← Frontier Inference Margins · all research reports
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].
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.
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.
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].
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.
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%.
| Provider and product | Published rate | Per card-hour | Notes |
|---|---|---|---|
| Tencent TI-ONE H800 HCC, 8-card monthly | ¥418,988–424,235/mo [DISCLOSED] | ¥71.74–72.64 / $10.55–10.68 | Doc updated 2026-06-03 |
| Tencent PNV6 H20 monthly (1 card) | ¥15,676/mo [DISCLOSED] | ¥21.47 / $3.16 | Lowest clear Tencent monthly H20 |
| Tencent HCC-PNV6 H20, 8-card monthly | ¥174,579/mo [DISCLOSED] | ¥29.89 / $4.39 | HCC fabric, big host bundle |
| Tencent PNV6 H20 on-demand | ¥30.48/card-hr [DISCLOSED] | $4.48 | CPU-heavy variant ¥34.25/$5.03 |
| Tencent HCC-PNV6 H20 on-demand | ¥212.63–340/8-card node-hr [DISCLOSED] | $3.91–6.25 | Bundle spread |
| Alibaba RDS Custom AI H20-96G | ¥50.32/card-hr [DISCLOSED] | $7.40 | RDS product, not plain ECS |
| Alibaba RDS Custom AI H20-141G | ¥72.24/card-hr [DISCLOSED] | $10.62 | |
| Huawei Cloud 910C private-cloud procurement | ¥8,500–11,200/card-month [CREDIBLY REPORTED public award] | ¥11.64–15.34 / $1.71–2.25 | Huatai Securities framework award (avg 192 one-card instances, 1 yr); Huawei Cloud won at ¥8,500. Best reproducible 910C price. |
| China Telecom Tianyi 910B monthly | ¥18,454/card-mo [DISCLOSED] | ¥25.28 / $3.72 | 910B public-cloud proxy |
| China Telecom Tianyi 910B on-demand | ¥38.45/card-hr [DISCLOSED] | $5.65 |
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).
| Hardware | Listing | Per card-hour | Label |
|---|---|---|---|
| H800 SXM, Cloud Engine | ¥59,000/8-card node-mo | ¥10.10 / $1.48 | COMMUNITY ESTIMATE |
| H800 SXM, 2026 market interview | ¥63,000–64,000/node-mo | ¥10.79–10.96 / $1.59–1.61 | CREDIBLY REPORTED (also: ~30% rise post-Spring-Festival; annual contracts dominant) |
| H800 SXM, APEtops | ¥82,000/node-mo | ¥14.04 / $2.06 | CE (full CPUs/RAM/9×400G) |
| H800 SXM, Luchen Cloud | from ¥12.70/card-hr | from $1.87 | CE (hourly retail) |
| H20 96GB, Cloud Engine | ¥30,000/node-mo | ¥5.14 / $0.76 | CE (lowest transparent) |
| H20 141GB, Cloud Engine | ¥40,000/node-mo | ¥6.85 / $1.01 | CE |
| H20 96GB, APEtops | ¥38,000/node-mo | ¥6.51 / $0.96 | CE |
| Ascend 910B2 | ¥15,000–22,000/node-mo | ¥2.57–3.77 / $0.38–0.55 | CE — 910B only, never transfer to 910C |
Rental values to encode (annual-commit/bare-metal/private-cloud):
| Accelerator | Central | Range |
|---|---|---|
| H800 SXM | $1.75/h | $1.47–2.06 |
| H20 96/141GB | $1.00/h | $0.76–1.12 |
| Ascend 910C | $1.95/h | $1.71–2.25 |
| Ascend 910B | $0.47/h | $0.38–0.55 |
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.
| Accelerator | Dense 8-bit | HBM | BW | Power | Interconnect |
|---|---|---|---|---|---|
| H800 SXM | 1.979 PF dense FP8 [DISCLOSED] | 80GB | 3.35TB/s | 700W max | 400GB/s NVLink |
| H20 96GB | 296 TF dense FP8 [DISCLOSED by operator] | 96GB | 4.0TB/s | 400W nominal (~460W max CR) | 900GB/s NVLink |
| Ascend 910C | 1.504 POPS INT8 [DISCLOSED]; 752 TF FP16 | 128GB | 3.2TB/s | 600W central (550–690 CE) | 784GB/s D2D; no native FP8 |
| Ascend 950DT | 1.0 PF FP8 / 2.0 PF FP4 [DISCLOSED] | 144GB | 4.0TB/s | ~650W (550–800 SPECULATION) | 2TB/s aggregate |
| Accelerator | Evidence | Use |
|---|---|---|
| H800 SXM | Exporter listings ~$29–34k installed; higher domestic quotes ~$44–47k [CE] — banned finite stock, provenance-dominated | $40k central; $30–48k |
| H20 | Reuters $10–12k/chip [CR]; full-node quotes imply ~$17.8k (96G) / ~$22k (141G) installed [CE] | $11k module / $20k installed central; $18–22k |
| Ascend 910C | CloudMatrix 384 ≈ RMB 60M [CR] ⇒ RMB 156,250 = $22,965/NPU at 2026-07-09 FX | $23k central; $22–29k (RMB 180–200k figure = older market estimate) |
| Ascend 950DT | No price; 950PR RMB 50k (DDR) / 70k (HBM) [CR] are PR prices | $18k installed central; $14–25k for mid-2027 modeling |
| Deployment | Condition | Per-chip | Label |
|---|---|---|---|
| DeepSeek H800 production | V3/R1 real traffic | 1,850 out tok/s/GPU | DISCLOSED→derived (14.8k/node) |
| DeepSeek H800 production prefill | mixed traffic incl. disk-KV hits | 9,212 in tok/s/GPU | DISCLOSED→derived |
| H800 optimized profile | R1, batch 128, ~50ms TPOT | 2,325 out tok/s/GPU | DISCLOSED (CloudMatrix comparison) |
| Ant/SGLang H20 Base | R1, TPOT <70ms, batch 48 + MTP | 714 out tok/s/GPU | DISCLOSED operator production |
| Ant/SGLang H20 Pro | TPOT <50ms, batch 32 | 675 out tok/s/GPU | DISCLOSED — best neutral H20 anchor |
| Ant/SGLang H20 Max | TPOT <30ms, batch 12 | 423 out tok/s/GPU | DISCLOSED — latency-sensitive preset |
| CloudMatrix 910C default prefill | R1 INT8, 4K input | 5,655 in tok/s/NPU | DISCLOSED (default EPLB) |
| CloudMatrix 910C idealized prefill | "perfect EPLB" | 6,688 in tok/s/NPU | DISCLOSED (idealized) |
| CloudMatrix 910C decode | R1 INT8, batch 96, 49.4ms TPOT, MTP@70% | 1,943 out tok/s/NPU | DISCLOSED vendor/operator — upper anchor |
| CloudMatrix 910C strict latency | 24.6ms / 14.9ms TPOT | 974 / 538 | DISCLOSED — latency slope |
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.
Effective decode MFU (37B-active convention, dense 8-bit peak denominator):
| Anchor | Derived MFU |
|---|---|
| H800 DeepSeek production 1,850 | 6.92% |
| H800 optimized 2,325 | 8.69% |
| H20 Base 714 | 17.85% |
| H20 Pro 675 | 16.88% |
| H20 Max 423 | 10.58% |
| 910C CloudMatrix 1,943 | 9.56% |
| 910C ~25ms TPOT 974 | 4.79% |
| 910C ~15ms TPOT 538 | 2.65% |
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.
Share of frontier-model inference accelerator-hours in mainland China:
| Operator | H800 | H20 | Ascend 910C/950 | Other domestic | Confidence |
|---|---|---|---|---|---|
| DeepSeek | 35% (25–40) | 10% (5–15) | 50% (45–65) | 5% | Medium-low |
| Zhipu / Z.ai | 20% (10–30) | 10% (5–20) | 50% (40–60) | 20% (10–35) | Low |
| Moonshot / Kimi | 55% (45–65) | 30% (20–35) | 10% (5–20) | 5% | Low |
| MiniMax | 35% | 25% | 20% | 20% | Very low |
| Alibaba / Qwen | 20% | 15% | 10% | 55% (T-Head PPU) | Low-medium |
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.
| Accelerator | Rent central (range) | Installed capex | Effective decode MFU | Power | Justification |
|---|---|---|---|---|---|
| H800 SXM | $1.75/h ($1.47–2.06) | $40k ($30–48k) | 7.0% (6–9%) | 700W | IDC quotes surround DeepSeek's $2/h; 7% reproduces 1,850 tok/s production |
| H20 96GB | $1.00/h ($0.76–1.12) | $20k installed ($18–22k; $11k module) | 17% (11–18%) | 400W (to 460) | 17% ⇒ ~680 tok/s, matching Ant's 675–714 production tiers |
| Ascend 910C | $1.95/h ($1.71–2.25) | $23k ($22–29k) | 7% (5–10%) | 600W chip-only or 1.45kW CloudMatrix-attributable | Huatai procurement midpoint; MFU between "60% of H100" and Huawei's optimized 1,943 |
| Ascend 950DT (mid-2027 projection) | $1.50/h ($0.90–2.40) | $18k ($14–25k) | 15% (10–20%) | ~650W | Native FP8, 144GB/4TB/s, decode-oriented; no independent data yet |
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.