Assembling compute capacity before selling it

A cluster can have sixteen idle GPUs without being able to accept a sixteen-GPU job. Eight may sit in one network domain and eight in another, while the workload needs a single approved fabric. A price list can describe those fragments accurately and still offer nothing the buyer can use. The missing service is sometimes an agreed rearrangement of existing allocations.

The U.S. broadcast incentive auction offers a useful historical case. Its distinctive institution combined payments for relinquishing rights with an engineering process that made the resulting spectrum usable. For compute, the relevant question is whether an operator can assemble a deliverable block by compensating customers willing to move, while preserving every commitment that remains.

The market included a rearrangement

Congress authorized the broadcast auction in 2012. Broadcasters could voluntarily relinquish spectrum usage rights, including through channel sharing or a move from UHF to VHF. The FCC could also reassign television channels, subject to statutory protections, to reorganize the remaining broadcast band. Voluntary sale and channel reassignment were different parts of the institution. 47 U.S.C. 1452, especially subsections (a) and (b).

The auction's results were announced on April 13, 2017. Gross proceeds were approximately $19.8 billion. The process repurposed 84 MHz, of which 70 MHz went to licensed use; the two quantities should not be treated as interchangeable. FCC results announcement, April 13, 2017; federal spectrum report, May 2019, pages 5 and 55.

The essential coordination problem was that a station's departure affected the possible locations of other stations. Acquiring a cheap set of relinquishments was useful only if the stations that remained could be assigned channels under the interference constraints. Those dependencies made the identity of the sellers matter alongside their prices.

The FCC's reverse auction used descending offers. Feasibility checks determined whether stations could be accommodated in their previous bands as bids were processed; an unsuccessful check could freeze a station's offer rather than reduce it. Forward bidding tested demand for the newly available licenses. If the final-stage conditions were not met, the design provided for another stage with a lower spectrum-clearing target. Final bidding procedures, October 14, 2015, sections V and VII.

The computational component was concrete. The FCC published SATFC, software for checking whether channel reassignments were feasible. It combined a logical representation of the constraints with specialized solving methods. FCC SATFC repository, release notes through March 29, 2016. This was infrastructure for deciding which combinations could be offered, not merely a way to display bids.

A hypothetical compute rearrangement

Consider one operator with two sixteen-GPU fabric domains, A and B, using compatible hardware. Each has eight idle GPUs. A contains an eight-GPU job whose customer is willing to checkpoint and restart elsewhere. B contains another eight-GPU job that must remain in place. A new customer wants sixteen GPUs together for eight hours.

These are invented conditions, not an available provider offer. Assume the flexible job can run in B, its customer approves the move, and the operator can preserve tenant isolation, data-location requirements and software permissions. Moving that job into B's idle half would leave all sixteen GPUs in A available for the new customer. The initial inventory contains enough idle devices, but the rearrangement makes them usable as one block.

The operator should first test the move: checkpoint duration, restart success, storage access and performance in the destination domain. The new eight-hour service window begins only after migration and acceptance checks. If the customer needs an immediate start, or the existing job cannot tolerate the pause, the proposed trade is infeasible regardless of the offered price.

Suppose the new customer offers $400 for the block. The incumbent accepts an $80 credit for moving. Migration and validation cost $40. Assume a further $240 covers delivery costs and all foregone alternative revenue, including use of the otherwise-idle space in B. The remaining contribution is $40. Every dollar figure is hypothetical; this arithmetic is a way to expose required inputs, not a compute price estimate.

A second move costing $60 would make the same arrangement unattractive under those assumptions. So would discovering that the $240 allowance omitted valuable demand for B. The test concerns the whole rearrangement, including displaced opportunities. Counting the new buyer's payment as the entire benefit would overstate the gain.

What the analogy can support

The inference for compute is that a marketplace may need a joint allocation plan before it can quote useful supply. An operator could collect conditional offers to shift jobs, combine them with a buyer's requirements, and commit only when the complete placement is technically feasible and commercially acceptable. Such a service would sell access to a particular workable configuration rather than sum nominal idle hours.

A small trial need not reproduce the FCC's auction. With one movable job, direct negotiation and a scheduler may be sufficient. Competitive bidding becomes more interesting when several feasible rearrangements exist and their costs are privately known. That possibility does not establish that an exchange would outperform an operator's existing scheduling process.

The authority differs sharply. The FCC operated under a special statute; a compute intermediary gains no power to relocate tenants from this analogy. Its permissions must come from actual contracts and consent. A solver also cannot manufacture checkpoint support, acceptable application performance or customer willingness to move. Those are inputs that require evidence.

The timescale matters as well. Spectrum reorganization involved a 39-month transition period ending in July 2020, with some extensions completed by September. FCC Order 21-111, released October 25, 2021, paragraph 6 and footnote 20. Compute availability can expire while a migration is still being prepared. A rearrangement that is eventually feasible may be worthless for a buyer's start deadline.

The practical experiment is therefore narrow: one operator, explicitly movable workloads and tested destination placements. Record incremental contribution, displaced revenue, migration failures, missed deadlines and useful work delivered. Include proposals rejected as infeasible. The useful lesson from spectrum repacking is that assembly costs and compatibility constraints belong inside the market's allocation decision. Whether paying for that assembly creates worthwhile compute supply remains an empirical question.