Phaidra

AI workloads quickly generate massive amounts of heat. Traditional controls react to temperature spikes, but intelligent ones should prevent them entirely. Cooling Distribution Units manage thermal control for AI factory infrastructure. GPU temperatures must stay below a specific thresh-hold. Overshoots risk system performance and equipment stress. Traditional control reacts when temperature spikes. By then, it's already a problem. AI agents trained on Class 3 and Class 4 data work differently. Class 3 data teaches the agent how the system behaves across operational conditions. The agent learns cause and effect. Class 4 data takes it further. When early signals emerge in production, the agent already knows what's coming. The result is proactive control. Not reactive. Not even fast reactive. Predictive. This is what Class 3 and Class 4 data enable: AI that understands system dynamics and anticipates what comes next. Full breakdown of how this data classification framework here: https://lnkd.in/eXDJu4pa

Hey folks! If you've wondered how agentic AI companies like Phaidra actually utilize digital twins to rapidly train / evaluate / iterate their AI agents — check out this blog post. The takeaway is simple: it's much faster to co-design #AIfactories (and the AI agents operating them) in simulation than on production systems. One does not replace the other course. They are complementary processes. In fact, we share data about how our bootstrapped liquid cooling AI agents (i.e. AI agents trained first in simulation) performed out-of-the-box before we further fine-tuned them on live data via online #reinforcement #learning. The result is AI agents that teach themselves to reduce the magnitude of thermal spikes (from synchronized AI workloads) by > 80% while meeting or improving SLAs. Not bad but this is just the beginning. You're going to see many more AI agents from Phaidra in the near future that are trained to perform specific tasks on behalf of our customers to help scale #datacenter capacity. These lessons learned and capabilities will be openly shared as part of our collaboration on NVIDIA's #Omniverse #DSX blueprint (i.e. a reference design for gigawatt-scale AI factories) so agentic AI companies can easily plug into AI factory infrastructure and vice versa. More details below 👇.

Inefficiencies are magnified at gigawatt-scale. If every 1GW AI factory represents both a $50B investment and a $200B revenue opportunity, every % of inefficiency represents billions in lost revenue How do we maximize the performance, as measured by tokens per watt, of these mission-critical assets? As Jensen Huang noted during the #GTC DC keynote, the answer is: extreme co-design. An AI factory cannot be designed as a collection of loosely-orchestrated components. It must instead be built and operated as a single integrated machine. Failing to orchestrate these components results in substantial underutilization of the asset. In the near future, AI factories should operate continuously at peak performance — aided by AI agents that vigilantly manage the complex infrastructure on a 24/7 basis. Read an update from co-founder & CEO, Jim Gao , about Phaidra’s work with NVIDIA and how agentic AI companies like Phaidra contribute to the NVIDIA #Omniverse #DSX #Blueprint: a comprehensive and open blueprint for designing and operating gigawatt-scale AI factories. https://lnkd.in/ehFwSHrp

"Extreme performance requires extreme co-design" - Jensen Huang from the NVIDIA #GTC keynote state in Washington, D.C. earlier today. Shortly after this statement, Jensen mentioned our engineering partnership with NVIDIA. We're incredibly proud to collaborate alongside the NVIDIA engineering team and the broader AI factory ecosystem for #Omniverse #DSX: a comprehensive blueprint for designing and operating gigawatt-scale AI factories. These computing systems are so large, interconnected, and complex that they must be operated as a single integrated system to maximize performance. Our contribution is the liquid-cooling AI agent that eliminates thermal spikes resulting from synchronized AI workloads, enabling an AI factory to run at higher TCS temperatures. This means less power required for cooling and more power available for revenue-generating GPUs. See more about NVIDIA's announcement on Omniverse DSX in the blog linked in the comments below

An AI could never find the words to describe the feeling of giving your all to a startup, and seeing your company logo in the NVIDIA keynote. So proud of Phaidra, this team, and so thankful for the opportunity to build the future.

Phaidra Leadership Team member spotted at NVIDIA #GTC

It was a privilege to share the stage for a fireside chat with Geoffrey Moore and Jim Gao at the Phaidra Summit this week. The talent of the Phaidra team is remarkable and the energy at the summit was off the charts. They have the right team at the right time and place to tackle the data center industry’s significant energy challenges. Phaidra is developing autonomous AI agents that continuously optimize the operation of compute facilities, from liquid cooling CDUs to massive chiller plants. They’re expanding their AI agent ecosystem beyond cooling into the complete AI factory compute stack, collaborating closely with NVIDIA on gigawatt AI factories, including NVIDIA's reference design and accompanying omniverse digital twins. I’m proud to be part of this extraordinary team’s journey! #AI #AIfactory#Datacenters #Phaidra #omniverse Katie H. Vedavyas Panneershelvam Christopher Vause Starshot Ventures Index Ventures Flying Fish Partners Mark Cuban firstminute capital Mustafa Suleyman firstminute capital Character Capital

Everyone wants autonomous AI. No one wants to do the work required to clean their data. That’s what makes industrial data so tricky. It looks complete. It sounds correct. But if you zoom in—just a little—you’ll see the cracks. → Drifted sensors still reporting as if correct → Values that break the laws of physics → Metadata that makes sense only to the original engineer And when that goes into your AI system? You get predictions that are precise —but totally wrong. This is where most agentic AI fails: it believes the story your data is telling. 1/ Bad data looks normal until it doesn’t. 2/ AI flags the wrong things—because it’s fed the wrong signals. 3/ Training models on broken baselines locks in your worst habits. 4/ Small sensor errors become large operational risks. 5/ Data cleaning isn’t just prep—it’s the whole job. 📍 You don’t need perfect data. But you do need to know what class of data you’re working with—and what to fix. Know what class of data you’re working with and what to fix: https://lnkd.in/ezXv87PD

Great summary of our recent Series B announcement from GeekWire Phaidra's AI agents for liquid-to-chip cooling systems will help AI factory operators maximize tokens per watt and minimize cooling power needed to do it. https://lnkd.in/eArDUaVa