# NVMe Bandwidth Is the Quiet Bottleneck Nobody's Pricing In

- Published: 15 September 2026
- Updated: 15 September 2026
- Author: RenderBob team
- Category: Technical
- Canonical: https://renderbob.io/blog/nvme-bandwidth-quiet-bottleneck-2026

This year's hardware conversation has been GPU and memory. Colibrì's own numbers point at a quieter bottleneck for hardware-constrained AI work: storage bandwidth.

Every hardware conversation this blog has had in 2026 has been about GPUs and GDDR7/HBM memory, the scarce resource everyone's fighting over, driving the price hikes covered repeatedly this month. Colibrì's own numbers point at a different bottleneck: for a whole class of memory-constrained AI workloads, the thing that actually sets your speed is not the GPU at all. It is NVMe read bandwidth.

The project states this plainly rather than burying it: decode is disk-bound on most machines running it, and the engine's own guidance is that a GPU only ever makes it faster. The baseline speed is set by your drive, because the experts are streamed from it. That is a different bottleneck profile than the rest of this blog's hardware coverage has focused on. The project's response is a set of techniques worth knowing even outside this specific tool: batching reads so each unique expert is fetched only once per batch of positions rather than redundantly; overlapping I/O with compute so the wait for one expert does not stall the ones already resident; prefetching a layer ahead based on measured routing predictability (reported around 71.6% predictable one layer ahead in this project's case); and, for anyone with a spare drive, striping reads across two SSDs at once, which measurably increases the aggregate bandwidth available to the streaming path.

This matters beyond one LLM engine because it is a preview of a constraint that is likely to bite harder as more hardware-constrained AI techniques lean on storage tiering the way Colibrì does. If disk-streaming and expert-offloading spread as a technique, and the existence of multiple independent projects in the space suggests it is not a one-off, then NVMe bandwidth stops being a background spec nobody thinks about and starts being a real capacity-planning variable, the same way VRAM already is. A studio evaluating hardware for any memory-constrained AI workload should start asking a question that would have seemed odd a year ago: not only how much VRAM the card has, but how fast the drive feeding it is, and whether that drive is the actual bottleneck.

If your studio is exploring any hardware-constrained local-inference approach, not necessarily Colibrì specifically, a fast NVMe drive, possibly two of them striped, is a cheap, unglamorous, currently undervalued lever compared with chasing an increasingly expensive GPU. It will not turn a memory-constrained workload into a fast one. In a year where every other hardware lever is getting more expensive and less available, a lever that is still reasonably priced and still readily purchasable is worth paying attention to.