A flexible data center starts with the customer contract
A workload may be technically movable while its customer deadline remains fixed. A credible flexible-compute service needs agreed completion windows, recovery capacity and a clear benefit for the customer accepting delay.
These are companion notes. The full essay lives on The Turing Pilgrim, my Substack publication.
Why this matters
The customer agreement defines some of the scheduler's operating boundaries. An overnight simulation might have room to start later, yet still need continuous execution and a firm morning deadline. A useful service makes those limits explicit before the operator offers the resulting flexibility to an energy partner.
Recovery is part of the commitment. Deferring a four-hour job from one in the morning to three consumes all the slack before a seven o'clock deadline. The operator must accommodate unfinished work alongside incoming jobs, account for the power ramp and keep enough capacity available for retries. A successful reduction during the grid event does not establish that the whole service worked.
A narrow trial can test whether the arrangement creates value for both parties. The provider needs to count customer incentives, recovery capacity, operating costs and service failures alongside energy benefits. Some customers may prefer predictable timing, and some workloads may offer too little slack to justify a separate service. The full article develops these tradeoffs through hypothetical examples and distinguishes research models from commercial evidence.
Key takeaways
- Agree completion deadlines, permitted start windows, interruption limits and data-location restrictions with customers before committing capacity.
- Carry the agreed terms into the scheduler so operational decisions reflect the service that was sold.
- Plan the recovery period alongside the reduction event, including new arrivals, retries and competing demands for spare capacity.
- Evaluate delivered work, deadlines met and net operating value alongside the electricity reduction, using an agreed measurement boundary.
- Start with willing customers and explicit limits. Reassess available flexibility when urgent work, outages or backlogs change the operating conditions.
Who should read it
Product, commercial and operations leaders designing compute services that can respond to grid conditions while keeping customer commitments.
Related field notes
The exception queue is the real interface for industrial AI
Make evidence and ownership available when an operating commitment needs human judgment.
The next AI shift is delegation
Define the boundaries around what software may do on our behalf.
When AI learns physics
Consider how physical constraints shape the decisions industrial models can support.
Working through this decision?
A short note on the AI bet you are weighing is enough to start. I usually reply within one business day.