We’re excited to share that Socratic Partners has led Delos Data’s Series A, and Christian Gallagher, General Partner and Co-Founder at Socratic Partners, will be joining the board.
We spent a long stretch of diligence getting to know Ed, Dan, and the broader team at Delos Data before writing this check and we’re delighted to be investing alongside Matrix Partners, Playground Global, Capricorn, Matter Venture Partners, IAG, DYNAMIQ, and other great investors.
For years, the network inside a data center was plumbing: necessary, but never the interesting part. That’s changed. As AI clusters scale from tens of GPUs to thousands, the connections between them now decide GPU utilization, and utilization has become the real driver of the economics behind every AI infrastructure dollar.
The Missing Middle
Operators building AI infrastructure today are stuck choosing between two imperfect options. Proprietary scale-up fabrics deliver the bandwidth and latency modern AI workloads need but cap the number of GPUs you can tie together and lock you into a single vendor’s stack. Standards-based networking preserves flexibility and vendor choice but wasn’t built for how tightly synchronized AI workloads move data, and it shows up as materially higher latency at cluster scale. As inference overtakes training as the dominant AI workload, that gap stops being a technical footnote. Training can absorb a stall and rerun the job. Inference can’t. Every inefficient hop shows up immediately as lost throughput.
Everyone’s Solving Half the Problem
This is not a market short on attempts. One camp keeps extending proprietary chip-to-chip interconnects outward toward larger and larger clusters. The other keeps pushing standards-based networking to behave more like a tightly coupled fabric. Both are running into the limits of assumptions that made sense at a smaller scale. What’s been missing is a layer purpose-built for this problem rather than stretched to fit it.
Why Delos, Why Now
Delos is building that layer: a silicon and software platform that sits between the GPU and the network and starts from a different assumption than either camp, that hardware at this scale will fail, so the system should be built to detect it, isolate it, and keep running rather than reset and start over. That is a meaningfully different posture than the industry’s current answer, which is largely to overprovision spare hardware and hope. The market backdrop is significant on its own: independent analyst estimates put the AI interconnect market on a path from roughly $10B today toward $100B by 2030, as spend shifts toward the infrastructure layers that determine utilization and throughput rather than compute alone.
Why We Bet on Ed, Dan, and the Delos team
Delos is founded and led by Ed Doe (CEO) and Dan Daly (CTO), who have been building networking silicon together for nearly two decades, dating back to Fulcrum Microsystems and continuing through Barefoot Networks and Intel. That kind of long-running technical partnership is rare, and it shows in how the team operates. They’ve taped out chips together before, been through an acquisition together, and have since built a stellar team drawing from Broadcom, Intel, Google, and Cisco. Every reference we ran during diligence came back the same way; this is one of the strongest teams in the space. They’re building out the team fast to match, so if this is a problem you know well, reach out, they’re hiring.
What’s Next
We’re proud to partner with Ed, Dan, and the Delos team as they build the layer AI infrastructure that is going to needed at every step from here. We believe the network, not the compute, is where the next real bottleneck shows up, and we’re glad to be backing this team as they solve a critical piece of that problem.