Thinking Session · May 2026

Edge Compute & AI Infrastructure
— What We Figured Out

Using Armada (armada.ai) as a concrete anchor to build reusable mental models about the economics, architecture, and future of distributed compute.

~80%
Earth's surface with no viable cloud alternative
$226M
Armada total raised (Founders Fund, M12, Shield Capital)
<5ms
On-site inference latency — mostly theoretical today
400 Gbps
InfiniBand bandwidth required for frontier AI training
Top 3 Insights
Insight 01
The real value is compute availability, not speed
The "edge is faster" narrative is mostly marketing. The defensible case is bringing GPU capacity to the ~80% of Earth where no cloud alternative exists. Bandwidth economics — processing 5 Gbps of camera video locally vs. routing it to Virginia — is the strongest near-term commercial argument.
Insight 02
Starlink dependency is the single biggest product risk
Atlas routes entirely through Starlink. Russia actively jams GPS and Starlink terminals in Ukraine. The unsolved problem: Galleon-to-Galleon mesh when all satellite links are simultaneously denied. goTenna solves the last mile; the node-to-node backbone in contested environments is still open.
Insight 03
Armada is an inference play — and that defines everything
Frontier training requires 10,000+ GPUs at 400 Gbps in one room. Armada can't compete there. But inference is stateless, parallelizable, and runs well on a Galleon. As AI shifts toward agentic, real-time, and autonomous applications, the edge inference market grows — and Armada's timing is right for that shift.
🏗️
Layer-by-layer data center anatomy: power, cooling, connectivity, compute, management
🧠
Training vs. inference: structurally different problems, not just different workloads
📡
Starlink as single point of failure; MANET gap for multi-Galleon denied environments
Backbone is not the bottleneck: the constraint moved to power grids and GPU scarcity
🏟️
Stress-testing use cases: sporting events, US metro inference, defense sovereign compute
🚀
Infrastructure-application timing: the killer apps come after the infrastructure, not before
Training
Orchestra: 10,000 musicians, one room, 400 Gbps sync, weeks-long performance
VS
Inference
Sheet music: anyone can play it, anywhere, independently, any time
Edge latency argument
50ms savings on a 3-second model response — imperceptible to humans today
VS
Bandwidth economics
5 Gbps of raw video locally → metadata trickle upstream — real money, real today
Networking bottleneck (2005)
Netflix solved routing hundreds of terabits of 4K video — backbone is overbuilt
VS
Today's bottleneck
Power grid capacity + GPU manufacturing scarcity + internal memory bandwidth
How We Worked
🔬
First Principles
Decomposed "data center" layer by layer rather than accepting the term as monolithic
🎯
Use Case Stress-Test
Separated claimed value drivers from actual value drivers for each scenario
📐
Diagrams for Clarity
Visual rendering forced precision — you can't draw a vague architecture diagram
🎓
Domain Depth as Lens
PhD MANET background surfaced the Galleon mesh gap that generic analysis would miss