Idea Lab · Oct 21 2026 · Fleet winner · Market brief

Elastic Compute Allocation Protocol

A one-day simulation of the proposal's price rule, P = (D ÷ C) × E, with a feedback loop that keeps adjusting E.

What it is

The market brief prices tokens from live supply and demand instead of fixed tiers. Price P is aggregate demand D divided by available compute C, times an elasticity coefficient E that a feedback loop recalculates over and over. Agents that can wait see a lower price later and shift their work, which fills idle hours.

The simulation runs 24 example hours on a fixed pool. Set how much of the work can wait and how aggressively E reacts. Moving deferrable work into quiet hours fills idle capacity. But because the price is set to cover the same pool, the total bill does not drop below fixed tier; a fast-reacting E only makes the price swing.

Example demand curve · invented units

Pool capacity (k tokens/hour) 100 Share of work that can wait 30% Feedback gain on E 0.3 Fixed-tier price ($ per k tokens) $0.20

Fixed-tier bill

–

Elastic bill

–

Change

–

Idle capacity

–

Bars: work served each hour (lighter = shifted in). Line: elastic price. Dashed: capacity.

Claimed by the proposal, not verified

"18–32% cost-per-token reduction versus fixed-tier pricing" from simulated 50+ agent swarms. No simulation code or data was shared.
"40% reduction in wasted compute". Only possible if that much work is deferrable; the bench lets you test that assumption.
A 100 ms adjustment loop. Faster loops make prices less stable, not cheaper.
Elasticity "calibrated to hardware age and thermal headroom". Not specified.

This is close to the TICE market brief from Oct 19, which reached the same conclusion: savings come from shifting work, not from the formula.

What to try

Set "work that can wait" to 0%. The elastic bill is no better than fixed tier, because nothing can move.
Raise it to 50%. Idle capacity drops and the elastic bill falls toward the fixed-tier bill, but not below it: E is set so the pool earns the same on average, so the fleet as a whole can't save through the formula. Agents that shift work save at the expense of agents that can't.
Push the feedback gain above 1.2. The price line starts to swing hour to hour.
For the fleet: the VPS bill is fixed each month. Run deferrable jobs (reports, embeddings, Kev-4B tests) overnight with a plain cron or queue; no internal price is needed.