TensorDock Pricing 2026: Cheapest H100 Rental Cost vs Spheron

TensorDock's H100 listing starts at $1.90/hr, which puts it near the front of almost every "cheapest GPU cloud" comparison table. But that number is a range, not a rate. TensorDock is a marketplace: independent hosts list their own hardware and set their own prices, and what you pay (and what uptime you get) depends on which host your job lands on. This post breaks down TensorDock's actual on-demand, spot, and reserved pricing tiers, what the marketplace model costs you in support and consistency, and where Spheron's managed per-minute rate is the better trade for production workloads. For the closest structural comparison on the blog, another marketplace with host-set pricing, see our Vast.ai pricing breakdown.

TensorDock's Marketplace Model: Host-Set Pricing and What That Means for Reliability

TensorDock says it plainly on its own site: "We're a marketplace of independent hosts who compete and set their own pricing." That's a structurally different model from a managed cloud. There's no single TensorDock rate card. Each host, whether a verified data center operator or a smaller colocation provider, lists an H100 at whatever price they choose, and TensorDock's platform layer handles billing, deployment, and search across that inventory.

Billing runs hourly by default, a la carte for extra CPU, RAM, and storage beyond the base GPU allocation. TensorDock deducts your account balance continuously while a server runs, and servers auto-delete once the balance hits zero, so there's no invoice cycle to manage but also no grace period if you forget to top up mid-job. Deployment is fast: TensorDock advertises servers launching in about 30 seconds with a minimum deposit of just $5, and the catalog spans 45 GPU models from consumer cards up to enterprise-grade hardware.

The marketplace model is also the source of TensorDock's biggest trade-off. Because pricing and hardware quality are host-set, two H100 listings on the platform right now can differ meaningfully in both price and the maintenance discipline behind them. TensorDock requires hosts to schedule maintenance at least two weeks in advance and claims a 99.99% uptime standard on paper. In practice, GPU Cloud List's TensorDock review measured observed uptime around 99.1% on TensorDock's data-center-verified hosts and closer to 97% on community hosts, a gap that matters a lot more on a 48-hour training run than on a five-minute smoke test.

TensorDock Is Now Part of Voltage Park, but the Marketplace Still Runs Independently

TensorDock was acquired by Voltage Park in March 2025. If you've evaluated TensorDock before and are checking whether that changes anything, the short answer is: not much, structurally. TensorDock founder Jonathan Lei moved into the role of GM of On-Demand at Voltage Park, while Melissa Du, formerly Voltage Park's Director of Customer Experience, became GM of TensorDock. The marketplace kept its own brand and its own host network.

Lei put it directly in the Voltage Park announcement: "Voltage Park will remain one of many participants on the TensorDock marketplace." Voltage Park's own H100 and H200 capacity now lists alongside third-party hosts rather than replacing them, which means the pricing dynamics you'd evaluate on TensorDock today, host-by-host variability, marketplace bidding, mixed reliability tiers, are unchanged from before the acquisition.

TensorDock H100 On-Demand and Spot Pricing (2026)

TensorDock splits H100 access into three pricing tiers: on-demand, marketplace bid, and reserved. Each targets a different commitment level, and the spread between the cheapest and most expensive tier is wide enough to change your effective cost by more than 30%.

On-Demand Rates: $1.90-$2.50/hr Staggered by Host

TensorDock's own H100 page lists on-demand pricing staggered from $1.90 to $2.50/hr per GPU, with $2.25/hr showing up as the rate most commonly quoted in comparisons and TensorDock's own marketing. The spread exists because on-demand listings are still host-set. A data-center-verified host running newer hardware in a well-connected region can charge toward the top of that range and still get picked; a smaller operator competing purely on price lists near the floor.

This matters for planning. If you're comparing TensorDock's advertised "$2.25/hr" against a fixed platform rate elsewhere, you're comparing a midpoint against a guarantee. The actual host you get assigned, and the price you pay, depends on what's available when you deploy.

Spot/Marketplace Bid Pricing: $1.30-$1.91/hr

TensorDock's marketplace bid tier is the spot-equivalent option, with a minimum bid around $1.91/hr and some listings reported as low as $1.30/hr depending on current supply. As with any bid-driven marketplace, the headline low end is real but not guaranteed to be available when you need it. TensorDock is explicit that "actual pricing fluctuates based on market conditions," which is the same caveat you'd apply to any spot market: the price you see on a quiet Tuesday isn't the price you'll get during a demand spike.

For workloads that checkpoint regularly and can tolerate interruption, this tier is where TensorDock's pricing gets genuinely competitive with the cheapest options in the market. For anything that needs a guaranteed instance for the full run, the on-demand or reserved tiers are the honest comparison points.

Reserved Pricing Tiers (Monthly, Annual, 3-Year Commit)

TensorDock's reserved tiers reward longer commitments with a predictable, lower rate: $2.00/hr on a monthly commitment, $1.90/hr on an annual term, and $1.50/hr on a 3-year commit, TensorDock's cheapest published H100 rate across any tier. TensorDock also sells bare-metal reservations at the full 8-GPU hostnode level, from $16.00/hr monthly down to $12.00/hr on a 3-year term (each per-GPU reserved rate corresponds to roughly 1/8th of the advertised node resources).

The trade-off with any reserved tier, on TensorDock or elsewhere, is that you're locking into a specific host's hardware and uptime track record for the length of the term. A 3-year commit at $1.50/hr is a strong rate on paper, but it only pays off if the host behind it stays online and doesn't change terms mid-contract.

TensorDock vs Spheron: On-Demand and Spot H100 Rates Side by Side

Spheron's live pricing for H100, pulled from the GPU rental marketplace on 23 Jul 2026: H100 SXM5 on-demand is $3.92/hr per GPU with spot at $2.91/hr per GPU, and H100 PCIe on-demand is $2.01/hr per GPU, all billed per minute.

Spheron's H100 PCIe on-demand rate at $2.01/hr lands inside TensorDock's own on-demand range, and below TensorDock's commonly quoted $2.25/hr figure, with a platform-level uptime guarantee instead of a host lottery. Spheron's SXM5 on-demand rate at $3.92/hr sits above every TensorDock H100 tier, and that's the fair comparison to make since TensorDock's H100 inventory is SXM5 across the board, not a mix of form factors. On the SXM5 tier specifically, TensorDock's marketplace pricing, and even its cheapest 3-year reserved rate of $1.50/hr, undercuts Spheron's fixed on-demand rate by a wide margin. Spheron's spot rate at $2.91/hr per SXM5 GPU narrows that gap but doesn't close it. If your workload can drop to the PCIe form factor, Spheron's PCIe rate is the closer apples-to-apples price point against TensorDock's on-demand range; see our H100 PCIe vs SXM5 form factor guide for what the interconnect difference actually changes for training and inference throughput.

Pricing fluctuates based on GPU availability. The prices above are based on 23 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

What the "Cheapest H100" Listing Doesn't Include

TensorDock's low headline number is real, but three factors consistently widen the gap between the advertised rate and what a production workload actually costs to run.

Hourly Billing Rounding vs Per-Minute Billing

TensorDock bills hourly by default. A 25-minute inference test or a short evaluation job still consumes a full billed hour. Run a handful of short jobs across a day and the rounding adds up in a way the $1.90/hr headline doesn't capture. Spheron bills per minute across every tier, so a 25-minute job costs 25 minutes, not 60.

Support Response Times and Documentation Depth

TensorDock support is email-only with 4-24 hour response times. That's workable for teams comfortable troubleshooting independently, but it's a real gap if a host-level issue interrupts a production job at 2am and you need someone to look at it before your next business day. Documentation is comparatively thin next to platforms with dedicated onboarding and account support.

Host-Level Uptime Variance (99.1% DC Hosts vs ~97% Community Hosts)

This is the number that matters most for anything beyond short-lived jobs. According to GPU Cloud List's TensorDock review, data-center-verified hosts measure around 99.1% observed uptime; community hosts run closer to 97%. A 3-percentage-point uptime gap sounds small until you map it onto a multi-day training run: at 97% hourly reliability, a simplified per-hour reliability model puts the odds of at least one interruption over a 72-hour run at roughly 89% (1 − 0.97^72), versus about 48% at 99.1% uptime over the same window. The same review clocks training throughput on H100 SXM at 97% of theoretical maximum on verified hosts, which is a genuinely strong number when you land on the right host. The catch is verifying which host you're on before you commit a long job to it.

TensorDock vs Other Budget H100 Providers (Vast.ai, Runpod Community)

TensorDock isn't the only marketplace-style option chasing the lowest H100 rate. In a market survey covering 15+ providers, TensorDock's commonly quoted $2.25/hr on-demand rate lands in the lower-mid tier: cheaper than Lambda Labs ($2.99/hr), Google Cloud ($3.00/hr), AWS ($3.90/hr), CoreWeave ($6.16/hr), and Azure ($6.98/hr), but not the single cheapest listing in the market. Vast.ai and NeevCloud marketplace rates undercut it on the low end.

TensorDock and Vast.ai run the same fundamental model, independent hosts, host-set prices, and comparable price bands. Our Vast.ai pricing breakdown (linked above) covers the same billing-rounding and host-reliability trade-offs in more depth, and our Vast.ai alternatives guide has a dedicated section on TensorDock's transparency around host reliability ratings relative to Vast.ai. Runpod Community Cloud is a third variant of the same marketplace pattern; the Runpod H100 pricing breakdown covers where that tier sits against Runpod's own managed Secure Cloud option, and our Runpod alternatives guide includes a TensorDock pricing entry for readers comparing multiple budget marketplaces side by side. For the full field including hyperscalers and other neoclouds, see the 2026 GPU cloud pricing comparison.

Choosing Between a Peer-to-Peer Marketplace and a Managed Neocloud

The decision isn't "TensorDock is cheap, Spheron is expensive." It's a trade between the lowest possible listed rate and a fixed rate with a platform standing behind it. Which one wins depends entirely on what you're running and how much an interruption actually costs you.

When TensorDock's Price Wins

Short experiments, prototyping, and dev work where a failed job just gets restarted. Budget-constrained testing where the cheapest available host is worth the reliability trade. Teams that can actively vet hosts, checking data-center-verified status and reliability history before committing a long job. TensorDock's reserved 3-year tier at $1.50/hr is also genuinely hard to beat if you can commit to a specific host for that long and are comfortable with the risk of a single-host dependency over three years.

When Spheron's Managed Model Wins

Production inference that needs consistent latency and can't tolerate a host going offline mid-request. Training runs long enough that a host-level interruption costs real, unrecoverable compute time. Teams that want per-minute billing instead of hourly rounding on variable-length jobs. And teams that would rather deal with one platform's support and uptime accountability than vet individual hosts themselves. Spheron aggregates capacity from 5+ providers on the backend, which gives it more inventory to route around any single provider's availability gaps, while still presenting one managed rate and one support relationship to you. For a deeper look at the managed-platform-vs-marketplace trade-off in general, Spheron vs Vast.ai walks through the same framework this post applies to TensorDock. Full setup and API details for deploying on Spheron are at docs.spheron.ai.

TensorDock's marketplace can beat Spheron's on-demand PCIe rate on a good day, but the price you actually get depends on which host you land on. If you need a fixed rate, per-minute billing, and platform-level uptime instead of a host lottery, Spheron's H100 instances are the straightforward alternative.

Check H100 availability on Spheron →