Chinese-market AI accelerators — GPT-5.6 Pro deep dive

Run 2026-07-09 · 57m45s · model gpt-5-6-pro (Pro-verified, completion_path: api, Ashitaorbis project) · conversation Headline for the calculator: 910C peak is 1.504 POPS INT8 (the CloudMatrix paper's "1,054" is a typo); annual-commit rents H800 ≈ $1.75/hr, H20 ≈ $1.00/hr, 910C ≈ $1.95/hr (Huatai procurement); neutral 910C decode MFU 7% (5–10%); H20 17% (Ant production tiers); and "one China rental rate" is a category error — the same H20 spans ¥5–72/card-hour by rental class. Archived verbatim below (citation-marker artifacts from the source renderer preserved).

Evidence snapshot: 2026-07-09. Dollar conversions use the SAFE midpoint of ¥6.8036/$ [DISCLOSED]. Monthly prices normalized at 730 hours/month [SPECULATION: calculator convention].

Evidence labels

• DISCLOSED — official specification, public rate card, operator disclosure, government document, or public procurement result.
• CREDIBLY REPORTED — Reuters, FT, or reporting with a reasonably defined source basis.
• COMMUNITY ESTIMATE — reseller listing, copied cloud-calculator output, open-source issue, or an anonymous market memo.
• SPECULATION — my modeling recommendation or projection.

Bottom line

1. Change the 910C peak from 1.054P to 1.504P INT8. Huawei says an eight-NPU Atlas 800T A3 has 12.0 POPS INT8 [DISCLOSED] ⇒ ~1.50P/NPU; the CloudMatrix paper's tables repeatedly use 1,504 TFLOPS [DISCLOSED]. The paper's isolated "1,054" sentence is internally inconsistent — a typo. Proof: reported prefill efficiency 6,688 ÷ 1,504 = 4.45 tok/s/TFLOP matches the paper's own table; 1,054 would give 6.35.

2. Do not use one generic "China rental rate." In June–July 2026 the same H20 ranges from ¥5–8/card-hour (long-term bare-metal IDC) to ¥21–30 (committed Tencent) to ¥30–72 (on-demand/bundled public cloud). Different products, not one market.

3. For marginal-serving presets use annual-commit IDC/private-cloud pricing: H800 ≈ $1.75/card-hour, H20 ≈ $1.00, 910C ≈ $1.95 [all SPECULATION←CR/CE/procurement].

4. H20's 714 tok/s/GPU is real production evidence but the relaxed <70ms tier. Neutral production anchor ≈ 675–700 (Ant Pro tier, <50ms); use 423 for <30ms latency-sensitive scenarios.

5. 910C: 7% effective decode MFU central, 5–10% range. CloudMatrix's optimized 1,943 tok/s = 9.56% MFU supports the upper end; the second-hand "60% of H100" claim implies 5.5–6.4%.

1. Rental rate cards

1.1 Hyperscaler / public cloud (keep as a separate rental class)

Alibaba: gn8v documented (96GB HBM3, 4TB/s, 900GB/s NVLink [DISCLOSED]) but mainland tariff is sales-gated; copied calculator outputs ≈ ¥22.8–24.1 monthly / ¥34.7–36.7 on-demand per card-hour [COMMUNITY ESTIMATE]; a PAI doc example bid ceiling of ¥33/8-card-node-hour (¥4.125/card) is an example, not a clearing price. No reproducible static Alibaba H800 tariff. Volcano: no reproducible static mid-2026 card; stale Feb 2025 quote ≈ ¥133k/mo per 8-H20 node (¥22.77/$3.35).

1.2 IDC / neocloud market

Rental values to encode (annual-commit/bare-metal/private-cloud):

DeepSeek's $2/H800-hour [DISCLOSED] was a Feb 2025 internal costing assumption, not an invoice — a sanity check near the top of the current IDC range.

2. Hardware inputs and owned TCO

2.1 Specification cross-check

2.2 Accelerator capex (installed per accelerator)

2.3 Electricity, PUE, facility, depreciation

• Western-zone delivered electricity: Ulanqab ~¥0.35/kWh [CR]; Zhongwei ~¥0.36 [DISCLOSED]; Qingyang ≤¥0.40 [DISCLOSED] → ¥0.38/kWh central; subsidized domestic-chip projects ¥0.30 (¥0.25–0.36); coastal scenario ¥0.70 (0.60–0.85).
• PUE: national caps 1.25 new large / 1.20 hub-cluster [DISCLOSED] → 1.20 central (1.12–1.30).
• Facility capex: China 50MW benchmark ≈ $5.61–8.61/W [CR→derived]; international ~$11/W → $7.8/W central ($6.5–9.5).
• Depreciation: Alibaba computer equipment 3–5 years [DISCLOSED] → 4 years central (3–5); facility 15 years (14–20). Chinese lessors' 5-year "payback model" is pricing practice, not depreciation economics.
• Chip-only electricity at ¥0.38 + PUE 1.2: H800 $0.047/h; H20 $0.027/h; 910C $0.040/h — depreciation and utilization dominate, not power.

3. Throughput anchors

Evidence gap: no credible hardware-normalized production tok/s/chip for DeepSeek V4 on 950-series, GLM-5 on H20/H800/910C, Kimi K2.x on any of them, or MiniMax — deployability evidence only. Huawei's Atlas 950 SuperPoD "19.6M tokens/s across 8,192 NPUs" (⇒2,393/NPU) is a marketing metric without model/precision/SLO — do not anchor on it.

4. The Ascend reality check

Effective decode MFU (37B-active convention, dense 8-bit peak denominator):

H20's high FLOP-MFU is real: little peak compute, excellent bandwidth — the 4TB/s memory subsystem does the economic work in decode.

Reconciling 1,943 with "60% of H100": the 60% claim [CR, unverified, no public benchmark] implies 1,110–1,303 tok/s ⇒ 5.5–6.4% MFU vs CloudMatrix's 9.56%. Both can be true: CloudMatrix is EP320 across a 384-NPU optical supernode with co-designed INT8 kernels, MTP, ~50ms TPOT, batch 96; a generic 910C cluster, older CANN, or stricter latency won't reproduce it. Neutral recommendation: 7% central; 10% optimistic; 5% conservative; 2.5–5% strict-latency tail. At 1.504P INT8, 7% ⇒ ~1,423 tok/s/NPU on 37B-active.

Power: CloudMatrix ~559kW total IT [CR] vs NVL72 ~145kW is 3.86× — but 384 vs 72 devices (5.33×). Per-token IT energy: CloudMatrix 0.749 J/token; chip-only 910C 0.309; H800 0.378; H20 0.560. Chip-only excludes CloudMatrix's CPUs/optics/switches. Use 600W chip-only or 1.45kW attributable IT/NPU for CloudMatrix-like deployments; never multiply 910C power by 3.9.

Honest error bars (generic 910C serving): throughput/MFU ±35–40%; chip power ±15%; whole-system J/token ±40–50%. The uncertainty is software/topology, not the INT8 peak.

5. Who runs what (SPECULATION priors — no audited fleet percentages exist)

Share of frontier-model inference accelerator-hours in mainland China:

Evidence: DeepSeek V3/R1 disclosed on H800; V4 optimized for Ascend 950 (Reuters); own ASIC early-stage. Zhipu: GLM-5 developed with domestic chips incl. Ascend; broad adaptations. Moonshot: H800-class deployment guidance historically; Ascend K2.5 guide exists but no migration disclosure. Alibaba: sizable internal PPU deployments reported + evaluating 950PR. Countervailing: a May 2026 rental-market interview claims 90–95% Nvidia utilization and low domestic-card use in the commercial rental channel — true of that channel, not of hyperscaler/lab internal fleets.

6. Forward 12 months (through mid-2027)

• Ascend 950PR/950DT: 950DT on Huawei Cloud ~Aug 2026 (CR; likely preview) then Q4 GA. 950PR: ~RMB 50k (DDR) / 70k (HBM) per card; ~750k units planned for 2026, H2-weighted [CR]. Cheaper domestic prefill; 950DT improves decode via native FP8 + interconnect. Risks: yield, HBM supply, software maturity.
• Ascend 960: 2P FP8 / 4P FP4, 2× memory/bandwidth, Q4 2027 [DISCLOSED] — outside horizon.
• H200 to China: approvals under discussion for <200,000 chips total [CR]; no confirmed deliveries as of Jul 9. Scenario: ~200k delivered ⇒ 5–12% reduction in H800/H20 annual-lease pricing by mid-2027 [SPECULATION].
• DeepSeek inference ASIC: early-stage [CR]; base case = no material serving capacity by mid-2027.

7. Recommended calculator presets (annual-commit / owned-fleet marginal serving)

Public-cloud sticker sanity checks (separate mode): H800 Tencent HCC ~$10.6/h; H20 Tencent $3.16–6.25/h; H20 Alibaba RDS $7.40–10.62/h; 910B Tianyi $3.72–5.65/h; 910C public retail not reproducible.

Other owned-TCO defaults: electricity ¥0.38/kWh (0.35–0.45), subsidized ¥0.30 (0.25–0.36); PUE 1.20 (1.12–1.30); facility $7.8/W (6.5–9.5); accelerator depreciation 4 yr (3–5); facility 15 yr (14–20).

Highest-value calculator improvement: a rental-class selector — (1) IDC/bare-metal annual, (2) enterprise private-cloud procurement, (3) hyperscaler monthly commit, (4) hyperscaler on-demand/elastic, (5) spot/subsidized. Without it, a "China H20 rate" can differ by >10× even though every quoted instance contains an H20.