Phaidra
AI agents that autonomously control cooling and power in data centers and AI factories, plus an AI assistant for facility operators.
Phaidra builds AI agents that operate the cooling and power infrastructure of data centers and AI factories, and has applied the same autonomous control to other mission critical facilities, including the chilled water plants at Merck's West Point pharmaceutical campus. Its founding team comes from Google and DeepMind's work on AI control of data center cooling.
The company sells two products. Phaidra Factory is a set of specialized agents that act directly on facility systems.
They include a Liquid Cooling Agent that watches GPU power to anticipate thermal spikes and adjusts coolant distribution units before temperatures rise, a chiller plant agent that manages staging, temperatures and pressures through BMS and SCADA integration to cut cooling energy, and Agentic Power Allocation, which shifts power between cooling and compute through NVIDIA's power APIs.
Phaidra Prism is an AI assistant for operators and technicians that analyzes live and historical equipment data, surfaces anomalies, and answers questions from operating data and system documentation.
The control agents run as a cloud service that writes setpoints into the site's existing control systems. Operators can hand control back to local logic at any time, agents can run in a recommendation mode before autonomous control is switched on, and multiple agents can be coordinated across interconnected plants. Phaidra holds SOC 2 Type 2 and ISO 27001 attestations.
Vendor details
Canonical URL
https://phaidra.ai
Category
Enterprise operations agent
Subcategory
Supply Chain — AI closed-loop industrial control
Use cases & customers
Primary use cases
Target customers
In practice
Synchronized GPU jobs spike your rack power in seconds and the coolant loop reacts too late. Phaidra's Liquid Cooling Agent reads the power ramp as a leading indicator and adjusts the CDUs before the heat arrives, so you can run warmer without throttling.
Your chiller plant runs a fixed sequence of operations tuned for the worst day. Phaidra's agent manages staging, temperatures and pressures through your BMS in real time, cutting cooling energy while operators can hand control back at any moment.
Technicians wade through thousands of alarms to find the one that matters. Phaidra Prism analyzes live and historical equipment data, flags the anomalies that need attention, and answers follow-up questions from your operating data and documentation.
Sources & related URLs
Agentic Index coverage score
7.5 / 14 capabilities · 54%
| Integrations & Tool Calling | Full |
|---|---|
|
Agents read from and write to named real systems. The PUE agent manages the chiller plant (chiller staging, evaporator temperatures, differential pressures) "via a BMS/SCADA integration", and the CDU agent "preemptively controls the CDU". Agentic Power Allocation "integrates directly with NVIDIA's Mission Control Domain Power Service (DPS) via the NVGrid API" to change compute allocation, and it updates power allocation policies using NVIDIA DSX Max-Q APIs. At Merck, the cloud service "updates the local BMS directly". Phaidra is also a named partner contributing to the NVIDIA DSX and Omniverse digital twin blueprints. Phaidra does not say how credentials are scoped or whether customers can add custom tools. Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/blog Merck case study (25 June 2024); readread 2026-09-15 |
|
| Workflow Orchestration | Full |
|
Coordination between agents runs under a coordinating agent. At Merck, four AI Virtual Plant Operators each run one chiller plant, and a fifth, the "AI Conductor", takes in campus level data and "coordinates each individual agent's production of chilled water to optimize load balance between all plants". Deterministic and autonomous steps mix. The agents write setpoints into the existing BMS, which keeps its hard coded sequence of operations, and a bumpless transfer hands control back to local logic. An agent can also send setpoint commands directly to a coolant distribution unit, while the coordination across plants comes from the AI Conductor. Phaidra names no authoring surface, versioning or reuse across teams, and the coordinator was built for one campus. Sourcephaidra.ai/blog Merck case study (25 June 2024), phaidra.ai/products/phaidra-factory; readread 2026-09-15 |
|
| Knowledge Grounding & RAG | Full |
|
Answers in Phaidra Prism are grounded in live data, historical trends and system documentation pulled at question time. Its assistant, Alfred, is there "to pull from live data, historical trends and system documentation to respond to questions, provide context and generate charts", and Prism "provides consistent analysis of historical and real-time data". Because documentation and operating history are pulled when a question is asked instead of trained in, new customer knowledge can be added without retraining. Phaidra does not say which sources are indexed, how they are refreshed or permissioned, or whether answers cite their sources. The control agents work differently. They compute setpoints from live sensor tags, and at Merck the agent was trained on the plant's historical data. Sourcephaidra.ai/products/phaidra-prism, phaidra.ai/blog Merck case study (25 June 2024); readread 2026-09-15 |
|
| Human Oversight & Guardrails | Full |
|
The service has three oversight mechanisms. Operators can pause and resume control, the agent can run in an approval mode, and structured guardrails hold it to system constraints. The service "has a built-in bump-less transfer to local control triggered by operators at any time", and the agents resume correctly after AI Off periods. Before autonomous control at Merck, the agent ran in a manual approval mode that "sent setpoint recommendations to on-site staff for their review and implementation". Staff gave feedback on rejected changes, and the move to autonomous control was the customer's decision. The agent was also "programmed to understand and respect all system constraints". Agentic Power Allocation keeps a conservative margin while "maintaining rigorous site-safety guardrails". All three mechanisms are built into the service, so the customer does not have to supply them. Phaidra does not say whether approvals can route into existing ticketing channels or whether the service explains why a checkpoint fired. Sourcephaidra.ai/blog Merck case study (25 June 2024), phaidra.ai/products/phaidra-factory; readread 2026-09-15 |
|
| Security, Identity & Governance | Partial |
|
SOC 2 Type 2 auditing and attestation is complete, with the report available to customers on request, and Phaidra "has obtained ISO/IEC 27001:2013 certification" through ANAB. Access controls cover Phaidra's own staff. Least privilege and role based access "enforced via the Google IAM service" govern access to Phaidra's production environment. Phaidra says "only specific employees with authorized credentials can access your data" and "all platform access is audit-trailed". It names no access controls that customers operate themselves. Data is encrypted at rest and in transit with GCP KMS. Sourcephaidra.ai/security; readread 2026-09-15 |
|
| Observability & Auditability | Partial |
|
As part of what it calls "AI transparency & explainability", Phaidra says it has developed ways "to illustrate not only what the AI is currently doing, but also how it's performed historically and what it's planning into the future". The aim is to turn the black box into a white box for plant operators. Phaidra names no run history, action log, trace view or audit trail of agent decisions that a customer can open. It does not say whether customers can inspect inputs, actions and outputs step by step, or whether audit logs are kept apart from runtime traces. The audit trail on platform access covers staff access to Phaidra's production environment, not the agents. Phaidra Prism charts chillers, alarms and consumption, which show the customer's equipment, not the agents. Sourcephaidra.ai/security, phaidra.ai/products/phaidra-prism; readread 2026-09-15 |
|
| Memory & State Persistence | Not documented |
|
Phaidra describes no agent memory that a customer can see or govern. It names no memory types, such as session, conversation, workflow or long term memory, and no way to review, edit, delete or scope memories by user, team or workspace. The agents learn instead. An agent trained in simulation improves further after several hours of live learning, and the reinforcement learning agent teaches itself to get better without human intervention. That learning changes the controller's policy. During AI Off times at Merck, the agents "still receive live operational data", so on reactivation they understand what changed. They keep taking in telemetry while control is paused. Sourcephaidra.ai/blog Merck case study (25 June 2024); readread 2026-09-15 |
|
| Deployment & Data Residency | Not documented |
|
The control agents run as "Phaidra's cloud-based service", which analyzes live operational data and updates the local BMS. The service runs on Google Cloud Platform, with GCP Key Management Service and Google IAM. Phaidra names no on premises, private cloud, VPC or edge option. No hosting region is published, and Phaidra describes no way for a customer to pin model traffic or data storage to specific regions or infrastructure. The cloud service writes setpoints into on site BMS and SCADA systems through integration. Sourcephaidra.ai/security, phaidra.ai/blog Merck case study; readread 2026-09-15 |
|
| Prebuilt Agents, Templates & Packs | Partial |
|
The two named products are Phaidra Prism and Phaidra Factory. Factory presents three named agent solutions, each with a stated challenge, solution and result. They are CDU control (the Liquid Cooling Agent), PUE optimization (a chiller plant agent via BMS/SCADA) and increasing IT capacity (Agentic Power Allocation). Phaidra calls these "just two of several Phaidra agents" and says more agents are on the way. The three solutions are presented together, with no individual entries. Each is built for a particular facility, not adopted ready made. At Merck, nearly a year of joint work went to "develop and deploy" an agent for each chiller plant, and the control agents are trained on each plant's history. Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/applications, phaidra.ai/blog Merck case study; readread 2026-09-15 |
|
| Triggers & Channel Coverage | Full |
|
The agents start work on their own, from telemetry events and from a continuous feed of live data. The Liquid Cooling Agent "monitors GPUs to predict thermal spikes" and, "when conditions suggest a spike is imminent", begins the cooling process, with response times under 10 seconds. The PUE agent proactively manages the chiller plant. At Merck, the cloud service continuously analyzes live operational data and updates the local BMS directly. The agents keep receiving live data during AI Off periods, so they respond correctly on reactivation. Phaidra Prism also alerts operators to outliers and anomalies from historical and real time data without being asked. Telemetry and anomaly events are the only starts Phaidra names. It names no schedule, webhook or messaging channel. Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/products/phaidra-prism, phaidra.ai/blog Merck case study; readread 2026-09-15 |
|
| Model Flexibility & Routing | Not documented |
|
On both products, Phaidra chooses the model. The control agents are Phaidra's own reinforcement learning models, trained for each facility, as at Merck. Phaidra calls Prism "the only LLM designed for data center operators", but it names no underlying model or provider and discloses no routing across models. It describes no way for customers to select a model or bring their own key. Sourcephaidra.ai/products/phaidra-prism, phaidra.ai/blog Merck case study; readread 2026-09-15 |
|
| APIs, SDKs & MCP Extensibility | Not documented |
|
Phaidra publishes no API, SDK or MCP surface for its platform, so outside software has no way to call or compose it. Agentic Power Allocation calls NVIDIA's NVGrid and DSX Max-Q APIs, so there Phaidra is the caller, not the one being called. Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/applications; readread 2026-09-15 |
|
| Testing, Debugging & Optimization | Full |
|
Before any autonomous control was allowed at Merck, Phaidra ran a controlled trial. In a month long demonstration, the agent's recommendations ran against the plant's existing sequence of operations, and Phaidra published a comparative performance analysis (Figure 1). The agent showed 16.2 percent energy savings against the local SOO, 70.5 percent better thermal stability and 50.9 percent less excess equipment runtime, after which the customer approved the change. After deployment, staff feedback on recommendations they did not implement was used to improve the agent. Agents are also trained in simulation before deployment. Phaidra runs the trial as part of its deployment process. It names no standing evaluation surface, test fixtures or datasets that a customer runs. Sourcephaidra.ai/blog Merck case study (25 June 2024); readread 2026-09-15 |
|
| Browser & Computer Use | Not documented |
|
Each action Phaidra names goes through a machine interface, and Phaidra describes no browser or computer control. The PUE agent works through a BMS/SCADA integration, the Liquid Cooling Agent controls the CDU, Agentic Power Allocation uses NVIDIA's NVGrid API, and in the Merck deployment the service updates the local BMS directly. Phaidra Prism generates charts for operators. Sourcephaidra.ai/products/phaidra-factory, phaidra.ai/blog Merck case study; readread 2026-09-15 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Pricing
Contact sales
flat
What is public
Phaidra (phaidra.ai - reinforcement-learning AI control system / 'virtual plant operator' agents for mission-critical industrial facilities, primarily AI-factory & cooling (liquid cooling, chiller plants, power, workload management), also refineries/pharma/steel; plugs into existing BMS/SCADA via BACnet/OPC-UA, no new hardware) is enterprise/quote-based - no public price tiers.
Billing mechanics
B2B SaaS subscription (sales-led), no hardware purchase required; priced per facility/deployment (custom); typically positioned so energy savings (~10-30%) offset the subscription within months. No public figures.
Additional watchouts
Custom enterprise pricing only (no public tiers); value framed as energy-savings ROI; mission-critical deployment requires work + shadow-mode onboarding
Sales call required
Yes, required for paid access
Commercial notes
Seattle, founded 2019 (Jim Gao ex-Google DeepMind Energy, Veda Panneershelvam ex-AlphaGo, Katie Hoffman ex-Trane); ~$108.5M+ total raised across multiple rounds - incl. $25M Series A (2022) and $50M+ Series B (Collaborative Fund, Oct 2025) with Helena, Index Ventures, NVIDIA, Sony Innovation Fund + individuals Mustafa Suleyman & Mark Cuban; deployments incl. NVIDIA, CoreWeave, ST Telemedia, Khazna, Merck; competes with Honeywell/Johnson Controls/Siemens/Schneider AI-controls
Support SLA / resale
Supervisory software layer over BMS/SCADA; shadow-mode training then closed-loop control; newer products Phaidra Prism (LLM for ops) + Phaidra Factory (monitoring agents); NVIDIA Omniverse DSX partner
Related vendors
- 11th Estate — Agentic platform whose AI engine scans markets, matches an…
- Abacum — FP&A platform for growing companies whose AI Modeler builds, edits…
- Adopt AI — Finance agents that categorize, reconcile and close the books inside…
- Aera Technology — Decision intelligence platform where an always on agent executes…
- Akro AI — On premise operational intelligence that automates document heavy…
- alfred_ — AI executive assistant that triages the inbox overnight and scores…
Alternatives to Phaidra
The closest documented capability profiles to Phaidra among enterprise operations agents tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Optimal Dynamics7.5 / 14Adds documented APIs, SDKs & MCP Extensibility
- Kyrok6.0 / 14A lighter documented profile than Phaidra
- Qureos7.0 / 14Adds documented Browser & Computer Use
- Unit219.0 / 14Adds documented APIs, SDKs & MCP Extensibility and Browser & Computer Use
- Basis7.5 / 14Fuller documented coverage on Security, Identity & Governance and Observability & Auditability
- Cambio6.5 / 14Fuller documented coverage on Prebuilt Agents, Templates & Packs
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded