Arbitraging spot and preemptible capacity.

This is a portability play, and it sits inside the broader question of where workloads should run. Here is how to capture the saving without buying instability.

What is the discount, and why is it a floating rate?

Every provider holds spare capacity it would rather sell at a discount than leave idle. AWS Spot, Azure Spot Virtual Machines, GCP Spot VMs, and OCI preemptible instances all sell that surplus at 60 to 90 percent below on demand list, in exchange for the right to reclaim it when paying demand returns. AWS gives a two minute interruption notice; GCP and Azure give roughly 30 seconds; OCI preemptible instances terminate on reclaim. These figures are indicative and move with supply.

Because the price tracks live supply and demand, it is a floating rate, not a committed one. That is the opposite of a Savings Plan or a Reservation, where you lock a rate by carrying utilization risk. Spot carries interruption risk instead. The two levers are complementary: cover your steady baseline with commitments, and run the elastic, interruptible top layer on spot.

Which workloads can take an eviction?

The single rule that governs spot economics: the cost of an interruption must be smaller than the discount it buys. That favours work that is stateless, fault tolerant, and cheap to restart.

• Strong fit. Batch processing, data pipelines, CI and build runners, video rendering, machine learning training with checkpointing, and Kubernetes worker pools behind a queue. Losing a node means the scheduler reschedules the task, and the only cost is a little wasted compute.
• Poor fit. Primary databases, stateful singletons, and latency sensitive request paths. An eviction mid request triggers retries, error budgets, and on call pages, and the operational cost dwarfs the rate saving.

The middle ground is large. Many production services run perfectly on spot when fronted by a queue, spread across instance families to dodge correlated reclaims, and backed by a small on demand or committed floor so the service never drops to zero capacity.

When does cross cloud arbitrage actually pay?

Moving a workload to whichever provider has the cheapest spare capacity sounds like free money, but the egress and engineering costs usually eat it. Data transfer out of a provider is billed per gigabyte, and rebuilding the same workload to run portably on four clouds is real engineering effort. The arbitrage pays only when the workload is already portable, the dataset it needs is small or already replicated, and a scheduler makes the placement decision automatically.

In practice the highest return version is single provider, multi pool: let a Kubernetes autoscaler or a managed spot service spread across many instance families and availability zones within one cloud. You capture most of the discount without paying egress every time the price shifts. True cross cloud spot is worth it for a narrow set of stateless, data light batch workloads where portability already exists.

A worked example

Make spot a deliberate layer, not a gamble

We help enterprises decide which workloads belong on spot, set the on demand and committed floors that keep them safe, and automate the placement so the discount holds without an engineer watching it. Our guarantee: we reduce your cloud spend or we reimburse our service fee. Pricing is either a Fixed Fee scoped up front or Gainshare, a share of verified savings with no retainer and no risk.