← Frontier Inference Margins · all research reports
Run 2026-07-09 · 46m58s · model
gpt-5-6-pro(Pro-verified, Ashitaorbis project) · conversation Verdict: Kimi K2.7 Code output-token marginal serving gross margin ≈ 81% central, 80% CI ≈ 55–91% at the $4/M list price (central COGS $0.75/M output). Note: the first Moonshot dive completed but its response expired from the MCP retention window before collection; this is the re-run. Archived verbatim below (DOM-extracted; math notation may render with minor duplication artifacts).
Evidence labels: DISCLOSED = first-party model/configuration/pricing documents, regulatory filings, or Moonshot-authored systems work; CREDIBLY REPORTED = reputable reporting but not audited disclosure; COMMUNITY ESTIMATE = reproducible external benchmark or public market proxy; SPECULATION = my modeled inference.
Moonshot's lineup as of July 9, 2026 is unusual: the newest model is a coding-specialist flagship, while the preceding model remains the recommended general-purpose workhorse.
| Role | Model | Architecture | Precision & acceleration | Serving-cost implication |
|---|---|---|---|---|
| Coding flagship | Kimi K2.7 Code | DISCLOSED — 1.0T total; 32B active; 61 layers; 384 routed experts; 8 selected + 1 shared; 256K context; 400M MoonViT encoder. Thinking-only; Moonshot recommends K2.6 for general use. K2.7 page · model list | DISCLOSED — native selective, weight-only INT4 (4-bit groupwise, group 32; attention/shared-expert/several MLP/head/vision modules excluded). num_nextn_predict_layers: 0 — no native MTP. config |
DISCLOSED — 595GB checkpoint. COMMUNITY ESTIMATE — vLLM verified minimum 8×H200. INT4 cuts weight residency/bandwidth; the 1T expert pool still has to live somewhere. vLLM recipe |
| Workhorse | Kimi K2.6 | DISCLOSED — same 1.0T/32B backbone, text+image+video, thinking and non-thinking. K2.6 card | DISCLOSED — native INT4, ~595GB. | Economically close enough to K2.7 that K2.5/K2.6 benchmarks inform K2.7. |
Qualifications: hosted precision is not disclosed (released-weights precision ≠ Moonshot's production path — any exact hosted-precision claim is SPECULATION). DISCLOSED — K2.7 uses ~30% fewer thinking tokens than K2.6 on reported evals — reduces cost per completed task, not cost per token.
Bottom line on INT4: economically significant because it makes a single high-memory 8-GPU replica feasible and reduces decode weight traffic; it does not remove the 1T fixed-memory footprint, routing communication, or long-context KV requirements.
Strongest evidence: Mooncake paper, USENIX FAST '25:
Owned vs rented: undisclosed; historical Volcengine relationship proves at least some rented capacity.
| GPU | 8-GPU monthly quote | Per chip-hour |
|---|---|---|
| H20 141GB | ¥40,000 — DISCLOSED vendor quote | ¥6.85 / $1.01 — COMMUNITY ESTIMATE |
| H800 | ¥59,000 | ¥10.10 / $1.48 |
| H100 | ¥68,000 | ¥11.64 / $1.71 |
| H200 | ¥80,000 | ¥13.70 / $2.01 |
Vendor quote — retail proxies, not Moonshot's basis. Cloud credits should still be valued at opportunity cost.
| Model/tier | Cached input | Uncached input | Output | Other |
|---|---|---|---|---|
| K2.7 Code standard | $0.19 | $0.95 | $4.00 | 262,144 context — all DISCLOSED |
| K2.7 Code HighSpeed | $0.38 | $1.90 | $8.00 | ~180 tok/s (to 260 short-context); "5–6× standard speed" — DISCLOSED |
| K2.7 standard Batch | $0.114 | $0.57 | $2.40 | Batch = 60% of standard — DISCLOSED, derived |
| K2.6 | $0.16 | $0.95 | $4.00 | 262,144 context — DISCLOSED |
K2.7 pricing · Batch · K2.6 pricing. Cache discount is 80% ($0.19 vs $0.95). Old kimi-k2-turbo-preview/thinking-turbo SKUs discontinued May 25, 2026 — DISCLOSED; the current speed tier is HighSpeed.
| Plan | Monthly | Annual-equivalent monthly |
|---|---|---|
| Moderato | $19 | $15 |
| Allegretto | $39 | $31 |
| Allegro | $99 | $79 |
| Vivace | $199 | $159 |
(All DISCLOSED.) Free Adagio + tiers include ~6/60/150/360/720 agent-credit equivalents + Kimi Code multipliers; Moonshot warns equivalents are approximate. Documentation is internally inconsistent on credit pools — subscription revenue cannot be cleanly divided into a realized token price. OK Computer is an Agent mode (Sep 26, 2025), not a model/API SKU. Membership pricing
K2.7 is thinking-only; the Alibaba channel states reasoning/CoT tokens bill as output — $4/M applies to reasoning tokens. The 30% thinking-token reduction improves price-per-task, not per-token margin.
Sophisticated serving-systems evidence, but no serving COGS/token, no inference gross margin, no GPU-hour basis, no audited product margin. Mooncake's real trace: ~12K-token average conversation inputs, ~40% prefix-cache ratio — DISCLOSED (predates K2.7).
128×H200 (32 prefill + 96 decode GPUs), 2,000-in/100-out requests, decode batch 480, 224K prefill / 288K decode tok/s aggregate, $2.30/H200-hr — all COMMUNITY ESTIMATE:
Limitations: decode-GPUs-only, original K2 not K2.7, short-context run, excludes reserves/CPU/network/storage/KV-infrastructure/ops, NVIDIA DGX Cloud infrastructure. An excellent decode accelerator floor, not a Moonshot COGS estimate.
8K-in/1K-out on single 8-GPU nodes; TCO $1.41/H200-hr, $1.95/B200-hr:
| Per-user speed | H200 INT4 | B200 NVFP4 |
|---|---|---|
| 32 tok/s | $0.413/M out | $0.140/M |
| 40 tok/s | $0.453/M | $0.154/M |
| 50 tok/s | $0.511/M | $0.177/M |
| 70 tok/s | $0.660/M | $0.244/M |
| 90 tok/s | $0.996/M | $0.347/M |
Facts demonstrated: per-user speed/batching moves cost >2× on identical hardware; hardware+precision moves it ~2.7–3.0× at matched interactivity; a Western B200/NVFP4 deployment can be far cheaper than a China H800/H20-style one on the identical model.
| Provider | Uncached in | Cached in | Output |
|---|---|---|---|
| Together | $0.95 | $0.19 | $4.00 |
| Fireworks | $0.95 | $0.19 | $4.00 |
| Novita | $0.95 | $0.19 | $4.00 |
| Baseten | $0.95 | $0.16 | $4.00 |
| DeepInfra | $0.74 | $0.15 | $3.50 (priority $1.11/$0.225/$5.25) |
| Groq | — | — | (only old K2-instruct at $1.00/$0.50/$3.00; not comparable) |
Reseller-ceiling interpretation (SPECULATION): DeepInfra's $3.50/M output is a weak long-run market-price ceiling, not a cost floor. The identical $4 pricing across four providers may reflect reference pricing/wholesale pass-through rather than four independent profitable-cost observations.
No public inference spend, gross profit, credit balances, depreciation, or contribution margin.
K2.7 Code standard output tokens:
Start from the most transferable benchmark: $0.45–0.51/M accelerator cost for H200 INT4 at 40–50 tok/s/user on the same architecture (COMMUNITY ESTIMATE) + ~$0.20–0.30/M for lower fleet utilization, online headroom, redundancy, CPU/network/storage/KV infrastructure and possibly less favorable hardware (SPECULATION) → $0.65–0.80/M, rounded to $0.75. Lower endpoint $0.35 stays above the $0.213 short-context decode floor; upper endpoint $1.80 allows an older/fragmented A800/H800/H20 fleet, low utilization, latency reserves, long-context and multimodal overhead — still materially below DeepInfra's $3.50 selling price.
Unit serving economics (accelerators, power, hosts, network/storage/KV, headroom, control plane, direct ops); excludes training/R&D/sales/corporate. Output-token estimate, not blended — a blended margin needs Moonshot's I/O ratio, cache-hit ratio, tier mix, batch share, subscription utilization, and discounts. Do not apply 81% to HighSpeed: 2× price for 5–6× user speed likely costs substantially more than 2× per token.