Frequently Asked Questions

On-demand list rates for the NVIDIA H100 80GB SXM run $3.99 to $12.29 per GPU-hour in October 2026, a median of $6.52, with RunPod Secure Cloud at the low end and Azure at the high end. The cheapest published consumer option is the RTX 4090 at $0.34/hour on RunPod Community Cloud. Marketplace providers such as Vast.ai publish no list rate because their hosts set prices, and spot or committed tiers run below every figure here. GYGO compares pricing across providers so teams find optimal cost-performance.

Major cloud GPU options include NVIDIA H100 80GB and H200 for large-scale training, A100 80GB for general AI workloads, L40S for inference, and RTX 4090 for cost-effective fine-tuning. AMD MI300X is emerging as a competitive alternative. Availability varies by provider and region.

Minimum rental periods vary by provider. On-demand instances start at per-second or per-hour billing with no commitment. Committed-use discounts are published as ceilings: CoreWeave offers ‘up to 60% discounts over our On-Demand prices for committed usage’, and AWS Savings Plans cut EC2 cost by up to 66–72% on one- to three-year terms. Lambda and the others quote reserved capacity on request. Tell GYGO your workload duration and we include both options in the request.

GYGO is a marketplace aggregator that compares pricing, availability, and specs across multiple GPU providers including Lambda Labs, CoreWeave, RunPod, Vast.ai, and major cloud platforms in one interface. This saves time versus checking each provider individually and helps find the best price-performance ratio.

NVIDIA and AMD publish no list price for datacentre GPUs and sell through allocation, so these are named reseller listings, not MSRP. H100 SXM5 is not sold as a loose card: a complete 8-GPU HGX H100 server (Supermicro SYS-821GE-TNHR with 8x H100 80GB SXM5) listed at $289,000 at Vipera on 13 October 2026, or $36,125 per GPU deployed, with two other resellers listing the same build within 4%. A single H100 80GB PCIe card was $32,000 at CDW when ServeTheHome checked in October 2023; no named reseller shows a fetchable 2026 price for it. An A100 80GB PCIe lists at $16,535 at SabrePC (shown out of stock on 13 October 2026; in-stock new cards list higher), and a new RTX 4090 was $4,399 on Amazon as tracked by GPUDojo the same day, against its $1,599 launch price, because the card is discontinued. Every tier-one channel quotes rather than lists, so treat any single figure as a starting point.

Yes. GYGO connects buyers with value-added resellers who quote volume pricing on multi-unit orders of enterprise GPU hardware. Custom server configurations and pre-built GPU clusters are also available through our reseller partners.

GPU purchases through GYGO's reseller network carry the OEM or reseller warranty of the system they ship in. NVIDIA publishes no standalone H100 warranty term (its warranty page says the period varies by product line and region), so ask the reseller for the term and for support-contract options. Value-added resellers provide additional support including installation assistance, firmware updates, and hardware replacement. Extended warranty options are available.

Buy if you have sustained workloads running 6+ months at high utilization, as ownership becomes cheaper than rental at that point. Rent for variable workloads, prototyping, or when you need different GPU types. GYGO helps evaluate both options based on your specific usage patterns and budget.

GPU colocation means placing your own GPU hardware in a third-party data center providing power, cooling, networking, and physical security. Unlike cloud rental where you pay per-hour for shared resources, colocation gives you dedicated hardware at a fixed monthly cost, about 69% cheaper on the 64-GPU H100 model this site publishes (see the colocation vs cloud ROI guide).

Modern GPU clusters require 20-100+ kW per rack depending on configuration. A single NVIDIA DGX H100 system draws approximately 10 kW. Data centers offering GPU colocation must provide high-density power, direct liquid cooling or rear-door heat exchangers, and sufficient cooling capacity per rack.

High-performance GPU colocation facilities offer InfiniBand NDR (400 Gbps) for training clusters, RDMA over Converged Ethernet for flexible deployments, and high-bandwidth spine-leaf network architectures. GYGO helps match your interconnect needs with facilities that support your cluster topology.

Size it to the sustained peak heat load, the rack kW your cluster actually holds through a training run. Don't size to nameplate power, which sizes busbar and breakers rather than a heat exchanger, and don't size to the duty-cycle average either. Checkpoint dips buy you no capacity credit, because a CDU's own coolant inventory rides through only seconds of full load and a dip has to outlast that to matter. What the dips do create is a part-load control problem, so ask about the bottom of the envelope and not just the top. Ask for the minimum controllable flow and the lowest rack load at which the unit still holds setpoint. Vendors publish maximum flow and rated capacity but almost never the low end, so you have to request it. The failure you're guarding against is the facility-side control valve losing authority as load falls away: below a minimum valve position it stops metering flow, the secondary loop drifts outside its band, and the unit can trip on under-temperature to avoid condensation. That gets worse the more the CDU is oversized, which is the real argument against buying too much capacity. Then ask for the design supply temperature on the secondary (technology cooling) loop, the band it is held to, and how the plant holds it through a step change. ASHRAE is explicit that IT vendors define the temperature and flow rate-of-change envelopes for stable operation, and that erratic load with low diversity makes supply-water temperature fluctuate. A synchronized training cluster is exactly that load. On redundancy, ask how the pumps are configured and at what operating point. N+1 is common but not universal, and some units run N in their highest-flow mode, so the honest question is whether a pump is still spare at your design flow. Pumps are not the only single point of failure: the heat exchanger and controller sit in the same box, so ask whether redundancy exists at the CDU level, meaning a spare unit that can pick up full load, and how long changeover takes. In a brokered colocation deal the partner facility owns and sizes the CDU, so what you hand over is the spec: your peak kW per rack and the duty cycle you expect. Then settle whether you are getting dedicated CDU capacity or a share of a larger unit. Row-based CDUs often run 800 kW to 2 MW and serve many racks, so shared capacity is common and not automatically the worse deal. Both models exist across GYGO's partner network, so name which one your deployment needs when you ask for quotes rather than finding out in the contract.

Published 2026 rate guides put air-cooled GPU colocation at $120–$350 per kW per month across the four primary US markets, depending on market, rack density, cooling type, and contract term. A standard 40 kW air-cooled GPU rack therefore runs about $4,800–$14,000 per month before cross-connects, remote hands and power billing. See the GPU colocation pricing page for ranges by market and the sources.

Yes. GPU investment programs let you purchase GPU hardware deployed in data centers and leased to AI companies. Returns depend on GPU model, utilization rates, and market demand. GYGO connects hardware investors with partner data center operators that report utilization and revenue.

Returns depend on rental demand for the model and the utilisation the operator actually achieves, which no operator publishes as a checkable figure; Aethir's own hosting economics, for example, assume 95%+ utilisation for an 8-GPU H100 node (see the GPU staking returns guide). Consumer GPUs like the RTX 4090 cost less up front and earn less per unit.

Key risks include hardware depreciation as newer GPU generations release, market demand fluctuations affecting utilization rates, operator reliability, and potential hardware failures. NVIDIA has shipped a new data-center GPU generation roughly every two years (A100 in 2020, H100 in 2022, B200 in 2024). Ask operators for their reported utilization before committing.

Minimum investments vary by program. A single enterprise H100 works out to about $36,125, one eighth of a complete 8-GPU server reseller-listed on 13 October 2026. Fractional ownership programs may offer lower minimums. GYGO's platform shows available investment opportunities across multiple operators with varying entry points.