Benchmark Report

Shared Reserved Inference

A third way to serve open models — one GPU split into equal, guaranteed lanes for a bounded cohort, billed by the hour instead of the token. Here's how it performed under 32 real users running DeepSeek‑V4‑Flash‑0731 for an hour.

Live run2026-08-18, 13:41–14:41 UTCDeepSeek-V4-Flash-0731 (full weights, 1M context)measured, not modeled

97.0%Prefix cache hit rate, real agentic load
32 / 32Concurrent users, each a guaranteed lane
7,562Requests served, zero failures
1.7–3.4×DeepSeek token value per user, off-peak → peak
01

A third way to serve a model

Hosted inference forces a bad choice. Per-token APIs once looked cheap and elastic — but rates now move without warning, providers throttle or retire models, there's no guaranteed throughput, and at real volume the metered bill quietly overtakes the cost of the GPU actually doing the work. The obvious fix — rent your own GPU— flips the problem: a dedicated node bills every hour whether you use it or not, which no individual or small team can keep busy enough to justify. So you're stuck between an API you don't control and hardware you can't afford to leave idle.

Shared Reserved Inference splits one GPU into equal, guaranteed lanes across a small, fixed cohort. A handful of users share the cost of one warm, well-utilized node — each paying a flat hourly fraction — while keeping the guarantees a shared API can't give.

 Per-token APIShared ReservedDedicated GPU
Who's on the GPUAnonymous shared poolA fixed cohort, equal lanesOne tenant, whole GPU
GuaranteeNone — best-effortGuaranteed lane + burstFull GPU
PricingPer token; rates shift, models retireFlat hourly, no surchargePer GPU-hour
Cost to userVolatile; overtakes GPU cost at volumeLow, fixed — a fraction of a nodeHigh — you pay for idle time
Warm cacheFragmented across replicasSticky — stays hotSticky

The result is dedicated-grade guarantees at shared-grade price: your lane is always there, the cache stays warm, and the bill doesn't move when the world gets busy.

02

The benchmark setup

We put 32 real users on a single reserved node for an hour, each with their own API key, running coding-agent workloads through a standard CLI harness.

ModelDeepSeek-V4-Flash-0731 — full weights, no sub-native quantization
Context window1M input tokens · 128K output tokens
AgentOpenCode CLI, OpenAI-provider config
Auth32 unique per-user API keys
Lane policy1 guaranteed request/user, extra requests queue in order
WindowOne hour — 2026-08-18, 13:41–14:41 UTC
03

What one hour looked like

7,562Requests served~125/min sustained, 0 failures
467.6MInput tokens processed~97% served from warm cache
7.03MOutput tokens generatedoutput-light workload (65:1 in:out)
97.0%Prefix cache hit rateheld steady under full load
~1.0sWarm time-to-first-tokencache-hit responses start fast
30 / 32Peak concurrent lanesexcess queued, never dropped
04

The cache is the engine

Agentic coding re-sends the same repo, system prompt, and history on every turn. A sticky lane keeps that context warm, so 97% of all input tokens were served from cache — skipping the expensive prefill step. That hit rate held steady across the whole hour, under full load, and warm requests started returning tokens in about a second.

What the cache saved: serving these 467.6M input tokens with no cache would cost ~$107 in raw prefill value at DeepSeek off-peak rates. With 97% hits, the same work priced out at ~$10.8 — a ~10× reduction. Warm context is what makes a shared lane economical: the tokens you re-send cost almost nothing to serve.
05

Value delivered per user

Every token each user consumed was priced at DeepSeek's own published API rates, so the numbers are directly comparable to what the same usage would have cost on the first-party API. Against a flat $0.20/hour lane:

Average per-user value for a $0.20 lane-hour

DeepSeek-equivalent token value consumed, one hour, per user — dashed line marks the $0.20 lane cost

Off-peak
$0.34 · 1.7×
Peak
$0.68 · 3.4×

DeepSeek doubles all token rates for ~7 hours a day; a flat lane never does — so the same usage is worth 1.7× the lane price off-peak and 3.4× at peak. Usage was evenly distributed across the cohort, from lighter to heavier users:

Per-user value distribution (32 users, off-peak basis)

Each bar is one user's DeepSeek-equivalent value consumed in the hour

heaviest · $0.52median $0.35lightest · $0.17

Average $0.34, median $0.35, range $0.17$0.52 off-peak (double those at peak). The heaviest user pulled 2.6× off-peak / 5.2× peak value from a single $0.20 lane.

06

The saturation ceiling

Across the live hour, the 32-lane cohort used only about a third of the node's capacity — real demand, with clear headroom to grow. To establish the full ceiling we saturation-tested the node separately at 64 concurrent streams with speculative decoding on. The result: a warm cache doesn't just cut cost, it raises throughput.

64-concurrent profileOutput tok/sPer userTotal tok/s
~98% cache hit (production-like)6,206~97107,000
Cache miss (fresh prompts)3,291~5129,600
Very-long-context, uncached (64:1)551~935,800

Warm cache nearly doubled sustained output (6,206 vs 3,291 tok/s) by freeing compute from prefill and redirecting it to generation. At the production-like ceiling, 64 lanes each sustain ~97 tok/s — and the live run's throughput sat at only ~30% of this, confirming the node carried its cohort with headroom to spare.

The core lever: the operator sets lanes-per-node to trade guaranteed per-user speed against price. A cohort splits one warm, well-utilized GPU running at 6,000+ output tok/s, instead of each paying for an idle dedicated one — and the sticky cache is what keeps that number high.
Methodology & honest caveats
  • Measured, not modeled. All figures are from live telemetry over the exact window 2026-08-18 13:41:49–14:41:49 UTC: per-key token and spend logs, and the serving engine's own cache/throughput counters. Token value is computed at DeepSeek's published off-peak rates ($0.22 input / $0.007 cached / $0.66 output per 1M); peak figures apply DeepSeek's 2× peak schedule.
  • One workload class. The cohort ran coding-agent traffic (large, stable contexts; short outputs), which is cache-favorable by nature. Cache hit rate and per-user value will differ for one-shot or low-reuse workloads.
  • The live run used ~30% of node capacity, so its per-user values reflect real demand, not the ceiling. Ceiling figures come from a separate 64-concurrent saturation benchmark (cache-miss and ~98%-hit datasets, speculative decoding on, driven on the node).
  • $0.20/lane-hour is an illustrative price, not a cost disclosure. Value multiples compare user-consumed DeepSeek-equivalent value to that price.