The discount is bigger than most people budget for
Check the live spread on H200 NVL, RTX 3090, V100, A100 PCIe, RTX 4090, A100 SXM 40GB, or MI300X and you’ll find community and spot tiers pricing more than 70% below secure, SLA-backed capacity on the same card, as of this writing. That’s not a rounding error you eat for convenience. On a month of steady inference traffic, that spread is the difference between a project that pencils out and one that doesn’t.
The reflex is to treat that discount as too good to trust. It isn’t a pricing mistake. It’s a different product: the provider reserves the right to reclaim the box on short notice, usually because a secure customer or their own reserved-capacity commitment needs it back. You’re renting the leftover, and the leftover is priced like leftover.
Stateless workloads collect the discount for free
Inference serving through vLLM or SGLang holds almost no state that matters between requests. A preempted node drops whatever’s mid-flight, the load balancer retries against another box, and the caller sees added latency on a fraction of requests. There’s no checkpoint to lose because there’s nothing to checkpoint.
This is why community and spot tiers are the correct default for serving open models like Qwen3 30B A3B or GLM-4.5 Air: the failure mode is a retried request, not lost work. Run the math on /calculator with your actual request volume against the secure-tier floor on /gpus, and the spot discount usually swamps the retry tax by a wide margin unless your traffic is unusually latency-sensitive.
Long-context agentic calls are the one serving pattern where this gets shakier. Kill a request halfway through a 40-step tool-use chain and it doesn’t retry cleanly; it restarts the whole chain from the orchestrator’s last saved state, which for most agent frameworks is the beginning of the task. If your workload is agents running long sequential tool calls rather than single-shot completions, weight the interruption risk higher than the raw discount suggests.
Training is where the objection lands
The honest counter-argument: training and fine-tuning runs carry optimizer state, gradient accumulation, and multi-GPU coordination state that a serving box never has to worry about. Lose the node mid-step and you lose everything since the last checkpoint, not just the in-flight request.
Consider the concrete failure. A team fine-tuning against a base like Llama 3.3 70B sets checkpoint intervals at 60 minutes to cut disk I/O and storage cost, reasonable on secure capacity where interruptions basically don’t happen. On a spot node with a preemption every 90-120 minutes, that 60-minute interval means losing up to an hour of compute on every reclaim, sometimes twice in a session. The discount evaporates once you’re paying for the same GPU-hours twice to redo lost epochs.
This is the boundary that matters: checkpoint overhead as a fraction of job runtime, not the sticker discount, decides whether spot pays off for training. If your checkpoint save-and-restore round trip runs under roughly 5% of the interval between expected preemptions, spot wins outright and the discount dominates. If it’s higher, the resume tax eats into or exceeds the savings, and secure tier is the better bet even at full price.
The B300/B200 trap
Tier pricing on newer silicon doesn’t move in lockstep with the generation label. B300 spot capacity has, on days we’ve tracked, priced below B200 secure, despite B300 being the newer part with more memory bandwidth on paper. Providers reclaim B300 capacity for internal training runs more aggressively than B200, which dumps more idle inventory onto spot markets at a steeper discount than the generational gap would suggest.
The trap is assuming “newer generation costs more” holds across tiers. It doesn’t. Pull the current numbers on /gpus before committing, and check /trends to see whether that inversion is a one-day anomaly or a pattern for the specific card you’re renting. Buy B300 spot on the assumption it’s automatically pricier than B200 secure and you can leave real savings on the table.
The rule
Serving traffic on open models: default to community or spot tiers, checkpoint concerns don’t apply, and the discount is close to free money as long as your orchestration retries cleanly. Multi-hour or multi-day training and fine-tuning: compute your checkpoint round-trip time as a percentage of your expected mean-time-between-preemption on that specific tier and card (check the tier explainer in the guides section for how providers define preemption windows). Under 5%, take the spot discount. Over that, pay for secure and stop re-running lost epochs.
What this can’t tell you is your provider’s actual preemption frequency this week. That number moves with whoever’s holding a reserved-capacity contract on the same rack, and it isn’t published anywhere consistent. Track your own interruption log for a week before committing a training budget to spot; the average on /gpus is a floor price, not a preemption forecast.