Most AI infrastructure is priced for a chatbot: instant answers, capacity held open around the clock, billed whether or not anything is running. A large share of real corporate AI work is nothing like that. Classifying a quarter of support tickets, extracting fields from ten thousand contracts, enriching a customer table, translating a documentation set — that work has a deadline, not a latency budget. MicroDC.ai queues it to GPUs that would otherwise sit idle, and bills you for the compute it consumed.
A dedicated GPU deployment bills every hour it exists. Industry utilization for self-hosted inference typically lands between 15% and 40%, so most of the invoice buys readiness rather than output.
Serverless inference APIs price in the guarantee of an instant answer. If your job runs overnight, you are paying a premium for a property the work does not use.
Self-hosting adds a GPU fleet to someone's job description — drivers, model weights, autoscaling, on-call. That salary line rarely appears in the build-versus-buy spreadsheet.
Not a projection. This is seven consecutive days of production traffic on the platform — 17,757 jobs, roughly 11.9M tokens across three open models — taken from the billing ledger and then repriced against AWS at published rates.
| Option | What you pay for | Cost for that week |
|---|---|---|
| MicroDC.ai | Tokens actually processed | $0.21 |
| AWS Bedrock (serverless) | Same tokens, published per-token rates | ≈ $2.35 |
| Self-hosted on EC2 | 69.3 GPU-hours of busy time on a g6e.xlarge | ≈ $129 |
The EC2 row is deliberately generous to EC2: it counts only the hours the GPU was working. A real deployment also pays for idle time between jobs, for a second instance kept for availability, and for the engineer who maintains it. Two of the three models are hosted natively on Bedrock under the same names, so the comparison needs no "roughly equivalent model" hand-waving for the bulk of the bill. The full methodology, including where the numbers favour AWS, is published: read the breakdown →.
Rates below are pulled live from the model catalog, so this reflects today's prices rather than a screenshot. Set the shape of one recurring batch and read off the monthly figure.
—
Inference cost only. Container and Docker workloads are metered per GPU-hour or CPU-core-hour instead — see pricing. Per-token rates move with network capacity and demand, so treat this as a planning figure, not a quote. Failed jobs are not billed.
Typical end-to-end time is 30 seconds to 5 minutes depending on model and queue depth. If that is a problem for your use case, this is the wrong tool and we would rather say so now.
Set a monthly ceiling for the organization and a separate ceiling per member. Jobs that would breach a cap are refused at submission — there is no overage invoice to dispute.
Credits belong to the organization, not the individual. Owner, admin and member roles; API keys pinned to the org that issued them; every charge attributed to the person who incurred it.
Balances are derived from an append-only transaction ledger rather than stored as a number. You can reconcile any balance back to the jobs that produced it.
Optional end-to-end encryption. Results are encrypted to your public key and cannot be read by MicroDC. Encrypted jobs route only to MicroDC-managed workers, never to third-party hardware. The precise guarantees →
Restrict your jobs to a named set of workers — hardware you have vetted, or machines you contribute yourself. Jobs outside the group never touch them.
Job content and uploaded files are purged automatically after 30 days; only billing metadata is retained. Encrypted payload keys are destroyed as soon as the job is acknowledged.
On certifications, plainly: SOC 2 and ISO 27001 are in progress and not yet awarded. We would rather tell you that than let a footer badge imply otherwise. If your procurement process requires a completed audit today, we are not there yet — talk to us about what your review needs and we will tell you honestly whether we can meet it.
One recurring batch job with a known current cost. Not your hardest problem — your most repetitive one. A number you already have on an invoice is what makes the result meaningful.
Keep your existing pipeline running. Submit the same work here on free credits and compare output quality side by side before anything is switched over.
Compare cost per document and turnaround against your baseline. The ledger gives you the per-job numbers, so the comparison is arithmetic rather than argument.
Expand, keep it to that one workload, or walk away. No contract, no minimum, and no capacity commitment to unwind if the answer is no.
Tell us what you run, how often, and what it costs you today. If the numbers do not work in your favour we will say so — that answer is cheaper for both of us than a pilot that goes nowhere.