Flume Fabric

For the last three years, the world has focused on the build: gigawatt campuses, transformer orders, $500B in announced capacity, and the race to stand up enough hyperscale capacity to feed the next generation of frontier models. But as the announcements pile up and the concrete stays unpoured, a different problem has come into focus.

The models know how to think. The grid can't deliver the power to run them.

Of 40 GW announced for U.S. AI infrastructure, only 1 in 3 of the 12 GW planned for 2026 is actually under construction. The bottleneck isn't capital or chips — it's transformers, switchgear, and a five-year average wait for grid interconnect in every metro that matters. Meanwhile, 85% of "inference" provider revenue is still training, and only 2% of providers can serve a Fortune 500 regulated customer at all.

The Problem: Inference Was Never Going to Live in the Desert

Hyperscale was built for training. One run, one location, one continent's worth of power. Inference is the opposite workload: continuous, latency-sensitive, regulated, and physically close to the user. A hospital running clinical models can't wait 200ms for a round-trip to Northern Virginia. A bank running fraud detection can't put customer data on a shared tenant. A robotics company doing real-time world models needs sub-10ms, not sub-second.

The current answer — wait five years for a substation upgrade, or rent shared GPUs from a hyperscaler — works for neither.

Inference doesn't need more data centers. It needs them in more places.

An Inference Delivery Network

Flume is building the inference layer the same way the internet built content delivery: distributed, redundant, and physically close to demand. Their thesis is that the buildings to host enterprise inference already exist. Someone just has to know how to wire them.

Class A commercial real estate was built for peak occupancy that no longer exists. Post-COVID hybrid work left 1–2 MW of stranded electrical capacity in the average office tower, sitting behind a 480V service that's already energized. Flume taps a single breaker, drops in a 500 kW air-cooled modular rack co-designed with Daikin, and turns up a site in 60 days. Their platform does three critical things:

1. Activates stranded capacity. No permits, no construction, no grid wait. A single breaker tap inside an existing building.
2. Delivers sovereign inference. Air-gapped, single-tenant, HIPAA / SOC 2 / FedRAMP-ready — purpose-built for regulated enterprise.
3. Routes intelligently across sites. BGP/ECMP overflow between metro nodes via private interconnects to AWS, GCP, and Azure, with sub-10ms latency to the customer.

This isn't theoretical. Flume already operates 600+ buildings under master lease across eight states — the legacy of Flume 1.0's dark-fiber business — and is converting those head-ends from optical networking gear to GPU racks. Their first two months of sales produced $4M in ARR, with deals signed at Axion Ray, MIT, Supertrace, and Espresso AI. NVIDIA brought them into the Inception program. Brennan Group signed on the real estate side; BXP, Tishman Speyer, DivcoWest, Hines, and Alexandria are in conversation. 45 MW of power is in the pipeline.

Distributed Inference, Now

We invested in Flume because three things had to be true at the same time, and they finally are:

• The demand is here. Inference is 80–90% of AI compute spend, and the regulated-enterprise slice — air-gapped, on-prem, compliant — is growing at ~45% CAGR. No incumbent serves it well.
• The hardware is here. Frontier-class models are collapsing in size. Qwen, Mistral, Gemma, and the open-source 8–30B class now do work that required 800B parameters a year ago. 500 kW sites are commercially viable for frontier-class inference.
• The gap is infrastructure. A greenfield data center costs $19M+ per megawatt and takes 3–5 years. Flume delivers the same megawatt for ~$2M in 60 days, because the building, the fiber, and the power are already in place.

Prashanth, Brandon, and the Flume team spent five years doing the unglamorous physical work of building a national fiber operation — site selection, utility relations, permitting, landlord trust. That operating muscle is exactly what every neo-cloud is now trying to learn from scratch. Flume is five years ahead.

The future of inference is distributed. Flume is the network.

‍