SPACE: The Cost of Keeping Paused AI Work

Read the cost of continuing work through an open runtime

On 15 October 2026, Perplexity open-sourced SPACE, an environment for AI code execution. Its description includes pausing and resuming work and recovering state.^[1][2]

Consider the release through a choice about preserving work already done. A calculation awaiting a person’s response, an interrupted investigation or a document whose next conditions are undecided can leave work that is inactive now but may need to continue later. Comparing only the time spent running misses part of that demand.

Keeping working state, retaining completed outputs alone and removing the environment to recreate it later leave different costs during the wait and different tasks on return. The comparisons that follow concern operating choices; they do not describe a particular service’s prices or retention policies.

An output and the state behind it have different uses

Imagine a job that has finished aggregating a table and now awaits instructions about the next comparison. The table can be read on its own. Continuing from the same point may require intermediate processing and the materials consulted. What is sufficient to keep depends on the next task.

If the purpose is to deliver a finished document, keeping that document may be enough. If the purpose is to continue the analysis under changed conditions, an output alone can leave preparation to repeat. Before comparing the size of stored items, establish where the progress worth retaining resides.

Checking whether an output is correct has a cost separate from preparing to continue the work. The conditions for carrying evaluation evidence across a change inform the decision to use that output. Even when it is suitable for the next step, the location of the materials needed to proceed can change the preparation involved.

Three retention choices distribute future work differently

Keeping working state preserves progress for resumption while accepting the burden of storage and management during the wait. Keeping only outputs protects readable results but leaves environment preparation until continuation is needed. Removing the environment reduces current retention work but may require the task to run again if it returns.

Declaring one choice always cheaper can mistake a transfer of costs for a reduction. A short calculation that is easy to repeat differs from an investigation requiring materials to be gathered again, even over the same waiting period. Keeping an unused environment and deleting useful progress can both be costly.

The relevant decision occurs before resumption is certain. How much burden should be accepted now for work that might never return? Looking only at jobs that did resume makes retention appear convenient. Including those that never returned allows the retention choice as a whole to be assessed.

Recreation costs more than repeating a calculation

Estimating recreation solely through repeated computing can bias the comparison. Finding materials again, aligning changed inputs and recovering intermediate decisions may also require work. A machine might repeat its processing quickly while a person still has to reconstruct the assumptions.

Conversely, when the original task has been organized into clear procedures and outputs, returning may require little investigation. The value of saved state depends on the future work deletion would require, rather than the total time spent in the past. Development costs already incurred should not simply be added to the benefit of retention.

Running again can also serve another purpose. If inputs need updating, preserving the former environment may not avoid that update. Separating the steps retention can save from those needed under every choice helps avoid both overstating and understating its value.

Where a review pauses affects what is worth keeping

If temporary work unnecessary to the next decision can be consolidated before a pause, less state may need to be retained. When the next question remains broad, reducing everything to outputs might instead create more investigation later. Costs depend on where work stops as well as how long it waits.

Defining the question in advance can therefore be valuable. Knowing whether the next decision concerns only a change in the population being aggregated or also new research helps identify progress needed for continuation. Between keeping everything and deleting everything lies a choice about useful stopping points.

It also costs time and effort to keep progress ready for handover while allowing people time to decide. Version and handover records help trace materials and conditions. A document version number does not necessarily preserve a job’s working state, so preparation should be compared for each item needed to continue.

Does a longer wait justify the same retention choice?

Keeping progress may be convenient for a brief response delay, while a long pause changes the comparison. Retention burdens continue and the task’s purpose or materials may change. Carrying the first day’s decision forward indefinitely can mean continuing to pay for progress that is no longer needed.

A falling likelihood of return does not automatically make deletion sensible. Rarely resumed work may still be worth keeping if preparation would be difficult to repeat. Even near-certain return may not justify retaining state when preparation is easy and outputs can carry the work forward.

A uniform short-versus-long cutoff therefore leaves out why the outlook changed. A postponed response differs from a project approaching cancellation, even if both add a week. Retention is spending intended to reduce future work, and changes in that future belong in its reassessment.

When the request changes, which progress still helps?

A new purpose on return can change the value of different parts of progress. Gathered materials may remain useful while an unfinished output needs replacement. If the subject changes, the ease of starting fresh may outweigh the convenience of returning to the former environment.

Deciding what to retain solely by the job’s name is therefore unhelpful. Consider which parts of collecting, organizing, calculating and explaining are affected by the change. The technically preservable range need not be the economically useful range.

Earlier conditions checked and rejected approaches can still inform the next decision. That knowledge need not require an entire environment to survive. Distinguishing knowledge transferable into documents and outputs from progress worth keeping in its working state can narrow retention without discarding useful learning.

The ability to recreate work changes the retention decision

Investment in making work reproducible can make deletion easier to choose. With organized materials and procedures, a business can move to its next task without retaining an old environment indefinitely. Improving how work is done thus affects retention costs through a route distinct from better storage features.

Preparing for recreation itself consumes time, however. Detailed rerun procedures for a short job that returns only once may cost more than keeping it. Requiring every task to be fully reproducible before it pauses can place too much work ahead of the wait.

Modifications that use existing designs connect past knowledge with the work needed to make a change. Mapping the materials retained from unfinished AI work to the steps needed to prepare it again likewise helps compare spending on retention and reproducibility. Expected repetition and the difficulty of preparation affect where to spend time first.

Restoring local state need not reverse external work

Suppose a job sent materials to another party midway through a calculation. Returning the local workspace to an earlier state would not erase the recipient’s receipt of those materials. Assessing continuation therefore requires considering both preserved progress and work already done elsewhere.

Repeating the delivery might create additional checks; sending a revised output might instead be necessary. The state from which computing resumes most quickly need not be the state from which business with the recipient continues with least effort. Keeping a record of completed handovers alongside outputs can reduce later duplication and confusion.

The ability to continue work may involve both restarting a process and connecting an output already handed over with the next one. An estimate limited to internal restoration leaves out checks by the receiving team and explanations that must be sent again. The benefit of retaining work should include these adjustments on both sides of the handover.

The number of retained jobs differs from the number running

A retention choice for one task does not yet establish the service-wide burden. Few jobs may be running while many await responses, accumulating saved progress. Conversely, a stream of short completed tasks can produce many executions with little state kept for long.

This is a distinction between an amount held at a point in time and an amount processed over a period. Combining retained, started and completed jobs into one usage count obscures which burden grew. The basics of matching a number to its subject and comparison period help separate state retention from demand for processing.

More retained progress proves neither greater value nor greater waste by itself. It could reflect a new ability to continue long tasks or a failure to clear work that will not return. Understanding which future jobs the state supports helps explain the next spending decision.

Responses arriving together create a different burden

The same inventory of saved jobs can create different execution demand when jobs resume in sequence or all at once. A meeting ending, or one approval enabling several tasks, can cause work that appeared dispersed while waiting to restart together.

Making retention cheaper can encourage more progress to be kept. If jobs previously terminated are retained instead, their future resumption demand still needs separate consideration. Better storage efficiency alone cannot establish an equal proportional increase in the number that can run together.

Businesses also have choices. Distinguishing work that must resume immediately from work that can queue may reduce the need for equal execution headroom behind every saved job. Promising internally that every retained task can restart instantly can instead accept a capacity burden larger than retention’s benefit.

How long resumption can wait changes the value

A job that can continue tomorrow and one whose continuation is needed immediately place different demands on operations, even if the same state survives. The former may allow execution to be spread over time. The latter may require capacity free from competing work when it returns.

Price comparisons should therefore consider acceptable time to continuation alongside retention duration. Cheap storage does not ensure a lower total burden if a later queue loses a valuable opportunity. Buying constant readiness for work that can wait can conversely mean paying for unused headroom.

Relating speed to task deadlines is more informative than arranging products along one speed ranking. A small group needing immediate continuation and a large group that can be processed later allow retention scope and execution commitments to be considered separately. This does not infer guarantees in any actual contract.

Separate the retention estimate from the resumption estimate

An operating estimate can include the burden of keeping work over time, the resources used when it returns and people preparing its continuation. Even if one monthly fee combines them, the underlying work does not respond identically to changes in use. Explaining which task grows first when use increases has value.

Applying the same estimate to users retaining many jobs and users repeatedly running short jobs can make one group bear some of another’s burden. This does not prescribe a charging method. A predictable flat price and a more precise reflection of usage differences each offer benefits and require administrative work.

The execution conditions and verification scope of advertising code show why the same label can cover different offerings. Estimates for continuing work also become easier to compare when they specify retention periods, the number of resumptions and the outputs to be handed over. Before combining these in one price, identify where the burden differs between users.

How far can different users share their waiting time?

When different users return to work at different times, common execution resources may be used in sequence. Pooling can offer benefits relative to each business maintaining the same spare capacity independently. More users do not, however, automatically mean more dispersed demand.

Shared deadlines, regional working hours or one external publication can bring resumption times together. Estimates based on smoothed counts hide that concentration. Resources can be idle on average and still face competition at the moment continued work is needed.

Provider scale is therefore difficult to assess solely through state counts or customer numbers. Understanding which tasks can shift in time and which return together is also valuable operating work. Users able to explain their deadlines help make capacity requirements concrete and can reduce unnecessary demands for immediacy.

Which progress is shared when several alternatives remain open?

When several approaches are compared within one investigation, repeating preparation for each differs from using progress common to them. Separating progress worth keeping before a choice from processing needed only afterwards may preserve alternatives while reducing unnecessary execution.

Keeping more alternatives is not always beneficial. If each holds different materials and assumptions, little may be shared, and identifying what was eventually selected adds work. More stored items do not establish whether more progress was reused or merely more alternatives required management.

Running alternatives together and trying them sequentially also create different execution peaks despite identical retention volumes. A short decision deadline may require parallel execution to obtain the benefit of preserving several approaches. Compare the cost of keeping alternatives with the cost of running them to make a choice over the same decision period.

Where does people’s work move when progress is handed over?

Preserving state can spare machine processing while leaving a person to reread earlier decisions. Conversely, an organized output alone may make continuation easier for another worker to understand. Keeping the whole environment and enabling someone to begin the next task should not be counted as identical benefits.

The recipient matters too. The original worker may remember context that another department needs explained. As retention lasts longer, the question of who can interpret unfinished progress becomes an economic condition alongside the burden of storage.

Expected productivity gains are not determined when one processing run is avoided. Include the work of preparing explanations, understanding continuation and using results when asking how much effort the same objective now requires. Retaining an intermediate environment and organizing outputs for handover allocate people’s time differently.

Retention prices can change what users keep

When users do not directly encounter the burden of retention, keeping things just in case may become more attractive. When even brief retention appears costly, users may remove progress that would be difficult to recreate. Prices recover costs but can also influence how much progress survives.

The same separation can occur when one team chooses what to retain and another pays. Convenience alone may drive retention, or a budget-led clearance may remove useful work. Explaining what preparation would be avoided on return helps both parties make the choice.

This does not imply that retention should always be metered finely. More detailed charging also creates estimation and checking work. A contract may preserve simplicity while making long-lived state and unusually demanding resumptions visible, improving comparisons without multiplying every small decision.

What should move when the provider changes?

A possible future provider change belongs beside the benefit of keeping an environment for a long time. Outputs may be transferable even when intermediate state cannot simply resume elsewhere. Availability of public code alone does not establish portability of every retained job.

Distinguish finished outputs, materials gathered so far and state needed to continue. Moving outputs and starting fresh differs from transferring intermediate progress in both migration effort and later preparation. What deletion would require recreating matters more than stored volume alone.

Relying on customized operations over time can replace some benefit of a common implementation with migration preparation. Customization may nevertheless be rational when its value to particular work exceeds that cost. Public availability does not determine one optimal retention arrangement for every business.

Read market effects through computing and retention demand separately

Execution counts alone can miss part of the demand connecting AI use with cloud markets. If more long tasks continue through pauses, the work of retaining progress and allocating execution when it returns may change too. Usage counts do not determine where the spending goes.

More efficient execution might encourage uses that keep more work. Conversely, retaining outputs alone may prove sufficient for growing numbers of executions, leaving long-term state retention behind. Following changes in the work kept is more concrete than turning a technical improvement automatically into either more or less computing demand.

How much continuity should a smaller user buy?

A user with limited operating expertise can benefit from purchasing preparation needed to continue long jobs. It need not demand the same continuity for every task. Recreating short calculations while preserving progress on preparation-heavy work may narrow the external services required.

A public implementation expands the choices available for a common foundation. Selecting retained jobs and coordinating their return during concentrated demand still require operations. Comparing in-house provision with purchased services needs to include continuing work during waits and resumptions alongside initial development.

Even when authority to order further investigation and the handover process are clear, the person continuing the work still needs materials. Smaller teams may face a greater burden when resumption depends on someone’s memory. Aligning the outputs handed over with the state retained could reduce dependencies that otherwise leave other work waiting for the decision maker’s return.

The strong objection: if recreation is easy, why keep state?

When rerunning is quick, materials stable and procedures clear, the benefit of carrying state may be small. Retaining useful outputs and beginning a new job could be simpler than managing an ongoing retention period. The availability of continuity technology does not imply that its use should always expand.

Some work can also carry old assumptions forward by returning to an earlier state. If gathering fresh materials makes the next request clearer, deletion offers more than a loss of progress: it can organize the next task. The benefits of beginning again belong in the retention comparison.

This objection weakens when recreation requires substantial preparation that retained progress genuinely avoids. Knowing the work likely to return makes selection more concrete. Neither rerun speed alone nor the time previously spent is enough to choose universally.

What comparison establishes the benefit of continuity?

One way to evaluate retention is to compare continuing the same objective with saved progress against preparing again from outputs. The relevant difference concerns repeated steps before continuation becomes useful, beyond startup time alone. Differences in tasks or inputs cannot simply be attributed to retention.

Counting only returning jobs omits the burden of keeping state that proved unnecessary. Allocating equal preparation costs to every waiting job instead hides differences between heavy and light work. Viewing retention periods, return rates and preparation tasks over a common interval helps identify where preservation was useful.

Values depend on conditions of use; a general feature description cannot establish an individual business’s savings. Deciding in advance which findings would support more state retention and which would support outputs alone helps avoid a policy sustained only by convenient success stories.

Three changes in use create different spending needs

First, more jobs may return quickly and reach completion using earlier progress. The benefit comes through preparation avoided, rather than the number saved. Follow how much repeated machine processing and human rereading actually decline.

Second, more jobs may never return, making the organization of stored work a central task. Instead of merely targeting lower volume, identifying progress that can be consolidated into outputs may reduce the burden without creating more recreation later. Differences in what is kept change the next spending need.

Third, many jobs may return together, making time to continuation the constraint. Additional retention features might then matter less than execution headroom or dispersing work able to wait. The same expansion in continuing use can move work toward storage, organization or execution, changing providers’ costs and the services on which they compete.

SG Group View: choose what to keep through the work it will save

SG Group assesses paused state as progress retained to begin a possible future continuation. Maximizing saved jobs or eliminating waits is not the objective itself. Connect keeping state, consolidating outputs and recreating work to the steps each can save and the retention burden accepted now.

This also clarifies when the ability to pause creates value in continued use. People gain time to make necessary decisions; if work can then continue without repeated preparation, the result offers a different use from short uninterrupted processing. Someone still needs to accept the storage and resumption work supporting that benefit.

Look for explanations of continuation, beyond feature counts

When reading a provider’s description, consider the conditions for continuing paused work and how they differ from handing over outputs alone. Actual retention periods, usable state and preparation on return help connect the service with the buyer’s tasks. A shared product name does not establish identical operating conditions.

For a provider’s business, useful evidence concerns the growth of continuing work and the spending accompanying it. More retained jobs, more resumed jobs and concentrated execution can create different costs. Alongside customer numbers and simple execution counts, understanding the work a service accepts helps explain revenue quality.

Public feature descriptions do not determine customer-specific prices, retention policies or preparation actually avoided. Future descriptions of use conditions and task-level comparisons could clarify when to retain more and when to rerun. Rather than relying on an unconfirmed publication schedule, identify which decision such evidence would change.

Final assessment: stopping execution leaves work to supply

Viewing AI work solely as a succession of short calls can conceal the burden of tasks continuing through human response delays. Putting the benefit of preserved progress beside the work of holding it makes the services required by continuity more visible.

An open runtime provides a starting point for that comparison. Relate the outputs kept, the duration of state retention and the timing of continuation to the business’s actual work, rather than choosing a purchase through one feature label. Changing that combination can also change providers’ revenues and responsibilities.

Market effects need not appear simply as faster processing or a uniform decline in computing demand. Businesses able to relate retained progress to work that will resume may reduce both unnecessary retention and unnecessary repetition. The cost connecting work before and after a pause provides a way to assess continuity’s value.