Model licences, speed commitments and inference margins

1. Read the exact model licence

Three published licences illustrate why downloadable weights do not imply identical commercial terms. These are selected conditions, not a complete licence review.

The two bespoke licences define Model as a Service as giving a third party access to inference or fine-tuning in a way that lets that third party exercise meaningful control over inputs, parameters or training data. They exclude end-user products whose model capabilities are solely embedded within specific features or harnesses, and mere relaying of requests to models hosted by others. Their Section 2 revenue thresholds cover the business and its affiliates, not only the named model's sales. Separate contracts may add conditions. This summary is not legal advice, and the K3 reading does not establish terms for K3.1 Pro.

2. Name the margin denominator

Let R be hypothetical token revenue per GPU-hour at a defined reference workload and price; U the fraction of allocated hours sold and served at that reference load; q an independent price multiplier; s a model-provider share of that revenue; and c the included cost per allocated GPU-hour, excluding the share. Revenue before the share is G = R × U × q.

This is a linear scenario with constant workload, billing mix and cost. A benchmark does not establish that its reference point is full capacity or that customers will fill it. U is not a device busy percentage; a rate already averaged over allocated hours must not be discounted for the same idle time again. In particular, v1.1’s R already averages over all allocated hours: when reusing that measured rate for the same window, set U = 100%.

The contribution G(1 − s) − c can be expressed against two different bases:

These are analytical ratios under the stated cost boundary, not a decision about accounting presentation. Include hardware, power, networking and other relevant service costs consistently; financing and taxes need separate treatment where excluded.

Hypothetical example: R = US$20, c = US$4, q = 1 and s = 30%. The 30% is chosen to illustrate the formula and is not attributed to any model licence. Break-even U∗ is 28.6%. At twice that load, 57.1%, the margin is 35% of before-share revenue, or 50% of retained revenue. Achieving 50% on the first basis takes 3.5U∗ = 100% in this example.

In general, a before-share margin target m requires U = c/[R × q × (1 − s − m)], with a positive denominator. If the required U exceeds 100%, the target is unattainable within this reference-load scenario. Revenue shares calculated on another contractual base require a different model.

3. Speed and cost must describe the same service

SemiAnalysis’s InferenceX GLM-5.3/B300 page reports the following frontier for its AgentX agentic coding workload, using Dynamo SGLang in FP4. The run window is 12–22 October 2026; the page was retrieved on 28 October. These are the tester's published measurements, not an independent reproduction. [4]

At the 200-token/s target the published aggregate rate is 46.2% of the rate at 100. That comparison describes this workload and tested configurations. Each point selects the best measured engine and precision for that target; the frontier is a dated snapshot and can change. Total tokens are not all output tokens, and an interactivity target does not by itself guarantee first-token latency, reliability or application quality.

The same source's cost tiers must also retain their labels:

These SemiAnalysis AI Cloud total cost of ownership (TCO) model tiers were retrieved on 28 October 2026. They are model inputs published by their author, not executable on-demand quotes or contracted service costs. [4–6]

For two systems A and B with the same billable mix, price, revenue share and service requirements, a conditional comparison is:

Here T is comparable throughput on the same system boundary. A cheaper GPU-hour can still require more paid load if throughput is lower. Different precision, quality, cache behaviour or test windows can invalidate the shortcut. This note makes no chip ranking.

4. Cache billing changes the value of a token stream

DeepSeek's official Flash price page maps the service to V4.1-Flash-1017. At peak rates, uncached input, cached input and output cost US$0.30 / US$0.006 / US$1.20 per million tokens, respectively. Off-peak rates are half those amounts. Peak periods are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. Retrieved 28 October 2026. [7]

This is DeepSeek’s own API tariff, used here to illustrate cache-differentiated billing; it does not establish another host’s selling price.

For a hypothetical fixed stream of one million input tokens plus 100,000 output tokens, using those peak prices:

At 90%, input charge is 0.1 × 0.30 + 0.9 × 0.006. The token stream is invented; the rates are the dated public observation. The example isolates token charges and assumes no other fees or credits. A hardware prefix-cache statistic cannot automatically be substituted for the billing counter. Higher cache reuse may free capacity, but the effect on margin also depends on cost and whether that capacity can be sold.

5. A proposed operating discipline

Before committing to a service price, record the exact licence and any separate agreement, benchmark the offered response-time requirement, join billable usage to the full allocated-capacity window, and compare the resulting revenue with costs on the same boundary. State the margin denominator beside every percentage.

The practical value of a benchmark is to expose which assumption needs testing next. Licence eligibility, an hourly cost tier and a token-rate headline each answer only part of the commercial question.

Sources and scope

Primary sources below were retrieved on 28 October 2026 (Asia/Singapore). Licence and price summaries are dated observations; performance figures are attributed test reports. Equations, the 46.2% comparison and both hypothetical examples are our calculations. The final section is a proposed method, not an empirical result.

1. DeepSeek V4.1 Flash-1017 licence
2. Z.AI GLM-5.3 licence
3. Moonshot AI Kimi K3 licence
4. InferenceX GLM-5.3 on B300
5. InferenceX GLM-5.3 on MI355X — cost tiers only
6. InferenceX GB300 NVL72 cost tiers
7. DeepSeek Models and Pricing

Related methodology: From token throughput to billable revenue, v1.1.