Who this is for: infrastructure leads, platform engineers, and capacity planners securing GPU compute for training or inference at scale. You are weighing cloud-managed vs. bare-metal, reserved vs. spot, hyperscaler vs. neocloud, and NVIDIA vs. alternative silicon — against real constraints like a fixed training deadline, data residency, and a unit-economics target. You need a sourcing framework, not a vendor listicle.
The ledger
This is a representative sourcing structure, not a live capacity or pricing feed. Each row maps a real, documented provider to the capability surfaces it exposes. ✓ = yes, ~ = partial / limited / via partner, — = no. Capacity counts, regional availability, wait times, and per-GPU-hour pricing are intentionally omitted — they change faster than a static page can track and must be verified on request.
| Provider | Cloud-managed | Bare-metal / dedicated | Spot / preemptible | Reserved capacity program | Multi-region | NVLink / NVSwitch fabric | Notes |
|---|---|---|---|---|---|---|---|
| AWS (EC2 P5 / Trainium2) | ✓ | ~ | ✓ | ✓ | ✓ | ✓ | P5 H100/H200 instances on NVSwitch; Capacity Blocks for ML reserve whole-instance blocks. Trainium2 is AWS-native silicon. |
| Google Cloud (A3/A4, TPU) | ✓ | ~ | ✓ | ✓ | ✓ | ✓ | A3 Mega/A4 VMs on NVSwitch; Future Reservations + capacity reservations. TPU v5e/v5p is Google-native. |
| Microsoft Azure (NDv5) | ✓ | ~ | ✓ | ~ | ✓ | ✓ | ND H100 v5 / H200 on NVSwitch; standard reserved-instance pricing, not capacity blocks. |
| Oracle Cloud (OCI) | ✓ | ✓ | ~ | ~ | ~ | ✓ | BM.GPU.H100.8 bare-metal H100; fewer AI regions than the top three. |
| CoreWeave | ✓ | ~ | ~ | ✓ | ✓ | ✓ | GPU neocloud; committed-use discounts; US + EU regions. |
| Lambda | ✓ | ~ | ~ | ~ | ~ | ✓ | GPU neocloud; on-demand + reserved H100/H200; smaller regional footprint. |
| NVIDIA DGX Cloud | ✓ | — | — | ~ | ~ | ✓ | Hosted DGX on partner clouds (AWS/GCP/Azure/OCI); capacity via partner programs. |
This ledger is a capability checklist, not a benchmark or inventory report. We do not publish live capacity, availability, utilization, or pricing data unless we have collected and verified it. Procurement decisions should follow your workload constraints and a current quote, not a static ranking.
How to decide
- Start from your deadline. A fixed training date forces a reserved-capacity program (AWS Capacity Blocks, GCP Future Reservations) or committed-use neocloud terms. If you have flexibility, spot/preemptible capacity cuts cost but cannot backstop a launch date.
- Decide cloud-managed vs. bare-metal. Cloud-managed wins on elasticity and managed storage/networking. Bare-metal (OCI, CoreWeave, self-hosted DGX) wins on host control, isolation, and long-cluster amortization. Pick the model before picking the vendor.
- Resolve regional constraints early. Data residency and latency may eliminate providers whose AI regions are thin. Neoclouds typically cover fewer regions than AWS, GCP, or Azure — confirm current regional coverage before designing around it.
- Price the alternative silicon honestly. Trainium2 and TPU can beat NVIDIA per-unit cost but trade portability and add migration work. Model the migration cost against the unit-economics gap, not the headline rate.
Get the deeper capacity framework
We are building a fuller, constraint-driven framework for GPU capacity sourcing — reservation timing, build-vs-buy-vs-lease economics, and a regional-availability query playbook — delivered through the biweekly Gen Alpha AI briefing. No spam, unsubscribe anytime.
Get the framework →Sponsor this coverage
This ledger sits in high buyer-intent territory — readers are mid-procurement on GPU capacity. If you sell cloud, neocloud, bare-metal, or accelerator capacity and want to reach these buyers with clearly labeled, editorially independent sponsorship, talk to us. No fabricated audience metrics; we share real analytics with serious sponsors.
View sponsor inventory →Need a sourcing decision, not a matrix?
If you are stuck choosing a GPU capacity strategy against a real deadline, budget, and residency constraint, a focused advisory session can resolve it. Bring your workload, your deadline, and your regional constraints — we hand you a written, prioritized procurement recommendation.
Book an advisory session →