Example projects on open models,
with the numbers worked out
Eight example projects built on open models: who the company is, what they build, how many GPUs it takes, what it costs against what it saves, and how the rollout runs. The companies are made up. The arithmetic is shown so you can redo it with your own numbers.
Pick the project closest to yours
None of these are customer stories. Each is a composite of a common kind of project, with its assumptions stated. Most real systems combine several: a support assistant usually sits behind a classifier and calls tools, for example.
Support Chat Assistant at a Regional Broadband and Mobile Provider
A broadband and mobile provider answers billing, outage and plan questions in its 9,000 daily support chats, and hands the rest to agents with a summary.
Policy and Procedure Assistant at a Regional Bank
Bank operations staff ask questions of 40,000 internal documents and get answers that cite the paragraph they came from, without the data leaving the bank's network.
A Self-Hosted Coding Assistant at a Payments Software Company
A payments software company with 400 engineers runs code completion, chat and pull-request review on open models inside its own network.
Claims Document Intake at a Regional Insurer
An insurer turns 2,500 claim documents a day into validated records, and people review only the ones that fail checks.
Deposition Summaries at a Litigation Firm
A litigation firm drafts deposition summaries with page:line citations overnight, and paralegals review drafts instead of writing them.
Shipping Exception Handling at an Industrial Parts Distributor
A parts distributor resolves 250 shipping exceptions a day with an agent that investigates and drafts, while a person approves every refund or order change.
Listing Screening at a Second-Hand Marketplace
A second-hand marketplace screens 400,000 new listings a day before they go live, and its reviewers see fewer, better flags.
Catalog Localization at a European Home Goods Retailer
A home goods retailer translates 25,000 new or changed products a week into five languages, with its glossary enforced and native speakers reviewing a sample.
From a project to a dedicated endpoint
Every project here runs on a dedicated endpoint: GPUs reserved for you, with their own worker layout and scaling limits. We buy the GPU capacity and run it with our Token Factory, built with NVIDIA DSX.
Profile your workload
We measure what your requests look like: prompt and output lengths, how many arrive at once, and which latency your users notice.
Optimize and benchmark
We choose the model and GPUs with you, decide whether prefill and decode share GPUs or run in separate pools, set the endpoint to favor latency, throughput or a balance, and benchmark it on your traffic.
Commit to it in the contract
We commit to the throughput and latency agreed for your workload, not just availability, backed by contractual remedies such as service credits where applicable.
Tell us what you are building
Bring us your model and a description of your traffic. We will tell you what it takes to profile the workload and what a performance commitment would cover. Self-service accounts can create inference endpoints to try the platform, but not on our top-tier GPU capacity.