's largest dies (4 of 5 die scrapped) vs ~60% for a mature leading-edge node. Cambricon competes with Huawei and every other domestic designer for the same scarce SMIC N+2 wafer allocation — a shared single-source bottleneck.
• Memory (HBM) — the second chokepoint: the 590 ships with 80GB HBM. HBM3/3e is export-restricted to China; domestic CXMT is "several generations behind" and China is largely stuck on HBM2e or high-speed GDDR6 workarounds. HBM/LPDDR shortages are explicitly flagged as a 2026 ramp risk.
• Advanced packaging: CoWoS-class packaging lead times run 20–30 weeks and capacity is scarce inside China.
• Downstream: ByteDance (dominant), Alibaba (prospective), plus Chinese foundation-model labs — DeepSeek, Qwen (Alibaba), Hunyuan (Tencent) — now targeting domestic hardware for inference.

Verdict on the chain: Cambricon's design is de-risked from US tools (it's Entity-Listed and coping); its manufacturing is entirely hostage to two domestic bottlenecks it does not control — SMIC N+2 wafer yield and domestic HBM — both shared with Huawei. Names or it didn't happen: SMIC, CXMT, ByteDance, DeepSeek, Alibaba, Tencent. ✅

Lens 3 · Competitive Advantages (moats)

The bull moat is not technology — the 590 openly "lags NVIDIA's A100 in speed and memory bandwidth" and the 910C/H100-class comparisons are aspirational. The real moats are three, all fragile:

1. Regulatory moat (rented, not owned): the entire thesis is export controls created a demand vacuum and Cambricon is a listed pure-play that fills it. This is a policy-granted moat — powerful while US controls hold and Beijing subsidizes self-reliance, but it is a moat the US and China governments own, not Cambricon.
2. Software / switching cost (the durable one, if it exists): Neuware is Cambricon's CUDA-analog. The genuine long-term moat in AI silicon is the software ecosystem + switching cost once a customer ports models to your stack. But industry observers put every Chinese CUDA-alternative "several years behind CUDA in optimization depth and library completeness", and Moore Threads' MUSA (with auto CUDA→MUSA translation) is widely judged more sophisticated than Cambricon's stack. So Cambricon's software moat is real vs a from-scratch entrant but weak vs Huawei's CANN and arguably vs Moore Threads.
3. Listed-liquidity / capital moat: Cambricon is public and profitable while MetaX and Moore Threads are still loss-making — it can raise ~RMB 4B at 1,195 yuan/share to pre-buy HBM inventory and fund the 690. In a supply-constrained race, a war chest to lock wafer/HBM allocation is an advantage.

Bargaining power: weak over customers (80% of revenue is one buyer who could dual-source to Huawei tomorrow), weak over suppliers (competes for scarce SMIC/HBM with a bigger rival). The only thing propping pricing power is that Chinese buyers have nowhere else to legally go — a geopolitical, not a competitive, advantage. Ground truth: this is a be-early policy trade, not a compounding-moat business (yet).

Lens 4 · Segments

FY2025 by product line:

Geography: effectively ~100% China — Entity-Listed, it sells domestically. No meaningful export segment. No segment-level data exists (no EDGAR); all figures. The concentration into a single product line into a single customer is the headline: Cambricon is a one-product, one-customer, one-country company right now.

---

Phase B — Measure performance

Lens 5 · Earnings Result

The FY2025 print (reported Mar 2026) — the moment the thesis went from narrative to numbers:

• Revenue RMB 6.497B, +453.2% YoY (~$900M ``)
• Net profit RMB 2.059B, +555.2% YoY — first full-year profit since founding (FY2024 was a RMB ~0.85B loss)
• Gross margin 55.15% (−1.56pp YoY — still a fabless-design-grade margin)
• Net margin 31.68%
• First-ever dividend: RMB 1.5/10 shares ≈ RMB 632M payout

Latest quarter — Q1 2026:

• Revenue RMB 2.88B, +160% YoY; net profit RMB 1.01B, +185% (recurring +239%)
• Operating cash flow RMB +834M vs −RMB 1.4B a year earlier — the single most important quality tell: earnings are now cash-backed, reversing the historic cash-burn.
• Gross margin held >54%
• Q1 2026 net income more than doubled Q4 2025 (RMB 455M) — accelerating sequentially.

Balance-sheet flag: inventory was ~RMB 2.76B by Q1 2025 — a deliberate build (pre-buying scarce HBM/wafers to secure the ByteDance ramp), but it is also the classic revenue-recognition/write-down risk if a customer slips. Watch inventory-vs-revenue in every print.

Market reaction: the stock is priced for perfection and reacts violently to the rate of acceleration, not the level — see Lens 8. Flag vs its own history: a company that did RMB 65M in H1 2024 revenue did RMB 2.88B in H1 2025 — a 44× jump. There is no clean comp for modeling a base rate off that.

Lens 6 · Earnings Calls (sentiment trend)

No transcripts on the research layer (Chinese filings). From reported management commentary, the tone arc:

• 2023–early 2024: defensive — Chairman Chen publicly conceded the Entity List "blocked the supply chain so that products are affected to some extent" and warned switching suppliers "may adversely affect operating performance". Survival tone.
• Mid-2025: inflection — management pivots to capacity: the bottleneck is how many they can make, not sell; DeepSeek/Qwen/Hunyuan optimizing for domestic hardware cited as structural demand.
• Late 2025–2026: ambition — publicly guiding ~500,000 accelerators in 2026 (incl. up to 300,000 of 590/690), a tripling of output. But note the self-issued risk warning (Jun 2026) flagging valuation, entity-list supply risk, and pullback risk — an unusually candid, regulator-prompted disclosure (see Lens 8/10).

Recurring phrases: "domestic substitution," "self-reliance," "supply-chain security." What they've stopped saying: the survival/impairment language of 2023. The tone has flipped from defense to capacity-scramble — bullish, but the Jun-2026 warning shows management itself thinks the stock has run ahead of the business.

Lens 7 · Comps

Peer table — Cambricon vs global AI-silicon leaders and Chinese pure-play peers.

Read: Cambricon trades at ~8× NVIDIA's P/E and ~4× its P/S while owning 1% of China's AI-chip market (2024) and displacing NVIDIA only inside China. Its market cap ($110B) is comparable to Intel's despite ~1.6% of Intel's revenue. Against its listed Chinese peers it is the only profitable one — which is precisely why it commands the premium, but a ~96× sales multiple prices in years of flawless execution.

Lens 8 · Stock-Price Catalysts (>5% moves, ~2 yr)

Pattern of what actually moves this name:

• Sep 2024 →: +562% run off China semiconductor self-reliance policy + "top A-share of 2024."
• Jan–Feb 2025 (DeepSeek): DeepSeek's efficient domestic models → domestic-hardware demand narrative → sharp re-rate; "record earnings boost from DeepSeek".
• Aug 2025 earnings (4,300% H1 rev): +14× quarterly-revenue headline; made Chen Tianshi briefly one of the world's richest; stock to ~1,522 yuan.
• Oct 2025: RMB 4B private placement priced at 1,195 yuan/share (capital catalyst).
• Apr 30 2026: hit daily +price-limit at RMB 1,700 on Q1'26 beat; single-session turnover >RMB 26B.
• Jun 30 2026: first STAR-Market stock to cross RMB 1 trillion — then the company's own risk-warning that evening → −RMB 140B in two sessions to ~RMB 862B.

What the tape reveals: this stock reacts to (1) policy/geopolitics (export-control headlines, self-reliance policy), (2) the rate of revenue acceleration at earnings, and (3) its own valuation warnings. It is a high-beta policy-and-momentum instrument. Earnings level matters less than second derivative. That makes it violently mean-reverting around round-number/psychological levels.

---

Phase C — Judge people & books

Lens 9 · Management

• Chen Tianshi (陈天石) — Chairman & CEO. Co-founder (2016), CAS-ICT researcher, child-prodigy background (USTC "gifted youth" program). Brother Chen Yunji is Chief Scientist — the pair are the technical + commercial core. This is a founder-led, technically-native company, not a professional-manager shop.
• Skin in the game: Chen owns ~28% directly (some cite ~29%), partly via Beijing Aixi L.P.. Net worth ~$21.5B (Bloomberg, Feb 2026) — essentially all Cambricon stock. Extreme alignment to the share price — for better (he wins if it compounds) and worse (paper-wealth incentive to feed the narrative).
• Track record: built the first Chinese cloud-AI-chip startup; navigated the Entity List without collapsing; delivered the first profitable year in 2025. Genuinely impressive execution under sanction. But the commercial track record is ~18 months long — the profitable era is brand new.
• Capital allocation: historically a cash-burner (years of losses funded by equity); now turning the corner — first dividend (RMB 632M) and a growth raise (RMB 4B) in the same cycle, i.e. returning a token of cash while raising far more to fund the 690/HBM. Reasonable for the stage. ROE/ROIC was negative for years; 2025 is the first positive read — too short a series to judge capital-allocation skill.
• Red flags: the CAS-ICT state lineage + Entity-List "close ties to the Chinese military/defense industrial base" designation is a governance/geopolitical flag by definition. No disclosed related-party or accounting scandal. Founder archetype at a policy-favored national champion → implies aggressive scaling and strong state backing, but also that the company is an instrument of state industrial policy, which cuts both ways.

Lens 10 · Forensic Red Flags

Accounting risk map (all `` — no audited EDGAR filings to inspect; this is the core provenance limitation):

• Revenue recognition / customer concentration: ~80% of revenue from one customer (ByteDance) is the single largest forensic risk — not fraud, but fragility. Any pull-forward, channel-stuffing, or timing shift around a mega-pre-order would swing reported growth. Cannot verify recognition policy without the Chinese annual report.
• Inventory: ~RMB 2.76B build (Q1'25) outrunning historic revenue — legitimate pre-buy of scarce HBM/wafers, but write-down exposure if the 590 is superseded or a customer slips. Inventory-to-revenue is the metric to watch.
• Cash vs earnings — the good tell: Q1'26 operating cash flow +RMB 834M vs −RMB 1.4B prior-year — earnings are now cash-converting, which reduces the "profitable on paper only" risk that plagues its peers. This is the strongest quality signal in the file.
• SBC / non-GAAP: recurring net profit (+239%) tracked reported (+185%) in Q1'26, so SBC is not flattering the number materially. Positive.
• Goodwill/intangibles: no material acquisition history flagged; largely organic R&D (~30% of revenue to R&D).

Regulatory findings (read regulatory/regulatory-findings.md, generated 2026-07-06):

• SEC (EDGAR EFTS — LR + AAER): 0 findings. Cambricon has no CIK and is not required to file with the SEC, so no EDGAR enforcement search is possible.
• US export-control action (material, non-SEC): Cambricon and seven affiliated entities (Anhui Cambricon, Cambricon HK, Kunshan, Nanjing, Xi'an IC, Jixingge Nanjing) were added to the US Commerce/BIS Entity List in December 2022 for allegedly seeking US-origin items "in support of China's military modernization". This is not a securities-fraud finding but is the single most material regulatory fact about the company — a permanent supply-chain constraint the company itself re-disclosed in its Jun-2026 risk warning.
• Non-SEC enforcement (FTC/DOJ/etc.): no material US civil/criminal enforcement hits beyond the export-control listing.
• China/CSRC: no disclosed CSRC enforcement action; the Jun-2026 "Stock Trading Risk Warning" was a proactive/regulator-prompted disclosure, not a penalty.
• Bottom line: No securities-fraud or accounting-enforcement findings via SEC EDGAR EFTS, web search, or available disclosure as of 2026-07-06 — but the Entity-List designation is a material, permanent regulatory overhang, and the absence of English-language audited filings means the accounting is unverified by this analyst (label everything unaudited-per-public-sources).

---

Phase D — Project & stress-test

Lens 11 · Forward Projection (FY2026 / FY2027 / FY2028 EPS)

Grounding: FY2025 actuals — revenue RMB 6.497B, net income RMB 2.059B, GM 55%. Sell-side consensus reportedly sees ~+58.6% rev / +61.9% EPS CAGR and ~RMB 28B FY2027 revenue. Share count is not cleanly sourced in English — Chen's ~28% ≈ his stake, total ≈ ~418M shares — so **EPS is derived, treat as with wide error bars.**

Scenario build (RMB; net income → EPS on ~418M shares ``):

``

Even the bull FY2028 EPS ~RMB 45 against ~RMB 1,353 price = ~30× FY2028 earnings — i.e. the current price already discounts the bull case ~3 years out. That is the whole valuation problem in one line.

(No forecast.ts create in --watchlist mode per skill rules. If committed: <688256> FY2026 net income >= RMB 4.0B, p≈0.55, resolves 2027-03-31.)

Lens 12 · Bull vs Bear

Bull case. Export controls have handed China's AI-inference market to domestic silicon, and Cambricon is the only listed, profitable, cash-generating pure-play on that shift. It has a signed demand signal (ByteDance's ~200k pre-order), a proven flagship (590) in mass production, a next-gen part (690) coming, a war chest (RMB 4B raise) to pre-lock scarce HBM/wafers, extreme founder alignment, and a national-champion tailwind with State Council backing. Neuware gives a nascent switching-cost moat. If NVIDIA stays banned and SMIC/HBM constraints ease, revenue compounds 50–60% for years and today's multiple deflates into itself.

Bear case (2–3 permanent-impairment risks).

1. Customer concentration is a loaded gun. ~80% of revenue is one buyer who is also a natural Huawei Ascend customer. If ByteDance dual-sources or Huawei wins the socket, revenue doesn't decelerate — it falls, and a ~96× sales stock cannot survive a revenue decline.
2. The moat is rented from two governments. The entire premium is the export-control vacuum. A US policy softening (H20/H200 re-admitted to China) or a Beijing decision to consolidate around Huawei as the national champion removes the moat overnight. Cambricon controls neither lever.
3. Manufacturing is capped by a bottleneck it doesn't own. ~20% SMIC yield + HBM scarcity means the 500k target may be physically unreachable; the gap between guided and deliverable shipments is where the stock breaks.

Pre-mortem (18 months out, thesis broken): most likely story — Huawei's Ascend 910C/920 ramp (600k+ units in 2026, better software/CANN ecosystem) takes ByteDance's incremental sockets; Cambricon's 590 looks dated vs the 690's delayed ramp; SMIC prioritizes Huawei's wafers; a quarter misses the acceleration bar and the momentum crowd exits a 250×-P/E stock, halving it. Or: US–China détente re-admits NVIDIA, and the substitution premium evaporates.

Are multiples too high? Unambiguously yes on any absolute basis (~96× sales, ~250–369× earnings). The only defense is optionality on a genuinely enormous TAM (China AI compute) that Cambricon might capture a large slice of. It is priced as a call option on China winning AI-hardware self-sufficiency, not as a business on cash flows.

Contrarian view (what the market refuses to see): the market treats Cambricon and Huawei as complements ("both win from substitution"). They are substitutes fighting over the same buyers, the same SMIC wafers, and the same HBM. Huawei is bigger, better-resourced, has a deeper software stack, and is Beijing's more likely anointed champion. The non-consensus risk is that Cambricon is the number-two domestic player being priced as the number one.

Lens 13 · Devil's Advocate (short-seller)

Dismantling the bull case.

• Revenue concentration: one customer ≈ 80%. This isn't a risk factor buried in a 10-K — it is the company. A single procurement decision at ByteDance is an existential event. Short thesis writes itself the day ByteDance confirms a Huawei order.
• Why the moat is weaker than bulls think: there is no technology moat (590 < A100), the software moat (Neuware) is judged behind both CUDA and Moore Threads' MUSA, and the "moat" that matters — the export ban — is exogenous and reversible. You are underwriting US and Chinese politics, not a business.
• Most dangerous competitor bulls underestimate — Huawei. Ascend has >75% of China's AI-chip production, a full-stack ecosystem (CANN), superior scale, hard state backing, and is doubling output to 600k+ 910C in 2026. If Beijing must pick one national champion, it is Huawei, not a CAS spin-out.
• Capital-allocation / incentive flags: founder's entire ~$21B net worth is the stock → structural incentive to sustain the narrative (guidance of "triple output" to 500k that the SMIC-yield reality may not support). The Jun-2026 self-issued risk warning is management effectively saying the price is detached — and the stock still trades near ~$110B.
• Assumptions that must hold for today's price: (i) NVIDIA stays banned from China for years; (ii) SMIC yields improve enough to hit ~500k units; (iii) HBM supply unlocks; (iv) ByteDance and new customers keep buying Cambricon over Huawei; (v) ~55% gross margin holds as domestic competition floods in. All five must be true. Miss any one and the ~250× P/E compresses.
• −20–30% growth shock: if FY2026 growth comes in at +80% instead of +130%, the stock is a candidate to halve — momentum unwinds faster than fundamentals.
• Single permanent-impairment scenario, plausibility: ByteDance standardizes on Huawei Ascend for its next inference build-out. Plausibility moderate — ByteDance is diversifying toward domestic silicon and may well split volume, and every incremental Huawei win is a direct Cambricon loss. This is the scenario to monitor above all.

Lens 14 · Management Questions (ordered by information value)

1. What share of FY2026 revenue is contracted to your single largest customer, and what is your customer-concentration target by end-2027?
2. Have you received binding orders from a second hyperscale customer (Alibaba/Tencent), and for what unit volume?
3. What is the deliverable 2026 shipment number given SMIC N+2 yield and HBM allocation — as opposed to the 500k target?
4. What die yield are you actually achieving on the Siyuan 590/690 at SMIC N+2, and what is the trajectory?
5. How is HBM sourced for 2026–27 — domestic (CXMT) vs stockpiled imported HBM2e/3 — and how many units of supply is secured?
6. Where does Neuware stand vs CUDA and vs Huawei's CANN on library coverage and model-port time, with third-party benchmarks?
7. If US export controls on NVIDIA into China were relaxed, what happens to your order book?
8. How do you compete for SMIC wafer allocation against Huawei, and do you have a guaranteed capacity contract?
9. What is your gross-margin floor as domestic competitors (Moore Threads/MetaX) scale and pricing pressure builds?
10. What is the 690's mass-production timeline, and how does it close the gap to NVIDIA H100 / Huawei 910C on training?
11. How do you plan to reduce reliance on a single foundry node — any second-source or alternative-node roadmap?
12. What portion of demand is genuine end-use vs strategic inventory pre-buying by customers hedging supply?
13. What is your inventory-write-down risk if the 590 is superseded by the 690 or a customer defers?
14. How do you view the risk that Beijing consolidates domestic AI-compute procurement around Huawei as the national champion?
15. What use of proceeds from the RMB 4B raise most de-risks the 2026 ramp — capacity, HBM, or software?

---