sequenced.ai
Articles/Models & infrastructure/Blueprint//8 min read

How San Francisco Compute lets AI teams resell unused GPU time

A reservation buys time on specified nodes. A sale returns credits only when another buyer takes a window, so the job schedule belongs next to the order book.

By Sequenced deskAI-assisted, source-led · how we work
Visit San Francisco Compute website ↗
Node-hoursTrading unitOrders specify nodes and a time window.
Compute poolsAllocationPurchased time can outlive an instance.
Market ordersResale mechanismAn unused window needs a buyer.
Account creditsSale proceedsA filled sale returns credits after the platform fee.
San Francisco Compute mark
San Francisco Computesfcompute.com · independent research

Represent this company? Verify your work email to access its workspace, or send the desk a factual correction.

A training team may know when it wants to start a large run and still have little idea when it will finish. The input pipeline could need another week, or a promising early result could bring the experiment to a close. San Francisco Compute gives the team a market for the accelerator time left on a reservation. The financial question is what the lab can afford to reserve even if nobody buys the spare time.

In brief
  1. 01What is traded A contract specifies a node type, region, quantity and time window. The operator provides the cluster behind the contract.
  2. 02What is separate The customer buys compute time for a pool and creates instances to use that time. Recreating an instance does not undo the purchase.
  3. 03What resale returns A filled sale removes the purchased time from the pool and returns account credits after fees.

01 / ProductA GPU cloud with an order book for time

San Francisco Compute builds and operates clusters and offers virtual machines, bare metal and managed Slurm. Its order book lets buyers and sellers express which capacity they need and when. A buyer acquiring a window is buying the ability to run work on the described infrastructure. The company also says it operates the underlying capacity, which matters when a node needs attention or a distributed run needs a particular network. SF Compute infrastructure Operating the compute market

Think of the reservation and the instance as two different records. One describes the resources available over a period. The other is a running or bootable environment that consumes part of that allocation. SF Compute's preview quickstart describes purchasing a SKU for a named pool, selecting an image and creating an instance against that pool. The instance can be terminated and recreated while the purchased time continues. Check the applicable interface for a particular account before writing internal instructions. Compute pool quickstart, preview

That distinction changes what counts as unused capacity. An empty job queue does not automatically remove a reservation from the bill. Equally, shutting down an instance for a software update does not by itself sell its remaining allocation. Someone who knows when the next job needs those nodes has to decide what time can be offered and at what price. Once another buyer acquires a window, the original team must work around it.

02 / AudienceWhich workloads can make use of a time window

A research group running a finite sequence of experiments has a promising fit. The group can describe an earliest start, a latest useful finish and the node count that the distributed job can actually use. Evaluation runs, data preparation and review often separate rounds of training. Those gaps become candidates for resale only when the scheduler and the scientists agree that the nodes will genuinely be free.

An always-on prediction service has a different obligation. An idle replica may be the reserve needed to survive a traffic spike. Reselling its time converts a buffer the team already owns into a dependency on finding fresh capacity. A service with a strict response deadline should first identify the minimum footprint it will keep at all times. A separate pool for flexible experiments makes an easier market trial than a first experiment with production headroom.

A small company renting one accelerator should also check the unit of sale. The documentation shows an eight-GPU node in its example. An attractive figure per GPU-hour can describe a whole node that the buyer must be able to use. Compare the hardware and minimum allocation for the actual SKU with the number of workers the code can keep busy. A job waiting for a slow data reader will leave processors idle regardless of the headline rate. Compute pool quickstart, preview

03 / WorkflowA proposed capacity plan for a second training run

Begin with the time from the first run. Mark the intervals spent reading data, training on multiple nodes, writing checkpoints and waiting for a researcher to inspect outputs. Decide which of those stages truly needs the interconnected cluster. For a comparable second run, estimate a narrow plan based on the expected path and a separate plan that leaves time for a restart. Keep both visible to the person authorising the purchase.

Match the requirement to a specific SKU, region, node count and schedule. SF Compute describes vetted cluster specifications and offers information on interconnects and support. Confirm the topology for the intended group of nodes. A distributed training job that exchanges large amounts of data between workers may complete less useful work in an hour on an unsuitable network, even if the individual cards have the expected memory. SF Compute cluster specifications

Inspect the offer for the full interval before committing. If the needed configuration has no executable offer at the approved price, a standing order can wait for a match. The scientific schedule should then describe capacity as requested, rather than booked. An order that fills tomorrow may still be worthless to a run that had to start today. Record a deadline for cancelling an outstanding order that no longer serves the experiment. Compute pool quickstart, preview

Once the run starts, keep a copy of the last recoverable checkpoint outside the instance lifecycle. Test bringing up the software and reopening that checkpoint before promising another team access to the remaining window. Compare the estimated restart time with the interval between a completed sale and the next scheduled job. This is a proposed operating rehearsal. The provider's infrastructure description does not establish how fast any particular training program will recover.

04 / PricingHow a matched order becomes a cost or a credit

The terms describe contract specifications that include price, region, instance type, start, duration and node count. Transactions are funded or paid by card subject to the provider's conditions. The seller needs the company's prior authorisation to resell unused time under the marketplace terms. Confirm the account's permissions early if the business case depends on selling. The examples in the CLI guides demonstrate mechanics and do not establish an available price. Marketplace terms

EventWhat changesHow to budget it
Buy a windowA matched order secures specified node timeMultiply the accepted node-hour rate by the hours and node count
Offer spare timeA standing offer waits for demandAllow for no sale before the window expires
Filled resaleSold time leaves the pool and account credits arrive after feesUse the fill and fee estimate rather than the advertised ask

Compute purchase and resale, consulted 21 November 2026. Marketplace terms Resale and platform fees, preview

For illustration, two nodes reserved for ten hours represent twenty node-hours. If only one node is genuinely free for the final four hours, the lab can make four node-hours available. The gross value of that sale would be four times the filled node-hour rate. The preview resale guide describes both a fixed charge per sold node-hour and a percentage of sale value. The actual terms can differ by organisation, so check the displayed fee estimate. An offer with no buyer returns nothing. Resale and platform fees, preview

05 / DistinctionsWhat changes compared with simply renting accelerators

The CoreWeave Blueprint is a useful companion when evaluating a dedicated AI cloud. Compare delivery for one real job: network, supported image, expected node count, support coverage and the earliest usable start. Then add a second comparison for what happens when the job finishes early. Capacity that can be offered to another buyer has a different exit path from a commitment that simply runs to term.

The Lambda Blueprint provides another angle on obtaining GPU infrastructure. A team with short, independent jobs may care most about finding a suitable instance when each job appears. A team consolidating several labs' demand may be able to reserve a longer block and allocate time internally. Track internal use and external sales separately. A busy internal pool can justify a reservation even when there is little outside demand for its occasional gaps.

06 / QuestionsWhat could make available time difficult to sell

A sell order specifies hardware and a period. Four spare nodes spread across short, disconnected windows can attract different demand from the same number of hours in one continuous block. The preview resale guide allows partial fills by default and says buyers can take portions of the offered time or node count. Inspect exactly what sold before promising the remaining nodes to a researcher. One large offer can leave several small gaps. Resale and platform fees, preview

Read the hardware-failure and support arrangements against the cost of repeating work. SF Compute discusses repacking onto healthy nodes and refund arrangements for failures. A credit for machine time and a recovered training state are two separate outcomes. Store enough information about images, checkpoints and data versions to estimate the second. That estimate will help determine how much capacity to retain as an operational buffer. SF Compute cluster specifications

07 / DecisionWhich capacity deserves a reservation

01

A lab with a dated research milestone

Identify the minimum continuous run and the latest feasible start. Obtain a real offer for both the nodes and interconnect required, then budget that purchase assuming zero resale.

Treat an executed purchase as the point at which capacity is secured.
02

A team with predictable downtime

Check resale authorisation, review market depth for the specific SKU and schedule, then offer one genuinely spare interval. Compare the credits received with the time surrendered.

Base the next reservation on actual fills after fees.
03

A service with variable production demand

Separate capacity required to meet the service commitment from experiments that can wait. Practice restoring a job across a planned gap in the flexible pool.

Sell only time the service can afford to give up.
What should we explore next?

A business worth understanding.

Suggest your business or one you find interesting. Tell us what you want to understand about its product, positioning, design or workflows.

Suggestions are free. Selection and publication stay with the desk.

Sources
Filed under Models & infrastructureCompany San Francisco ComputeNot affiliated with San Francisco ComputeRequest a correctionRequest a refresh by email

Continue reading

All in this category