DeepSeek V4 and GLM 5.1 Hosting: 4x Price Spread in 2026

Cost per token rises down that list at low utilization and falls at high utilization. That single sentence is the whole decision, and the rest of this post is the arithmetic that tells you which side of the crossover you are on.

The most common search on this topic asks how many monthly calls Ollama Cloud Max gives you for DeepSeek V4 Pro, Kimi K2.6 and GLM-5.1. The answer is that the number does not exist.

Ollama Cloud is Free at $0, Pro at $20 per month or $200 per year, and Max at $100 per month, with Team listed as coming soon. Concurrent cloud models are 1, 3 and 10 respectively. Quota is published relatively: Free is "light usage", Pro is "50x more than Free", Max is "5x more than Pro". Limits reset "every 5 hours and 7 days", so there is no monthly window to count calls in. The API returns prompt and eval token counts plus nanosecond timings, but exposes no quota headers and no quota endpoint, and the documented cloud limits page 404s. A feature request in the project's public issue tracker, opened April 2026, states that a user has to log into the website or wait for a warning email to find out where they stand, and that neither path works from a CLI tool.

Two more things a shopper should know before pulling out a card. Max is marked "new signups paused" as of 2026-07-27. And the catalogue does carry everything the query asks about: DeepSeek V4 Pro, DeepSeek V4 Flash (284B total, 13B active), Kimi K2.6, GLM-5.1 and GLM-5.2, plus GPT-OSS, Qwen 3.5 and Nemotron 3.

Compare that to the one subscription host that publishes an absolute number. Cerebras Code Pro is $50 per month for up to 24 million tokens per day, and Max is $200 per month for up to 120 million per day. Fully consumed, that works out to $0.069 and $0.056 per million tokens, which undercuts every per-token endpoint in this post by 6x or more. Both tiers are marked sold out.

What $100 buys elsewhere, so you can size the trade yourself:

DeepSeek first-party, 10:1 input:output, no caching
  (10 x $0.435 + $0.87) / 11 = $0.4745 / 1M blended
  $100 -> 211M tokens

Same, 60% cache-hit rate on input
  -> $0.239 / 1M blended
  $100 -> 418M tokens

DeepInfra V4 Pro: (10 x $1.30 + $2.60) / 11 = $1.418  -> 71M tokens
Together V4 Pro:  (10 x $1.74 + $3.48) / 11 = $1.898  -> 53M tokens

So Ollama Cloud Max at $100 is a good deal above roughly 70M to 200M blended tokens per month, which at ~50k tokens per agentic coding task is something like 1,400 to 4,000 tasks. That is a modeled envelope, not a published one, because Ollama publishes no token quota. Buy it for the single endpoint, the flat bill and the stated no-logging policy, not because you priced it against per-token billing. On the local versus hosted question underneath all of this, see Ollama versus vLLM for serving open-weight models.

This is the table most of this post exists for. Live per-provider values for DeepSeek V4 Pro pulled from public aggregator endpoint listings on 2026-07-27, converted to dollars per million tokens. Two rows cross-check against the providers' own pricing pages exactly: Together at $1.74 / $3.48 / $0.20 cached, and DeepInfra at $1.30 / $2.60 / $0.10.

Read the last three columns before the first two.

"1M context" is not what three of these providers actually serve. DigitalOcean and BaseTen cap at 262,144, and Together caps at 512,000, on a model whose own window is 1,048,576. The output side is worse: Venice serves the full million of context but caps output at 32,768, and DeepInfra, one of the cheapest credible rows, caps max output at 16,384 tokens, which silently truncates long agentic runs and refactors rather than erroring cleanly.

Quantization is not uniform. FP4 at five providers, FP8 at seven, and six reporting unknown (including DeepSeek's own endpoint). You are not buying the same model at the same quality across rows. Worth knowing: NVIDIA's own NVFP4 build of GLM-5.1 measures the damage at close to zero against an FP8 baseline (SciCode 47.14 to 47.34, AIME 2026 unchanged at 96.67, GPQA Diamond 85.61 to 85.02, IFBench 76.56 to 76.33), with only the MoE expert linear weights quantized and the shared expert left alone. FP4 is not the problem. Not knowing which precision you got is.

Reliability varies too, and it does not track price. In the one-day uptime figures published alongside these endpoints, Ionstream sat at 78.5% and Fireworks at 92.4%, while SiliconFlow, Novita and BaseTen were all above 99.6%. Pick on price alone and you will be on the cheap row during an outage.

For the other two models, the pattern inverts and repeats:

The recommendation, stated plainly: for DeepSeek V4 Pro, buy first-party if jurisdiction allows it. Only two other endpoints land within 60% of its output price, and everything past those is 2.6x or more. If jurisdiction does not allow it, StreamLake and GMICloud at FP8 are the value rows and DeepInfra is only cheap if you never need more than 16k of output. For GLM-5.1 and Kimi K2.6, there is no first-party discount, so pick on quantization, served context and uptime.

Before pricing an hour, check whether the model fits. DeepSeek V4 Pro's released checkpoint is 862 GB in mixed FP4/FP8. Allow roughly 127 GB of an H200's 141 GB for weights after activations and graphs and you need 7 cards for the weights alone, so 8 in practice.

The SGLang project's day-0 benchmarks independently landed on exactly this: V4 Pro at tensor-parallel 8 on B200, V4 Flash at tensor-parallel 4 on H200. NVIDIA's NVFP4 GLM-5.1 card launches with tensor-parallel 8.

On a full 8x H200 node (1,128 GB), 862 GB of weights is 76% of the box, leaving about 266 GB for KV cache and activations. MLA-family attention runs on the order of 70 KB per token (measured for DeepSeek V3, where 576-dimension compressed KV across 61 layers gives 70,272 bytes, versus 192 to 328 KB for GQA models, per arXiv 2601.19910). V4 uses a hybrid compressed-sparse attention rather than plain MLA, so treat that as an order of magnitude and not a measurement. At 70 KB per token, 266 GB is about 3.8M tokens of KV: roughly 14 concurrent 256K sessions, or 3 concurrent sessions at the full 1M window. That is the real concurrency ceiling of a single node, and it is much lower than people expect.

Hourly rates for that class of hardware span 2.9x on identical silicon. On 2026-07-27 an H200-hour ran $2.69 on DeepInfra's dedicated tier, $3.59 and $4.39 on RunPod community and secure, $5.99 on Together's GPU clusters, $7.00 at Fireworks (which charges the same for an H200 as for an H100, so you get 141 GB for the price of 80 GB), and $7.91 per GPU-hour on AWS p5en on-demand list pricing, with p5en spot at $3.41. We priced the full ownership worksheet, including idle-GPU tax and headcount, in the self-host versus API break-even once idle GPU and headcount are priced in; this section only converts an hour into a token.

State the assumptions or the number is worthless:

At 100% that node produces 45.6 billion output tokens per month. Here is what an output token costs at various duty cycles:

Three conclusions, and they are not symmetric.

You cannot beat DeepSeek's own API by renting GPUs unless you keep the node about 39% busy every hour of the month (17.8B output tokens). On AWS on-demand, breakeven is 52.4B tokens, which exceeds the node's theoretical capacity: it is arithmetically impossible, not merely expensive. Renting to save money against a $0.87 endpoint is a losing trade. You rent for residency, determinism and rate-limit control.

Against Western serverless the picture inverts. At $2.60 to $3.48 per million output tokens, a dedicated node pays off at 10 to 21% utilization, or roughly 4.5 to 10 billion output tokens per month. That is a mid-size product, not a hyperscaler.

The utilization gap is the entire argument. A single user on a well-tuned node sees roughly 200 to 266 tokens per second of decode (the SGLang day-0 numbers: 199 tok/s for V4 Pro on B200, 266 tok/s for V4 Flash on H200, degrading under 10% out to 900K of context). A saturated node sees ~2,200 per GPU. That 10x gap is why renting a node for one developer is indefensible and renting it for a fleet is obvious.

Market estimates put an integrated 8x HGX H200 node at $320,000 to $420,000, call it $370,000, of which $75,000 to $90,000 is chassis, CPU, NVLink, networking and integration. That is a market estimate, not an OEM quote, so treat it as a planning figure.

Power is not a rounding error. Eight H200 SXM at up to 700 W configurable is 5.6 kW of GPU plus roughly 2.4 kW of host, and at PUE 1.4 that is about 11.2 kW of facility draw, or 8,064 kWh per month. At $0.08 to $0.20 per kWh that is $645 to $1,613 monthly. Amortize $370,000 over 36 months and add power and you land near $11,250 per month before colo, spares, and the engineer who owns the box.

Payback against renting is 25.5 months versus the cheapest dedicated node and 15.2 months versus a secure marketplace tier. Set that against a model half-life that in 2026 is well under 18 months and the trade is clear: you buy hardware for control, not for savings.

The consumer-adjacent path got worse, not better. The RTX PRO 6000 Blackwell 96 GB was listed at $13,250 in mid-2026, up from an $8,565 launch price in March 2025, a 55% increase in about 16 months on GDDR7 supply. Eight of them is 768 GB for $106,000 of GPUs at 4.8 kW, which does not fit DeepSeek V4 Pro's 862 GB. You would need ten cards (960 GB, $132,500, 6.0 kW), or step down to V4 Flash or an NVFP4 GLM-5.1. A four-card box at 384 GB runs either of those comfortably. For the full owned-hardware path on GLM specifically, see what it takes to run GLM-5.2 on hardware you own.

Look again at the DeepSeek V4 Pro provider list. Several of the operators serving it sit under PRC jurisdiction, and the same is true for GLM-5.1 and Kimi K2.6 pools. Aggregators sort by price by default. That means default routing will place EU or US regulated traffic on infrastructure your compliance reviewer has never heard of, and it is the single most under-discussed fact in "cheapest open-weight API" content. We covered the jurisdiction failure mode for hosted endpoints in the cost per resolved ticket, hosted API versus self-hosted comparison, and the broader compliance argument in cloud versus on-premise AI security and cost.

The control exists, but it is opt-in. The major routing aggregators attach a structured data policy to every endpoint, distinguish "may train" from "does not train" and "zero retention", and let you restrict an individual request rather than only the whole account. Constrain every request:

{
  "model": "deepseek/deepseek-v4-pro",
  "provider": {
    "data_collection": "deny",
    "only": ["parasail", "coreweave", "deepinfra"],
    "allow_fallbacks": false
  }
}

allow_fallbacks: false is the line people omit. Without it, an outage on your compliant endpoints silently reroutes you to whatever is cheapest and available. If you route across several providers at all, do it through a gateway in front of multiple inference providers so the policy lives in one declarative config instead of scattered request bodies.

Two further gaps worth naming. First, quantization opacity is a compliance problem before it is a quality one: if a third of the endpoints for your model report precision as unknown, you cannot document which weights served a regulated request. Second, the public pricing pages of two major Western platforms carry no retention, zero-data-retention, SOC 2, HIPAA or BAA language at all; those terms come from a contract, so do not assume them.

Ollama is unusually clear for a consumer-priced tier: prompts and responses are never logged or trained on, partners are contractually required to run no-logging and zero-retention, and compute is "primarily in the United States" with routing to Europe and Singapore for capacity. That last clause is the one a data-residency reviewer will stop on.

The escalation ladder, cheapest to strictest: shared serverless, retention-filtered serverless, single-tenant dedicated endpoint, rented GPUs in a chosen region, hardware in your own facility. A business-hours workload with a 3:1 peak-to-mean ratio sits in the 10 to 20% duty-cycle band, and that band is exactly where an on-premise deployment stops reading as a premium and starts being the cheaper option that also passes review. That is the arithmetic behind most of the private-AI work we do at Particula Tech for regulated buyers: the compliance case and the cost case point the same direction more often than people expect.

Yes, on the prefix-cache axis, and only there. Both models officially support vLLM and SGLang (GLM-5.1 pins SGLang v0.5.10+ and vLLM v0.19.0+; Kimi K2.6's card names vLLM, SGLang and KTransformers). vLLM's automatic prefix caching hashes KV blocks from the parent hash plus block tokens and caches only full blocks. SGLang's RadixAttention keeps an LRU radix tree of KV across requests and matches at token granularity, and the original RadixAttention paper reports up to 6.4x higher throughput on agent control, few-shot, JSON decoding and RAG pipelines.

The practical rule: token-level radix matching wins when prefixes are long, shared and not block-aligned, which is what multi-turn agents and RAG-with-a-fixed-header look like. When prefixes are short or unique the two converge, so choose on operational grounds. We benchmarked both in SGLang vs vLLM for production inference.

One cost nobody models: rent GPUs and you inherit the engine-version treadmill. Day-0 support for a new checkpoint is a real selection criterion, and serverless providers absorb that maintenance for you. Price a few engineer-days per model release into any rented-GPU plan.

Under 5 billion output tokens per month, use per-token serverless and stop thinking about it. Buy DeepSeek first-party if jurisdiction allows, otherwise an FP8 endpoint with a named quantization and full context, pinned with an explicit provider allowlist and fallbacks disabled. Between 5 and 20 billion, rent a dedicated node and only then if your duty cycle clears 20%. Above that, or the moment a reviewer asks where the weights physically ran, move to hardware you control and stop optimizing per-token pricing entirely. Subscription tiers are a developer-seat product, not an infrastructure product: buy Ollama Cloud Max for a workstation, not for a workload.