Section B · Operations

Hardware & Infrastructure

RunPod's GPU inventory spans consumer and datacenter cards across both products. The Secure Cloud datacenter footprint is small relative to enterprise neoclouds but growing.

GPU mix

Across both products, RunPod lists or operates:

  • Consumer NVIDIA: RTX 3090, 4090, 5090.
  • Workstation: RTX A6000, RTX 6000 Ada.
  • Datacenter: A100 (40 and 80GB), L40S, H100 (80GB), H200, occasionally B200.
  • Limited AMD (MI300X experimental).

The mix is similar in shape to Vast but the absolute count is smaller. Where Vast might have hundreds of H100 listings at any moment, RunPod might have dozens-to-low-hundreds.

Community Cloud hardware

Community Cloud hardware is whatever providers list. The mix tilts more toward small-fleet operators than Vast's more hobbyist-heavy supply:

  • More multi-GPU machines (4x and 8x configurations more common).
  • More datacenter-grade cards (A100, H100) and fewer pure-consumer 3090s.
  • More consistent bandwidth (providers are often colocated, not residential).
  • Slightly more reliable on average than Vast's average listing.

This shape is consequence of RunPod's slightly higher curation bar — the marketplace deliberately discourages purely-residential listings.

Secure Cloud hardware

Secure Cloud is RunPod-procured GPUs in RunPod-controlled or RunPod-partnered datacenters. The hardware mix is more concentrated on datacenter-grade cards:

  • Heavy on A100, H100, H200.
  • Some L40S for inference workloads.
  • B200 capacity coming online as supply allocations arrive.
  • Multi-GPU configurations (typically 8x H100 per node) for cluster workloads.

The configurations are more uniform than Community Cloud — same hardware spec across all listings, predictable performance.

Datacenter footprint

RunPod's Secure Cloud spans multiple regions across North America and Europe. The specific datacenter count is small compared to enterprise neoclouds (Crusoe has multi-hundred-MW sites; RunPod's footprint is more typically tens of MW across its sites).

The strategy is to grow footprint where customer demand justifies it rather than building speculative capacity. This is more conservative than CoreWeave's strategy of pre-building for committed customers.

Network

Secure Cloud networking includes:

  • High-bandwidth inter-pod networking within a region (10-100 Gbps).
  • InfiniBand for multi-node training in specific cluster offerings.
  • Standard cloud egress to the public internet.

Multi-node distributed training is possible on Secure Cloud but the cluster sizes are smaller than what CoreWeave or Crusoe offer. Tens of GPUs work; thousands probably don't.

Scaling supply

RunPod can scale supply two ways:

  • Recruit more Community Cloud providers. Asset-light; happens organically; bounded by how aggressively RunPod competes with Vast on supply-side terms.
  • Buy more GPUs for Secure Cloud. Capital-intensive; requires fundraising or revenue-based cap-ex investment; takes 6-12 months to deploy new capacity in a new region.

The blended approach lets RunPod grow capacity faster than a pure cloud (because Community Cloud scales without cap-ex) but more slowly than a pure marketplace (because Secure Cloud's growth is capital-constrained). Net: middling growth rate by neocloud standards.

Takeaway

RunPod's hardware footprint is meaningful but not enormous. The mix is shaped by the dual-product strategy — broader and more variable on Community Cloud, uniform and dedicated on Secure Cloud. The next chapter looks at who's actually buying.