BRICKY pins on‑prem breakeven 4–6 months
- Analyst account $BRICKY argued this week that sustained H100 demand now tilts sharply toward owned clusters, with on‑prem economics beating cloud surprisingly fast. - His worked example put five years of owned H100 capacity near $1 million versus roughly $3 million to $6 million rented in cloud. - Falling cloud GPU rates changed the timing, not the logic — steady utilization still makes rented H100s the expensive default.
The object here is H100 compute — the Nvidia GPU capacity companies rent to train models and serve inference. The stakes are simple: if you use a lot of it, the wrong deployment choice can burn millions. The gap is that people talk about “cloud versus on-prem” like it’s ideology, when it’s mostly a utilization math problem. What changed is that $BRICKY laid out a very blunt version of that math this week, arguing that for sustained H100 use, owned clusters can pay back in roughly 4 to 6 months. ### What is he actually comparing? He’s comparing two ways to get the same class of AI compute. One is renting H100-backed instances by the hour from cloud providers. The other is buying the hardware stack up front — GPUs, servers, networking, and the rest — then spreading that cost over years of use. New H100s still sell in roughly the $25,000 to $40,000 range per GPU, and full 8-GPU DGX-class systems land around the low-to-mid six figures. Cloud rental, meanwhile, has come down a lot, but on-demand H100 still commonly runs around $3 to $4 per GPU-hour at specialist clouds and hyperscalers, with Azure often higher. ### Why does utilization matter so much? Because owned hardware is mostly a fixed cost, while cloud is almost perfectly variable. If a cluster sits idle, on-prem looks dumb fast. But if the GPUs stay busy, every extra hour is almost free relative to renting. That’s why breakeven can arrive much earlier than people expect. At around 20% utilization, one H100 is more than $5,000 a year in rental cost per GPU before you add storage, networking, and margin. ### Does the 4–6 month claim make sense? Basically, yes — if the workload is steady and the hardware estimate is realistic. Take an 8-GPU box. If cloud pricing is roughly $24 to $32 per hour for eight H100s, running that box continuously costs about $17,000 to $23,000 a month. Over a year, that becomes a serious fraction of the purchase price of a computer to keep those GPUs loaded with training jobs or high-volume inference. The exact month count depends on financing, power, support, and whether you’re comparing against hyperscaler list prices or cheaper GPU clouds. ### So why doesn’t everyone buy? Because the cloud is selling more than silicon. It gives instant capacity, no lead-time risk, no datacenter work, and a clean way to handle bursty demand. If your usage spikes for a launch and then collapses, renting is still the right answer. The catch is that many AI teams say they are “bursty” when they actually have a pretty stable base — own the floor, rent the peaks. ### Why is latency part of this? Inference is not just a cost problem. It is also an operations problem. If you serve a steady stream of tokens all day, predictable local capacity helps with latency, queueing, and placement. On-prem also avoids some cloud egress and data-movement friction. That matters more once a workload graduates from experimentation into production. ### Has cloud gotten cheaper anyway? Yes — and that’s important context. H100 cloud prices are much lower than the panic-era rates from 2023 and early 2024. AWS p5 pricing now starts around $40,179 a month for an 8-H100 instance, or about $3.90 per GPU-hour, while Google A3 pricing works out in a similar band and Lambda lists H100 SXM at $3.99 per GPU-hour. So the argument is not that cloud is absurdly overpriced now. It’s that even after price cuts, steady rented H100 still compounds into a very expensive habit. ### What’s the real takeaway? This is less a hot take than a reminder. Cloud wins flexibility. On-prem wins repetition. If a company knows it will keep feeding H100s month after month, the breakeven window can be shockingly short
