It raised prices on its two most popular cards, the RTX 4090 and RTX 5090, across every paid priority tier. That's the news. The more useful story is what sits underneath it: SaladCloud's headline rate comes from a distributed network of idle consumer GPUs rather than dedicated data center capacity, and understanding that model is what actually tells you whether a $0.16/hr RTX 4090 is comparable to a dedicated-capacity rate or a different product entirely.
This post breaks down what changed, SaladCloud's full 2026 rate card by GPU class and priority tier, and how it stacks up against Spheron's RTX 4090 GPU rental pricing. If you're new to how GPU rental marketplaces work in general, our guide to what a GPU cloud is covers the basics this post builds on.
TL;DR: SaladCloud Pricing 2026 and RTX 4090 Cost vs Spheron
- What changed: raising RTX 4090 and RTX 5090 rates the same day.
- RTX 4090: SaladCloud High priority rose to $0.330/hr and Lowest (Batch) held at $0.160/hr as of 12 Sep 2026; Spheron's RTX 4090 on-demand rate is $0.48/hr as of 28 Sep 2026.
- RTX 5090: SaladCloud High priority rose to $0.500/hr and Lowest held at $0.250/hr; Spheron's RTX 5090 on-demand rate is $0.78/hr.
- The catch: SaladCloud's on-demand catalog has no A100 or H100, and its community-fleet nodes carry no formal SLA, unlike Spheron's.
- Verdict: SaladCloud's cheapest tier wins on raw dollars for retryable batch jobs; compare both on the GPU pricing page before choosing.
SaladCloud Pricing 2026: What Changed on September 12
SaladCloud publishes its rate changes directly, and the September 12, 2026 update is a clean before/after: most of the catalog got cheaper, and the two GPUs everyone actually rents got more expensive, according to SaladCloud's own announcement of the changes. That combination is the whole story for anyone comparing SaladCloud pricing in 2026, so it's worth separating the two halves before looking at the full rate card.
The 15 GPU Classes That Got Cheaper (16-38% Cuts)
Fifteen GPU classes dropped in price, with an average cut of 25% at High priority and individual reductions ranging from 16% to 38% depending on the class, per SaladCloud's pricing changes post. The affected list covers the RTX 5080, 5070 Ti, 5070, 5060 Ti, 5060, 5090 Laptop, 4080, 4070 Ti Super, 4070 Ti, 4070, 4060 Ti, 3090 Ti, 3090, and 3080 Ti. The RTX 5080's Lowest-tier rate is a concrete example: it fell from $0.180/hr to $0.150/hr in the same update, a discount that updates the $0.219/hr figure this site's own RTX 5080 vs RTX 5090 comparison cited earlier this year, before this repricing.
SaladCloud didn't publish a per-class old/new breakdown for every card on that list beyond the RTX 5080 figure above, so the specific dollar move for the other 14 classes isn't independently verifiable from the announcement alone. The 16-38% range and 25% average are what SaladCloud reported for the group as a whole.
Alongside the GPU changes, SaladCloud's CPU-only compute pricing moved too: vCPU-hour pricing rose from $0.004 to $0.005, while RAM pricing held steady, according to the same announcement.
Why RTX 4090 and RTX 5090 Went Up Instead
The RTX 4090 and RTX 5090, SaladCloud's two most heavily used cards, moved the opposite direction from the rest of the catalog. SaladCloud's announcement puts it plainly: "we've had to raise High, Medium, and Low Priority prices on these two classes," without giving a stated reason for the divergence from the broader discount trend. What's verifiable is the size of the move, covered in full in the rate card section below.
SaladCloud's Distributed GPU Model: Why the Price Is Low and What You Give Up
SaladCloud's prices sit well below what a dedicated data center GPU marketplace charges, and the reason is structural rather than promotional: the underlying hardware is a network of individually owned consumer machines, not provisioned server racks. That's also the reason the pricing model comes with priority tiers instead of a single flat rate.
Priority Tiers: High, Medium, Low, and Lowest (Batch)
SaladCloud sells the same GPU at four different price points depending on how much preemption risk a workload can absorb. High priority is never preempted, though it can still be interrupted by a node disconnecting, in which case SaladCloud automatically reallocates the workload to another node. Medium and Low priority can be preempted for higher-priority work, with a roughly 30-minute minimum runtime guarantee, and Lowest priority is built specifically for retryable batch jobs, according to SaladCloud's pricing page. Billing is per second, and it only starts once an instance is actually in the running state, meaning allocation, image download, and container startup are free.
That tier structure is why the same RTX 4090 can cost $0.160/hr or $0.330/hr depending on what you're willing to risk, a spread that doesn't exist on a dedicated-capacity platform with a single on-demand rate.
The Reliability Numbers: Uptime, Reallocations, and No Formal SLA
The nodes behind SaladCloud's Community Cloud are individually owned consumer machines, and they can disconnect for reasons that include the owner voluntarily reclaiming the device, power or network or hardware failure, or SaladCloud proactively reallocating a node to a higher-priority workload during a demand spike, per SaladCloud's own service performance documentation. There's no formal SLA attached to that model; reliability is a property of the fleet's statistics, not a contractual guarantee.
Those statistics are genuinely good, and SaladCloud has published them. In its own multi-day benchmark, daily node reallocations declined from more than 45 per day to fewer than 15 per day, and mean node availability, the share of requested instances actually running, exceeded 99% with a measured minimum of 94%, according to SaladCloud's benchmarking report. Average uptime across all instance runs in that benchmark was 60 hours, and even runs that were eventually interrupted averaged 35 hours before that happened, per the same report. That's a fleet that behaves well in aggregate. It's just a different kind of promise than a contractual uptime SLA, and the gap between the two is what a reader is actually paying for when they choose one model over the other.
SaladCloud GPU Pricing 2026: Full Rate Card by Class
Here's the confirmed rate card for SaladCloud's highest-traffic consumer cards after the September 12, 2026 update, all figures from SaladCloud's pricing changes announcement and pricing page.
Consumer RTX Cards, Lowest to Highest Priority
| GPU | Lowest (Batch) | Low | Medium | High |
|---|---|---|---|---|
| RTX 4090 | $0.160/hr | $0.217/hr | $0.273/hr | $0.330/hr |
| RTX 5090 | $0.250/hr | $0.333/hr | $0.417/hr | $0.500/hr |
| RTX 5080 | $0.150/hr | Not published | Not published | Not published |
RTX 4090 and RTX 5090 rates above are the post-September 12 figures across all four tiers. The RTX 5080's Lowest-tier rate reflects the same update; SaladCloud's announcement didn't break out its Low, Medium, and High tier changes individually.
Pricing fluctuates based on GPU availability. Spheron rates in this post are live as of 28 Sep 2026; SaladCloud rates reflect its September 12, 2026 published announcement and may have changed since. Check current GPU pricing → for live rates from both.
Why There's No A100 or H100 on the On-Demand Catalog
SaladCloud's self-serve, on-demand catalog is consumer RTX GPUs only. Data center accelerators like the A100 and H100 aren't listed for self-serve rental at all; they're only available through SaladCloud's separately contracted "Dedicated" product, priced individually through its sales team rather than a published hourly rate, according to ComputePrices' breakdown of Salad's GPU catalog. For anyone comparing SaladCloud pricing against a marketplace that lists A100 and H100 as bookable, self-serve SKUs, that's a real gap in the product, not a missing line in a rate table.
GPU Cloud Pricing Comparison: SaladCloud vs Spheron by GPU Class
The direct comparison only really works GPU-by-GPU, since the two platforms don't sell the same catalog. Here's how SaladCloud's published rates line up against Spheron's live on-demand pricing for the cards both platforms actually offer.
| GPU | SaladCloud High priority | SaladCloud Lowest (Batch) | Spheron on-demand | What you're paying for |
|---|---|---|---|---|
| RTX 4090 | $0.330/hr | $0.160/hr | $0.48/hr | Spheron: 99.9% uptime SLA, no preemption risk. SaladCloud Lowest: cheapest raw rate, retryable jobs only |
| RTX 5090 | $0.500/hr | $0.250/hr | $0.78/hr | Same trade-off one tier up the consumer stack |
| A100 / H100 | Not on self-serve catalog | Not on self-serve catalog | $1.43/hr / $2.64/hr | Spheron: bookable, self-serve SKU. SaladCloud: sales-quoted Dedicated contract only |
Pricing fluctuates based on GPU availability. Spheron rates above are live as of 28 Sep 2026; SaladCloud rates reflect its September 12, 2026 published announcement and may have changed since.
For a workload that genuinely tolerates a node disappearing mid-run, SaladCloud's Lowest tier is a legitimately lower dollar-per-hour number than any dedicated-capacity platform, Spheron included; that's not a claim worth dressing up. Where Spheron's marketplace earns its higher rate is the other side of the trade: a 99.9% uptime SLA.
This is one slice of a much wider market; our broader GPU cloud pricing comparison covers rates across 15+ providers for readers who want the full picture beyond just these two. If SaladCloud and Spheron aren't the only two providers on your shortlist, our Runpod vs Vast.ai pricing comparison covers the other pair of marketplace-style clouds readers usually put next to SaladCloud, and Vultr's GPU pricing is worth a look too if you need data center GPUs SaladCloud simply doesn't carry on-demand.
When SaladCloud's Model Fits (Batch, Fault-Tolerant) and When It Doesn't
The pricing comparison above only answers half the question. The other half is whether your workload can actually use SaladCloud's cheapest tier, because the Lowest-priority rate and the High-priority rate are effectively two different products wearing the same GPU name.
Good Fits: Retryable, Stateless, Interruption-Tolerant Jobs
SaladCloud's Lowest (Batch) tier is a strong fit for work that's naturally retryable: rendering queues, dataset preprocessing, model evaluation sweeps, offline batch inference, and any job architecture where a checkpoint-and-resume pattern already exists. Those workloads absorb a reallocation as a retried task, not a failure, which is exactly what the tier is built for and priced for.
Bad Fits: Production Inference, SLA-Bound Workloads, and Anything Needing A100/H100
Production inference behind a user-facing request, anything with a contractual uptime commitment, and any workload that needs an A100, H100, or larger data center GPU on a self-serve basis all fall outside what SaladCloud's on-demand catalog is built to do. A reclaimed node mid-inference is a dropped or delayed response to a real user, and there's no SLA on SaladCloud's side to fall back on when that happens. For those cases, a dedicated-capacity marketplace with a published uptime commitment and a data center GPU catalog, Spheron's RTX 4090 rental and RTX 5090 instances among them, is the more comparable product even at a higher hourly number.
SaladCloud's cheapest tier is genuinely the right call for retryable batch work; for anything that needs a card that won't disappear mid-run, Spheron's live GPU pricing is the place to compare rates before committing.
Frequently Asked Questions
By raw sticker price, SaladCloud's Lowest (Batch) tier is close to the floor: RTX 4090 at $0.160/hr and RTX 5090 at $0.250/hr as of SaladCloud's September 12, 2026 repricing, with some legacy consumer cards advertised from $0.015/hr. Those rates come from a distributed network of individually owned machines with no formal SLA, so they're cheapest only for work that can tolerate a node disappearing mid-run. For dedicated capacity with an uptime SLA, Spheron's RTX 4090 on-demand rate is $0.48/hr as of 28 Sep 2026, its cheapest listed GPU.
It depends heavily on the card and the reliability tier. Consumer RTX cards on community-fleet providers like SaladCloud run roughly $0.15-$0.50/hr depending on priority tier, while dedicated-capacity marketplaces charge more for a guaranteed node: Spheron lists RTX 4090 on-demand at $0.48/hr and RTX 5090 at $0.78/hr as of 28 Sep 2026. Data center GPUs like H100 or A100 cost more again and aren't available at all on SaladCloud's self-serve catalog. Check a live pricing page before budgeting since marketplace rates move with supply.
There's no single best GPU independent of the job. For retryable, stateless, fault-tolerant batch work, a cheap preemptible tier on a community-fleet provider is genuinely the best economic choice. For production inference or anything with a latency or uptime commitment, a dedicated-capacity marketplace with an SLA is the better fit even at a higher hourly rate, since a reclaimed node mid-inference costs more in broken requests than the price difference saves.
As of SaladCloud's September 12, 2026 update, its Lowest tier holds RTX 4090 at $0.160/hr and RTX 5090 at $0.250/hr. On the dedicated-capacity side, Spheron's on-demand rate for RTX 4090 is $0.48/hr as of 28 Sep 2026. Because both figures are marketplace rates that move with availability, verify them on the provider's live pricing page before committing budget.
