Spot GPU Instance Walkthrough: Launch One on Spheron (2026)

You don't need a tutorial to click "deploy" on a cloud console. You need one when a single wrong checkbox means your instance gets reclaimed mid-job with no warning and you lose three hours of training. This spot GPU instance walkthrough is the exact click path for launching one on Spheron, including the one checkbox that's easy to miss and the price-cap control that turns an open-ended risk into a bounded one.

TL;DR: Spot GPU Instance Walkthrough for Spheron

Sign up, fund your balance, add an SSH key, then deploy a spot-labeled GPU, accepting spot terms once.

• Funding first: Billing requires a minimum $15 (card) or $50 (crypto) before deploying.
• One extra click: your first spot deploy requires ticking a spot terms checkbox; every deploy after shows the same terms panel above the Deploy button.
• Safety net: Spheron ES's max-price control (v1.31.5, 13 Oct 2026) auto-stops an instance when spot price spikes past your cap, keeping disk and IP.
• Discount range: Spheron says spot runs up to 50% off.
• Decision deferred: whether spot fits your workload is covered in Spot GPU Instances in 2026: How Much You Save (and What Breaks).

What a Spot GPU Instance Actually Is on Spheron

A spot instance on Spheron is the same NVIDIA hardware you'd get from a dedicated deploy, listed in the same catalog, at a price that moves instead of staying fixed. In the Deploy GPUs catalog, spot offers are labeled as interruptible and show a spot price sitting next to the regular dedicated price for the same GPU, so you're comparing both numbers before you commit to either, per Spheron's instance-types docs.

The mechanic that makes spot cheaper also makes it reclaimable: you're renting capacity the provider isn't currently selling at the dedicated rate, and it can be taken back. On Spheron ES specifically, that spot price isn't static. It's set per GPU type and region and can reprice as often as every 15 minutes, a live-repricing mechanism that started on 8 October 2026, documented in Spheron's changelog. The one constraint that holds through all of that movement: a spot instance is never priced above the dedicated rate for the same machine in the same region. A running spot instance bills at whatever the current rate is as it moves, not the rate you started at.

Coverage isn't uniform across the aggregated marketplace, and this matters before you pick a GPU to deploy. Spheron AM's AMD MI300X capacity (v1.29.0, 10 Sep 2026) and the newest Spheron AI capacity added in Noida and Chennai, H100 SXM, H200 SXM, RTX PRO 6000, and L4 (v1.31.0, 26 Sep 2026), shipped on-demand only, with no spot tier at all. If a GPU you want doesn't show a spot price in the catalog, that's why. There's also a split in what spot supports once it's running: spot machine types on Spheron AI can't be stopped, while Spheron ES spot instances support full stop and start, including the price-cap auto-stop covered below. Check which backend you're deploying onto before you plan around either capability.

Spot GPU Instance Walkthrough: Step-by-Step Launch Guide

Here's the full flow, start to finish, based on Spheron's getting-started docs and the instance-types docs for the spot-specific steps.

1. Sign up at app.spheron.ai using Google, GitHub, or email.
2. Add credits under Billing. The minimum is $15 by card or $50 by crypto through CopperX, capped at $5,000 per transaction. No instance, spot or dedicated, deploys without a funded balance.
3. Add an SSH key in User Settings, before you deploy. Keys attach at deploy time only; you can't bolt one onto an already-running instance later.
4. Go to Compute > Deploy GPUs and browse the catalog.
5. Pick a GPU and choose the spot option instead of dedicated. The catalog shows the spot price next to the dedicated price for that same GPU and region, so you see both before deciding.
6. Tick the spot terms checkbox. This appears once, on your first-ever spot deploy. Every spot deployment after that shows a spot terms panel directly above the Deploy button instead, as a standing reminder rather than a one-time gate.
7. Configure region, OS, and storage. Ubuntu 22.04 LTS is the recommended OS; storage defaults are fine for most jobs unless you know you need more.
8. Add an optional startup script if you want the instance to run setup commands automatically on boot.
9. Review the Order Summary. This shows the hourly cost at the current spot rate, your remaining balance, and the full configuration before you commit.
10. Click Deploy Instance. Provisioning takes about 30-60 seconds.
11. Connect once it's live. The instance dashboard shows the IP, username, and SSH command. Run ssh <user>@<ip>, then nvidia-smi to confirm the GPU is actually there and visible.

That's the whole flow. Steps 1-4 and 7-11 are identical to a dedicated deploy; steps 5 and 6 are the only places spot diverges from dedicated at all.

Setting a Max Price Cap So a Spike Doesn't Drain Your Balance

A floating spot price is fine when it's falling and a problem when it isn't, since a running instance keeps billing at whatever the current rate becomes. Spheron ES addressed this directly with a spot max-price control, shipped in v1.31.5 on 13 October 2026, per the changelog.

You can set a maximum hourly price either during deploy or later, from the instance's Set max price or Edit control. The options are no limit (the default), 10%, 25%, or 50% above the current spot price, or a custom figure you type in yourself.

If the live spot price rises above whatever cap you set, the instance auto-stops. The pause is recorded with pauseInitiator: "price-cap", it keeps its disk and IP exactly where they were, it bills the stopped rate instead of continuing to run, and you get an email. Starting the instance back up is blocked for as long as the live price stays above your cap, which is the point: the control is there to stop the bleeding, not to let you restart straight into the same spike.

This is the feature that makes a Spheron ES spot deploy meaningfully different from a hyperscaler spot auction, where a price spike just costs you more with no equivalent self-set ceiling. It's worth repeating that this is a Spheron ES mechanic specifically: it isn't available on spot machine types through Spheron AI, which can't be stopped at all, price-triggered or otherwise. If you're deploying on Spheron AI's spot capacity, treat reclaim risk as unmitigated by this control and lean harder on checkpointing instead.

What Happens on Preemption, and How to Checkpoint Around It

There's a real difference between the price-cap stop above, which is self-triggered, documented, and preserves your disk, and a provider-initiated reclaim, which isn't the same kind of event. Spheron's spot-instances feature page doesn't publish a grace period or warning window for a reclaim, and recommends regular checkpointing on exactly that basis: there's no grace period to count on.

That's a meaningfully more conservative posture than even the already-short notice windows hyperscalers publish. AWS gives exactly a 2-minute Spot Instance interruption notice, delivered through EventBridge and instance metadata, before stopping or terminating a Spot Instance, per AWS's own documentation. Two minutes is already tight for a shutdown hook to flush state cleanly. With no published window on a Spheron reclaim at all, the only safe assumption is zero warning, same as if you were running on a spot type with no lifecycle controls whatsoever.

A couple of lifecycle details matter here regardless of how the stop happens. Every instance on Spheron carries a 20-minute minimum billed runtime by default, and only running time counts toward it, stopping early doesn't serve out the rest. And a hard restart (reboot) loses anything that only lived in memory, so the docs' advice is to write checkpoints and flush open files before restarting rather than relying on a clean in-memory handoff, per Spheron's instance-lifecycle docs.

Practically, that means:

• Checkpoint on a timer, not on a signal. Since a reclaim notice isn't documented, don't write your checkpoint logic to wait for one. Save state every N minutes or N steps, whichever is shorter for your job.
• Write checkpoints to somewhere that survives the instance, not just the disk the GPU was attached to. Spheron's persistent volumes, covered in Spheron's April 2026 product update, are the obvious target: they're detachable from a terminated instance and reattachable to a fresh one.
• Know the difference between Stop and Terminate before you pick either. A Stop (including the price-cap auto-stop above) keeps your disk; a Terminate doesn't. Spheron's Stop and Start guide covers the full billing and dashboard mechanics of the Stop path in detail.
• Treat async, retryable jobs as the natural fit for spot, and anything a user is waiting on synchronously as a worse one, regardless of which provider you're on.

Spot vs On-Demand vs Reserved: Picking the Right One Per Job

This walkthrough is deliberately just the deploy flow. The "should I actually use spot for this job" question, including the savings math and failure modes across providers, gets the full treatment in the companion post linked in the TL;DR above. The short version worth carrying into your decision: spot is for work that can restart cleanly from a checkpoint, on-demand is for anything that can't, and reserved capacity is for workloads with high, predictable utilization where paying for guaranteed access beats paying the spot-vs-dedicated spread hour by hour.

If you've already landed on spot and want to go further than a single manual deploy, the next step is routing jobs to whichever provider's spot pool is cheapest at a given hour instead of sitting on one. GPU Spot Instance Arbitrage: Bidding, Failover, Forecasting covers the cross-provider scheduling and failover mechanics for that, once a single spot instance stops being enough. And if spot turns out to be one lever among several you're pulling on a GPU bill, The GPU Cloud Cost Optimization Playbook places it alongside instance selection, checkpointed training, and idle-GPU cleanup.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 10 Nov 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.

Launching a spot instance takes the same four setup steps as a dedicated one, plus a checkbox. The price cap and checkpointing habits above are what keep a reclaim from costing you more than the discount saved.

Deploy a spot GPU on Spheron →