Pebble for Neoclouds
Power is your largest controllable cost and your hardest input to scale. Pebble caps every GPU at its efficiency knee — cutting demand charges today and fitting more revenue-generating GPUs into the same power envelope. Higher margin per GPU-hour, same facility, same utility contract. Validated live on AMD Instinct™ MI350X.
Instinct™ MI350X
MiTACComputing
infrastructure partner
+86.5%
tokens/sec at −17.5% power on the live AMD MI350X pilot
+2 GPUs
of revenue capacity per 6 deployed, in the same kW envelope
30–70%
of an industrial power bill is the demand charge Pebble shaves
0%
SLA impact · zero application changes · behind your firewall
Your facility is capped at a fixed number of megawatts. Every watt a GPU wastes as heat is a watt you can't sell — and a demand-charge spike you pay for anyway. Pebble turns that wasted power back into sellable capacity.
01 · Margin per GPU-hour
Sonar caps each GPU at the operating point where tokens-per-watt peaks. Same throughput, less power drawn — the cost side of every GPU-hour drops, and it lands on the next utility invoice, not a five-year capex plan.
02 · Revenue per megawatt
The reclaimed 20–30% of stranded power becomes room for additional GPUs inside the same megawatts. On the AMD pilot that was +2 GPUs of capacity per 6 deployed — more inventory to rent from the facility you already pay for.
03 · Demand charges
Utilities bill on your single highest 15-minute peak. One synchronized GPU burst can reset your rate for the whole month. Pebble clips that peak in real time — where 30–70% of the bill lives — with no batteries and no SLA hit.
“A very clever, very interesting product, where they actually utilize AI and have our GPUs consume less power — but the performance per watt is increased.” — Ted Marena, AMD · Plug and Play San Jose State, May 2026
0:53
Pebble deploys inside your environment and runs against your existing serving stack — vLLM, SGLang, TRT-LLM — with zero application changes. AMD Instinct and NVIDIA GPUs both supported. The optimization happens entirely within your infrastructure.