Feature · capacity

Dedicated AI capacity for production

Dedicated capacity is not just a bigger rate limit. It is a different failure domain, a different cost curve and a different contract. Here is how to size it without overbuying.

Short answer. Dedicated capacity reserves GPU, memory and sometimes network, with custom limits and a private or region-locked data plane. Plugsky Enterprise offers dedicated GPUs, VPC, on-premises and air-gapped deployment, and custom SLAs. Minimum commit and price are quoted per deployment; model the break-even before you commit.

Shared vs dedicated capacity

Shared capacity is a pool. You get fair-use limits, burst protection and a low entry price, but a noisy neighbour or spike can change latency. Dedicated is a reservation: isolation, predictable throughput and a private failure domain, but you pay for idle time. The question is whether p95 latency and compliance justify the commitment.

DimensionSharedDedicated
TenancyMulti-tenant poolReserved GPU, memory and sometimes network
Noisy neighbourPossible at peakIsolated from other customers
LimitsPublished RPM, concurrency, tokens/dayCustom rate limits and quotas
BurstBest effort within fair usePlanned headroom or overflow agreement
SLAPlan-level uptime targetCustom SLA, often per deployment or region
CostFlat or per-token, low entryCommitment plus utilization risk
OperationsVendor-managedVendor-managed, but you own capacity planning

GPU and region isolation

Isolation has layers. NVIDIA Multi-Instance GPU (MIG) can partition supported GPUs into isolated instances with dedicated compute and memory and guaranteed performance. A whole GPU or node pool avoids sharing. A private endpoint keeps traffic off the public internet. A region-locked data plane keeps data and inference in one jurisdiction. Ask which layers are included, not just whether the offering is called “dedicated.”

Region isolation is a contract and an architecture. Confirm where the control plane, data plane, logs and backups run, and what happens during failover. A control plane in one region can manage another, but its outage may still block new deployments or key changes. Get the region and failover maps in writing, and test a cross-region restore.

Burst, queueing and admission control

Burst is a policy, not free capacity. Most APIs use a token bucket or a concurrency limit, then queue or reject excess. A 429 means slow down; a 503 means the service cannot serve you now. Retrying a 429 without backoff makes the queue worse; retrying a 503 may be appropriate with jitter. Ask for retry guidance and the queue timeout.

Negotiate burst explicitly: ceiling, duration, refill rate and overflow path. Measure p95 queue time, not just tokens per second. A 500 tok/s model with a two-second queue feels slower than 300 tok/s with no queue. For batch work, use a separate queue or batch tier.

SLA, failure domains and multi-region

Ask what the SLA measures: cluster, region, account or end-to-end API. A cluster can be up while the router, control plane or upstream provider is down. Ask whether it covers time to first token, successful completion rate, or only availability, then get the credit schedule, exclusions and claim process. Enterprise terms are custom; put them in the order form.

Define RTO and RPO for inference, not just the database. How long does a regional failover take, and what can you lose? If it depends on DNS TTL, model warm-up or a manual runbook, test it. A multi-region design that has never failed over is a diagram. Run a game day with a canary and keep the results in the QBR.

Minimum commit and the cost model

Dedicated pricing usually has a monthly commitment, a ramp, a true-up and an overage rate. Plugsky does not publish a minimum commit; it is quoted per deployment. Model the break-even: if the reservation costs R per month and on-demand is H per GPU-hour, break-even is R / (H × 730) GPU-equivalents. At 50% utilization you pay for idle capacity; at 80% the reservation usually wins.

break_even_gpus = monthly_reservation / (on_demand_gpu_hour × 730)

Example only, not a Plugsky price:
R = $10,000 / month, H = $3.00 / GPU-hour
break_even = 10,000 / (3.00 × 730) = 4.57 GPU-equivalents

Include the costs that are easy to miss: storage and egress, observability, support, a second region for failover, idle capacity for burst, and the engineering time to run capacity planning. A dedicated contract can be cheaper per token at steady utilization and more expensive during a launch, a seasonal peak or a proof of concept. Ask for a ramp that matches your forecast, not a flat commitment from day one.

Migration, cutover and the RFP

Migration is a capacity project, not a URL change. Inventory models, context lengths, concurrency and peak patterns. Build a capacity model from measured tokens per request and requests per second. Run evals on the new deployment, shadow production traffic, compare quality, latency and cost, canary a small percentage, keep the shared path as fallback, and define the rollback trigger before cutover.

Put these in the RFP: isolation layer; region map and residency; burst ceiling and overflow; queue timeout and retry guidance; SLA metric and credits; minimum commit, ramp and true-up; model catalogue and change notice; security evidence; migration and exit support. Ask for a reference architecture and a sample bill at your forecast. The pricing page shows the self-serve plans; rent GPU covers raw GPU options; managed API vs building your own stack compares the build option; Enterprise covers the commercial packaging.

FAQ

What is the difference between shared and dedicated AI capacity?

Shared capacity is a multi-tenant pool with fair-use limits. Dedicated capacity reserves GPU, memory and sometimes network for your workload, with custom limits and a private or region-locked data plane. Dedicated costs more at low utilization.

Does dedicated capacity guarantee a GPU?

It depends on the isolation layer. Ask whether you get a whole GPU, a MIG partition, a dedicated node pool or only a reserved throughput pool. Each has different isolation and failure characteristics.

How does burst work on dedicated capacity?

Burst is a policy, not free capacity. Negotiate a ceiling, duration, refill rate and overflow path, and measure p95 queue time. A burst allowance without an overflow plan still fails at the peak.

What is the minimum commit?

Plugsky does not publish a minimum commit; it is quoted per deployment. Ask for the ramp, true-up, overage rate, renewal and termination terms, and model the break-even utilization before signing.

How do we migrate without downtime?

Inventory models and traffic, build a capacity model, run evals, shadow production, canary a small percentage, keep the shared path as a fallback, and define a rollback trigger. Test the cutover before the contract start date.

Sources

Primary sources checked 20 November 2026. Provider terms, limits and prices change; verify the current version before you commit.

  • NVIDIA Multi-Instance GPU user guide — describes partitioning supported GPUs into isolated instances with dedicated compute and memory and guaranteed performance.
  • Amazon EC2 Reserved Instances — explains reservation terms, discounts, capacity reservation in a specific Availability Zone and the trade-off between commitment and flexibility.
  • vLLM paged attention design — explains KV-cache paging and memory management, which underpin throughput and concurrency planning for LLM serving.

Reviewed 20 November 2026. Dedicated capacity, minimum commit and Enterprise SLA terms are quoted per deployment. Confirm the current order form and capacity plan before committing.