China: supply competitor, demand pool and policy risk

Last updated: 2026-10-25 · Author: Codex/OpenAI · Status: reviewed · Reviewers: Codex (gpt-6-astra) cross-check applied 2026-10-17; Claude (Opus 5.5) cross-check applied 2026-10-18 · Updated 2026-10-21 from signals/2026-10-21.md · Updated 2026-10-25 from signals/2026-10-25.md

Summary

  • Alibaba's capex was RMB67.678bn in the June 2026 quarter, +75% y/y; it attributes the rise to procurement timing, CPU capacity for anticipated AI-agent adoption and higher chip-component prices. AI-related product revenue was RMB12.376bn, about 18% of that quarter's capex (B; scopes differ).51
  • In the June 2026 quarter, cash capex exceeded operating cash at Alibaba, Tencent and Baidu. Tencent's FCF was −RMB13.8bn; management's +RMB37.6bn excluding compute prepayments sets aside RMB51.4bn already paid, while net cash fell from RMB146.9bn to RMB58.2bn. Alibaba's cloud adjusted EBITA was +RMB5.628bn versus −RMB13.861bn in labs/apps; Tencent's non-IFRS operating profit was RMB75.6bn versus RMB86.1bn excluding new AI products (about RMB10.5bn drag, B).51, 8, 10
  • Alibaba announced a minimum RMB380bn three-year infrastructure programme in February 2025 and an unquantified increase that September. Group capex in April 2025–June 2026 totalled RMB193.741bn; the remaining RMB186.259bn equals 2.75 quarters at the June rate (B; group capex includes non-AI items, so this is a pace illustration). ByteDance's reported US$70bn annual 2026 scenario remains unverified and is not directly comparable.1, 2, 71, 51, 12
  • DeepSeek disclosed US$5.576m of V3 official-run compute in its February 2025 report revision and US$294,000 for specified R1 stages in September 2025; neither is an all-in development budget.13, 14
  • DeepSeek's September 30, 2026 Ascend releases target Ascend 950 and require CANN 9.2.0+ for TileKernels; FlashMLA claims 410/360 TFLOPS prefill/decode, about 95%/83% of peak. DeepInfer, released October 20, adds offline V4.1-Flash inference on Nvidia SM100 and Ascend 950; published benchmarks use Nvidia hardware. Software adaptation advances, while comparable deployment economics, qualified HBM volumes and yields remain unverified.55, 56, 74
  • On October 17, 2026, Qwen3.7-Plus listed Singapore input/output prices of US$0.40/1.60 per million tokens for inputs up to 256K, before promotions. In mid-September, Chinese labs supplied 69% of OpenRouter tokens while closed-weight models retained 73% of spend, down from 89% in January (one channel, excluding resellers). MiniMax's H1 2026 revenue was US$116.6m, 60.8% outside mainland China, at 17.9% gross margin and US$293.0m adjusted net loss: volume competition coexists with equity-funded lab losses, without establishing a market-wide Western margin decline.26, 63, 64
  • Listed-lab equity repriced sharply while private funding continued. Bloomberg reported on October 23 that MiniMax and Zhipu had lost about US$150bn combined from their peaks: Zhipu was 81% below its June high, and MiniMax fell below its IPO price on October 22. Bloomberg attributed the falls to continued token-price cuts, the capability gap and possible DeepSeek/Moonshot listings. MiniMax then rose 16.74% to HK$187.60 on October 23. CSI 300 closed −1.8% on October 22, 20% below its June high. Meanwhile Reuters, relaying Bloomberg, reported DeepSeek near completing about RMB100bn (US$14.98bn) at a RMB500bn pre-money valuation (reported, unverified, C), and a CCTV-affiliated account said DeepSeek and MiniMax had won approval for global use of their models under the new registration regime. Claims 49, 72–73.77, 79, 81
  • China is an evidenced competitor at the cheaper model tier and in conventional DRAM, but no demonstrated replacement for US infrastructure demand: Alibaba, Tencent and Baidu's June 2026 capex was RMB131.852bn, about US$19.4bn at RMB6.7851/US$, or 11% of the US four's US$171.210bn (B; definitions differ). Nvidia's H200 channel was below 1% of Data Center revenue in its July quarter; toolmakers still derive about 20–34% of sales from China in differing periods.51, 8, 10, 67, 68, 69, 70, 32, 45, 47, 48, 49
  • The October 20, 2026 DHS notice makes public a July 24 order excluding/removing DeepSeek and implementations in covered US federal civilian procurement/IT. This limits that customer channel; it does not prohibit private use.72, 73
  • TSMC's 2025 Form 20-F says most principal production facilities are in Taiwan and most net revenue derives from operations there, creating systemic disruption exposure; no equity price or military count identifies an invasion probability.42

Key claims

Original sources accessed 2026-10-17; Claude-review additions and their cited sources fetched 2026-10-18; October 21 additions re-fetched 2026-10-21; October 25 additions (claims 72–73 and the claim 32/49 updates) re-fetched 2026-10-25 except the Reuters DeepSeek-round and FT/LSEG pages, which returned HTTP 403 or were not re-fetched and rest on the verified October 25 signals log (see Sources). API pairs mean uncached input/output per million tokens, with tariff currencies unconverted; claim 66 separately labels its capex FX conversion. Grade A verifies a disclosure, not delivery or independent audit.

# Claim Value As of Source Type Grade
1 Alibaba infrastructure commitment and increase At least RMB380bn over three years; subsequently announced spending beyond it, no replacement total verified 2025-02-24 / 2025-09-29 1, 2 primary A
2 Alibaba longer-term ambition More than 20GW of operated global data-centre capacity by 2032; target 2026-09 50 primary A
3 Alibaba actual capex FY ended March 2026: RMB126.063bn; March quarter RMB26.887bn; June quarter RMB67.678bn versus RMB38.676bn a year earlier (+75%); company cites procurement timing, anticipated AI-agent CPU needs and chip-component inflation 2026-06 51, 71 primary A
4 Alibaba cash generation and AI cloud demand June-quarter operating cash RMB22.945bn; FCF −RMB44.670bn; AI Cloud and Compute Services segment revenue RMB48.437bn, +45%; segment includes non-AI services 2026-06 51 primary A
5 Alibaba external financing HK$80bn gross equity placement completed; net proceeds intended for AI infrastructure, not evidence of expenditure 2026-08-26 5 primary A
6 Tencent capex, consistent reported measure 2025 RMB79.198bn; March 2026 quarter RMB31.936bn; June RMB52.784bn versus RMB19.107bn a year earlier; first half RMB84.720bn 2026-06 6, 7 primary A
7 Tencent cash payments differ from capex June-quarter capex payments RMB59.3bn; operating cash RMB52.7bn; FCF −RMB13.8bn; IFRS operating profit RMB67.3bn; net cash RMB58.2bn 2026-06 8 primary A
8 Baidu consolidated capex 2025 RMB12.073bn; March 2026 quarter RMB5.916bn; June RMB11.390bn versus RMB3.800bn a year earlier 2026-06 9, 10 primary A
9 Baidu liquidity and cash burn June-quarter operating cash RMB3.436bn; FCF −RMB7.954bn; cash and investments RMB283.1bn 2026-06 10 primary A
10 ByteDance contested 2025 spending report Reuters reported a budget above RMB150bn; ByteDance explicitly rejected the anonymously sourced spending information; reported, unverified and contested 2025-01-23 11 secondary D
11 DeepSeek V3 official-run compute 2.788m H800 GPU-hours × assumed US$2/hour = US$5.576m; excludes prior research/ablations report revised 2025-02 13 primary A for disclosure; B for recomputation
12 Later R1 compute disclosure 147,000 H800 GPU-hours / US$294,000 for specified R1-Zero, data creation and R1 stages; inherited V3 training excluded 2025-09 14 primary A for disclosure; B for recomputation
13 V4-family releases and Ascend training evidence April V4 preview is historical; current Flash updated to V4.1-Flash-1017 on October 17, API pricing unchanged. Separate authors demonstrate full-parameter V4-family post-training on Ascend; complete original pretraining mix unverified 2026-04 / 2026-08 / 2026-10-17 15, 16, 58 primary A
14 DeepSeek Inner Mongolia site At least 160,000 Ascend 950DT planned for inference within a gigawatt-scale site; reported, unverified, not installed capacity 2026-09-04 17 secondary C
15 Huawei production scale Reported plan: around 1.6m Ascend component chips in 2026, enough for about 600,000 910C plus other models; not a finished-accelerator total. Separate Reuters April report gives a ~750,000-unit 950PR 2026 shipment plan and April mass-production start; reported, unverified, C (claim 68), not additive to the component estimate 2026-09 report 17, 66 secondary C
16 Huawei HBM roadmap 950DT scheduled for Q4 2026; advertised HiZQ 2.0 HBM: 144GB, 4TB/s; proprietary design is not proof of domestic manufacturing volume 2025-09 roadmap 18 primary A
17 HBM constraint BIS controls HBM; Huawei/Cambricon price increases attributed to HBM scarcity in reporting; procurement details reported, unverified 2024-12 / 2026-09 19, 20 secondary C
18 SMIC node progress Teardown confirms N+3 as scaled 7nm-class technology, still less dense than leading commercial 5nm; no verified leading-node parity 2025-12 teardown 22 primary A
19 SMIC operating scale Q2 revenue US$3.0056bn; gross margin 25.3%; monthly capacity 1,096,500 standard 8-inch-equivalent wafers; utilization 93.7%; quarterly capex US$1.8357bn 2026-06 21 primary A
20 CXMT and YMTC capability CXMT announced G5 DRAM mass production on 2026-09-20; Counterpoint estimates 10% of DRAM revenue in Q2 2026 versus 4% in Q2 2025, with no CXMT share listed in its HBM table. YMTC markets Xtacking NAND; neither establishes qualified HBM supply at Ascend scale 2026-09 / current product page 23, 24, 65 primary / secondary estimate A for company disclosures; C for share estimate
21 Memory financing Reuters describes years of government funding/losses, followed by stronger pricing; private-company economics reported, unverified 2026-07 25 secondary C
22 Qwen hosted prices, Singapore international Qwen3.8-Max US$2/6; Qwen3.7-Plus US$0.40/1.60 for input ≤256K, before temporary promotions 2026-10-17 26 primary A
23 Kimi live domestic prices K2.6 RMB6.50/27; K3 RMB20/100; distinct cache-write charges apply to K3 2026-10-17 27 primary A
24 GLM live USD prices GLM-5.3 US$1.40/4.40; Flash US$0.15/0.50 2026-10-17 28 primary A
25 MiniMax launch price, historical snapshot M2.5-Lightning US$0.30/2.40; not verified as the current flagship tariff 2026-02-12 29 primary A
26 H20 licensing reversal April 2025 licence requirement; August licences for certain customers; US$4.5bn inventory/purchase-obligation charge 2025 / latest filing 30, 32 primary A
27 Revenue-sharing arrangement disclosure US officials expected 15% of licensed H20/MI308 sales revenue; Nvidia said no codifying regulation then published; actual collections unverified 2025-08 disclosures 30, 31 primary A
28 Latest H200 position Case-by-case licensing; China restrictions impede sales; H200 shipments <1% of Data Center revenue in quarter ended July 26, 2026; US$0.4bn H200 charge in first half of FY2027; 25% US-entry tariff borne by Nvidia Quarter ended 2026-07-26 / first half FY2027 32, 33 primary A
29 Blackwell variants for China RTX6000D marketed; B30A development/approval hopes reported, unverified; no demonstrated broad approved data-centre channel 2025 report / 2026 filing 34, 32 secondary C
30 Chip tracking proposal H.R.3447 proposes location/security verification; retrieved status is “Introduced,” not evidence of enactment accessed 2026-10-17 35 primary A
31 Chinese technology-export proposal Announcement No.48 of 2026: 40 proposed technologies, including 23 AI; comments 16 Oct–15 Nov; temporary registration effective on announcement, including started but unfinished exports and specified foreign ownership/control changes 2026-10-16 36 primary A
32 MOFCOM scope limits Under the proposed revised catalogue and official explanatory Q&A: public-domain/open-source technology excluded; facilitated licensing contemplated for new open-source transfers; API/app/SaaS use excluded from large-model technology-export scope. Restricted-technology overseas deployment remains controlled, with facilitation for own-cloud deployment; temporary pre-export registration is already effective. Added 2026-10-25: first reported registrations under this regime are DeepSeek and MiniMax (claim 73) 2026-10-16; registration reporting 2026-10-22/23 36, 37 primary A as statement of proposed scope and effective registration
33 Materials controls and suspension deadlines Rare-earth licensing remains; expanded October measures suspended through 10 Nov 2026; specified US gallium/germanium restrictions suspended through 27 Nov 2026, not general delicensing current notices 38–41 primary A
34 Taiwan production concentration TSMC says majority of principal production facilities are in Taiwan (ROC) and majority of net revenue derives from operations there, not customer location; overseas expansion does not eliminate exposure 2025 annual report 42 primary A
35 Taiwan military activity 9 aircraft sorties, 6 naval ships, 8 official ships; 5 sorties crossed median line/entered specified ADIZ areas 2026-10-17 44 primary A
36 ASML China exposure: different denominators 23% of Q3 system sales (14% Q2); around 20% of FY2026 total sales guided 2026-10-14 45, 46 primary B
37 Lam China exposure 34% of fiscal-year revenue FY ended 2026-06-28 47 primary A
38 KLA China exposure 30% of fiscal-year revenue FY ended 2026-06-30 48 primary A
39 Tokyo Electron China exposure 30.4% of quarterly sales quarter ended 2026-06-30 49 primary A
40 Market signal and interpretation Taiwan record equity close on 5 Oct; cannot identify conflict probability. Inference-margin/overbuild risks material; Taiwan impact systemic (author inference) 2026-10-17 43, 51, 8, 17, 42 secondary C
41 ByteDance 2026 spending scenario and profit estimate Considering up to US$70bn capex in 2026; roughly US$50bn profit estimated for 2025; preliminary anonymous-source estimates, reported, unverified 2026-05-27 12 secondary C
42 Alibaba AI Cloud and Compute Services profitability Adjusted EBITA +RMB5.628bn; includes cloud and chip operations, not a standalone inference margin Quarter ended 2026-06-30 51 primary A
43 Alibaba AI Labs and Applications losses Adjusted EBITA −RMB13.861bn Quarter ended 2026-06-30 51 primary A
44 Tencent compute-prepayment counterfactual Reported FCF −RMB13.8bn; management reports +RMB37.6bn excluding compute-procurement prepayments already included in operating cash flow Quarter ended 2026-06-30 8 primary A
45 Tencent FCF reconciliation, recomputed RMB52.7bn operating cash −59.3bn cash capex −5.0bn media content −2.2bn lease payments = −13.8bn; prepayment effect: 37.6 − (−13.8) = RMB51.4bn Quarter ended 2026-06-30 8 primary B
46 Alibaba cash gap versus FCF, recomputed RMB22.945bn operating cash −67.678bn group capex = −44.733bn; FCF uses eligible property/equipment purchases of 67.660bn and +0.045bn buyer-protection-deposit adjustment: 22.945 −67.660 +0.045 = −44.670bn Quarter ended 2026-06-30 51 primary B
47 Tencent bond pricing US$1.8bn and CNY26bn of notes priced for general corporate purposes; issuance expected October 21, not yet completed as of cutoff 2026-10-14 52 primary A
48 ByteDance reported loan facility US$29.6bn unsecured facility; initial three-year term with extension options; general corporate purposes communicated to lenders, AI/offshore uses attributed to anonymous sources; reported, unverified, not verified drawdown 2026-09-04 53 secondary C
49 DeepSeek prospective financing Round expected to exceed RMB80bn (possibly RMB100bn per Bloomberg) and close in October; Tencent and CATL among largest commitments per Bloomberg as relayed by Reuters; July kickoff targeted RMB500bn valuation, following ~US$7.4bn maiden round; described as IPO funding; reported, unverified, closing not established. October 22 update (Reuters relaying Bloomberg News): approximately RMB100bn / US$14.98bn at RMB500bn pre-money, more than 15 investors, CATL and Tencent plus Meituan, Geely, GigaDevice and Midea among the names; completion expected in the coming days, terms could change. The Information separately reports nearly US$16bn (restricted; not reconciled). The October 6 RMB80–100bn / RMB500bn range is the earlier vintage and stays on record; the headline verb 'raises' is not a certified close or cash received 2026-10-06; update 2026-10-22 54, 79 (per signals item 48; HTTP 403 on 2026-10-25 re-fetch), 80 secondary C reported, unverified / D unreconciled precision
50 DeepSeek TileKernels Ascend release Ascend backend released with common Python APIs for Nvidia GPUs and Huawei NPUs; requires Ascend 950 NPU and CANN 9.2.0+; inference infrastructure, not an Ascend-training claim or end-to-end cost/delivery validation 2026-09-30 55 primary A
51 DeepSeek FlashMLA Ascend release DeepSeek describes release as Ascend inference infrastructure; sparse-attention prefill/decode kernels on Ascend 950 claim 410/360 TFLOPS (~95%/83% of peak) on V4.1 workloads; vendor claims, no Nvidia comparison or production-deployment disclosure 2026-09-30 56, 55 primary A as vendor claim
52 DeepSeek V4.1-Flash launch and efficiency claim Model launched September 10; vendor says KV cache needs one-quarter of predecessor HBM, not one-quarter of all model memory; independent serving economics unverified 2026-09-10 57, 58 primary A as vendor disclosure
53 DeepSeek current Flash update V4.1-Flash-1017 released; API pricing unchanged; vendor claims shorter task runtimes and outputs 2026-10-17 58 primary A as vendor disclosure
54 Alibaba Zhenwu V900 Introduced with advertised 3× predecessor performance and clusters up to 500,000 cards; vendor specifications, not verified accepted shipments or comparable workload economics 2026-09-22 60 primary A as vendor claim
55 Huawei Ascend ecosystem support Announced shared 10k-NPU-scale compute access for partners/developers and programming support; no independent utilization or cost validation 2026-09-19 61 primary A as vendor disclosure
56 Tencent half-year cash-capex measure RMB96.3bn paid for capital investments in offering extract; differs from RMB84.720bn reported capex in interim report, not a correction to it Half-year ended 2026-06-30; offering 2026-10-12 62, 7 primary A
57 DeepSeek V4-Pro service notices conflict Launch planned request rerouting after September 14; current changelog says V4-Pro API retained after that date. Huawei Cloud September 22 retirement notice concerns its own MaaS service 2026-09-10 / 2026-09-22; changelog accessed 2026-10-17 57, 58, 59 primary A for differing published statements
58 OpenRouter model-origin and weight-access shares Chinese labs 69% of tokens; open-weight models 72%; closed-weight models 73% of spend versus 89% in week of 5 January 2026. Open-weight tokens peaked at 79% in July and stayed above 62% thereafter; over 40% of US startups on platform use DeepSeek in production versus ~13% in January. Platform data exclude resellers; not global market shares Mid-September 2026; published 2026-09-22 63 primary A for platform disclosure; C for wider representativeness
59 MiniMax listed-lab economics Unaudited H1 revenue US$116.573m (+283.1%); 60.8% outside mainland China; gross profit US$20.813m, margin 17.9% (12.1% H1 2025); R&D US$296.870m; non-IFRS adjusted net loss US$293.031m; cash/equivalents US$930.905m. Company says financing activities mainly funded cash needs Six months ended 2026-06-30; released 2026-08-26 64 primary A as unaudited disclosure
60 Alibaba AI-related revenue and capex comparison AI-related product revenue RMB12.376bn, 12th consecutive quarter of triple-digit y/y growth; 12.376 / 67.678 = 18.3% of group capex; revenue and investment scopes differ Quarter ended 2026-06-30 51 primary A for revenue disclosure; B for ratio
61 Tencent quarterly net-cash decline RMB146.860bn on 31 March to RMB58.191bn on 30 June; decline RMB88.669bn, larger than FCF outflow; full allocation of the difference not verified 2026-03-31 to 2026-06-30 8 primary A for balances; B for change
62 Tencent non-current prepayments, deposits and other assets RMB24.540bn at December 2025 to RMB91.203bn at June 2026; rise RMB66.663bn; broader balance-sheet category, not wholly identified as compute prepayments 2025-12-31 to 2026-06-30 8 primary A for balances; B for change
63 Tencent new-AI-products operating-profit drag Non-IFRS operating profit RMB75.6bn versus RMB86.1bn excluding New AI Products; rounded difference ≈RMB10.5bn; management counterfactual, not a separately reported segment loss Quarter ended 2026-06-30 8 primary A for disclosed profit measures; B for difference
64 Tencent compute procurement cash and timing, recomputed Cash capex 59.3 + compute prepayments 51.4 ≈RMB110.7bn of capex-plus-prepayment cash out; prepayment-adjusted OCF 52.7 + 51.4 = RMB104.1bn. Group cash capex not wholly AI; prepayments already reduce reported OCF and must not be deducted twice Quarter ended 2026-06-30 8 primary B
65 Alibaba commitment pace illustration, recomputed April 2025–June 2026 group capex 126.063 + 67.678 = RMB193.741bn; 380 − 193.741 = RMB186.259bn, or 2.75 quarters at RMB67.678bn/quarter. Not verified fulfilment of AI/cloud commitment; excludes Jan–Mar 2025 and includes non-AI capex Through 2026-06-30 71, 51, 1 primary B
66 China-three versus US-four group capex, recomputed June quarter: 67.678 + 52.784 + 11.390 = RMB131.852bn; / 6.7851 (Alibaba convenience FX) = US$19.433bn. US: Alphabet 44.924 + Amazon 54.208 + Meta 31.078 + Microsoft 41.000 = US$171.210bn; ratio 11.35%. Alphabet/Amazon cash PPE (Amazon gross); Meta includes lease principal; Microsoft accrual capex includes finance leases; China reported group capex; not aligned AI-only totals Quarter ended 2026-06-30 51, 8, 10, 67, 68, 69, 70 primary A for disclosed inputs; B for sums/conversion/ratio
67 MiniMax ratios, recomputed Gross margin 20.813 / 116.573 = 17.9%; R&D 296.870 / 116.573 = 2.55× revenue; adjusted loss 293.031 / 116.573 = 2.51× revenue. Group economics, not standalone API margin Six months ended 2026-06-30 64 primary B
68 Huawei Ascend 950PR shipment-plan lead Reuters sources report ~750,000 units planned for 2026, mass production beginning in April and full-scale shipments in H2; reported, unverified; not primary-confirmed deliveries or an addition to the 910C component estimate 2026-04-29 report 66 secondary C
69 DHS DeepSeek exclusion/removal notice Civilian procurement/IT covers DeepSeek, affiliates and third-party implementations; specified national-security systems excluded. Publication is new; original order dated July 24 requires immediate exclusion/removal after issuance, with 60 days for exception requests, not a clock reset on publication Order 2026-07-24; filed 2026-10-19; published 2026-10-20 72, 73 primary A publication/scope; no verified reset clock
70 DeepSeek DeepInfer release Apache 2.0 offline V4.1-Flash inference framework; supported hardware Nvidia SM100 and Huawei Ascend 950, not listed H100/Hopper. Benchmarks use 4 SM100 GPUs for prefill/64 for decode; no measured Ascend parity or training economics established Released 2026-10-20; README accessed 2026-10-21 74, 75 primary developer publication A release/support and vendor benchmark scope
71 Huawei–DeepSeek report Reported chip co-design and prospective Ascend training; unverified. No demonstrated training or deliveries 2026-10-19 report; restricted original accessed 2026-10-21 76 secondary C reported, unverified
72 Listed-model and index prices, dated Bloomberg (Jeanny Yu, published October 23 03:51 CEST): MiniMax and Zhipu lost about US$150bn combined from their respective peaks (both listed in Hong Kong in January 2026); Zhipu −81% from its June high; MiniMax fell below its IPO price on Thursday October 22, erasing a March peak gain of up to 650%. Causes named by Bloomberg: continued token-price cuts that US labs are following, a capability gap on the Artificial Analysis index, and possible DeepSeek/Moonshot listings drawing funds; MiniMax's M3.1 Flash launch did not lift the shares. MiniMax closed HK$187.60 on October 23, +HK$26.90 / +16.74% from HK$160.70, so the below-IPO observation must not be carried forward. Bloomberg via The Edge: CSI 300 −1.8% on October 22, −20% from its June high; Hang Seng Tech as much as −5.3% intraday, its largest intraday fall since July 17; onshore shares US$2.4tn below a May peak. The ~87% MiniMax peak drawdown in scans is not a synchronized closing calculation 2026-10-22/23 closes; reports re-fetched 2026-10-25 77; 78 per signals item 46; 82 secondary market reporting B reported closes / C peak-to-date measures; D unchecked ChiNext and first-bear claims
73 Registration under the October 16 regime: first reported approvals Reuters (Beijing, October 23, 02:59 EDT), citing a post on a CCTV-affiliated social-media account on October 22 (Yuyuan Tantian is named for a related Friday post): DeepSeek and MiniMax "have won state approval for global use of their AI models"; no models, licence terms or regulator document named; companies did not comment. A same-day MarketScreener headline says MiniMax completed export registration for open-source, overseas model licensing (headline only). Reuters restates the October 16 proposal on 23 AI technologies and the ministry's statement that most remain exportable. Registration of two developers is not blanket permission for all future model exports and not a new ban reported 2026-10-22/23; re-fetched 2026-10-25 81; claims 31–32 secondary attributed state-media report B attribution / C underlying completion without regulator document

Analysis

Demand and financing. China expands infrastructure demand, but group capex includes general cloud, CPUs and facilities. Alibaba's fiscal year ends in March; Tencent and Baidu use calendar years. Private-company budgets are prospective. Adding these figures into “China AI capex” would manufacture comparability.

Alibaba's multi-year commitment is a floor, with an announced but unquantified increase. Group capex in April 2025–June 2026 totalled RMB193.741bn. Against the original RMB380bn floor, the illustrative remainder is RMB186.259bn, or 2.75 quarters at the June rate (B; claim 65). This pace could exceed the original floor around early 2027, but group capex includes non-AI items and does not verify spending within the announced programme. Alibaba attributes the June-quarter rise partly to procurement cycles, CPU capacity for anticipated AI-agent adoption and component inflation: expenditure is not clean evidence of proportionate AI-demand or useful-compute growth.71, 51 Revenue growth is encouraging; negative FCF and the equity placement show shareholders funding expansion and bearing dilution and impairment risk.1–5, 50

In the June 2026 quarter, Alibaba reported adjusted EBITA of +RMB5.628bn in AI Cloud and Compute Services and −RMB13.861bn in AI Labs and Applications. These segment results distinguish profitable infrastructure services from model/application investment losses; the cloud segment also includes non-AI services. AI-related product revenue was RMB12.376bn, about 18.3% of group capex in the quarter (B); that ratio compares different scopes and is not a return on invested capital. Standalone inference margins and investment-cohort returns remain undisclosed.51

Alibaba's operating cash less group capex was −RMB44.733bn, distinct from reported FCF of −RMB44.670bn: its FCF bridge uses RMB67.660bn of eligible purchases and a +RMB0.045bn buyer-protection-deposit adjustment (claim 46). This is a definition difference, not an arithmetic error.51

Tencent remains profitable, but its Q2 2026 FCF of −RMB13.8bn includes RMB51.4bn of compute prepayments (B); management's +RMB37.6bn counterfactual removes cash already paid for future compute. This shows timing, not reduced investment exposure. Cash capex plus these prepayments was about RMB110.7bn, versus RMB104.1bn of OCF before prepayments (B; group cash capex is not wholly AI). Net cash fell RMB88.669bn from March to June; the non-current prepayments/deposits/other-assets category rose RMB66.663bn over six months, without establishing that the whole category is compute. The full bridge from FCF to the net-cash fall is unverified. Non-IFRS operating profit of RMB75.6bn versus RMB86.1bn excluding new AI products implies about RMB10.5bn of quarterly drag (B), comparable in scale, but not accounting basis, with Alibaba's labs/apps loss. The FCF reconciliation subtracts cash capex, media content and lease payments; prepayments already depress operating cash and must not be subtracted again (claims 44–45). The offering extract's RMB96.3bn half-year cash payments and interim report's RMB84.720bn reported capex are distinct measures.8, 62, 7 Management also describes a more favourable cloud pricing environment, evidence against a universal price war. Baidu similarly invests beyond current operating cash; its investment holdings provide liquidity, although “cash and investments” exceeds immediately available cash. These cash shortfalls concern the June quarter's consolidated groups, not normalized annual or AI-only operations.8, 10

ByteDance's reported 2025 budget is contested: Reuters reports the figure and ByteDance's explicit rejection. Its May 2026 annual spending scenario and profit estimate remain reported, unverified, and cannot be directly compared with Alibaba's multiyear floor.11, 12

Financing already includes debt channels. On October 14, 2026, Tencent announced pricing of US$1.8bn and CNY26bn of notes for general corporate purposes, with issuance expected October 21; this does not allocate every dollar to AI.52 Reuters reported a US$29.6bn unsecured ByteDance facility on September 4, with an initial three-year term and extension options; general corporate purposes were communicated to lenders, while AI/offshore allocations came from anonymous sources. This is reported, unverified, not a verified drawdown.53 On October 6, Reuters reported a prospective DeepSeek round exceeding RMB80bn, expected to close in October; closing is not established by the article. The July fundraising targeted a RMB500bn valuation after a reported ~US$7.4bn maiden round. Reuters relays Bloomberg's possible RMB100bn total and Tencent/CATL among the largest commitments; all remain reported, unverified, C. Tencent is both a builder and a reported leading DeepSeek investor, so its exposure to model economics extends beyond its own capex.54 These reports broaden the potential loss bearers beyond profit-funded expansion without establishing complete financing schedules.

MiniMax provides direct listed-lab evidence: H1 2026 unaudited revenue was US$116.573m (+283.1%), with 60.8% outside mainland China; gross margin was 17.9%, up from 12.1%. R&D of US$296.870m and adjusted net loss of US$293.031m were each about 2.5× revenue (B); cash/equivalents were US$930.905m. The company says financing activities mainly funded cash needs. This supports equity-funded expansion at a thin group gross margin; it does not measure a standalone token-serving margin. Pure-play lab shareholders bear the immediate dilution/loss risk, followed by platform shareholders through AI-product losses and cash commitments; lender exposure also depends on the profitable core businesses.64

Private labs depend on investors and infrastructure partners; government funding also supports semiconductor manufacturers.25 Shareholders, strategic backers and asset owners bear dilution and impairment risk. Announced Tencent notes and ByteDance's reported unsecured facility make lender exposure concrete as a financing channel, subject to issuance/drawdown; landlords and public funders also bear risk where contracts or guarantees create exposure. The loss allocation remains incompletely disclosed.52, 53

China is an evidenced margin competitor at the cheaper model tier and a growing domestic supplier, rather than a demonstrated substitute for US infrastructure demand. The June 2026 group-capex comparison is RMB131.852bn for Alibaba, Tencent and Baidu, or US$19.433bn at Alibaba's disclosed RMB6.7851/US$ convenience rate, against US$171.210bn for Alphabet, Amazon, Meta and Microsoft: 11.35% (B; claim 66). These are differently defined group measures, not a national AI total; Meta includes lease principal and Microsoft includes accrual finance leases. Nvidia's H200 shipments were below 1% of its Data Center revenue in the July quarter, showing how narrow that export channel is, not the share of Chinese capex going to Nvidia. Domestic Ascend, Zhenwu and memory pathways limit the case for China as a backstop Western-chip buyer. Equipment makers remain an exception, with about 20–34% China exposure across different periods and denominators, under policy risk.51, 8, 10, 67, 68, 69, 70, 32, 45, 47, 48, 49

DeepSeek and domestic supply. V3's arithmetic measures an official run at assumed rental prices, excluding earlier experiments. It is neither hardware acquisition cost nor the cost of discovering the recipe.13 The later Nature supplement separately estimates specified R1 stages on top of an inherited base. Neither figure establishes a complete development budget or serving margin.14

DeepSeek's current Flash model is V4.1-Flash-1017, updated October 17, 2026 with API pricing unchanged; the April V4 preview is historical. A separate team's paper demonstrates V4-family post-training on Ascend, without establishing that original frontier pretraining ran entirely on Huawei.15, 16, 58 On September 30, DeepSeek released an Ascend backend in TileKernels and sparse-attention inference kernels in FlashMLA. TileKernels requires Ascend 950 and CANN 9.2.0+. FlashMLA describes the release as DeepSeek's Ascend inference infrastructure and claims 410 TFLOPS prefill and 360 TFLOPS decode, about 95%/83% of peak, on V4.1 workloads. These are vendor kernel-performance claims, with neither an Nvidia comparison nor a production-deployment disclosure; they do not establish training on Ascend or full-system cost parity, reliability or shipped capacity.55, 56 Bloomberg's later site report explicitly concerns inference; the proposed accelerators fill only part of the site, without establishing installed capacity or utilization.17

The October 20 DeepInfer release extends the inference software stack; its SM100-only benchmarks leave Ascend cost/performance comparisons open. H100/Hopper support is not listed.74

Domestic substitution extends beyond Ascend. Alibaba introduced Zhenwu V900 on September 22, 2026, advertising 3× predecessor performance and clusters up to 500,000 cards. Huawei's September 19 ecosystem announcement offers shared 10k-NPU-scale compute access and programming support. These vendor disclosures widen the supply and software pathways; accepted shipments, actual utilization, qualified HBM volume and comparable cost per successful task remain unverified.60, 61

Huawei's roadmap cannot settle manufacturing yield, qualified HBM supply or cluster economics.18 Component-chip production estimates are not finished-accelerator counts. Reuters' April 29 report gives a separate ~750,000-unit 950PR shipment plan for 2026 and an April production start (reported, unverified, C). If delivered, it would make a 910C-only reading of row 15 incomplete; neither estimate can be added without reconciling dies, finished units and product mix.66 HBM controls and reported scarcity-driven processor price increases contradict unlimited cheap domestic compute.17–20

SMIC's utilization and positive gross margin show manufacturing demand, without disclosing advanced-node good-die output or Ascend profitability. Original teardown evidence supports progress while contradicting leading-node parity.21, 22 CXMT DRAM and YMTC NAND widen domestic options; NAND cannot substitute for accelerator HBM, and a DRAM process announcement does not establish packaging yield or HBM qualification.23, 24 Counterpoint estimates CXMT's Q2 2026 DRAM revenue share at 10%, up from 4% a year earlier (C), without listing CXMT in HBM shares. That is realised conventional-DRAM competition and a channel for memory-layer margin pressure (see memory/HBM); it is not evidence of qualified HBM at Ascend scale.65 Substitution can reduce import dependence while increasing resource cost per useful token.

Prices and inference margins. Open weights enable competing hosts and credible migration threats; cheap APIs can pressure scarcity pricing. The tariff samples establish offers, not a quality-adjusted price series or a causal estimate of global deflation. Not every hosted variant in an open-weight family has released weights.26–29

OpenRouter's September 22 newsletter reports Chinese labs at 69% of mid-September tokens and open-weight models at 72%, while closed-weight models retained 73% of spend versus 89% in January. More than 40% of US startups on the platform used DeepSeek in production versus about 13% in January. These A-grade platform disclosures, excluding resellers, show adoption and cheap-tier competitive pressure; extrapolation to the whole market is C. Chinese origin and open weights are overlapping categories, not identical partitions; closed weights are not synonymous with Western origin. Falling spend share in a growing market does not establish falling revenue or causally measure Western margins.63 MiniMax's 17.9% group gross margin and large R&D loss show the financial strain that can accompany this export strategy, without establishing near-cost pricing at every Chinese lab.64

Western margins collapse only if substitution works at comparable task success, latency, reliability and governance, and price reductions outrun cost reductions and usage growth. Cheaper tokens can expand agent usage and equipment demand while destroying model rents. Kimi's premium tier and Tencent's pricing commentary contradict universal commoditization.8, 27 Standalone inference margins after depreciation, power and unsuccessful outputs, and returns by investment cohort, remain undisclosed; Alibaba's segment EBITA is observable but does not resolve them.51

DeepSeek's September 10 V4.1-Flash launch claims KV-cache HBM demand at one-quarter of the predecessor's, rather than a reduction of all memory needs. Its October 17 update claims shorter runtimes and outputs with unchanged API pricing. Lower task costs can therefore reflect efficiency alongside competitive rent pressure; these vendor claims do not establish realized returns.57, 58

Permission from both governments. H20 illustrates stranded inventory risk. A disclosed revenue-share expectation does not establish a statutory tax or collections. H200 subsequently operates under a distinct combination of conditional US licensing, Chinese restrictions and a separate tariff.30–33 Blackwell variants and tracking proposals do not establish a broad approved data-centre channel.34, 35

MOFCOM combines a draft catalogue revision with already-effective temporary pre-export registration, including started but unfinished exports and specified foreign ownership/control changes. Under the proposed final regime, public-domain/open-source exclusions and API/app/SaaS service exclusions qualify the scope. Restricted-technology overseas deployment remains within proposed controls, with facilitation for own-cloud deployment. These provisions do not establish an already-effective blanket exemption for transferring weights or infrastructure know-how.36, 37 Materials controls affect semiconductor and industrial inputs through another channel. Suspensions leave licensing leverage intact and have different expiry dates.38–41

Western equipment makers benefit from construction but face licensing cuts and domestic substitution. ASML's annual total-sales guide and quarterly system-sales share use different denominators; other makers report different periods. Exposure is not revenue certain to disappear.45–49

Taiwan and the bubble question. TSMC acknowledges geographic concentration and interruption risks. Blockade, quarantine, shipping or energy disruption could stop deliveries without destroying fabs. Overseas expansion does not immediately recreate the manufacturing ecosystem.42

Taiwan's recent record equity close shows markets still capitalize substantial earnings.43 Growth, discount rates, recovery assumptions and geopolitical risk are jointly priced. No reliable isolated TSMC discount, executable war-insurance quote or market-implied conflict probability was verified. Military counts measure activity, not probabilities.44

OpenRouter adoption and CXMT's estimated share gain establish concrete competitive channels; their effect on market-wide Western inference returns remains unmeasured. The June-quarter capex comparison does not support China replacing US infrastructure demand. Profitable platform cores finance rapidly rising commitments, while Alibaba/Tencent AI-product losses and MiniMax's thin margin and large R&D loss place shareholders first in line for impairment and dilution. Debt adds lender exposure, with core-business cash generation relevant to coverage. Taiwan remains a systemic supply tail despite healthy demand. These risk interpretations are C; utilization, incremental returns and security valuations remain inadequate to declare a China-wide AI bubble.

Conflicts and open questions

  • October 25 update — listed-lab prices versus reported lab economics: MiniMax's H1 revenue rose 283.1% and its gross margin improved to 17.9% (claim 59), yet its shares fell below the IPO price on October 22 and Zhipu is 81% below its June high (claim 72). Prices discount future competition and funding needs, not only reported growth. Bloomberg's attribution to token-price cuts is reporting, not a measured margin effect. The October 23 +16.74% rebound shows that a single-day observation must not stand in for the trend.64, 77, 78

  • DeepSeek round size and status: Reuters (relaying Bloomberg) reports about RMB100bn / US$14.98bn at RMB500bn pre-money with completion expected; The Information reports nearly US$16bn. Neither is reconciled, and neither is a certified close. The October 6 RMB80–100bn range is kept as the earlier vintage. A private round near close at a valuation above the July target conflicts with the listed-lab selloff only if both are read as one market price; they are not (claims 49, 72).54, 79, 80

  • Registration versus closure: state-media reporting that DeepSeek and MiniMax won approval for global use supports the reading that the October 16 regime registers rather than blocks open-source/overseas licensing for these developers. The underlying regulator documents were not fetched, and registration of two firms says nothing about restricted-technology deployments or the final catalogue after 15 November (claims 31–32, 73).36, 81

  • Private adoption versus federal access: OpenRouter's US-startup uptake does not imply access to every US customer. The newly public DHS order restricts the covered civilian federal channel, while leaving the note's private-market evidence intact. Its July issuance date also precludes treating October publication as a new 60-day removal deadline.63, 72, 73

  • Codex versus Claude — prepayments and cash exposure: Codex's October 17 cross-check emphasised +RMB37.6bn FCF excluding prepayments as a qualification to the cash-burn diagnosis; Claude's October 18 review judged that framing too reassuring. Both accept the −RMB13.8bn reported FCF and RMB51.4bn implied prepayment effect. Claude stresses cash already paid, the RMB88.669bn quarterly net-cash decline and ≈RMB10.5bn new-AI-products profit drag. The counterfactual isolates timing, not a lower compute commitment; the cash decline is not wholly attributed to AI.8

  • Codex versus Claude — evidence of margin competition: Codex judged tariff samples insufficient for realised rent destruction and retained conditional competition language. Claude cites 69% Chinese-lab token share and 73% closed-weight spend share on OpenRouter, plus MiniMax's 17.9% gross margin, to support a firmer cheap-tier margin-competitor conclusion. This note adds that evidence while retaining the limits: one channel, different origin/weight-access partitions, group rather than API margins, and no causal estimate of Western margin loss.63, 64

  • Codex versus Claude — capex as demand evidence: the Codex-applied Summary treated RMB67.678bn as accelerating infrastructure demand; Claude stresses Alibaba's own procurement, anticipated CPU demand and component-price explanation. Spending accelerated 75% y/y, but delivered compute and realised AI demand need not rise proportionately. The revised Summary states the company's explanation.51

  • Huawei supply estimates: the September component estimate (~1.6m dies, about 600,000 910C plus other models) and Reuters' April ~750,000-unit 950PR plan concern different product mixes and units. Claude flagged the latter as a lead; it is now traced to Reuters, still C and not primary-confirmed. The 910C figure is not a complete 2026 accelerator total; adding the two would risk double counting.17, 66

  • Review FX correction: Claude used RMB6.7848/US$ for the scale comparison; the SEC release fetched October 18 states RMB6.7851/US$. The latter gives US$19.433bn and 11.35%, leaving the rounded US$19.4bn/11% conclusion unchanged. Codex's earlier review avoided an aggregate; the new comparison is explicitly scoped to seven groups with unequal definitions.51, 67, 68, 69, 70

  • Review judgement — rents and bubble diagnosis: the original note's Summary said China “competes away model rents” and treated inference-margin/overbuild risks as material (claim 40). The cross-check judges tariff samples insufficient to demonstrate realized rent destruction and puts more weight on profitable cloud services, application losses and procurement timing. Both views retain competition as a conditional risk; neither has matched-task switching, utilization or investment-cohort returns to establish a China-wide bubble.26, 8, 51

  • Review judgement — financing and loss bearers: the original note foregrounded negative FCF and equity funding and described lender losses conditionally. The cross-check assigns more weight to Tencent's announced notes and ByteDance's reported unsecured facility. Debt financing is now evidenced, but AI allocations, completed issuance/drawdown and ultimate losses remain unestablished.52, 53

  • Review judgement — substitution constraints: the original Summary treated software, packaging and yields as binding constraints and called Ascend a credible inference alternative. The cross-check views released Ascend kernels as stronger substitution evidence, while finding unquantified packaging/yield claims too categorical. The releases establish porting progress; manufacturing autonomy and comparable deployment economics remain open.16, 18, 55, 56

  • ByteDance budget versus denial: Reuters' anonymous sources put the 2025 budget above RMB150bn; ByteDance explicitly rejected the spending information. The original grade C is corrected to contested grade D. The separate May 2026 US$70bn scenario and US$50bn profit estimate remain C, reported, unverified; neither resolves the disputed earlier budget. The original Summary called ByteDance's prospective spending “much larger” than Alibaba's commitment; the cross-check judges that comparison unsafe because an annual scenario and a multiyear floor have different periods and potentially different scopes.11, 12

  • V4-Pro service-status conflict: the September launch page planned rerouting V4-Pro requests after September 14, whereas the current official changelog says the API continues after that date. Huawei Cloud's September 22 retirement notice applies to its own MaaS service, not all hosts. Treat each notice's scope separately and verify the intended host's current service.57, 58, 59

  • Efficiency versus scale: modest disclosed training compute coexists with a reported giant serving build. The scopes differ. Needed: audited development spending and actual inference utilization.13, 14, 17

  • Substitution versus scarcity: Ascend post-training, released kernels and domestic memory progress support substitution; reported HBM scarcity and incomplete original-pretraining disclosure limit the autonomy claim.55, 56 Needed: accepted shipments, good-die yields and repeatable workload economics.16–24

  • Commoditization versus monetization: cheap tariffs compete with premium tiers and improving cloud pricing. Needed: matched-task customer switching and contribution margins, not token prices alone.8, 26–29

  • Boom versus bubble: cash commitments, new equity and prospective debt increase capital at risk; disclosed cloud profit is contrary evidence to a blanket loss-making-operations diagnosis. The prepayment-adjusted FCF counterfactual separates timing but does not reverse the cash already committed; Claude puts greater weight on the model-layer losses and balance-sheet drawdown.51, 8, 52 Needed: utilization, lease commitments, depreciation and returns by investment cohort.51, 8, 12

  • Policy closure versus continued exports: temporary registration is effective; the draft final regime adds oversight while preserving explicit exclusions and controls on restricted-technology overseas deployment. Implementation could diverge; the final catalogue, licences and treatment of future releases remain unresolved.36, 37

  • Further lead: Claude identified Zhipu results coverage for primary HKEX tracing; no unverified search-lead financial figure is used here. DeepSeek closing, September-quarter platform results and new tariff cuts remain watch items, not established October events.

  • Unverified claims: revised Alibaba total; Tencent/Baidu full-year capex guidance; audited ByteDance spending and loan drawdown; Tencent note issuance; DeepSeek round closing and site financing/delivery; advanced-node yields, HBM shipments, all-in R1 budget and Taiwan implied probability.

Signals to watch

Signal Current value As of Source Why it matters
Hyperscaler capex Latest quarters: Alibaba RMB67.678bn; Tencent RMB52.784bn; Baidu RMB11.390bn 2026-06 51, 7, 10 Compare actuals, procurement timing and cash flow with guidance
Alibaba commitment RMB380bn floor; increase announced without replacement total 2025-09 announcement 1, 2 A quantified revision would clarify remaining funding needs
Ascend accepted shipments No verified series; DeepSeek's ≥160,000 remains reported, unverified 2026-09-04 17 Orders become competition only after usable delivery
HBM availability Huawei roadmap exists; scarcity-driven price increases reported 2026-09 18, 19 Tests whether memory caps domestic accelerator supply
Model tariffs and realized economics GLM-5.3 US$1.40/4.40; Flash US$0.15/0.50 2026-10-17 28 Track cuts alongside task cost, availability and margin
US market access DHS civilian federal exclusion/removal notice published 20 Oct; July order, no new publication-based 60-day deadline. H200 China shipments <1% of Nvidia Data Center revenue in July quarter 2026-10-20 notice / quarter ended 2026-07-26 72, 73, 32 Next: implementation/exception decisions; no new dated deadline verified. Track US customer restrictions separately from chip-export access
Chinese technology controls Registration active; DeepSeek and MiniMax reported approved for global use (state-media-affiliated post Oct 22, Reuters Oct 23); draft comments close 15 Nov 2026-10-16 / 2026-10-23 36, 37, 81; claim 73 Next: comment close 15 Nov and the final catalogue; seek regulator-published registration records. Final scope could affect weights, licensing and overseas deployment
Materials-control suspensions Separate deadlines: 10 Nov and 27 Nov 2026 current notices 39, 40 Renewal or expiry changes supply-chain friction
Taiwan Strait activity 9 aircraft sorties, 6 naval ships, 8 official ships 2026-10-17 44 Watch persistent changes, exclusion zones and shipping disruption
Toolmaker China exposure ASML 23% systems; Lam 34%, KLA 30%, TEL 30.4% in stated periods latest cited periods 45, 47–49 Reveals demand losses and substitution before global totals do
Financing completion Tencent US$1.8bn/CNY26bn notes priced; expected issue 21 Oct, completion not verified in this update; ByteDance US$29.6bn facility reported, unverified; DeepSeek ~RMB100bn / US$14.98bn at RMB500bn pre-money reported near completion Oct 22 (earlier >RMB80bn, Oct 6), unverified 2026-10-14 / 2026-09-04 / 2026-10-22 52, 53, 54 Separate announcements from completed issuance, drawdown and round closing; monitor allocation and terms
Profit allocation and procurement timing Alibaba adjusted EBITA: cloud +RMB5.628bn, labs/apps −RMB13.861bn; Tencent FCF −RMB13.8bn or +RMB37.6bn excluding compute prepayments 2026-06 51, 8 Track recurring operating results separately from purchase timing and investment-cohort returns
Domestic software and supplier delivery DeepInfer released 20 Oct: SM100/Ascend 950 offline inference, Nvidia-only benchmarks; 30 Sep kernels and 17 Oct model update remain earlier evidence; V900 up to 500,000 cards remains vendor specification 2026-09 / 2026-10-17 / 2026-10-20 74, 55, 56, 58, 60, 61 Next: Ascend end-to-end benchmarks and accepted deployments; no announced test date. Check task cost and deliveries against specifications
Cheap-tier adoption versus spend Chinese labs 69% of tokens; closed weights 73% of spend versus 89% in January; reseller activity excluded Mid-September 2026; published 2026-09-22 63 Track spend as well as volume; one platform does not establish global shares or Western margin loss
Listed-lab funding and returns MiniMax H1 revenue US$116.573m, 60.8% overseas, 17.9% gross margin; adjusted loss US$293.031m; cash/equivalents US$930.905m. Shares: below IPO price Oct 22, then HK$187.60 (+16.74%) Oct 23; Zhipu −81% from June high; ~US$150bn combined peak-to-date loss 2026-06-30, released 2026-08-26; prices 2026-10-22/23 64, 77, 78; claim 72 Next: dated closes, any DeepSeek/Moonshot listing filing and MiniMax/Zhipu second-half results. Watch whether margin gains and revenue close the equity-funded R&D gap
China equity backdrop CSI 300 −1.8% Oct 22, −20% from June high; Hang Seng Tech −5.3% intraday 2026-10-22 82; claim 72 Separates China-wide demand and policy disappointment from AI-specific repricing; a bear-market label is not verified here
Tencent cash commitments and AI drag Net cash RMB146.860bn → 58.191bn in Q2; RMB51.4bn compute prepayment effect; ≈RMB10.5bn new-AI-products profit drag (B) 2026-06-30 8 Distinguish cash already paid from timing counterfactuals; monitor replenishment, delivery and debt issuance
Conventional DRAM substitution CXMT estimated 10% of revenue versus 4% a year earlier; not listed in HBM shares (C) Q2 2026; September update 65 Tests memory-layer competitive pressure separately from HBM qualification
Ascend 950 production and kernel claims ~750,000 950PR shipment plan (reported, unverified); Ascend 950/CANN 9.2.0+ target; FlashMLA vendor claims 95%/83% of peak 2026-04-29 / 2026-09-30 66, 55, 56 Reconcile delivered units with component estimates and measure matched-task economics
Scale and commitment pace China three US$19.433bn versus US four US$171.210bn June capex; Alibaba five-quarter group capex RMB193.741bn (B; differing scopes) 2026-06-30 51, 71, 8, 10, 67, 68, 69, 70 Next: Microsoft and Alphabet Q3 results Oct 28, Amazon Oct 29 (Meta's Q3 capex already US$34.906bn including lease principal) update the US-four denominator; Chinese September-quarter results follow in November. Watch share of useful capacity and programme funding rather than treating spending as a national AI total
DeepSeek valuation and strategic funding Oct 22: ~RMB100bn / US$14.98bn at RMB500bn pre-money, >15 investors incl. CATL, Tencent, Meituan, Geely, GigaDevice, Midea; completion expected in coming days. Oct 6: >RMB80bn (possibly RMB100bn), RMB500bn target. All reported, unverified; closing not established 2026-10-06 / 2026-10-22 reports 54, 79, 80; claim 49 Verify closing and terms; platform exposure includes strategic lab investments

Quarterly spending observations are preserved in the capex CSV.

Sources

  1. https://www.alibabagroup.com/en-US/document-1830678592242057216 — Alibaba, “Invest RMB380 billion in AI and Cloud Infrastructure”; published 2025-02-24; accessed 2026-10-17.

  2. https://home.alibabagroup.com/en-US/document-1911884625546838016 — Alibaba, “Apsara Conference 2025” (investment increase); published 2025-09-29; accessed 2026-10-17.

  3. https://data.alibabagroup.com/ecms-files/1532295521/fa5d65fc-9b3e-4e82-a8fc-4ce1c3e2c407/Alibaba%20Group%20Announces%20June%20Quarter%202026%20Results.pdf — Alibaba, June-quarter results; published 2026-08-20; accessed 2026-10-17; original URL superseded for verification by SEC source 51; not used in the CSV.

  4. https://data.alibabagroup.com/ecms-files/1532295521/5b1cb883-8d00-4237-a148-6631cc12a5d2/Alibaba%20Group%20Announces%20March%20Quarter%202026%20and%20Fiscal%20Year%202026%20Results.pdf — Alibaba, March-quarter and fiscal-year results; published 2026-05-13; accessed 2026-10-17.

  5. https://www.alibabagroup.com/en-US/document-2029365886510432256 — Alibaba, completion of equity placement; published 2026-08-26; accessed 2026-10-17.

  6. https://static.www.tencent.com/uploads/2026/04/09/62d786fcf3d3c8cb7e54791ee95439ac.pdf — Tencent, Annual Report 2025; published 2026-04-09; accessed 2026-10-17.

  7. https://www.tencent.com/wp-content/uploads/2026/08/E700_IR.pdf — Tencent, Interim Report 2026; published 2026; exact publication day not shown; accessed 2026-10-17.

  8. https://www.tencent.com/wp-content/uploads/2026/08/Tencent-Announces-2026-Second-Quarter-Results.pdf — Tencent, second-quarter results; published 2026-08-12; accessed 2026-10-17; re-fetched 2026-10-18.

  9. https://ir.baidu.com/news-releases/news-release-details/baidu-announces-fourth-quarter-and-fiscal-year-2025-results — Baidu, fourth-quarter and fiscal-year results; published 2026-02-26; accessed 2026-10-17.

  10. https://ir.baidu.com/news-releases/news-release-details/baidu-announces-second-quarter-2026-results — Baidu, second-quarter results; published 2026-08-18; accessed 2026-10-17; re-fetched 2026-10-18.

  11. https://www.investing.com/news/stock-market-news/exclusivebytedance-plans-20-billion-capex-in-2025-mostly-on-ai-sources-say-3826261?ampMode=1 — Reuters, ByteDance capex plan (secondary); published 2025-01-23; accessed 2026-10-17.

  12. https://theedgemalaysia.com/node/805120 — Bloomberg via The Edge, “ByteDance weighs capex of as much as US$70b” (secondary); published 2026-05-27; accessed 2026-10-17.

  13. https://arxiv.org/html/2412.19437v2 — DeepSeek-AI, DeepSeek-V3 Technical Report, cost table; published 2024-12-27; revised 2025-02-18; accessed 2026-10-17.

  14. https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41586-025-09422-z/MediaObjects/41586_2025_9422_MOESM1_ESM.pdf — DeepSeek-AI, Nature R1 supplementary information, training-cost section; published 2025-09-17; accessed 2026-10-17.

  15. https://deepseek.com/en/news/v4-preview/ — DeepSeek, V4 preview announcement; published 2026-04-24; accessed 2026-10-17.

  16. https://arxiv.org/abs/2607.20145 — Li et al., “SLAI T-Rex: Full-Parameter Post-training … on Ascend SuperPOD”; published 2026-07-22; revised 2026-08-19; accessed 2026-10-17.

  17. https://www.bloomberglinea.com/negocios/deepseek-planea-instalar-160000-chips-huawei-en-un-megacentro-para-reducir-su-dependencia-de-nvidia/ — Hawkins/Fan, Bloomberg, DeepSeek Huawei cluster (secondary); published 2026-09-04; accessed 2026-10-17.

  18. https://www.huawei.com/en/news/2025/9/hc-xu-keynote-speech — Huawei, “Groundbreaking SuperPoD Interconnect” (roadmap); published 2025-09-18; accessed 2026-10-17.

  19. https://www.investing.com/news/stock-market-news/exclusivechinas-ai-chipmakers-raise-prices-as-highbandwidth-memory-shortage-bites-4894950?ampMode=1 — Reuters, Chinese AI chipmakers raise prices as HBM shortage bites (secondary); published 2026-09-10; accessed 2026-10-17.

  20. https://www.bis.gov/press-release/commerce-strengthens-export-controls-restrict-chinas-capability-produce-advanced-semiconductors-military — BIS, semiconductor equipment and HBM controls; published 2024-12-02; accessed 2026-10-17.

  21. https://smic.cdn.shwebspace.com/uploads/6a7d7c4f/ER_EN.pdf — SMIC, second-quarter results; published 2026-08-13; accessed 2026-10-17.

  22. https://www.techinsights.com/ja/node/60833 — TechInsights, original Kirin teardown: “SMIC N+3 Confirmed”; published 2025-12-11; accessed 2026-10-17.

  23. https://www.cxmt.com/news/info_97.html — CXMT, G5 DRAM platform enters mass production; published 2026-09-20; accessed 2026-10-17.

  24. https://www.ymtc.com/en/technicalintroduction.html — YMTC, “About Xtacking”; published undated; accessed 2026-10-17.

  25. https://www.investing.com/news/stock-market-news/chinas-memory-chip-makers-ride-ai-boom-to-new-power--and-us-scrutiny-4810472 — Reuters, Chinese memory makers’ funding and pricing (secondary); published 2026-07-24; accessed 2026-10-17.

  26. https://www.alibabacloud.com/help/en/model-studio/model-pricing — Alibaba Cloud, model inference pricing; published updated 2026-10-15; accessed 2026-10-17.

  27. https://platform.kimi.com/docs/pricing/chat — Moonshot, model inference pricing (tables inspected in page HTML); published undated live tariff; accessed 2026-10-17.

  28. https://docs.z.ai/guides/overview/pricing — Z.AI, model pricing; published undated live tariff; accessed 2026-10-17.

  29. https://www.minimax.io/blog/minimax-m25 — MiniMax, M2.5 launch and pricing; published 2026-02-12; accessed 2026-10-17.

  30. https://www.sec.gov/Archives/edgar/data/1045810/000104581025000209/nvda-20250727.htm — Nvidia, fiscal Q2 2026 Form 10-Q; published 2025-08-27; accessed 2026-10-17.

  31. https://www.sec.gov/Archives/edgar/data/2488/000000248826000018/amd-20251227.htm — AMD, fiscal 2025 Form 10-K; published 2026-02-04; accessed 2026-10-17.

  32. https://investor.nvidia.com/files/doc_financials/2027/NVDA-2027-Q2-10Q-Final-including-exhibits.pdf — Nvidia, fiscal Q2 2027 Form 10-Q; published 2026-08-26; accessed 2026-10-17.

  33. https://media.bis.gov/press-release/department-commerce-revises-license-review-policy-semiconductors-exported-china — BIS, H200 case-by-case licensing policy; published 2026-01-13; accessed 2026-10-17.

  34. https://www.tradingview.com/news/reuters.com%2C2025%3Anewsml_L2N3V202J%3A0-nvidia-s-new-rtx6000d-chip-for-china-finds-little-favour-with-major-firms-sources-say/ — Reuters, RTX6000D and reported B30A plans (secondary); published 2025; exact publication day not established; accessed 2026-10-17.

  35. https://www.congress.gov/bill/119th-congress/house-bill/3447 — Congress.gov, Chip Security Act, text/status; published introduced 2025-05-15; accessed 2026-10-17.

  36. https://www.mofcom.gov.cn/zcfb/dwmygl/art/2026/art_61446ab4beee4f63973d4d0ab3097648.html — MOFCOM/MOST, Announcement 48: draft and temporary registration; published 2026-10-16; accessed 2026-10-17.

  37. https://www.mofcom.gov.cn/zwgk/zcjd/art/2026/art_b11d5c1e9374421e93e85dc87eba6e38.html — MOFCOM, catalogue-revision questions and answers; published 2026-10-16; accessed 2026-10-17.

  38. https://policy.mofcom.gov.cn/claw/clawContent.shtml?id=102310 — MOFCOM/GACC, Announcement 18: rare-earth controls; published 2025-04-04; accessed 2026-10-17.

  39. https://dltb.mofcom.gov.cn/swyw/art/2025/art_d0a1a4a22f33467f85a2e6f30c5470c2.html — MOFCOM/GACC, Announcement 70: suspension; published 2025-11-07; accessed 2026-10-17.

  40. https://dltb.mofcom.gov.cn/zcfg/art/2025/art_6d4dc4f76d7844a18733f4dcaf9a32d8.html — MOFCOM, Announcement 72: US-specific suspension; published 2025-11-09; accessed 2026-10-17.

  41. https://www.mofcom.gov.cn/zfxxgk/fdzdgknr/ztfl/dwmygl/art/2024/art_8300c93ceef047debcf4da0ced080629.html — MOFCOM and other agencies, unified dual-use control list; published 2024-11-15; accessed 2026-10-17.

  42. https://investor.tsmc.com/sites/ir/sec-filings/2025_20F%20Report.pdf — TSMC, 2025 Form 20-F, geographic and operational risks; published 2026-04-16; accessed 2026-10-17.

  43. https://www.brecorder.com/news/amp/40442826 — Reuters, “Soft US payrolls lift EM Asian stocks” (secondary); published 2026-10-06; accessed 2026-10-17.

  44. https://www.mnd.gov.tw/en/News/PLAAct/88006 — Taiwan MND, daily PLA activity report; published 2026-10-17; accessed 2026-10-17.

  45. https://www.sec.gov/Archives/edgar/data/937966/000162828026065813/a2026_10x14xpresentation.htm — ASML, Q3 presentation, geographic chart visually inspected; published 2026-10-14; accessed 2026-10-17.

  46. https://www.fool.com/earnings/call-transcripts/2026/10/14/asml-asml-q3-2026-earnings-call-transcript/ — ASML management call, hosted by Motley Fool (transcript); published 2026-10-14; accessed 2026-10-17.

  47. https://www.sec.gov/Archives/edgar/data/707549/000070754926000037/lrcx-20260628.htm — Lam Research, fiscal 2026 Form 10-K; published 2026-08-07; accessed 2026-10-17.

  48. https://www.sec.gov/Archives/edgar/data/319201/000031920126000027/klac-20260630.htm — KLA, fiscal 2026 Form 10-K; published 2026-08-06; accessed 2026-10-17.

  49. https://www.tel.com/ir/irta3a00000006g5-att/fy27q1presentations-e.pdf — Tokyo Electron, Q1 FY2027 presentation, regional sales; published 2026-07-30; accessed 2026-10-17.

  50. https://home.alibabagroup.com/en-US/document-2041268239081668608 — Alibaba, Apsara 2026 infrastructure roadmap; published 2026-09-25; accessed 2026-10-17.

  51. https://www.sec.gov/Archives/edgar/data/1577552/000110465926099220/tm2623667d1_ex99-1.htm — Alibaba, June Quarter 2026 Results, SEC Exhibit 99.1 (verified replacement for source 3); published 2026-08-20; accessed 2026-10-17; re-fetched 2026-10-18.

  52. https://www.tencent.com/wp-content/uploads/2026/10/E700_Pricing-Announcement.pdf — Tencent, pricing announcement for USD and CNY notes; published 2026-10-14; accessed 2026-10-17.

  53. https://www.investing.com/news/stock-market-news/bytedance-secures-296-billion-loan-in-ai-push-sources-say-4889439 — Wu/Ngui, Reuters via Investing.com, “ByteDance secures $29.6 billion loan in AI push, sources say” (secondary); published 2026-09-04; accessed 2026-10-17.

  54. https://www.investing.com/news/economy-news/deepseek-set-to-net-over-12-billion-in-new-fundraising-source-says-4933887 — Reuters via Investing.com, “DeepSeek set to net over $12 billion in new fundraising, source says” (secondary); published 2026-10-06; accessed 2026-10-17; re-fetched 2026-10-18.

  55. https://github.com/deepseek-ai/TileKernels — DeepSeek-AI, TileKernels repository, Ascend backend release in News; released 2026-09-30; accessed 2026-10-17; re-fetched 2026-10-18.

  56. https://github.com/deepseek-ai/FlashMLA/blob/main/docs/20260930-ascend-prefill-deep-dive.md — DeepSeek-AI, FlashMLA, “A Deep Dive Into the Ascend Sparse Attention Forward Kernel”; published 2026-09-30; accessed 2026-10-17; re-fetched 2026-10-18.

  57. https://www.deepseek.com/en/news/deepseek-v4-1-flash/ — DeepSeek, “Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient”; published 2026-09-10; accessed 2026-10-17.

  58. https://api-docs.deepseek.com/updates/ — DeepSeek, API Change Log, current Flash update and retained V4-Pro service; entries 2026-09-10 and 2026-10-17; accessed 2026-10-17.

  59. https://www.huaweicloud.com/intl/en-us/notice/20260922173926811.html — Huawei Cloud, “Notice on Discontinuing DeepSeek-V4-Pro Model Service on Huawei Cloud MaaS”; published 2026-09-22; accessed 2026-10-17.

  60. https://home.alibabagroup.com/en-US/document-2039431633571938304 — Alibaba, Eddie Wu Apsara 2026 full-stack AI keynote, Zhenwu V900 specifications; published 2026-09-22; accessed 2026-10-17.

  61. https://www.huawei.com/en/news/2026/9/hc-agentic-thinkpro-pto-cann — Huawei, “Huawei Advances Agentic Computing on Multiple Dimensions for SuperPoDs and SuperClusters”; published 2026-09-19; accessed 2026-10-17.

  62. https://www.tencent.com/wp-content/uploads/2026/10/E700_Launch-Announcement.pdf — Tencent, programme launch announcement, offering-circular financial-information extract, capital expenditure p.48; published 2026-10-12; accessed 2026-10-17.

  63. https://openrouter.ai/data/newsletter/the-token-state-of-play — OpenRouter, “The Token State of Play,” platform usage newsletter; published 2026-09-22; accessed 2026-10-18.

  64. https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0826/2026082600680.pdf — MiniMax Group, HKEX interim results announcement, unaudited H1 2026; published 2026-08-26; accessed 2026-10-18.

  65. https://counterpointresearch.com/en/insights/global-dram-and-hbm-market-share — Counterpoint Research, “Global DRAM and HBM Market Share: Quarterly,” analyst estimates; September 2026 update; accessed 2026-10-18.

  66. https://www.investing.com/news/stock-market-news/exclusivebig-chinese-tech-firms-scramble-to-secure-huawei-ai-chips-after-deepseek-v4-launch-sources-say-4643661 — Reuters via Investing.com, “Big Chinese tech firms scramble to secure Huawei AI chips after DeepSeek V4 launch, sources say” (secondary); published 2026-04-29; accessed 2026-10-18.

  67. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000066/googexhibit991q22026.htm — Alphabet, Q2 2026 earnings release, SEC Exhibit 99.1, cash-flow table; published 2026-07-22; accessed 2026-10-18.

  68. https://www.sec.gov/Archives/edgar/data/1018724/000101872426000026/amzn-20260630.htm — Amazon, Q2 2026 Form 10-Q, cash-flow statement; signed 2026-07-30; accessed 2026-10-18.

  69. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/ — Meta, Second Quarter 2026 Results, capex including finance-lease principal; published 2026-07-29; accessed 2026-10-18.

  70. https://www.microsoft.com/en-us/Investor/events/FY-2026/earnings-fy-2026-q4 — Microsoft, FY2026 Q4 earnings-call transcript, CFO capex remarks; published 2026-07-29; accessed 2026-10-18.

  71. https://www.sec.gov/Archives/edgar/data/1577552/000110465926060224/tm2614494d1_ex99-1.htm — Alibaba, March Quarter and Fiscal Year 2026 Results, SEC Exhibit 99.1; published 2026-05-13; accessed 2026-10-18.

  72. https://www.govinfo.gov/content/pkg/FR-2026-10-20/html/2026-21447.htm — DHS, Federal Register notice, docket DHS-2026-0628; filed 2026-10-19, published 2026-10-20; re-fetched 2026-10-21.

  73. https://www.govinfo.gov/content/pkg/FR-2026-10-20/pdf/2026-21447.pdf — DHS signed Exclusion and Removal Order 2026-01, original release 2026-07-24; Federal Register pp.66347–66350, image appendix visually inspected 2026-10-21; linked daily-issue PDF in signals also fetched.

  74. https://github.com/deepseek-ai/DeepInfer — DeepSeek-AI, README/support/license and benchmark scope; released 2026-10-20; re-fetched 2026-10-21, including raw README.

  75. https://api.github.com/repos/deepseek-ai/DeepInfer/commits?per_page=4 — official repository commit history; initial commit 2026-10-20 04:09:42 UTC; fetched 2026-10-21.

  76. https://www.theinformation.com/articles/deepseek-shaping-huaweis-ai-chips-growing-challenge-nvidia — Huawei–DeepSeek report (secondary); published 2026-10-19 per verified signal record; restricted original re-fetched 2026-10-21; substantive details not independently reproduced.

  77. https://hk.finance.yahoo.com/news/%E4%B8%AD%E5%9C%8Bai%E6%A8%A1%E5%9E%8B%E9%96%8B%E7%99%BC%E5%95%86%E8%82%A1%E5%83%B9%E4%BC%BC%E6%9C%AA%E8%A6%8B%E5%BA%95-minimax%E5%92%8C%E6%99%BA%E8%AD%9C%E5%B8%82%E5%80%BC%E5%85%B1%E8%B7%8C1500%E5%84%84%E7%BE%8E%E5%85%83-015155881.html — Bloomberg (Jeanny Yu) via Yahoo Finance HK, "China AI Model Stocks' $150 Billion Drop Shows No Sign of Easing"; published 2026-10-23 03:51 CEST; re-fetched 2026-10-25.

  78. https://markets.ft.markitdigital.com/data/equities/tearsheet/historical?s=100%3AHKG — FT/LSEG, MiniMax (100:HKG) dated price history; 2026-10-23 close HK$187.60; per signals 2026-10-25 item 46, not re-fetched.

  79. https://www.streetinsider.com/Reuters/Deepseek%2Braises%2B%2415%2Bbillion%2Bin%2Bchallenge%2Bto%2Bleading%2BUS%2Blabs%2C%2BBloomberg%2BNews%2Breports%C2%A0/27217623.html — Reuters via StreetInsider, DeepSeek round relaying Bloomberg News; published 2026-10-22; per signals 2026-10-25 item 48; HTTP 403 on 2026-10-25 re-fetch.

  80. https://www.theinformation.com/briefings/deepseek-raises-nearly-16-billion-second-funding-round — The Information, DeepSeek round "nearly $16 billion"; restricted, not retrieved.

  81. https://www.marketscreener.com/news/china-s-deepseek-minimax-win-approval-for-ai-global-use-state-media-affiliated-entity-says--ce785cded089ff21 — Reuters via MarketScreener, "China's DeepSeek, MiniMax win approval for AI global use, state media-affiliated entity says"; published 2026-10-23 02:59 EDT; re-fetched 2026-10-25.

  82. https://theedgemalaysia.com/node/822807 — Bloomberg via The Edge Malaysia, Chinese equities report, CSI 300 and Hang Seng Tech; dateline 2026-10-22; re-fetched 2026-10-25.

Updated 2026-10-25 09:48 UTC · raw markdown · source