The energy layer for AI datacenters
Pebble reclaims it — with zero SLA impact — and lets you spend it your way: deploy more GPUs, cut your power bill, or sell it back to the grid. Proven live on AMD Instinct™ MI350X.
Read-only assessment. No code changes. Deploys in minutes.
Partnered with
When AMD's Ted Marena gave an enterprise-AI keynote at Plug and Play San Jose, he spent several minutes of it walking the room through our joint POC on AMD Instinct™ MI350X at MiTAC Computing — and read the numbers off the slide himself. Watch the clip, then see the experiments behind it.
Instinct™ MI350X
MiTACComputing
infrastructure partner
Experiment 1 · Llama 3.1 70B · 8 GPUs
Efficiency knee: 1,000 W → 825 W per GPU.
Experiment 2 · Llama 3.1 405B · oversubscription
Same strict 6 kW envelope: 6 GPUs → 8 GPUs.
Zero application changes. Zero hardware changes.
0:53
Joint announcement with MiTAC Computing · June 26, 2026
AI growth isn't gated by chips anymore — it's gated by power. The megawatts you need to grow aren't available, and the ones you already have keep getting pricier and less predictable.
95%+
datacenter occupancy in major markets — little room left
+75%
jump in grid power prices in a single year
20–30%
of GPU power burned past the efficiency knee for almost nothing
The fastest megawatt you'll ever get is the one you stop wasting.
Every GPU has an efficiency knee — the power level where extra watts stop buying real performance. Run past it and you're burning 20–30% of your power for almost nothing. Pebble profiles each workload in real time, finds that knee, and caps exactly there — continuously, per GPU, per job. Throughput holds. SLAs hold. The wasted power comes back as headroom you control.
01 · Profile
Pebble runs alongside your serving stack (vLLM, SGLang, TRT-LLM) and continuously measures power draw, throughput, latency, and SLA compliance — per job, per GPU, per region.
02 · Cap at the knee
Every workload, on every hardware config, has both. The efficiency knee is where tokens-per-watt peaks. The SLA threshold is the lowest cap that still hits SLA. Pebble computes both continuously — they drift as your traffic and model drift.
03 · Reclaim
Cap each GPU at its efficiency knee → reclaim 20–30% of stranded TDP → fit more GPUs into the same kilowatt envelope. Or operate inside the SLA threshold → release power to the utility on grid signal, take it back later, no SLA breach.
Pebble's thesis isn't a contrarian bet — it's where the AI economy is heading. From the people running it.
“I actually think before we hit that, you're going to run into energy constraints.”
“In the future, data centers are going to be thought of … as AI factories. Their goal in life is to generate revenues, in this case, intelligence.”
“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.”
Deploy more GPUs in the megawatts you already have.
Fit +2 for every 6 — capacity you'd otherwise wait years for. The reclaimed power becomes room for more revenue-generating GPUs inside your existing envelope.
Cut your largest controllable cost.
Pebble shaves the demand-charge peaks behind your “random” power-bill spikes — utilities bill on your single highest 15-minute peak, where 30–70% of the bill hides.
Turn idle headroom into revenue.
When the grid signals, Pebble flexes within your SLA bounds in seconds — qualifying you for flex-power contracts that pay you back through the grid.
Pebble is the only energy layer that does all four:
Kubernetes, Slurm, or Docker — a controller plus a node-level agent, installed in one command. Dry-run mode shows you the savings before you enforce a single cap. mTLS by default. Everything stays behind your firewall. Works with vLLM, SGLang, and TRT-LLM, on NVIDIA and AMD.
helm install pebble/pebble
helm install — provisions agents and controller across the cluster in a single command.SonarPolicy CRD with dryRun: true to preview recommendations before enabling live caps.ansible-galaxy install pebble.agent
systemd daemon installs on every compute node. Auto-starts on boot — no manual restarts, no job disruption..sif Singularity image for air-gapped sites.docker compose up -d
docker-compose.yml to each GPU node, or use the install script to push and configure across all nodes.docker compose up -d. Agent and controller start with health checks, auto-restart, and mTLS between components.systemd unit wraps the stack for boot persistence. Set the controller address once per node — no further config needed.Pebble watches power and SLA from outside your serving loop — no fork of vLLM, no model rewrite, no special drivers.
At Pebble, our team is driven by a shared passion for solving hard engineering problems. We believe cloud infrastructure should be efficient, intelligent, and cost-effective. Our diverse team of AI, DevOps, and data center experts brings decades of experience to deliver a platform that drives real financial results.
Pradeep Gaddam
CEO / CTO
Chet White
Chairman
Keval Shah
Chief AI Officer
Carl Page
Sustainability Advisor
Philippe Cousteau
Sustainability Advisor
Michael Tessler
Technical Advisor
Ravi Chandra
Technical Advisor
Tim Profeta
Technical Advisor
No. Power capping operates within the GPU's published voltage / frequency envelope — you're not undervolting beyond spec or pushing thermals. In practice, capping to the efficiency knee runs each die cooler than the default cap, which is the more common driver of long-term wear.
Independent studies on dynamic GPU power management have not found accelerated degradation from frequency / power-cap adjustments compared to running at TDP. See: Patel et al., POLCA (ATC '24) · Stojkovic et al., 2024 · Krzywaniak et al., SC '23. Vendor documentation (NVIDIA DCGM, AMD ROCm SMI) also describes power capping as a supported, in-spec operating mode.
The headline figures (+86.5% tokens/sec, +82.4% tokens/watt at −17.5% power) come from a live pilot on 8× AMD Instinct MI350X at MITAC Computing, serving Llama 3.1 70B FP8 with vLLM — comparing the same vLLM configuration at the hardware-default 1,000 W cap vs. Pebble's 825 W efficiency-knee cap. The vLLM configuration is held constant between runs, so the gain is attributable to power management, not a serving-config change. Gains vary by model, batch profile, GPU SKU, and traffic mix — production deployments typically see 15–30% tokens/watt improvement with no SLA regression.
We run a representative job on your stack as part of every demo so the number you cite internally is your number, not ours.
No code changes. Pebble runs alongside vLLM, SGLang, TensorRT-LLM, and Hugging Face TGI as a Kubernetes operator, Slurm daemon, or Docker sidecar. It reads GPU telemetry (DCGM for NVIDIA, rocm-smi for AMD) and writes power caps via the same vendor APIs your ops team already trusts.
Pebble Flex profiles every workload's SLA-breach cliff continuously — the lowest power level at which the workload still hits its latency targets. When a utility curtailment signal arrives, we never cap below that cliff. The fleet curtails toward it, holds, and restores within seconds of the signal lifting. Time-to-first-token may tick up briefly; SLAs do not break.
NVIDIA Hopper (H100, H200) and Blackwell families. AMD Instinct MI300X and MI350X. Kubernetes (GKE, EKS, AKS, on-prem), Slurm, and bare-metal Docker. Full integration matrix and quick-start commands live in the deployment guide.
Sonar produces a complete efficiency-knee atlas for a workload in roughly 10–30 minutes of profiling. Flex requires a few hours of normal traffic to build SLA-cliff baselines across the fleet. Time-to-first- curtailment-event: typically the same day.
More questions? Email engineering or book a working session.
Grow your fleet without waiting years for the grid. Reclaim the power your GPUs waste and turn it into more billable capacity inside the megawatts you already have.
Tame your power bill and protect every point of margin. Pebble shaves the demand-charge peaks that quietly set your monthly rate — no batteries, no SLA hit.
Run Pebble read-only for two weeks. We'll show you your efficiency knee, your reclaimable headroom, and exactly what it's worth — in more GPUs, a lower bill, or grid revenue.