Using 16 to 32 SSDs in RAID0 is not just geek showmanship. We think it signals something more important first: the cost of running large models locally has become high enough that users are starting to use storage as a substitute for memory. The Reddit discussion puts the question bluntly: if VRAM and system memory are both too expensive, can a large pool of SSDs trade capacity for bandwidth and squeeze bigger open-source models into a local machine?
What this is
The post comes from r/LocalLLaMA. The core idea is simple: SSD capacity is far cheaper per unit than memory, so if enough high-speed SSDs are combined in RAID0, the aggregate bandwidth could theoretically approach system memory. Then, by moving parameters in and out aggressively, a machine might barely sustain inference for very large models. The hard part here is not whether the data can be read at all, but latency. High bandwidth does not mean fast response time, and large-model inference usually demands both bandwidth and low latency.
Industry view
Behind this discussion is a straightforward reality: open-source models keep getting larger, while consumer hardware is getting harder and harder to scale with them. The industry has effectively been following two paths. One is to buy more expensive GPUs and more memory. The other is to rely on quantization, tiered loading, and hybrid storage to squeeze as much as possible out of existing hardware. Supporters will argue that the SSD approach at least offers one direction for low-cost experimentation.
But the objections are just as clear. First, RAID0 stacks sequential throughput; it does not eliminate SSDs’ inherently high latency. Second, inference requires frequent data movement, and system overhead could easily consume the theoretical performance gains. Third, if you really want to deploy 16 to 32 SSDs, you also need enough PCIe lanes and a workstation-class motherboard, which may not be cheap at all. Our view is that this looks more like an edge-case experiment than a repeatable deployment path.
Impact on regular people
For enterprise IT: this is a reminder that the bottleneck in local large-model deployment is still hardware cost. Procurement cannot focus only on parameter count; it also has to evaluate the full storage and memory architecture.
For individual professionals: the threshold for running large models fully offline on local hardware has not meaningfully fallen. In the near term, the more practical route is still smaller models, local toolchains, and hybrid use with the cloud.
For the consumer market: if this kind of discussion keeps gaining traction, it could push the “AI workstation” category into further specialization. But the biggest beneficiaries may not be SSD vendors. More likely, they will be the companies that can turn storage, memory, and inference scheduling into an integrated system-level product.