For developers
Own your compute.
Dedicated bare-metal GPUs in the city where your users are, reached over a private link from your own cloud. You know where your models run and who else is on the hardware: nobody.
What owning it means
Renting a slice of a shared region leaves the important questions to someone else. Here they have plain answers.
- Whose hardware
- The GPUs you reserve serve your workloads only. Your capacity does not rise and fall with another customer's traffic.
- Where it is
- You choose the city, and you are told which site your hardware is in. Data that has to stay in one place stays there.
- How you reach it
- Over a private link from the cloud accounts you already have, so traffic between your services and your models stays off the public internet.
- What stays yours
- Your model weights, prompts, and outputs. Trace does not use them to train models and does not share them with other customers.
- What it costs
- You reserve hardware for a term at an agreed rate, so your bill does not move with other customers' demand. We quote on request.

Where we are
Tell us where your users are and we will place you at the nearest site.
- Montreal
- New York City
- Houston
- Seven other major cities in North America
Two ways to use it
Bare metal. Whole machines, single tenant, with no hypervisor between your software and the GPU. Memory is pinned, scheduling follows the machine's layout, and the network path bypasses the kernel, so the hardware you pay for is the hardware you get. You bring your own serving stack and models.
A managed API. Trace serves the model for you behind an API that follows the OpenAI request format. In most cases you change the base URL and the key, and the rest of your code stays as it is. You can run an open model or bring your own weights.
Private networking either way. Both connect by private interconnect to AWS, Google Cloud, Azure, Oracle, or your own network. Data moves between your cloud accounts and your Trace hardware on dedicated bandwidth, without crossing the public internet.
What keeps it running
Capacity is spread across many sites in each city and not gathered in one large hall, so there is no single building, feed, or region to lose.
Swipe sideways to see the whole drawing.
- Power and cooling
- Every site has more power and cooling than its load needs, so the loss of one unit does not take the site down.
- Spare GPUs
- Each cluster holds spare GPUs that are already powered and ready. A failed card is replaced by a spare, not by a ticket.
- Fiber
- Sites are reached over more than one fiber path, so a cut cable reroutes traffic and does not stop it.
- Routing
- Requests are sent to the nearest site with capacity. If a site drops out, traffic moves to the next one in the same city without a change on your side.
- Service levels
- Availability targets, and what you are owed if they are missed, are written into your agreement.
Reached over a private link from your own cloud.
Why distance matters
A worked example. The hardware is identical in both rows; only its distance from the caller changes.
- GPU across the country
- 70 ms per round trip. 1,000 calls in sequence spend 70 seconds on the network.An assumed figure for a round trip of roughly 2,400 miles.
- GPU across town
- 3 ms per round trip. The same 1,000 calls spend 3 seconds on the network.An assumed figure for a round trip within one city.
- GPU across the country
- 70 s
- GPU across town
- 3 s
The task is an agent that runs 20 turns and makes 50 model or tool calls in each, 1,000 calls in all, one after another. If every call needs 750 milliseconds of computing, the distant GPU finishes in 820 seconds and the nearby one in 753, and the difference is hard to notice.
Now make the model faster. As the computing in each call falls toward nothing, the nearby task falls toward 3 seconds, but the distant one stops at 70. Even a model that answered instantly would leave the distant task about 23 times slower, because the time left is all travel.
That is where things are heading. The length of task an AI agent can complete has been doubling about every seven months, so tasks contain more calls. The cost of running a model at a fixed level of quality fell about 280-fold in two years, so the computing in each call keeps shrinking. The trip does not shrink with it.
The example counts calls that run in sequence, which is how an agent works when each step depends on the last. Calls that can run side by side pay the trip once, not once each.
The figures are arithmetic on assumed round-trip times, not measurements of a Trace site. Sources for the trends: METR, Measuring AI Ability to Complete Long Tasks, 2025 (task length). Stanford HAI, AI Index Report 2025 (cost of a GPT-3.5-level model, November 2022 to October 2024).
Who it is for
Closeness and dedicated hardware pay off most in these cases.
- Agents and tool use
- Coding agents, research agents, and anything else that makes many calls in a row, where each step waits on the last and every round trip is paid in full.
- Live voice and video
- Conversation, translation, and generation that has to answer while someone is waiting. Inside one city a round trip takes single-digit milliseconds.
- Robots and vehicles
- Systems that act on what the model returns and cannot wait on a region a thousand miles away.
- Regulated industries
- Banks, hospitals, and public bodies whose data has to stay in a known city, on a private network, on hardware nobody else shares. Ask us for our current security documentation.
- Inference providers and neoclouds
- Wholesale bare metal in the middle of major cities, for providers who want to serve from close to their users without finding, fitting out, and running the sites themselves.
- Batch work
- Fine-tuning, evaluation, and bulk inference that can run in the hours when your reserved hardware would otherwise sit idle.
Security
Security here starts with the layout, before any software is involved.
- Your hardware is not shared, so there is no other customer on the same machine to be isolated from.
- Traffic reaches it over a private link from your own cloud account, not across the public internet.
- You know which city and which site your data is in, and it stays there.
- Each site is a locked service room in a managed commercial building, entered only by Trace staff under the building's access rules.
Request access
Tell us where your users are and what you want to run, and we will reply with what is available.