Skip to content
Seldon

Seldon Inference Cloud

Frontier-quality inference,
at one fifth of the price.

We serve open-weight frontier models on sparse mixture-of-experts architectures, speculative decoding, and FP4 kernels, across a GPU fleet sourced continuously from the cheapest available capacity. The quality is comparable. The bill is not.

420.0B tokens served daily140 teams in production99.97% uptime

Seldon Prime 1
$2.10
blended per 1M tokens
Claude Opus 4.8
$10.00
blended per 1M tokens
Difference
4.8x
lower cost per token
drop-in migration
from openai import OpenAI client = OpenAI(    base_url="https://api.seldon-ai.org/v1",    api_key=SELDON_API_KEY,) response = client.chat.completions.create(    model="seldon-prime-1",    messages=[{"role": "user", "content": prompt}],)

Two lines change. The rest of your application does not.

In production at

Customer stories
  • Developer tools / Series BAI pair programming across large monorepos421.5B tok/moDedicated capacity
  • Data infrastructure / Series ASchema inference and pipeline repair96.4B tok/moShared endpoint
  • Legal / Series BContract review and obligation extraction188.9B tok/moCustomer VPC
  • Customer support / Series AAutonomous resolution for logistics operators42.7B tok/moShared endpoint
  • Healthcare / Series CClinical documentation and coding support134.2B tok/moCustomer VPC
  • Fintech / Series BTransaction enrichment and risk narrative generation209.8B tok/moDedicated capacity
  • Data infrastructure / Series ARobotics fleet telemetry summarisation63.1B tok/moDedicated capacity
  • Customer support / EnterpriseMultilingual product and order assistance311.4B tok/moCustomer VPC
420.0B
Tokens served daily
140
Teams in production
99.97%
Uptime, trailing 90 days
210ms
Median time to first token

01The problem

Inference is not a line item. It is your cost of goods.

A traditional software company pays for hosting and keeps roughly 80 cents of every dollar. An AI-native company pays a per-token toll on every unit of value it delivers, and that toll scales exactly with usage. This is the structural reason AI companies trade at lower multiples than the software companies they otherwise resemble.

At published frontier rates, a coding assistant at 1,000 developer seats runs a 14% gross margin. That is beneath the floor investors use to identify a healthy AI-native business, and far beneath the 75 to 85 percent that defines a software company. The product is fine. The arithmetic is not.

See the full margin model

The question is not whether the model is good enough. It is whether you can afford to serve the answer.

$155,835
Monthly inference

AI coding agent, 1,000 developer seats, at published frontier rates.

14.1%
Gross margin

Against a 75 percent healthy-SaaS benchmark.

02How

Eight published techniques, assembled into one stack

None of this is proprietary research. Sparsity, speculative decoding, FP4 kernels, KV compression, disaggregated scheduling, and routing are all in the literature. The work is making them coexist in a single serving path under production traffic, and then buying the hardware well.

USD per million output tokens

Llama 3.3 70B dense, FP8 weights, one rented H200 at $4.39/GPU-hr, 1K in / 1K out chat, 50 tok/s/user, stock vLLM with continuous batching, PagedAttention and chunked prefill, no prefix cache, 100% utilization.

USD per million output tokens
StageMultiplierRunning cost
Stock stack1.00x0.70
Sparsity0.52x0.36
Speculation0.70x0.25
Precision0.64x0.16
KV compression0.81x0.13
Serving stack0.82x0.11
Disaggregation0.77x0.08
Routing0.78x0.06
Capacity0.63x0.04
Seldon0.058x0.04

Factors are discounted for overlap between techniques. The undiscounted product would be 83.9x, which would not be honest.

research/tech-batching-economics.md section 6.1; RunPod live pricing 2026-07-17; SemiAnalysis InferenceMAX H200 throughput

3.3xSparsityRoute each token to a small subset of expert feed-forward blocks, so parameter count and per-token compute decouple.
1.7xSpeculationA cheap drafter proposes several tokens and the target model verifies them all in one forward pass, converting idle FLOPs into throughput.
1.8xPrecisionStore and compute weights at four to eight bits instead of sixteen, cutting the bytes that decode is bound by and the FLOPs that prefill is bound by.
1.4xKV compressionShrink the bytes each sequence occupies in HBM, because KV cache size is what caps batch size and batch size is the denominator of cost per token.

03The tradeoff

Up and to the left

Cheap models that cannot do the work are not cheap. The only comparison worth making plots capability against price, and the only claim worth defending is a position on that frontier.

Composite quality index against blended price per million tokens

Log price axis. The quality index is banded in steps of 10 because the 2026 model generation shares no common public benchmark, so finer resolution would imply a precision that does not exist.

  • Seldon
  • Closed frontier APIs
  • Open-weight, other hosts
Composite quality index against blended price per million tokens
ModelCategoryBlended $/1M tokensQuality index
Seldon Prime 1Seldon$2.1080
Seldon Core 1Seldon$1.8080
Seldon Vector 1Seldon$0.2150
Seldon Horizon 1Seldon$1.2980
Claude Fable 5Closed frontier API$20.0090
Claude Opus 4.8Closed frontier API$10.0090
Claude Sonnet 5Closed frontier API$4.0080
Claude Haiku 4.5Closed frontier API$2.0060
gpt-5.6-solClosed frontier API$11.2590
gpt-5.6-terraClosed frontier API$5.6380
gpt-5.6-lunaClosed frontier API$2.2560
Gemini 3.1 ProClosed frontier API$4.5090
Gemini 3.5 FlashClosed frontier API$3.3870
Grok 4.5Closed frontier API$3.0080
DeepSeek V4-ProOpen-weight, other host$1.6380
DeepSeek V4-FlashOpen-weight, other host$0.1160
GLM-5.2Open-weight, other host$1.4580
Kimi K2.7-CodeOpen-weight, other host$1.4370
Qwen3.5-397B-A17BOpen-weight, other host$1.0980

We publish the derivation rather than a single headline score, because a composite index is only as good as its method. If you disagree with the weighting, every underlying figure is cited.

Pricing from vendor rate cards, July 2026. Quality index derivation is documented in the source tree.

04What it changes

The same product, on the right side of the margin line

Identical token volume, identical output quality requirements, identical everything except the rate card. This is what moving inference COGS does to a profit and loss statement.

Monthly inference spend by application archetype

Token volumes and cache hit rates are modelled per archetype with every assumption published. The frontier comparator is Claude Sonnet 5 at introductory pricing.

  • Seldon
  • Closed frontier
Monthly inference spend by application archetype
CategoryWorkloadSeldonClosed frontier
AI coding agent1,000 developer seats$106K$156K
Customer support agent100,000 conversations per month$8.3K$12K
Document and RAG analysis product10,000 monthly active users processing 200,000 documents per month$21K$36K
High-volume classification and extraction pipeline50,000,000 items per month$48K$68K

Sonnet 5's introductory rate expires 31 August 2026, after which the same comparison uses $3 input and $15 output, widening the gap.

Assumptions and derivation in the economics module

+25%
Margin points gained

AI coding agent, moving from a closed frontier model to Seldon Prime 1. From 14% to 39% gross margin.

The honest caveat

This does not help everyone equally. For a support agent, the same switch cuts the inference bill 9.6x and buys only 5.7% of margin, because inference was 6.3% of revenue to begin with. That was the entire prize. Inference was never the binding constraint there, and we tell you if you are in that case before you switch.

Read the counterweight

What changed

The same argument, from the other side of it

Every figure quoted below is computed, not asserted: each is the same cost function the calculators run, applied to that customer's published workload shape, for the specific pairing named beside it. All four assume a move from gpt-5.6-luna to Seldon Vector 1. The margin calculator fixes its Seldon leg to Seldon Prime 1, so it will not return these figures directly, and your own volumes on your own tier are the number that actually matters.

Inference was 40 percent of revenue and it was the only line that scaled with every seat we sold. Moving the completion path over took an afternoon because the API is wire-compatible. The margin change was the difference between raising a bridge and not needing one.

VP EngineeringDeveloper tools company, Series BMoved from a closed frontier tier, 2026-01
+35.9 ptsgross margingpt-5.6-luna to Seldon Vector 1

We were paying a long-context surcharge on every contract we ingested. Seldon does not charge one, and the prefix cache holds across a document set, so a hundred-clause review costs what the first clause used to.

Head of PlatformLegal company, Series BMoved from a closed balanced tier, 2026-03
10.4xlower inference COGSgpt-5.6-luna to Seldon Vector 1

The routing layer is the part that surprised us. Roughly seventy percent of our classification traffic never needed a frontier model, and nobody had been separating those requests out. We did not change our quality bar to get the saving.

Staff EngineerData infrastructure company, Series AMoved from a closed fast tier, 2025-11
9.1xlower cost per itemgpt-5.6-luna to Seldon Vector 1

They told us upfront that inference was only six percent of our revenue and that switching would not fix our margins, which is not what a vendor normally says in a sales call. We moved anyway for the latency, and the honesty is why we expanded.

CTOCustomer support company, Series AMoved from a closed fast tier, 2025-09
+5.7 ptsmargin, as predictedgpt-5.6-luna to Seldon Vector 1

05Capacity

The identical GPU trades at a wide spread

An H100 is an H100. The hourly price for one, on non-interruptible on-demand terms, varies by 6.2x between the cheapest and the dearest major provider. Even inside a single hyperscaler, the spread across regions and commitment terms reaches 5.4x.

That dispersion persists because capacity is sold on contract structures rather than spot commodity terms. Reservations, committed-use discounts, and interruptible tiers all price the same silicon differently depending on who carries the utilization risk. We buy across all of it continuously and carry that risk ourselves, which is a large part of why our floor sits where it does.

How the fleet is sourced
6.2x
On-demand spread

Cheapest to dearest published rate for an H100 on like-for-like terms.

2.7x
Standardized spread

Both legs normalised to identical terms by an independent index.

5.4x
Within one cloud

Same provider, same GPU, different region and commitment term.

$0.041
Cost floor

Per million output tokens after the full stack, on sourced capacity.

Runs on

  • NVIDIA H100
  • NVIDIA H200
  • NVIDIA B200
  • NVIDIA GB200 NVL72
  • AMD MI355X
  • InfiniBand NDR

Accelerators the serving stack is built and tuned for. Kernel work targets each generation directly rather than relying on a portability layer.

Drop-in for

  • OpenAI API format
  • Anthropic Messages format
  • LangChain
  • LlamaIndex
  • Vercel AI SDK
  • OpenTelemetry

Wire-compatible surfaces. Existing application code moves over by changing a base URL and a key, with no SDK rewrite.

Deploys into

  • AWS
  • Google Cloud
  • Azure
  • Customer VPC
  • On-premises
  • Air-gapped

Target environments for dedicated and private deployments, including customer-controlled infrastructure.

06Trust

We publish what we run, not a wall of badges

Seldon holds no certifications. Rather than imply otherwise, our trust center lists every framework we are working toward with a real status and a real date, alongside the controls that are implemented today.

A badge tells a technical buyer nothing. TLS 1.3 in transit, AES-256 at rest, tenant isolation, zero-retention inference with its carve-outs stated plainly, SSO and SCIM, and audit logging tell them something they can check.

Read the trust center
SOC 2 Type IIin progress
ISO/IEC 27001:2022planned
ISO/IEC 42001:2023planned
HIPAAin progress
GDPRin progress
CCPA / CPRAin progress

No item on this list is presented as achieved. Each report goes to customers under NDA once it exists, and not before.

Run the numbers on your own workload.

Send us a week of traffic shape, input and output token mix, cache hit rate, and latency targets. We will return a modelled cost comparison against your current provider, with the assumptions written out so you can check them.