In stealth · 2026

Inference for agents that run for hours, not seconds.

We have identified an infrastructure path to the cheapest place to run an agent that can wait.

Email jon@weinfer.ai Founder on LinkedIn

Agents run long enough to be expensive. Not yet long enough to be worth it.

Value per dollar against how long a single agent run lasts Value per dollar is reasonable for runs of a few seconds, falls to its low point at runs of five to ten minutes where you pay for the run and still have to sit through it, and recovers once runs last hours and the work can be delegated. We are here Seconds 5 to 10 minutes Hours Cheap, but small You pay and still wait You delegate and leave What you get per dollar
A run today is too long to watch and too short to walk away from, metered the whole way at prices set for someone who is watching. You pay premium rates for latency you cannot use. The return only arrives once a run is long enough to hand over.

Every layer already made this choice. All of them picked speed.

Where each layer of the inference stack sits between latency and throughput Six layers of the stack, from the API down to the data centre. Every one of them currently sits at the latency end. Work that runs for hours needs the throughput end at all six. Tuned for latency Tuned for throughput APIRequest routingInference engineKernelsSiliconData centre Where the stack is Where long runs need it
Latency and throughput trade against each other at every layer, and you do not get both. The industry picked latency, correctly, when a person waited on every token. Work that runs unattended wants the opposite, six times over.

The discount for waiting exists, but agents cannot reach it.

Google's four Gemini service tiers, price against how long you will wait Priority costs 75 to 100 percent above list price for a response in seconds. Standard is list price. Flex is 50 percent off for a one to fifteen minute target. Batch is 50 percent off for up to 24 hours. Price Priority+75 to 100%secondsStandardlist priceseconds to minutesFlex50% off1 to 15 minutesBatch50% offup to 24 hours How long you will wait
Batch pricing requires every request up front. An agent never has them: it does not know its next prompt until the last one returns. The cheapest tier is closed to it by construction. This is the gap we built for. Published tier pricing, Google Gemini, 2026.

We designed for all six. Nobody else takes more than two.

Every layer, the other way

API
hand the task off, never hold the connection open
Request routing
follow idle capacity, not the caller
Inference engine
batch depth over token speed
Kernels
tuned for occupancy, not first-token time
Silicon
more than one kind, chosen per deadline
Data centre
wherever capacity is cheap that hour

We serve open models. Your agent calls us one request at a time, each request free to depend on the answer before it, which is exactly what a batch endpoint cannot accept. You mark the deadline; the work lands wherever it is cheapest to finish by then. The saving is not one clever trick at one layer, it is the same choice taken six times.

We are in stealth. We still answer email.

If you build inference hardware, run capacity, or invest in inference, get in touch.

Email jon@weinfer.ai Founder on LinkedIn