Congratulations to Alon, Omer, Ben, Eyal, and the entire team at Atero.ai on their acquisition by Crusoe. The combination of Atero’s Unified Memory architecture with Crusoe’s energy and data center infrastructure will enable the next generation of efficient LLM platforms for training, inference, and reasoning. It also highlights that racking and stacking GPUs is not enough; system software is vital to automate operations and the key differentiator in the tokens per $ “tokenomics”.
I’ve been in venture capital for a long time, but this is the first time I’ve been able to combine a “This is why I invested” post with a “Congrats on the acquisition” post in a single blog. Atero was the third investment from Encoded, but my relationship with Alon goes back further, to his prior company Qubex. When we met, I knew I wanted to work with Alon. He is the exact kind of technical polymath and concert-level violinist that I look for. I’ve had good luck with high-level musicians! However, anyone in venture capital has experienced the feeling of liking the founder more than the idea. While Qubex had a novel approach for auto-scaling to solve the “Taylor Swift crushed my website” problem, the market seemed limited. But Alon and I stayed in close touch.
Given my prior investments in Habana (acquired by Intel) and Cumulus Networks (acquired by Nvidia), along with my long friendship with the Cerebras team, I had a strong intuition about the requirements for efficiency in large-scale GPU clusters. In the spring and summer of 2024, based on many conversations with people scaling LLM infrastructure, I began to codify my thoughts in a working paper called, “Where my FLOPS at?!?” TL;DR: FLOPs are abundant; getting them the data they need, when they need it, in the right place (level of memory and storage hierarchy) is the real bottleneck.
I reconnected with Alon in late spring of 2024. At that point, he had the vision of combining a low-level understanding of the memory hierarchy with abundant in-cluster RDMA networking bandwidth to build a distributed memory and caching framework. He had also assembled a world-class team. Omer brought deep, low-level system thinking from the Israeli military; Eyal has been defining and building RDMA protocols since they were popularized in the early InfiniBand days at Mellanox; and Ben brought large-scale GPU cluster architecture and operational experience from five years during the foundational days at OpenAI.
I teamed up on some of the diligence with my friends Brian and Tobias at Primary Ventures, who led the seed round. I also got to build stronger relationships with some of the angels who I knew but hadn’t worked with, especially Asaf Ezra, Haleli Barath, and reconnect with old friends Julian Levy, and the OG, Mark Leslie. Progress was swift, and before too long, Atero began working with some major design partners. Technical progress was impressive, benchmarks were very strong. Soon, one thing led to another, culminating in today’s acquisition.
For founders, selling their company is always a difficult and often emotional decision. Atero was no different. But what made this combination so exciting was the company across the table, Crusoe. They have demonstrated the ability to execute massive-scale power acquisition and data center buildouts. They have earned the trust of OpenAI, Oracle and are at the center of the massive Stargate project. Crusoe has also proven its ability to attract many of the most sophisticated investors in the data center ecosystem, which is critical in this capital-intensive business. I’m proud that Encoded is now part of that list.
I look forward to working with the Atero team for many years to come.
Alon, Omer, Nir, and myself at the airport in Tel-Aviv heading to Nvidia GTC
