Yield Control for the Gigawatt AI Factory.
A GPU cluster is a financial instrument with cooling fans. Every 5% yield gap on a 1GW campus is $15B/year in unrealized revenue. The LiquidMetal Yield Control Plane recovers it.
Download LiquidMetal AIStatic binary. No phone-home. No registration. No telemetry. Free up to one rack.
sha256:0000000000000000000000000000000000000000000000000000000000000000
All raw data stays on your hardware. The cloud side runs on summaries, not telemetry. It's entirely optional — the datacenter product runs fully independently.
Deployment
Static binary on your hardware. All raw data stays local. Cloud side is entirely optional — the datacenter product runs fully independently.
- No phone-home
- No registration required
- Works air-gapped
Pricing
$0
Up to 1 Rack
All features enabled. No gates, no time limits, no trials.
$X / GPU / month
Above 1 Rack
Automatic volume discounts at checkout. Buy any quantity with a credit card or ACH — no conversation needed.
The software still runs and all features still work above one rack. It never degrades, crashes, or stops producing data. A banner indicates the deployment is out of compliance with the license.
Cloud side: $Y per 1K tokens. No platform fees, no seat licenses. Set a monthly budget cap — critical alerts always flow. No surprise bills.
The Slow Loop
Agentic Fleet Reasoning
The cloud side is entirely optional. The datacenter product (Physion2) runs fully independently with no cloud connection required. Cloud agents operate on compressed summaries and yield signals — raw data never leaves your facility.
Yield Narrative
Translates yield signals into P&L. Tells you exactly how much Tuesday's thermal event cost in dollars.
> User: How much did Tuesday's thermal event cost in dollars?
> Agent: Checking Agent Memory...
> Agent: Checking Compressed Summaries... Thermal throttle on Cluster 2, 02:03-04:47 UTC.
> Agent: Investigating Cluster 2 Datacenter via Secure Tunnel...
> Response: $2.3M in unrealized revenue. $890K in SLA penalties. Total impact: $3.19M.
Incident Postmortem
Autonomous root-cause analysis. Correlates events across time and fleets to prevent cascades.
> User: GPU 47 thermal-throttled for the third time in 6 weeks — hardware failure or firmware divergence?
> Agent: Checking Agent Memory... Two prior events on GPU 47, both post-firmware v3.2.1.
> Agent: Checking Compressed Summaries... No similar pattern on pre-v3.2.1 clusters.
> Agent: Investigating firmware correlation across fleet...
> Response: Firmware-driven divergence. 73% probability. Recommend rollback of v3.2.1 on affected racks.
Capacity Planning
Projects physical limits against growth curves to predict saturation weeks in advance.
> User: When does Cluster 2 hit saturation at current growth rates?
> Agent: Checking Agent Memory... 8% week-over-week growth in dynamic inference.
> Agent: Checking Compressed Summaries... Power headroom: 12%. Thermal headroom: 8%.
> Agent: Modeling growth curves against physical limits...
> Response: Thermal ceiling in 6 weeks at current trajectory. Not power, not scheduling — cooling.
Fleet Logic
Unified view across multiple GW deployments. Spots firmware-driven divergence.
> User: AMD side is generating more dollars per GPU than NVIDIA this week — workload placement win or scheduling artifact?
> Agent: Checking Agent Memory... AMD yield: 94.2%. NVIDIA yield: 91.1%.
> Agent: Checking Compressed Summaries... NVIDIA divergence started post firmware v3.2.1.
> Agent: Cross-referencing workload placement and firmware versions...
> Response: Firmware-driven, not placement. v3.2.1 causing 3% yield divergence on NVIDIA racks. Rollback recommended.
Conversational Cluster
The natural language interface to your cluster intelligence. Uses a three-tier query flow: Agent Memory → Compressed Summaries → Datacenter Tunnel.
> User: Why did utilization drop at 2am?
> Agent: Checking Agent Memory... No scheduled maintenance.
> Agent: Checking Compressed Summaries... 23% drop on Cluster 2 starting 02:03 UTC.
> Agent: Investigating Cluster 2 Datacenter via Secure Tunnel...
> Response: GPU 47 thermal throttle triggered at 02:03 UTC. Third occurrence in 6 weeks. High probability of imminent failure.
Model Performance
The agent that watches the watcher. Detects model drift and flags for retraining.
> User: The auditor agent says yield baselines are drifting — is the model degrading or has the definition changed?
> Agent: Checking Agent Memory... Baseline recalculated after firmware v3.2.1 update.
> Agent: Checking Compressed Summaries... Pre-v3.2.1 baselines differ by 2.1% from current.
> Agent: Comparing model predictions against recalibrated ground truth...
> Response: Definition change, not degradation. Firmware update shifted rated capacity. Model is sound — baselines need recalibration.
The Fast Loop
The Edge Binary
Phase-Aware DVFS
Reclaims power during memory-bound decode phases.
Topological Steering
Reallocates GPUs between TP and Wide-EP at batch edges.
Thermal Headroom as Capacity
Biases cooling loops ahead of correlated prefill bursts.
Works With Your Existing Stack
Outputs data in formats compatible with Prometheus and Grafana. We're not observability — we're a better data source for the tools you already run. Point your existing stack at us like anything else.
Fully Independent
The datacenter product runs with zero cloud dependency. No phone-home, no registration, no telemetry. It always works, even air-gapped.
Architecture
Data flows from the datacenter binary through compressed summaries to cloud agents and back. Raw data never leaves your facility.
Results
Results and references from design partners will appear here.