Spheron is a marketplace. It aggregates GPU capacity from a network of certified partner data centers, publishes the cheapest live rate for each card, and bills per minute. It is not a decentralized compute network, whatever the older framing suggests, and it should be compared as a broker rather than a DePIN project.
Teams look elsewhere for one structural reason. The listed rate is the cheapest supplier with capacity right now, which means the rate can move between reading the page and deploying, placement depends on who has capacity, and the operator is a partner rather than Spheron. That is a fine trade for a training run you start once, and a harder one for production inference.
Six alternatives below, split by whether they run their own hardware. Competitor prices carry the date they were verified. Runpod rates pull live.
What to check before you switch
Whether the number is a floor or a rate. Spheron and TensorDock both publish “from” or cheapest-live prices. Everyone else here publishes what you pay. That difference matters more than a few cents.
Who answers when something breaks. On a marketplace, the hosting partner operates the hardware and support routes through them. That is the trade you are making for the lower headline number.
Whether you can pick a region. Marketplace placement depends on available capacity, which affects latency and data residency.
Spot against on-demand. Never compare a spot rate to an on-demand rate. Spot capacity can be reclaimed. If you want a lower-cost option that is not preemptible, that is a different product.
The billing increment. Per second, per minute and pay-as-you-go against a deposited balance are three different things on a bursty workload.
1. Runpod
Runpod is the strongest option if you want the marketplace price range without the marketplace variability.
Runpod operates its own fleet across 31 global regions and publishes 21 NVIDIA cards from the RTX A5000 up to the B300, plus AMD MI300X at $2.39/hr on Secure Cloud, verified 31 August 2026. The published rate is the rate, on hardware Runpod runs, in a region you select with no surcharge for choosing one.
Pods bill per second, which is finer than Spheron’s per-minute increment and matters on short inference calls. Serverless runs inference that scales to zero with sub-200ms cold starts via FlashBoot, which Spheron does not publish an equivalent for. Clusters provision multi-node self-serve up to 64 GPUs, against Spheron’s 24 to 48 hour brokered turnaround.
Community Cloud is the lower-cost tier, starting at {{gpu:rtx-a5000:community}}/hr for an RTX A5000. It is not preemptible, which makes it a different proposition from a spot tier. Runpod is SOC 2 Type II certified and HIPAA and GDPR compliant, with reports, BAAs and DPAs available through the Runpod Trust Center. On a brokered marketplace, compliance posture belongs to whichever partner hosts you.
Where it stops: no public spot tier, and no GB200, GB300 or GH200. Spheron publishes all four.
Best for: production workloads where knowing the operator and holding the rate is worth more than the lowest listed number.
Full detail in our Runpod vs Spheron comparison.
2. TensorDock
The closest structural match to Spheron, and the right comparison if a marketplace is what you actually want.
Their own FAQ is direct about it: a marketplace of independent hosts who compete and set their own pricing. Every published figure is a floor, and the same GPU carries multiple prices depending on host, location and redundancy. H100 SXM5 from $2.25/hr, A100 SXM4 from $1.80, RTX 4090 from $0.35, entry consumer GPUs from $0.12, all verified 20 August 2026.
45 GPU models across 100+ locations in 20+ countries, KVM virtualization with root access and Windows support, and no ingress or egress fees. Hardware is a mix of Tier 3 and Tier 4 data centers and, in their own words, converted mining rigs.
Where it stops: pay-as-you-go runs against a deposited balance and servers are automatically deleted when the balance reaches $0. That is a real operational risk worth designing around. They hold hosts to a 99.99% uptime standard, which is a standard applied to hosts rather than an SLA offered to you.
Best for: the widest hardware selection at floor prices, if you can tolerate host variability.
3. CloudRift
An operated fleet at marketplace-adjacent prices, billed per second with no minimum.
On-demand rates verified 20 August 2026: RTX PRO 6000 $1.29, RTX PRO 6000 Max-Q $1.55, A100 SXM4 80GB $1.05, RTX 5090 $0.65, L40S $0.63, RTX 4090 $0.39, MI350X $3.65. Reserved tiers are selectable in-console at 5% off for a week, 10% for a month, 15% for three months. No egress, ingress or API-call charges, and local storage is included in the hourly rate.
Read their VRAM column before quoting anything. Their H100 row lists 48GB rather than the usual 80GB, and the RTX 4090 and RTX PRO 6000 rows list 96GB, which is not a standard 4090 configuration. Their FAQ also says H100 and H200 are not orderable in the console and require contacting them.
Where it stops: the H100 and H200 are reserved-only and not self-serve, which is the opposite of what most people leaving a marketplace are looking for.
Best for: RTX PRO 6000 and consumer-class cards at low rates from an operator rather than a broker.
4. Verda, formerly DataCrunch
The best answer if what you liked about Spheron was the spot tier.
Published spot rates at roughly half the on-demand rate on every card, verified 31 August 2026: H100 SXM5 $3.25 on-demand against $1.63 spot, H200 SXM5 $4.00 against $2.00, B200 SXM6 $6.11 against $3.06, A100 SXM4 80GB $1.79 against $0.895. Reserved discounts run to 25% at two years. European, SOC 2 Type II and GDPR compliant, an NVIDIA Preferred Partner, with instant clusters and serverless containers alongside instances.
Storage is $0.20 per GiB per month, which is high once converted.
Where it stops: on-demand rates sit above Spheron’s marketplace floors on most overlapping cards, and the storage pricing is expensive for data-heavy work.
Best for: a real spot tier with a named operator and European data residency.
5. Thunder Compute
Small, simple and cheap on the cards it carries.
Billed per minute, with only four GPU models published, verified 20 August 2026: H100 PCIe 80GB $2.19, A100 80GB $1.09, L40 48GB $0.79, RTX A6000 48GB $0.35. 1x through 8x configurations. First 100GB of storage included while running, then $0.03 per 100GB per hour. Additional vCPUs $0.04 each per hour. No egress fees.
Where it stops: four GPUs is the whole catalog. No B200, no H200, nothing above an H100, and no serverless inference.
Best for: A100 and A6000 workloads where the rate is the only thing that matters.
6. Voltage Park
H100 only. $1.99/hr on Ethernet, $2.49/hr with 3200 Gbps InfiniBand, verified 20 August 2026. Self-serve, provisioned within 15 minutes, no minimum term, scaling 1 to 1,016 GPUs on demand. No ingress, egress or support charges.
Where it stops: there is no second GPU, and nothing under $1.99.
Best for: H100 capacity at scale, bought without a sales call.
Frequently asked questions
Is Spheron a decentralized compute network?
Not as currently positioned. Their pricing page describes a marketplace aggregating live rates across partner providers, with supplier matchmaking and certified data centers. USDT and USDC payment remains, but the capacity being sold is commercial data center supply.
Which alternatives are marketplaces rather than operators?
TensorDock is explicitly a marketplace of independent hosts. Runpod, CloudRift, Verda, Thunder Compute and Voltage Park all operate their own fleets, so the published rate is what you pay and the operator is who you contact.
Which has a real spot tier?
Verda publishes spot at roughly half its on-demand rate on every card. Spheron publishes spot on several cards at 13% to 48% off. Runpod has no public spot tier; Community Cloud is the lower-cost option and is not preemptible.
What is the finest billing increment available?
Runpod and CloudRift both bill per second. Spheron and Thunder Compute bill per minute. TensorDock is pay-as-you-go against a deposited balance.
Which has the most GPU models?
TensorDock publishes 45 across 100+ locations. Runpod publishes 21 NVIDIA cards plus AMD MI300X. Spheron lists 10+. Thunder Compute publishes four and Voltage Park one.
Does anyone charge for egress?
None of the platforms above publish egress fees. Runpod, TensorDock, CloudRift, Thunder Compute and Voltage Park all state there are none.
Get started
If the marketplace variability is what you want to leave behind, the test is your own workload against a published rate on hardware one company operates. Runpod bills by the second with no minimum and no quota request. See current pricing or deploy a Pod.
