Phaidra and the cooling layer
Phaidra optimizes the thermal domain. Synestra coordinates across all of them.
Sophisticated buyers will ask about Phaidra directly. The answer is not that we compete. The answer is that we operate at different layers of the same stack — and a facility running both gets more from each.
Synestra is the operational intelligence layer for AI factories.
The short answer
Phaidra prevents GPU throttling. Synestra answers what throttling cost — and closes the gap between infrastructure potential and delivered economic output.
These are different problems at different layers. The thermal telemetry Phaidra produces is exactly the kind of signal Synestra's EA model needs from the cooling domain. Not competing. Complementary.
Who Phaidra is
A serious company with a real capability in a real domain.
Phaidra was spun out of Google DeepMind in 2022. They apply reinforcement learning to cooling system optimization — specifically liquid cooling CDU (coolant delivery unit) management for GPU clusters. They are NVIDIA-backed, have raised over $50M, and are deployed at CoreWeave and Applied Digital. Their system controls CDU behavior in real time to prevent GPU junction temperature from reaching throttle thresholds. That is genuinely valuable in a world where liquid-cooled AI factory construction is accelerating faster than thermal management expertise.
Where each system sits
The stack is not flat. Synestra is the coordination layer above all domains — including cooling.
Each domain layer is managed by a system that sees only its own domain. Synestra sits above all of them and reads cross-domain consequence chains that no individual layer can observe.
How Phaidra’s output becomes Synestra’s input
Thermal intelligence becomes economically consequential inside the coordination layer.
Phaidra prevents throttling. That is the thermal layer working correctly. Synestra answers the question that follows: what did the throttling event that almost happened cost in compute yield? What is the economic value of the thermal headroom Phaidra created? How does that thermal headroom interact with current power cost, network congestion, and workload SLA requirements to determine where the next training job should run? Those questions cannot be answered from inside the cooling domain. They require the coordination layer above it.
The direct comparison
Different optimization objectives. Different scopes. Different value.
The case for deploying both
A facility running Phaidra and Synestra gets more from each.
Phaidra's CDU control produces better thermal telemetry — more precise, higher frequency, better correlated to GPU state — than passive BMS polling alone. That higher-quality thermal signal improves the precision of Synestra's EA model in the cooling domain. At the same time, Synestra's cross-domain consequence intelligence can surface the economic value of Phaidra's interventions in terms that inform infrastructure investment decisions: the PUE improvement Phaidra creates has an EA value that shows up in the EA index. A thermal optimization decision can be evaluated not just as a cooling efficiency metric but as a measurable EA improvement across the full campus.
The two systems answer different questions. That is why deploying both creates more value than either creates alone.