TL;DR: How Spheron's GPU Stock Alerts Work
- Scope: Each alert tracks the exact GPU count and region that failed, and starting the instance turns its own alert off.
- One email, not five: Alerts on the same returning GPU batch into a single email, capped at one per region every 30 minutes.
- Cross-region hint: Under a location filter, a sold-out card also shows where the GPU is still in stock, e.g. "In stock in USA and Norway."
- Spheron: H100 SXM5 on-demand runs $4.19/hr as of 11 Nov 2026. Check H100 availability before setting an alert.
Pricing fluctuates based on GPU availability. Spheron rates above are live as of 11 Nov 2026; other providers reflect their most recent published rates and may have changed. Check current GPU pricing → for live rates.
GPU Shortage 2026: Why a Start Gets Refused in the First Place
A refused Start isn't a bug. It means the region you picked has zero bookable units of that GPU count left, and on the current market that's a common state, not a rare one. CoreWeave's CFO said on the company's Q1 2026 earnings call, on 7 May 2026, that capacity was largely sold out, with 2027 now being allocated. Nebius has reported being sold out of available capacity in Q3 2025, Q4 2025 and Q1 2026, and its management has said four or more customers typically compete for every GPU it brings online.
Those are two of the GPU-first providers everyone points to as the deep-pocketed alternative to the hyperscalers, both telling shareholders the same thing: demand is ahead of supply. Spheron's own breakdown of the 2026 GPU shortage covers why, CoWoS packaging and HBM output are the structural ceiling right now, not GPU die yield, and why Blackwell-class parts like B200 run tighter than Hopper-class H100. None of that changes what a sold-out catalogue card means for your deployment, though: it blocks a new Start in that region, nothing more. An instance you already started keeps its allocation and keeps running; it isn't reclaimed because the catalogue later reads zero.
What's missing from most writing on this shortage, including pages that rank for the term, is the cloud side of it: what a renter is actually supposed to do with a sold-out card beyond refreshing it. CoreWeave, Lambda, and Runpod don't publish anything resembling a notify-me mechanism for capacity that comes back. A stock alert doesn't fix the underlying shortage. It just means you don't have to be the one refreshing a tab to catch it easing.
GPU Sold Out, Notify Me: Setting Up an Alert Step by Step
Setting an alert takes one click from either of two places: a sold-out catalogue card, or an instance card where Start just failed. Spheron then emails you the moment that exact GPU count and region is bookable again, no refreshing the catalogue and no waiting on a sales reply.
- Open the GPU catalogue and find the card marked "No stock right now."
- Click Notify me on that card. This sets the alert for the GPU type and region shown, with no extra configuration needed.
- Or, if you already tried to deploy, click Start on your configured instance. If the GPU is sold out in that region, the card shows a bell instead of launching; click it to subscribe to that exact configuration.
- Keep working. If you start the instance manually before the alert fires, Spheron turns the alert off for you, so you won't get a follow-up email for a GPU you already have.
- Watch your inbox. The email arrives once stock exists at the count and region you asked for, batched with any other matching alerts.
Say you need 2x B200 in the US region for a weekend benchmark run and the catalogue shows "No stock right now" on Friday afternoon. Clicking Notify me there takes less time than typing a support ticket, and it means you can close the laptop instead of reloading the catalogue every hour over the weekend. If the same 2x B200/US configuration is what someone else is also waiting on, your alert and theirs land as a single batched email when stock returns, not two separate pings competing for your attention.
What Triggers the Email, and Why You Only Get One
The alert fires the instant the catalogue shows stock at the GPU count and region you subscribed to, not on a polling schedule you have to guess at. Two rules keep it from turning into spam. First, when stock comes back for a GPU you have several alerts on, you get one email covering them instead of one email per alert.
Spheron's guide to reserved GPU capacity bidding covers how reserved slots and their floor prices work, if you're weighing one against the other.
Where Else the GPU Is in Stock: Reading the Cross-Region Hint
Waiting isn't always the right move, and Spheron's catalogue tells you that directly. Filter by location or provider, and a sold-out card says where it is sold out and where it is still in stock, something like "In stock in USA and Norway." If your workload isn't pinned to a specific region for latency or data-residency reasons, that line is often faster than any alert: move the deployment to wherever the hint points instead of waiting for your original region to free up.
It's also the honest limit of the alert itself. A Notify me email tells you stock exists; it doesn't hold it. If several other people get the same email and one of them clicks Start first, you're back to refused. The cross-region hint is the fallback for exactly that race: instead of refreshing a sold-out region, check whether the GPU you need is already open somewhere else.
That fallback matters more on tighter SKUs than looser ones. A100 and L40S tend to show up in stock across most regions at once, so a sold-out card for either is usually a short wait. B200 and other Blackwell-class capacity are the ones where one region can read zero while another sits open, which is exactly the gap the location filter is built to surface instead of leaving you to discover it by trial and error.
When to Alert vs When to Reserve Instead
A stock alert is the right tool when you need a specific GPU count and region for a bounded job, a benchmark run, a short fine-tune, a demo, and you can afford to wait hours or days for it to free up. It's email-only: no SMS, Slack channel, or webhook option, so it has to land somewhere you actually check.
It's the wrong tool when the requirement is guaranteed access rather than a chance at freed-up stock: a training run with a hard start date, or a launch that can't slip because the region you needed happened to sit at zero that week. For that, the on-demand catalogue isn't the product to lean on; a reserved capacity block or a bidding slot is.
| Need | Right tool | Why |
|---|---|---|
| Short job, flexible start, one region | Notify me alert | Free, one click, email when that exact count and region frees up |
| Short job, flexible region | Cross-region hint | Faster than waiting if the GPU is already open elsewhere |
| Hard start date, can't slip | Reserved capacity block | Guaranteed access, not a chance at freed-up stock |
| Want a slot that's currently taken | Reserved bidding subscription | Same email-on-open mechanic, applied to committed capacity |
Spheron's guide to negotiating a GPU cluster reservation contract covers the terms worth pushing back on, minimum commitments, overage rates, and capacity-guarantee SLAs, before you sign one. Either way, compare current on-demand rates across the catalog before deciding whether waiting on an alert is worth it or a reservation pays for itself faster.
A bell on a refused Start beats refreshing the catalogue, but it's still a race once stock reappears. Check B200 on Spheron if capacity's already open, or set the alert if it isn't yet.
Frequently Asked Questions
Click Notify me on the sold-out catalogue card, or on the bell that appears when a Start gets refused for a sold-out region. Spheron emails you once that exact GPU count and region is bookable again, and starting the instance first turns the alert off automatically.
No. Setting a Notify me alert costs nothing, you only pay the instance's normal hourly rate once you actually start it, $4.19/hr for H100 SXM5 on-demand as of 11 Nov 2026.
Nothing. A sold-out catalogue only blocks new Starts in that region; an instance you already started keeps running on its existing allocation and isn't reclaimed because the catalogue later reads zero.
Not yet. Spheron's stock alerts are email-only as of this writing, with no SMS, Slack, or webhook option, so set one on an address you actually check.
