⬡viabandwidth
Buying GPU Services buyer guide cover
872
Candidate domains reviewed
722
Retained active candidates
4,511
Scoped price representations
0
Required follow-ups open

ViaBandwidth buyer guide

This field manual turns GPU labels, headline rates and infrastructure claims into a defensible, workload-matched purchase decision. The cheapest GPU is the wrong question until the workload, data path and service boundary are defined.

01 · WorkloadStart with the workload

Write the requirement before opening a provider comparison. Training, fine-tuning, inference, rendering, VDI and research create different requirements for memory, topology, persistence, scheduler and utilization.

Workload questionWhy it changes the buy
Training, fine-tuning, inference, rendering, VDI or research?Changes model choice, interconnect, persistence, scheduler and utilization pattern.
Single GPU, multi-GPU or multi-node?Determines whether node topology and east-west fabric dominate.
Burst, steady or deadline-driven?Changes the value of on-demand, reservation, spot and dedicated capacity.
Data size and sensitivity?Changes storage, egress, sovereignty, private connectivity and security requirements.
Framework and image constraints?Changes driver/CUDA compatibility, managed-platform value and migration effort.
Human support tolerance?Changes the balance between self-service price and managed delivery.
Requirement template. We need [workload] on [GPU/model class] with [GPU count/topology], [CPU/RAM/storage/network], in [region], for [duration/utilization], with [security/support] and a successful run defined as [time, throughput, quality and cost].

02 · Supply modelChoose the right supply model

Direct cloud, marketplace capacity, dedicated servers, colocation and managed platforms solve different problems.

On-demand GPU cloudFast access and elasticity. Watch egress, storage and the instance bundle.
Reserved / committedPredictable utilization and capacity assurance. Watch prepayment and region lock.
Dedicated server / clusterStable bare metal and topology. Watch lead time and replacement SLA.
Marketplace / brokerBroad inventory discovery. Verify the actual seller, host and SLA.
Managed platformDeployment and operations. Separate platform fees from compute.
Colocation / own hardwareEquipment control. Account for power, cooling, transit, spares and operations.
Identity check. Ask who signs the contract, who operates the hardware, where the hardware sits, who controls the network and who handles a failed GPU. Brand, seller, operator and facility can be four different entities.

03 · ComparabilityMake prices comparable

Preserve the native offer first; normalize only inside a matched cohort.

FieldRecord exactlyCommon failure
GPUModel, variant, memory/profile, quantityTreating variants or MIG slices as equivalent
Sold objectGPU, VM, server, node, pod, endpoint or output unitDividing a node into a synthetic GPU rate
ResourcesCPU, RAM, storage, network, software/platformComparing component-only with all-in
MoneyNative amount, currency and tax statusAssuming every “$” is USD or tax-inclusive
TimeBilling unit and minimum durationUsing 730 hours without matching commitment
CommitmentOn-demand, spot, prepaid, reservation, termComparing three-year pricing with on-demand
PlaceRegion/facility and residency boundaryIgnoring regional price and availability
EvidenceURL, date, excerpt and card/table boundaryJoining unrelated fields from separate rows
Total workload cost = compute + required base resources + storage + data movement + platform/support + tax + engineering/operations impact. Divide by successful completed runs or useful output, not headline GPU-hours alone.

04 · QuoteInterrogate the quote

AskA good answer contains
What exactly is allocated?GPU count/variant, tenancy, topology, CPU/RAM/storage/network and software layer.
Is capacity available now?Quantity, region, start date, lead time and reservation mechanism.
What is the commercial unit?Native unit, minimums, commitment, prepayment, renewal and cancellation.
What changes the invoice?Storage, egress, public IP, support, platform, tax and overage rules.
Who operates each layer?Seller, hardware operator, facility, network and escalation owner.
What happens on failure?Replacement time, credits, persistence, checkpoints and support response.
How do we exit?Export, image portability, deletion, notice, fees and termination assistance.
Unknown is not zero. If ownership, stock, tax or a fee is not stated, record it as unknown and ask.

05 · Proof of valueRun a proof of value

  1. Use a representative model, dataset, image and framework version.
  2. Measure queue and provisioning time separately from runtime.
  3. Record throughput, latency, utilization, memory headroom and scaling efficiency.
  4. Include checkpoint, storage and data-transfer time.
  5. Trigger a restart or failure scenario and observe persistence and support response.
  6. Reconcile measured usage with the invoice or usage export.
  7. Repeat enough runs to detect variability and throttling.
  8. Document the exact configuration so the result can be contracted and reproduced.
MetricTargetMeasuredNotes
Provisioning / queue time________________________
Time to first useful output________________________
Throughput / latency________________________
GPU utilization / memory________________________
Scaling efficiency________________________
End-to-end workload cost________________________
Failure recovery________________________
Invoice reconciliation________________________

06 · ShortlistScore the shortlist

Weight evidence quality and operational fit alongside price. Mandatory security, region, model compatibility, capacity date and commitment requirements remain pass/fail gates.

CriterionWeightScore 1–5Evidence
Workload performance / fit20%_______________
All-in cost20%_______________
Capacity assurance / lead time15%_______________
Reliability / failure handling10%_______________
Data, security and sovereignty10%_______________
Network / storage path10%_______________
Support / operations10%_______________
Portability / exit5%_______________

07 · WorksheetProcurement worksheet

RequirementBuyer entry
Workload and success metric________________________________
GPU model / acceptable alternatives________________________________
GPU count / topology / tenancy________________________________
CPU / RAM / storage / network________________________________
Region / sovereignty________________________________
Start date / duration / utilization________________________________
Security / compliance________________________________
Support / SLA / replacement________________________________
Budget / native currency / tax________________________________
Exit / portability________________________________

Final offer record

FieldSelected offer
Seller / contracting entity________________________________
Operator / facility / upstream________________________________
Exact configuration________________________________
Native amount and unit________________________________
Commitment / minimum / cancellation________________________________
Extras and exclusions________________________________
Availability and provisioning date________________________________
Evidence URL / quote version / date________________________________
Proof-of-value result________________________________

08 · ReferenceRed flags and quick glossary

Red flagWhy it mattersBuyer action
GPU name without quantity/profileThe sold allocation is ambiguousRequire exact physical or virtual allocation
Price detached from its card/tableConfiguration binding may be wrongRequest a bound quote or order summary
Future facility presented as liveLifecycle mismatchRequire an orderable date and commitment
“Dedicated” without layerTenancy may apply only to VM or GPUDefine device, host and cluster tenancy
Marketplace brand hides hostCounterparty is unclearRecord seller, operator, region and SLA
Very low rate with missing extrasHeadline may be conditionalBuild all-in workload cost
No exit termsMigration can erase savingsContract export, deletion and termination
Final rule. Keep the original offer intact. Every conversion, allocation assumption and workload factor should be visible, dated and reversible.

viabandwidth is the verified directory of GPU compute, colocation and carrier networks.

This buyer guide uses a finite market audit and first-party provider evidence. Native offers, configuration boundaries and unknowns must be preserved; public prices do not establish stock or contract terms.

© 2026 viabandwidth (Steven Higashi). All rights reserved. The figures and charts in this paper may be quoted with clear attribution to viabandwidth.

Cite as: ViaBandwidth, Buying GPU Services, viabandwidth.com/reports/buying-gpu-services.

viabandwidth.com · [email protected]