The working framework

AI is not a software story, a hardware story, or a networking story.

It is the accelerating cycle between all three — and the cycle is governed by economic laws that compound, not converge.

The framework the weekly publication uses to filter signal from noise. Each issue tests it; each turn of the cycle revises it. Read this if you want to know why we’re tracking what we’re tracking, what we treat as noise, and what evidence would change our minds.

The flywheel

Three lenses. Four laws. One cycle.

Most coverage picks one lens. Models get most of the attention; hardware gets some; networking gets almost none. Read separately, the three lenses produce predictable misreads — efficiency is bullish for compute (or bearish), open weights threaten NVIDIA (or don’t), bandwidth is the bottleneck (or isn’t). Read across the loop and the contradictions resolve: a software breakthrough drives chip demand; chip ramps need fabric ramps; fabric ramps unlock new software.

Software → Hardware: new model capability creates new use cases that consume more compute. (Jevons: efficiency creates demand, not less of it.)
Hardware → Networking: more compute means more nodes, which means the value of the connecting fabric scales as the square of the nodes. (Metcalfe.)
Networking → Software: denser, higher-bandwidth interconnect makes new model architectures viable — multi-data-center training, distributed inference, hybrid deployments. (Gilder: bandwidth grows ~3x faster than compute.)
Hardware itself: GPU performance doubles faster than Moore’s Law. (Huang.)

The events that actually matter touch at least two of the three lenses. The events that don’t are noise. Each turn of the cycle is faster than the last, which is why the publication ships weekly — the doubling cadence outruns a monthly read.

The four laws

What to watch. What to ignore.

Each law tells you what signals to track, what noise to discount, and what evidence would invalidate it. The weekly issues track all three.

Jevons Paradox.

Software lens

When the cost of using a resource falls, total consumption rises, not falls.

Origin: William Stanley Jevons, 1865, observing UK coal use after steam engines became more efficient.
Why it matters in AI: When a lab cuts pricing or ships a smaller efficient model, treat it as a leading indicator for hardware demand, not a competitive threat to GPU revenue. The story most coverage gets wrong: 'efficiency hurts NVIDIA.' The story to track instead: how many net-new workloads are attempted in the 90–180 days after the cut, and what they consume. Cheaper inference expands the addressable surface; more workloads means more aggregate compute, not less.
Predictive value: Watch: per-token API pricing changes, open-weight efficient model releases, on-device inference benchmarks. Ignore: claims that efficiency reduces aggregate compute spend. Within 90–180 days of a major price cut, expect aggregate tokens served (and the GPU revenue tied to them) to rise.
Falsifiability: If a major price cut produces a measurable, sustained drop in aggregate tokens served, the law fails. Not yet observed at any meaningful scale.

Huang's Law.

Hardware lens

GPU performance for AI workloads doubles every 1–2 years on a faster cadence than Moore's Law.

Origin: Wired (2020), describing the cumulative gains in NVIDIA's AI silicon since the early 2010s.
Why it matters in AI: The doubling story most coverage tracks (per-GPU FLOPS) is the wrong metric for 2026. The story to track: per-rack power and per-rack throughput, because the binding constraint moved from silicon to grid interconnect and HBM. A platform that doubles FLOPS but only marginally moves rack-level power is a generation of marketing, not capacity. The effective cost-per-task is still falling an order of magnitude every 18–24 months when both lenses are combined — but only if you're measuring the right thing.
Predictive value: Watch: per-rack power deltas, HBM4 supply qualification status, custom-silicon share of incremental compute. Ignore: GPU FLOP claims that don't translate to rack-level throughput. A next-gen platform delivering <1.5x compute over its predecessor flattens the doubling slope; custom silicon above 30% of incremental compute compresses merchant pricing power.
Falsifiability: If a major next-generation platform delivers <1.5x compute over its predecessor, the law's slope flattens. Watch Rubin, MI400 series, and TPU v8 against the doubling trajectory.

Metcalfe's Law.

Networking lens

The value of a network scales as roughly n² (the square of the connected nodes), not n.

Origin: Bob Metcalfe, founder of 3Com, formalized in the early 1980s.
Why it matters in AI: Most coverage tracks raw colo capacity (megawatts) as the proxy for data-center health. The story to track for hybrid AI: cross-connect and fabric revenue at the operators that connect cloud, neocloud, and on-prem. Capacity scales linearly; interconnect value scales as the square of connected nodes — and that's where pricing power compounds. Hybrid deployments don't dominate because operators picked them; they dominate because each new connected node multiplies the value of every prior node.
Predictive value: Watch: cross-connect, fabric, and IX revenue line items at colo operators (Equinix, CoreSite, Iron Mountain, others), quarter over quarter. Ignore: raw megawatt absorption as a standalone indicator of category health. If interconnect revenue growth doesn't outpace compute revenue growth across two consecutive quarters, the law fails for AI workloads.
Falsifiability: If interconnect revenue growth lags compute revenue growth quarter after quarter, the law fails for AI workloads. Currently the opposite is observed.

Gilder's Law.

Networking lens

Bandwidth grows roughly 3x faster than compute power.

Origin: George Gilder, Telecosm (2000).
Why it matters in AI: Coverage often treats networking as a hardware afterthought to compute scaling. The architecture story to track: bandwidth is being delivered ahead of the compute it serves, which means hybrid AI architecture decisions made in 2026 should plan around fabric headroom, not fabric scarcity. Designs constrained by the assumption that the network is the bottleneck are misreading the curve. The binding constraint of distributed AI has already shifted from 'can we move data fast enough?' to 'do we have the compute to use it?'
Predictive value: Watch: 1.6 Tbps and co-packaged optics shipment volume, optical transceiver supply ratio against HBM, multi-DC training architecture announcements. Ignore: bandwidth-as-bottleneck framing in vendor pitches. If bandwidth growth stalls (optical packaging supply, fiber buildout) while compute keeps doubling, hybrid AI architectures stall — that's the falsifiability test.
Falsifiability: If bandwidth growth stalls (e.g., optical packaging supply, fiber buildout) and compute keeps doubling, hybrid AI architectures stall. Watch HBM vs optical transceiver supply ratio.

Why three lenses

The flywheel-spanning events are the only ones that matter.

Single-lens reads produce predictable misreads. Capex stories that ignore the workloads they’re funding. Model releases that ignore the chips and fabrics they need. Fabric stories that ignore the workloads they enable. The events that actually move the industry touch at least two lenses; the rest is noise dressed up as motion. Tracking the flywheel means tracking the connections.

Software → hardware: DeepSeek V4 (open weights crossing frontier on coding) → enterprise on-prem GPU procurement signal.
Hardware → networking: NVIDIA’s $2B Marvell investment + NVLink Fusion → cross-vendor scale-up fabric standard.
Networking → software: Spectrum-X + 1.6 Tbps co-packaged optics → multi-data-center training viability.
All three together: Anthropic’s 8.5+ GW commitment across NVIDIA + Trainium + TPU + Broadcom networking — the largest single-counterparty stack-spanning commitment ever.

The five hypotheses

Held until disproven.

Each issue tests these. Two consecutive issues of counter-evidence and the corresponding hypothesis is revised in writing here.

The cycle is accelerating, not slowing.

Each turn of the software–hardware–networking flywheel happens faster than the prior turn. The doubling-time cadence — capability doubling, density doubling, bandwidth tripling — is shortening.

Evidence as of inaugural issue: Open-weight models crossed the closed-frontier coding line in 6 months (April 2026), not the 18 months a 2024 baseline would have predicted. NVIDIA shipped Vera Rubin samples to customers within 12 months of Blackwell Ultra GA — half the prior generation gap.
What would change our mind: Two consecutive quarters where any one of the three lenses' doubling-time slope flattens (e.g., next-gen GPU at <1.5x; bandwidth growth stalling).

Capital is concentrated, returns are diffuse.

AI capex is concentrating in a handful of hyperscalers, neoclouds, and frontier labs — but returns are diffusing across enterprise software, productivity, and on-prem deployments. The capex bet doesn't pay off where the capex is spent; it pays off elsewhere.

Evidence as of inaugural issue: ~$700B aggregate hyperscaler capex 2026 vs ~$120B AI-attributable revenue. CoreWeave $66.8B backlog (+4x YoY) without proportionate revenue. Meanwhile, every F500 has Copilot or equivalent deployed.
What would change our mind: Hyperscalers disclose AI-segment-specific operating margin approaching non-AI cloud margins. Or capex revisions turn negative.

Networking is the durable layer.

Models commoditize. Chips margin-compress as competition intensifies. Networking and interconnect is the layer where pricing power holds longest because Metcalfe's Law makes each new connected node multiply existing value.

Evidence as of inaugural issue: Equinix Fabric Intelligence launch (April 15, 2026) acknowledges AI-native networking as a product category. NVLink Fusion + UALink standardization. Cross-connect and IX revenue growth at colo operators.
What would change our mind: Interconnect revenue growth lags compute revenue growth for two-plus consecutive quarters. Or a wireless / optical breakthrough that bypasses physical interconnect.

Open weights pull the floor up.

Open-weight models from China, Europe, and US labs are not primarily a competitive threat to closed labs — they are a demand catalyst for enterprise on-prem and sovereign-cloud GPU consumption. They re-route compute demand without reducing it.

Evidence as of inaugural issue: DeepSeek V4 (April 22–24, MIT-licensed, 1.6T MoE, 1M context) crosses the closed-frontier coding line. Sovereign and national programs surge: UAE 1 GW Stargate, Germany National DC Strategy, Mistral-Sweden, GMI Japan, IndiaAI ramp, UK AI Growth Zones aggregating $38.5B.
What would change our mind: Open-weight performance gap reopens >10 points behind closed for two consecutive quarters. Or regulatory capture restricts on-prem deployment.

Power is the binding constraint for the next 24 months.

The constraint on AI compute deployment is no longer chips, capital, or even land. It is electricity — generation, transmission, time-to-energization. This shows up as PPAs, behind-the-meter generation, and grid interconnection queues.

Evidence as of inaugural issue: Meta 6.6 GW nuclear, AWS 1.92 GW Susquehanna, Microsoft-Chevron $7B Texas gas in talks, 75% of new on-site DC power = natural gas, PJM 2026/27 capacity auction at $329/MW-day (vs $29 prior), interconnect waits 36–48 months in busiest US markets.
What would change our mind: A material grid expansion materially reduces interconnect queue length within 12 months. Or a breakthrough in efficient AI inference (e.g., 10x lower-power chips) collapses demand growth.

How the thesis evolves

Each issue may revise one or more hypotheses based on accumulated evidence. The change log at the bottom of this page notes what shifted and why.

Over the next 12–24 issues we expect to strengthen Hypothesis 3 (networking durability) with disclosed enterprise three-tier architecture data; refine Hypothesis 2 (capex math) as Q1, Q2, Q3 hyperscaler prints reveal margin trajectories; test Hypothesis 4 (open-weight catalyst) by tracking F500 self-hosted announcements; and revise Hypothesis 1 (acceleration) if any lens’ doubling time flattens. Hypothesis 5 (power) is the most evidence-rich today and the most likely to remain stable.

Edits by issue

2026-04-25 (v1, inaugural): All five hypotheses authored. No revisions. Flywheel diagram and four laws established as the conceptual spine.

Subsequent issues append entries here.