The Cheaper Fuel That Can Run Out

Opening Scene

Some ports sell surplus fuel at a steep discount, but with one significant condition: the supply can run out or be reclaimed at any moment, without meaningful advance warning, if a higher-paying buyer needs it. This fuel is genuinely useful for a voyage that can pause and wait when supply runs short, but genuinely dangerous for a voyage that absolutely cannot afford an interruption. Spot and preemptible instances in the cloud follow this exact same logic.

In Plain English

Spot instances (called preemptible instances on some clouds) are spare compute capacity sold at a steep discount, often 60-90% below on-demand pricing, in exchange for the cloud provider’s right to reclaim that capacity with little or no warning when it’s needed elsewhere. Spot instances are genuinely valuable for workloads that can tolerate interruption gracefully, and genuinely risky for workloads that cannot.

The Old Way

Before spot instances were a well-established, deliberately used tool, deeply discounted but interruptible capacity wasn’t yet a widely available pricing option:

• Organizations paid on-demand or reserved rates uniformly, without a meaningfully discounted option available for interruption-tolerant workloads.
• There wasn’t yet a well-established practice of architecting workloads specifically to gracefully handle sudden compute interruption.
• Deeply discounted, interruptible capacity, when it existed at all, was often treated as too risky to use for genuinely production workloads.

Uniform, non-discounted pricing without an interruption-tolerant, deeply discounted alternative is what spot and preemptible instance pricing directly addressed.

What’s Changing (and Why AI Is the Reason)

1. Organizations increasingly architect specific workloads — batch processing, fault-tolerant distributed training, non-urgent data pipelines — deliberately to run on spot instances, capturing significant savings.
2. This connects directly to the autoscaling patterns covered in Article 9, which often work in tandem with spot instance strategies to handle interruption gracefully.
3. As AI training workloads increasingly use checkpointing to save progress periodically, more training workloads have become genuinely tolerant of spot instance interruption, expanding where this discount can safely apply.

The Metaphor, Fully Extended

The Ship’s Engineer	Cloud FinOps Concept
Surplus fuel sold at a steep discount at some ports	Spare compute capacity sold at a steep discount as spot instances
The supply reclaimed at any moment without meaningful warning	Capacity reclaimed by the cloud provider with little or no warning
Genuinely useful for a voyage that can pause and wait	Genuinely useful for workloads that can tolerate interruption gracefully
Genuinely dangerous for a voyage that cannot afford interruption	Genuinely risky for workloads that cannot tolerate sudden loss of capacity

Quick Recap

• Spot instances offer steep discounts in exchange for the cloud provider’s right to reclaim capacity with little warning.
• This trade is genuinely valuable for interruption-tolerant workloads and genuinely risky for those that aren’t.
• Checkpointing and graceful interruption handling expand where spot instances can safely be used.
• Spot instances are a deliberate architectural choice, not a universal cost-saving default.

Where This Fits in the Series

Article 8 covered the tradeoff behind deeply discounted, interruptible capacity. Article 9 turns to a related practice: throttling engine output up and down as demand actually shifts.