Does one month of token sales cover the bill?

1. What the month costs

Add 3% of principal once per year for five years, 15% in total, added once, not compounded, then spread over 60 months. Rack 670 yuan, power 2,419 yuan, and labor 1,330 yuan are the same at every amount. Rack and labor do not shrink when the server is busy only 60% of the hours. Labor is listed on its own, apart from tax. Tax is outside this page.

Power scales with load: 7 kW × 60% × 24 hours × 30 days × 0.8 yuan = 2,419 yuan. 7 kW is an upper-bound estimate, not a meter reading. An idle server still draws power, so the real bill runs a little higher. Full power for 24 hours is 4,032 yuan, and that is the ceiling.

2. What the month sells

Sales depend on speed and price, not on whether the equipment is counted at 1.6 or 2.2 million yuan. The rows below are with acceleration on, 1K tokens in and 1K out. Zhipu is 64 concurrent requests. The MiniMax row in the table is also 64 per server. Price is half the published list, a month is 30 days, then folded by 60% of the hours.

Zhipu’s vendor package already includes speculative decoding. “Acceleration on” here means our production setting. On random chat it differs from the vendor package by 3.2%, inside the report’s roughly 9% retest band, so treat it as flat. The three reuse rows are a separate measurement. The 1.5× rows are a forecast on top of measured random chat. They are not measured, and they do not replace the reuse rows.

3. How long the equipment takes to recover

Each cell says whether this month’s bill is covered, then how many months the equipment takes. Months = amount to recover ÷ cash left after rack, power, and labor, rounded. At 1.6 million yuan, Zhipu random chat is 1,840,000 ÷ 48,221 ≈ 38 months.

This is a different sum from “pay the installment first, then split whatever is left.” Do not count both. The two 1.5× rows are not measured; if a later test misses them, drop them. The revisit row is a burst cap, not a pace the server holds every day. An October 2026 retest saw no hit on the over-pool revisit path. Treat that row as unchecked, and retest it before a contract.

Swipe sideways for all four amounts.

If a later test does hit 1.5×, Zhipu random chat would land on 25, 28, 31, 34 months from 1.6 to 2.2 million yuan. That is a forecast. The shared 32K document and the revisit burst are already measured. Do not replace them with this row.

4. How this page is calculated

Monthly sales

(input tokens/s × input price + output tokens/s × output price + hit tokens/s × hit price) × 86,400 × 30 ÷ 1,000,000, then × 60%, rounded to the yuan. Prices are yuan per million tokens.

Monthly bill

Amount to recover = principal × (1 + 3% × 5). Installment = that amount ÷ 60. Bill = installment + rack 670 + power + labor 1,330. Each step rounds to the yuan.

Months to recover

Amount to recover ÷ (sales − rack − power − labor), rounded to a month. No discounting, no interest. A month is 30 days.

Real chat

MiniMax real chat on one server at 64-way concurrency measured 1,532 output tokens/s. The report did not publish input tokens/s, so monthly sales assume input equals output, about 48% of random chat. At 16-way, speculative decoding only moved 627 to 785 tokens/s, up 25%. That figure is for the size of the gain. It is not scaled up to 64-way.

Price

List prices were checked on 2026-09-30. MiniMax marks a permanent half off the strikethrough price, and the amount actually charged is that half. This page does not halve it again. Zhipu publishes no such permanent half, so the half used here is a modeling assumption, not a published settlement price and not an agreed purchase price. DeepSeek has no such half either. This page does not put DeepSeek in the tables.

Where the speeds come from

• Zhipu random chat: 1,817 output tokens/s per server, and 2,102.5 input tokens/s that miss the cache. That is half of a two-server measurement.
• Reuse rows, also per server: 8K writes 1,398, hits 10,630.5, misses 1,181; 32K writes 694, hits 21,916.5, misses 2,435; revisit writes 362.5, hits 46,355, misses 0.
• MiniMax random chat: 3,198.5 output tokens/s per server. At 1K in / 1K out, input is taken as equal to output. Real chat measured 1,532 output tokens/s at 64-way on one server, and input is assumed equal to output.
• A same-configuration pair on Oct 1–2 read lower (Zhipu 3,503.5, MiniMax 6,134.2 tokens/s across two servers). This page still uses the higher compiled pair, halved per server, so leave some room.
• These test records have no official seal and no prepared-by, reviewed-by, or approved-by signature. The figures are only for this working model. They are not an acceptance record.

Glossary

token

A token is one piece of text the model reads or writes. Prices are per million tokens.

tok/s

Tokens per second. This page also says tokens/s.

Concurrency

How many conversations are served at once. 64-way means 64 at the same time.

MTP

Speculative decoding. The model drafts a short run of tokens, keeps the draft when it is right, and retries when it is wrong. Zhipu’s vendor package includes it.

Cache hit

The prompt was computed before and is still stored, so it barely uses compute. Only Zhipu’s reuse rows use the hit price.

Utilization

The share of hours the server is busy. This page uses 60%. Power follows it. Rack and labor do not.

This page is a working model. It is not an offer and not a quotation. The equipment amounts, Zhipu’s half price, and the two 1.5× rows are not contract prices. Binding terms are whatever a written contract later says.