← Frontier Inference Margins · all research reports
Run 2026-07-09 · 50m23s · model
gpt-5-6-pro(Pro-verified) · conversation Verdict: GLM-5.2 marginal serving gross margin ≈ 60% central, 80% CI ≈ 35–77% at direct API list price. Anchored by audited HK filings: FY2025 cloud/API gross margin 18.9% (H1 2025: −0.4%). Archived verbatim below (DOM-extracted; math notation may render with minor duplication artifacts).
Evidence cutoff: July 9, 2026. "Marginal serving cost" below means long-run incremental inference COGS: accelerator service, host/network/KV-cache resources, and directly attributable serving overhead. It excludes training, research, sales, general administration, and financing.
Confidence labels
DISCLOSED: first-party model documentation, provider price sheet, or audited/regulatory filing.
CREDIBLY REPORTED: Reuters or an independent/vendor benchmark with an identifiable methodology.
COMMUNITY ESTIMATE: a technically credible but unaudited self-report.
SPECULATION: my assumption, calculation, or inference.
| Model | Architecture | Precision evidence | Serving interpretation |
|---|---|---|---|
| GLM-5.2 | DISCLOSED — 744B total / 40B active (5.38% activation). Open release supports a 1M-token context window. | DISCLOSED — BF16 and FP8 open checkpoints. CREDIBLY REPORTED — Baseten created a custom NVFP4 checkpoint from the FP8 release and reported no material BFCL degradation. Zhipu's own production precision is not disclosed. | Frontier model and headline API product. The 40B active count understates memory/network burden because the full 744B MoE weight set must be resident or distributed. |
| GLM-4.7 — workhorse | DISCLOSED — 355B total / 32B active (9.01% activation). | DISCLOSED — BF16 and FP8 open checkpoints. First-party API serving precision not disclosed. | Z.ai's Coding Plan documentation recommends GLM-4.7 for routine work and GLM-5.2 for harder tasks — the best public candidate for the economical workhorse/routing tier. |
| GLM-4.7-Flash | DISCLOSED — 30B total / 3B active. | DISCLOSED — BF16 checkpoint. | International API price is free — acquisition/routing/ecosystem role. |
Sources: official GLM-5 repository, official GLM-4.5/4.7 repository, Z.ai Coding Plan overview, Baseten's GLM-5.2 serving write-up.
Two additional architecture claims matter for cost:
DISCLOSED, company claim — IndexShare reduces per-token attention FLOPs by 2.9× at a 1M-token context, compared with the model's non-IndexShare baseline.
DISCLOSED, company claim — multi-token prediction increased accepted-token length by up to 20%.
These are real cost levers, but neither directly states production dollars per token; realized benefit depends on context distribution, speculative-decoding acceptance, batch shape, and serving implementation.
Baseten's result is useful but should not be transferred to Zhipu's domestic fleet without adjustment: CREDIBLY REPORTED — more than 280 output tokens/s/user on Blackwell using NVFP4, with prefill/decode disaggregation, MTP, KV-aware routing, and prefix caching (disaggregation approximately doubled throughput on its observed workload mix). Evidence the model can be served economically on Western Blackwell — not evidence that Zhipu uses Blackwell.
DISCLOSED — Zhipu says GLM-5.2's online inference runs across multiple domestic platforms, explicitly naming Huawei Ascend, Alibaba T-Head, Moore Threads, Cambricon, Baidu Kunlunxin, MetaX, Hygon, Biren, and Iluvatar CoreX — described as production deployments supporting throughput, latency, and concurrency. Establishes domestic diversification; does not reveal traffic share by vendor or model. Zhipu GLM-5.2 release
For the closely related GLM-5 architecture:
DISCLOSED — one Huawei Atlas 800T A3 server (8× Ascend 910-series NPUs) can hold the ~750B model using W8A8 for attention/ordinary MLP and W4A8 for MoE experts.
DISCLOSED, company claim — Zhipu/Huawei say one domestic node can approach the throughput of two foreign-GPU nodes and lower long-context deployment cost by 50%. Vendor claims without disclosed workload/latency/price — not used directly in the central cost calculation. Zhipu GLM-5 technical report · Huawei GLM-5 deployment note · Atlas A3 specifications
No disclosed evidence ties Zhipu's first-party production GLM-5.2 service to a particular stock of H800s or H20s. The official repository contains H20-related examples (implementation/training examples, not fleet inventory). Baseten's Blackwell deployment and @_xjdr's GB300 run are independently useful serving benchmarks; neither establishes Zhipu ownership, leasing, or production use.
The audited procurement disclosures point toward a service-heavy cost basis:
DISCLOSED — 2025 capital expenditure was RMB74.7 million, down from RMB462.3 million in 2024.
DISCLOSED — in 2024 Zhipu primarily used equipment leases to acquire computing resources; in 2025 it primarily procured computing services, supplemented by equipment leases.
DISCLOSED — cloud-computing supplier agreements generally run for one to four years.
DISCLOSED — prepayments for computing services were RMB166.9 million at year-end 2024 and RMB134.0 million at June 30, 2025.
DISCLOSED — purchases from an anonymized "Supplier E," a Beijing cloud/IT subsidiary of a large Chinese internet company, were RMB394.8 million in 2024 and RMB123.5 million in H1 2025.
A rental/reserved-service model is therefore more appropriate than assuming Zhipu buys GPUs and depreciates them itself. It also makes accounting gross margin somewhat more informative about unit economics than for a provider carrying a large owned fleet — although contracts may still include minimum commitments and idle-capacity risk. HKEX FY2025 results · HKEX prospectus
These are not Zhipu's confidential rates — public CTyun rates useful for bounding the market. USD conversions use RMB6.9074/USD (March 31, 2026 reference).
| Accelerator | Public rate | USD equivalent | Confidence and interpretation |
|---|---|---|---|
| Ascend 910B, ordinary CTyun pool | RMB36.36–41.29/card-hour | $5.26–$5.98/card-hour | DISCLOSED by CTyun, retail/on-demand, not Zhipu contract pricing. |
| Eight-card Ascend server at the same rate | RMB290.9–330.3/server-hour | $42.1–$47.8/server-hour | Arithmetic; excludes host/network/storage/discount uncertainty. |
| Ascend 910B, special research pool | RMB9.60/card-hour | $1.39/card-hour | DISCLOSED by CTyun, probably not representative of unrestricted commercial capacity. |
| Nvidia H20 96GB | RMB17.05/card-hour | $2.47/card-hour | DISCLOSED by CTyun public research-pool rate. |
| Nvidia H800 PCIe | RMB28.56/card-hour | $4.13/card-hour | DISCLOSED by CTyun public research-pool rate. |
| Nvidia H800 NVLink | RMB34.00/card-hour | $4.92/card-hour | DISCLOSED by CTyun public research-pool rate. |
Sources: CTyun ordinary Ascend pricing · CTyun research-pool pricing
Best representation of Zhipu's undisclosed basis: SPECULATION — $2.0–$4.8 per contracted accelerator-card-hour, central prior $3.2/card-hour; 65–85% useful serving occupancy, central 75%. Not performance-normalized units — used only as a sanity check; the final token-cost model is anchored more heavily to financial margins, reseller prices, and measured serving economics.
Utilization evidence is directional: DISCLOSED/CREDIBLY REPORTED — Zhipu described a compute shortage beginning February 2026, demand exceeding supply, API requests rising by more than 400%, and paid-token consumption increasing.
All DISCLOSED — USD per 1M tokens (official pricing). Cache-storage charges temporarily free; cached-token processing billed at the cached-input rate.
| Model | Uncached input | Cached input | Output |
|---|---|---|---|
| GLM-5.2 | $1.40 | $0.26 | $4.40 |
| GLM-5.1 | $1.40 | $0.26 | $4.40 |
| GLM-5 | $1.00 | $0.20 | $3.20 |
| GLM-5-Turbo | $1.20 | $0.24 | $4.00 |
| GLM-4.7 | $0.60 | $0.11 | $2.20 |
| GLM-4.7-FlashX | $0.07 | $0.01 | $0.40 |
| GLM-4.7-Flash | Free | Free | Free |
| GLM-4.6 / GLM-4.5 | $0.60 | $0.11 | $2.20 |
| GLM-4.5-Air | $0.20 | $0.03 | $1.10 |
Two economically important observations: GLM-5.2 output is priced at 3.14× uncached input and 16.9× cached input — decode throughput is likely the largest direct determinant of gross margin. The cached-input discount is 81.4% — Z.ai passes a large portion of prefix-cache savings to customers.
No fixed token allowance; enforceable limits are prompt windows, weekly caps, and model-specific quota multipliers. All DISCLOSED (overview · transition):
| Plan | Monthly | Prompts per 5 hours | Prompts per week | Quarterly | Annual |
|---|---|---|---|---|---|
| Lite | $18 | 80 | 400 | $48.60 | $172.80 |
| Pro | $72 | 400 | 2,000 | $194.40 | $691.20 |
| Max | $160 | 1,600 | 8,000 | $432.00 | $1,536.00 |
Z.ai estimates one coding prompt ≈ 15–20 underlying model calls. GLM-5.2 and GLM-5-Turbo ordinarily consume quota at 3× during the 14:00–18:00 UTC+8 peak and 2× off-peak; a promotion reduces off-peak consumption to 1× through September 2026 — a powerful tool for controlling realized compute consumption without changing advertised prompt caps.
Official documentation makes two claims that do not reconcile cleanly:
DISCLOSED, company claim — monthly API-equivalent quota is approximately 15–30× the subscription fee.
DISCLOSED, company claim — the plan provides "tens of billions" of tokens at approximately 1% of ordinary API pricing (implying ~100× list-value leverage).
At the more conservative 15–30× claim, cash realization is 3.3–6.7% of list value on all three plans. Using SPECULATION — 8 input tokens per output token and the COMMUNITY ESTIMATE — 41% cache-hit rate (from @_xjdr's event), GLM-5.2 list ≈ $1.318 per million total tokens, implying (value-equivalent, NOT allowances): Lite ≈ 205–410M, Pro ≈ 820M–1.64B, Max ≈ 1.82–3.64B total tokens/month.
The plan therefore has obvious potential for usage leakage. Using the central cost model in Section 5, serving cost ≈ 39.8% of API list value; a user genuinely consuming 15–30× the fee in GLM-5.2 API value would incur ≈6.0–11.9× the subscription fee in serving cost. That does not mean the plan necessarily loses money overall — its economics must depend on user breakage and weekly caps; the 2×/3× GLM-5.2 quota multipliers; off-peak utilization of otherwise idle capacity; routing ordinary tasks to GLM-4.7 or cheaper variants; shorter actual prompts; and "API-equivalent value" as a marketing construct.
Evidence Zhipu has already tightened realization: DISCLOSED — legacy plans without weekly caps were phased out; Coding Plan prices rose 30% in February 2026; the introductory discount was removed; ~242,000 paying Coding Plan developers.
Status correction: DISCLOSED/CREDIBLY REPORTED — Zhipu listed in Hong Kong on January 8, 2026. As of July 9 it was pursuing a Shanghai STAR Market dual listing; no accepted STAR prospectus found. The primary financial source is therefore its HKEX prospectus and FY2025 results. Reuters on the dual-listing plan
| FY2025 item | Amount | Share / margin |
|---|---|---|
| Total revenue | RMB724.3m | 100% |
| Cloud-based model/API services | RMB190.4m | 26.3% |
| On-premise solutions | RMB534.0m | 73.7% |
| Enterprise agents | RMB165.7m | 22.9% of revenue |
| Enterprise models | RMB365.7m | 50.5% |
| Technology services | RMB2.5m | 0.35% |
| Cloud/API gross profit | RMB36.0m | 18.9% gross margin |
| Total cost of revenue | RMB427.7m | Company-wide |
| Net loss | RMB4.718bn | Not relevant to marginal serving margin |
The most informative sequence:
DISCLOSED — H1 2025 cloud/API gross margin was −0.4%.
DISCLOSED — FY2025 cloud/API gross margin recovered to 18.9%.
DISCLOSED — management attributed the improvement to inference-efficiency gains, economies of scale/declining marginal cost, price increases, and programming-subscription growth. The prospectus says H1 2025 was depressed by price reductions undertaken to gain market share.
This 18.9% is not the requested marginal GLM-5.2 margin — it includes reserved/idle capacity, personnel, support, promotions, model mix, subscriptions, older-model traffic. But because Zhipu increasingly purchases compute as a service, it is a more relevant anchor than consolidated company gross margin.
Zhipu reported API call prices rose 83% relative to end-2025, while requests rose more than 400%. Holding FY2025 unit cost and mix constant: 1 − (1−18.9%)/1.83 = 55.7%. SPECULATION — a price-only counterfactual puts cloud/API gross margin at 55.7% — a valuable independent anchor for a central marginal estimate in the high-50s or low-60s.
| Provider | Uncached input | Cached input | Output | Caveat |
|---|---|---|---|---|
| Baseten | $1.40 | $0.26 | $4.40 | Same as Z.ai list; optimized Blackwell/NVFP4 stack. |
| Fireworks Standard | $1.40 | $0.14 | $4.40 | Lower cache price. |
| Fireworks Batch | $0.70 | — | $2.20 | 50% discount, asynchronous scheduling. |
| Together | $1.40 | $0.26 | $4.40 | FP4 endpoint, 262K context. |
| DeepInfra Standard | $0.93 | $0.18 | $3.00 | Lowest public real-time price found. |
Sources: Baseten pricing · Fireworks pricing · Together GLM-5.2 · DeepInfra GLM-5.2
These are conditional cost ceilings: if DeepInfra's standard endpoint has positive contribution margin, its direct serving cost must be below $3.00/M output. Assuming a 10–30% reseller margin: implied output-cost ceiling $2.10–$2.70/M; Fireworks Batch implies $1.54–$1.98/M. For the Section 5 representative workload, the DeepInfra retail bill is $7.98 vs $11.86 at Z.ai list — a conditional lower-bound margin of ~32.7% at Z.ai's list price before granting DeepInfra any profit.
CREDIBLY REPORTED — Baseten exceeded 280 output tokens/s/user on Blackwell/NVFP4.
COMMUNITY ESTIMATE — @_xjdr reported ~232 output tokens/s/GPU and a 41% cache-hit rate during a GB300 NVL72 free-use event. Post
CREDIBLY REPORTED — SemiAnalysis' interpolated GLM-5.1 benchmark produces ~$0.41–$4.14/M output tokens across B200/GB300 configurations and 43–91 tok/s/user latency targets. InferenceX comparison — the very wide range shows how interactivity, batch, and hardware generation dominate a single "cost per token."
GLM-5.1 is not GLM-5.2, and a Western FP4 deployment is not Zhipu's domestic service. Still, the benchmark supports an output-cost prior centered around the low-$2 range.
SPECULATION — central marginal serving gross margin at direct GLM-5.2 API list: 60%.
SPECULATION — subjective 80%-confidence interval: 35%–77%.
8 billed input tokens per output token (SPECULATION); 41% input-cache hit rate (COMMUNITY ESTIMATE); long-run incremental serving cost.
| Token class | Z.ai list price | Central serving cost | Wider stress range | Central component margin |
|---|---|---|---|---|
| Uncached input | $1.40/M | $0.50/M | $0.22–$0.95/M | 64% |
| Cached input | $0.26/M | $0.08/M | $0.03–$0.18/M | 69% |
| Output | $4.40/M | $2.10/M | $1.20–$3.20/M | 52% |
Revenue = 4.40 + 8(0.59×1.40 + 0.41×0.26) = $11.8608. Cost = 2.10 + 8(0.59×0.50 + 0.41×0.08) = $4.7224. Gross margin = 1 − 4.7224/11.8608 = 60.2%.
Equivalently: blended list revenue $1.318/M total tokens; blended marginal serving cost $0.525/M; blended gross profit $0.793/M.
The central output cost of $2.10/M is deliberately close to the low end of the DeepInfra reseller-margin inference and the high end of a profitable Fireworks Batch inference. The overall 60% result is close to — but modestly above — the 55.7% price-only financial counterfactual, reflecting better utilization and inference optimization after FY2025.
Holding component costs fixed: 2:1 input:output → ~56%; 8:1 → ~60%; 20:1 → ~62%. The wider component stress ranges mechanically generate ~30–80% margin; the stated 80% interval narrows to 35–77% because the most extreme component assumptions are unlikely to co-occur.
Production decode throughput and latency target: ±15–20 margin points. GLM-5.2 output pricing is high, but decode becomes expensive with highly interactive latency, low MTP acceptance, or poor batching.
Contracted compute price and actual chip mix: ±10–15 points. Public Ascend/H20/H800 card-hour prices span more than 4× before performance correction; Zhipu's confidential 1–4-year service rates and traffic shares are unknown.
Useful occupancy and treatment of reserved idle capacity: ±7–12 points.
First-party serving precision and fallback behavior: ±6–10 points. Clean W4/NVFP4 expert paths are far cheaper than frequent FP8 fallback, quality-based rerouting, dual execution, or low-batch expert imbalance.
Workload distribution: ±5–8 points.
Non-accelerator serving overhead and reseller-evidence reliability: ±3–6 points.
Bottom line: for direct, undiscounted GLM-5.2 API traffic, 60% marginal serving gross margin is the central estimate, with 35–77% as the defensible 80% range. The most likely miss is not parameter count; it is assumed production output throughput or Zhipu's confidential compute-service price. Coding Plan margin can be dramatically lower — and negative for a fully saturating GLM-5.2-heavy user — because realized revenue is only a small fraction of nominal API-list value.
DISCLOSED — GLM-5.2 is a 744B-total, 40B-active MoE with BF16 and FP8 open checkpoints and a 1M-token context window. Repository
DISCLOSED — international GLM-5.2 list price is $1.40/M uncached input, $0.26/M cached input, and $4.40/M output. Z.ai pricing
DISCLOSED — Zhipu says first-party online inference uses multiple domestic hardware platforms, including Huawei Ascend, Cambricon, Kunlunxin, Moore Threads, MetaX, Hygon, Biren, T-Head, and Iluvatar. GLM-5.2 release
DISCLOSED — Zhipu moved toward primarily purchasing computing services in 2025; capex fell from RMB462.3m to RMB74.7m. FY2025 results
DISCLOSED — FY2025 cloud/API revenue was RMB190.4m and cloud/API gross margin was 18.9%, after a −0.4% margin in H1 2025. FY2025 results · prospectus
DISCLOSED — third parties sell GLM-5.2 below Z.ai list: DeepInfra at $0.93/$0.18/$3.00 and Fireworks Batch at $0.70 input/$2.20 output — conditional ceilings on profitable third-party serving cost.
Production chip mix and traffic allocation. Many supported domestic platforms; no disclosure of what fraction of GLM-5.2 tokens runs on each.
Confidential compute-service price. Filings disclose procurement form, suppliers, prepayments, aggregate spending — no card-hours, node-hours, token throughput, or unit rate.
Actual production precision. Open FP8 checkpoints and domestic W8A8/W4A8 recipes do not establish first-party serving precision.
Production throughput at a defined service-level objective. No output tokens/s/card, TTFT, batch size, context distribution, or MTP acceptance on Zhipu's own fleet.
Realized customer mix. Cache-hit rate, input/output ratio, reasoning length, GLM-4.7 routing, enterprise discounts, API vs Coding Plan division — undisclosed.
Subscription breakage and quota enforcement. The 15–30× API-value claim, "1% of API pricing" claim, prompt caps, and model multipliers do not reveal how much compute the median or top-decile subscriber actually consumes.