DIMAGGI AI · Infrastructure Research

Turning GPU capital into usable compute

An open, reproducible research program on the layers that decide how much of an AI cluster actually does useful work — and why fixing them usually beats buying more GPUs. Every result is seeded, tested, adversarially reviewed, and honest about what a skeptic should attack.

usable = nominal × network × scheduling × recovery × placement, under a governance layer

The series

Each repository isolates one factor, in the field's own metrics, with a runnable artifact.

GPU Cluster Networking

Past a scale threshold, interconnect bandwidth — not GPU count — caps AI training throughput.

GPU Cluster Scheduling

Reclaiming stranded capacity across Slurm, Kubernetes, and multi-cloud — allocation is where capacity is realized or lost.

Chaos Fidelity Standard standard · pip

A machine-checkable standard for which fault injection tests which layer. A pod kill is not an XID; tc/netem is not an InfiniBand flap.

Reliability Economics

Which recovery policy wins in which failure regime — a phase diagram in ETTR and dollars, validated against Meta's published reliability numbers.

Governed Autonomy pip

An autonomous control plane over clusters and networks: separated planes, a referee, an autonomy ladder earned by chaos experiments, and a sourced latency hierarchy from human intent to nanosecond in-ASIC reflexes.

Compute↔Power Placement

When does moving a job to cheaper power beat staying? Energy is ~3% of GPU-hour cost — so schedule against scarcity, don't arbitrage the spread. A fleet model prices the scheduler against itself with its ranking removed: ranking the curtailment hours by price wins in every simulated price year, and the headline saving is bound by the assumed spike ceiling — 4.8% at $5,000/MWh, 2.7% at ERCOT's $2,000 real-time cap.

Edge↔Neocloud Placement pip

Where AI workloads belong on the operator's footprint — tower, hub, metro PoP, central factory. The tower is never the recommended tier; latency is the weakest edge argument; fronthaul (~20 km), data gravity, and sovereignty do the actual forcing.

AI-RAN ↔ Neocloud Resiliency pip

The last mile priced in nines: a single-fiber edge site is a ~2.7-nines site; multi-transport bonding helps only the tenant classes; the generator — not a third link — buys the decisive nine.

Cooling & the PUE Ladder pip

Heat out, priced in PUE, megawatts, and seconds: a 120 kW rack leaves only two rungs standing; a PUE point converts to ~1.9 MW of unqueued capacity per 10 MW feed; and one adoption mechanism reproduces the survey "plateau" and hyperscale ~1.1 fleets at once.

Tools you can install

The standards ship as commands, not just prose — run them in your own CI.

Chaos-fidelity linter — reject a wrong-layer chaos experiment:

pip install ai-cluster-chaos-fidelity
chaos-fidelity my-experiment.yaml

Autonomy promotion gate — refuse an unearned autonomy promotion:

pip install governed-autonomy-gate
promotion-gate my-record.yaml

Placement gate — block an edgeward placement no gate forces:

pip install edge-continuum-placement
edge-placement evaluate -w frontier-pretrain -t tower  # exit 1

Resilience ladder — price the next nine before buying it:

pip install airan-neocloud-resiliency
airan-resiliency ladder

Cooling density gate — block a rack no rung can cool:

pip install cooling-pue-ladder
pue-ladder gate --density 120 --rung contained-air  # exit 1

How the claims are checked

Every repository ships a validation program that CI runs on each push, and every finding is filed under one of three kinds, labeled in the output: calibrated — pinned to a published figure, where passing proves the model has not drifted from its citation, not that it predicts; emergent — results the model was not tuned to produce, required to survive independent reseeding; and sanity — properties of the model's own structure, citing no evidence and claiming none. The newer registries add two things: they print the anchors they do not provide, so an uncalibrated input cannot pass for a result, and their test suites break the model on purpose — delete a placement gate, delete the power chain, switch off the price spikes — and require the registry to go red.

That is not decoration. In the most recent pass it caught a generator threshold chosen after seeing the number (it failed on reseeding, and was replaced by an ordering with nothing to tune), a tie between transport options stated as a maximum rather than a mean (same fix), a latency registry that passed with three of its four gates deleted, and a concentration figure that was identical under every scheduling policy — a fact about the electricity market being reported as a result about the scheduler. Each was rewritten in public rather than quietly kept.

The method

Machine-checkable where possible; calibrated to primary sources and cross-checked against published fleet data; conditional claims with a stated "what a skeptic should attack"; and each result adversarially reviewed before release. The goal is not a slogan but a body of work an engineer can run and a reviewer can trust.