AWS put its idle servers up for auction rather than let them burn money

the problem

Idle compute perished by the hour

background

what everyone would do

Offer a reserved-capacity discount for customers willing to commit upfront, or simply accept idle racks as the unavoidable cost of being provisioned for peak load -- the two standard responses to overcapacity, neither of which extracts any revenue from capacity nobody has committed to buying in advance.

what they saw

Idle compute isn't unsellable, it's just unsellable at the fixed on-demand price -- a server sitting empty for an hour is inventory that perishes the moment that hour ends, exactly like an empty airline seat or a hotel room nobody books for tonight, and those industries had already solved this with auction-based yield management decades earlier.

the move

Spot Instances: auction idle compute capacity that perishes every hour it sits unused.

why it works

Splitting demand into two markets -- a fixed-price tier for customers who need guaranteed availability, and a fluctuating-price auction for customers who can tolerate interruption -- lets AWS sell the exact same physical capacity to two different willingness-to-pay segments without discounting the main product for everyone. Workloads like batch jobs, crawls and transcoding don't care when exactly they run, only that they eventually finish, so they can absorb whatever slack the on-demand market leaves unsold; every dollar collected from an interruptible workload running on capacity that would otherwise sit empty is close to pure margin, since the server's fixed cost was already being paid for the reserved and on-demand tiers.

the payoff

Became the standard mechanism for monetising slack across cloud providers.

where it breaks

The mechanism only works for workloads genuinely tolerant of interruption -- anything requiring guaranteed uptime or continuous execution can't use the discount tier at all, which caps how much idle capacity the auction can actually monetize. It also depends on price transparency the customer can act on: AWS's own 2017 shift from a sealed-bid auction to an undisclosed-algorithm retail price, which research found came with a broad price increase rather than the drop a pure reliability improvement should have produced, shows that once the auction mechanism itself is removed, the same interruptible-capacity idea can quietly become a price lever working against the customer instead of a genuine two-sided market.

what came after

Google later offered its own reduced-price preemptible tier on the same interruptible-discount structure, and spot markets are now a standard way large-scale batch and ML workloads buy cheap cycles. The "auction" itself did not survive: in November 2017 Amazon replaced the sealed-bid market entirely with a retail price it sets and does not disclose the algorithm for, and researchers who compared prices before and after found the change was accompanied by a broad price increase — averaging roughly 37% to 61% across AWS's North American regions — not the drop a pure reliability cut should have produced.

references

1. [1]Amazon Web Services Announces Spot Instances for Amazon EC2Amazon (press release), 2009press.aboutamazon.com
2. [2]Analyzing AWS Spot Instance PricingIEEE IC2E / UC Santa Barbara, 2019sites.cs.ucsb.edu

keep it

same kind of clever

• 1985 · American AirlinesAmerican Airlines, bled by no-frills discounters, learned to sell the same seat at five prices
• 1991 · MarriottMarriott took the airlines' playbook and priced every room by the night it would die
• 1911 · Aalsmeer Flower Auction (Royal FloraHolland)Aalsmeer sells its dying flowers by racing a price downward instead of bidding one upward