docs
// Platform

Cost Comparison

What a month of inference costs on our accelerators, on yours, and in your own datacenter, and what it would take to build and run the platform yourself.

Three Ways to Run

The three ways to run and what each bills
Way Runs on Billed
Hosted Our accelerators Per token, by the endpoint's tier
Self-Hosted Your accelerators, behind our router Per EIM-hour connected, plus per uncached token; cached input free
Sovereign Your datacenter, the whole stack Per node, per month; six-month minimum

Hosted and Self-Hosted usage is paid from prepaid credits; see Credits. A sovereign license is billed per node by contract. The tier ladder and each tier's limits are on Service Tiers.

Rates are behind sign-in. This page shows how each cost is computed. Open an account and it shows the rates and runs the calculator, or email contact@erebine.ai.

Break-Even

One EIM on Self-Hosted bills its hourly rate for every hour it is connected, whether it serves one token or a billion, plus the Self-Hosted rate on each uncached token it serves; cached input tokens are free. A hosted tier bills per token only. Each token moved onto the EIM saves the difference between the two token rates; the break-even is the monthly volume at which those savings pay the EIM's hourly fee. A hosted tier whose rate is not above the Self-Hosted token rate has no break-even.

formula
break-even tokens per month = (hourly rate * 730 hours) / (hosted rate per 1M tokens - Self-Hosted rate per 1M tokens) * 1,000,000

It compares against the cheapest hosted tier at interactive scheduling priority. Self-Hosted requests are scheduled ahead of every hosted tier, so a tier that waits behind others is not a like-for-like comparison, and the cheapest rate in the band gives the largest break-even volume. List rates are used, with no cached tokens; a plan's discount lowers both token bills and raises the break-even.

Calculator

The calculator takes your monthly tokens, the hosted tier you would use, how many EIMs you would run and what their accelerators cost you per hour, and returns the monthly bill for each way to run. It loads the rates for signed-in accounts: open an account, or sign in. The formulas below compute the same figures by hand.

How It Is Computed

Hosted
Tokens per month / 1,000,000 * the tier's rate per 1M tokens. Input and output tokens bill alike; cached input tokens bill at the tier's cached-token multiplier, and a plan's discount scales the total. See Cached Token Pricing and Plans.
Self-Hosted
EIMs * 730 hours * (the Self-Hosted hourly rate + your accelerator cost per hour) + tokens per month / 1,000,000 * the Self-Hosted rate per 1M tokens. Cached input tokens are free on Self-Hosted, and a plan's discount scales the token part as on Hosted. An EIM bills its hourly rate only while it is connected; see Uptime Billing.
Sovereign
Frontend nodes * the frontend rate + router nodes * the router rate + EIMs * the EIM rate, per month, plus EIMs * 730 hours * your accelerator cost per hour. Six months minimum. EEMs are included.
Storage
The same on every way to run that Erebine hosts: GB stored * the platform rate per GB-month. See Storage Billing.

Building It Yourself

What a team would build and operate to match what Erebine ships today, line by line. The ranges are our estimate, in engineer-months of experienced engineers; each line states what it assumes. They exclude accelerators, hosting, power, security audits and certifications.

Estimated engineering effort to build each part of the platform
Part What it covers Effort Assumes
OpenAI-compatible API Chat completions, Responses, conversations, embeddings, rerank, files, uploads, batches, images, audio and models, with streaming and the error shapes the SDKs parse. See SDK parity. 12 to 24 engineer-months Request and response shapes are reimplemented and tested against the official SDKs on every release.
Router and scheduler Scoring on sixteen signals, weighted random selection, fairness, credit holds, queueing, SLO tracking, retries, circuit breakers. See The Router. 12 to 24 engineer-months One team designs, load-tests and tunes it; no prior router to start from.
Inference mesh EIM images for NVIDIA, AMD, Intel, Apple silicon and CPUs; enrollment with join keys; a mutually authenticated CurveZMQ mesh with key rotation; leases and streaming backpressure; model upload, versioning and catalog. See EIM. 15 to 30 engineer-months Each accelerator family is built and validated on real hardware.
Metering and credit ledger Per-token, cached-token and hourly metering; an atomic ledger with holds; plans and payments; invoices; budget alerts; usage dashboards and exports. See Usage Tracking and Billing. 9 to 18 engineer-months A payment processor handles cards; the ledger and metering are built in-house.
Multimodal Image and video input, audio input, transcription and translation, speech, realtime transcription, image generation. See Audio and Images. 6 to 12 engineer-months Open models and engines exist for each; the work is serving, routing and billing them.
EEM execution plane Execution managers on your hosts, approvals by risk level with escalation, webhooks, tool catalogs for cloud, Kubernetes and OpenStack, an operations dashboard. See Approvals. 12 to 24 engineer-months Tools are wrapped one at a time and each is tested against a live system.
Safeguards AES-256-GCM at rest under per-project keys with rotation; project isolation; rate limits; scoped, expiring API keys with IP lists; roles and two-factor sign-in; sessions and CSRF protection; an audit log; zero data retention per workspace; image prompt screening; DMCA and NCII takedown intake; retention and purge rules. See Encryption, Authentication and Audit Trail. 9 to 18 engineer-months Built in from the start; no external audit or certification included.
Clients A CLI, a desktop app, an MCP server, and editor integrations. See erectl and MCP. 6 to 12 engineer-months One client per surface, kept in step with the API.
Total To build 81 to 162 engineer-months About 7 to 14 engineer-years.

In dollars

The median annual wage for software developers in the United States was $135,980 in May 2025 [1]. Wages and salaries were 70.0 percent of employer compensation costs for private industry workers in June 2026 [2], so one engineer-year costs about $194,257 with benefits, before equipment and office. At that cost the build is about $1.3M to $2.6M.

Running it

A platform that serves production traffic needs someone on call at every hour. Our assumption, with the arithmetic: two engineers on call at once, a primary and a backup, in week-long shifts, and no engineer on call more than one week in four, so three weeks in four go to engineering. Two on call across a four-week cycle is eight engineers, each spending a quarter of their time on call. On-call alone is then at least two engineer-years a year, about $390,000 a year, before the time those engineers spend keeping up with new models, engines and SDKs.

Sources

  1. U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, Software Developers, Quality Assurance Analysts, and Testers: Pay. bls.gov/ooh/computer-and-information-technology/software-developers.htm
  2. U.S. Bureau of Labor Statistics, Employer Costs for Employee Compensation, June 2026. bls.gov/news.release/ecec.nr0.htm