On June 1, 2026, independent researcher Jarid Shaub published a paper titled "Artificial Intelligence or Substrate Drift? A Falsifiable Model of Physical Limits to Intelligence Scaling." It introduces the Substrate-Limited Intelligence Hypothesis (SLIH) — a formal, falsifiable model of the exact constraint that GridReadiness was built to navigate.
This is not a blog post. It is a scientific paper with a formal model, falsification conditions and a Monte Carlo robustness layer. And its core claim aligns precisely with what practitioners, investors and grid operators are discovering in the field.
THE CORE CLAIM — IN ONE EQUATION
Shaub defines the effective substrate as:
And the Omega Substrate Pressure Index as:
Ω(t) = D_I(t) / S_eff(t)
Where D_I(t) is intelligence-support demand (AI compute load) and S_eff(t) is the weakest physical layer sustaining it.
When Ω(t) > 1: substrate pressure exists. The system is constrained.
When Ω(t) → ∞: deployment collapses.
The minimum operator is the critical insight. A data center cannot operate on abundant capital but insufficient firm power. It cannot use abundant power if the grid connection queue is 10 years long. It cannot use the grid connection if transformers are on 5-year backorder. Every layer must be present simultaneously. The slowest layer determines everything.
WHAT THE MODEL FINDS
AI boom scenario: Ω = 28.45
Combined constraints: Ω = 45.60
Fast nuclear buildout: Ω = 16.50 — "helps but insufficient"
Efficiency rebound: Ω = 22.10 — worsens pressure (Jevons Paradox confirmed)
Infinite substrate null: Ω → 0 — model is falsifiable
Monte Carlo (40 runs): mean Ω = 16.85, std = 1.92 — all runs show pressure
The dominant bottleneck in most scenarios: P_firm — firm power. The second bottleneck: G_grid — grid capacity. This is exactly the hierarchy GridReadiness documents monthly: transformer procurement determines whether the grid connection can be used; the grid connection determines whether the power can flow.
THE BOTTLENECK MIGRATION — WHY SINGLE SOLUTIONS FAIL
Section 5.8 of the paper contains the most operationally important finding for data center developers:
"A system constrained by the minimum sustaining layer does not become unconstrained because one layer improves. If power is solved, grid can bind. If grid is solved, cooling can bind. If cooling is solved, materials or capital can bind. The bottleneck migrates."
This is why the BESS (battery energy storage) thesis that emerged in May 2026 — "the bottleneck has shifted from power to storage" — is incomplete. Adding BESS capacity addresses the cooling and grid smoothing layer. It does not address transformer procurement, RTE connection timelines, or firm power availability. The minimum operator still applies. The constraint simply migrates to the next weakest layer.
GridReadiness's Grid Deployment Risk Audit evaluates all six substrate layers for a specific site and project timeline — not just the most visible constraint.
WHY "MITIGATION IS NOT A SLOGAN"
The paper's classification of mitigation pathways is deliberately unforgiving:
Fast nuclear buildout: "Helps but insufficient" (Ω = 16.50)
Private power buildout: "Insufficient" (Ω = 16.80)
Efficiency breakthrough: "Helps but insufficient" (Ω = 12.40)
Efficiency rebound: "False salvation" (Ω = 22.10 — worse than baseline)
Classification rule: mitigation succeeds only when it reduces Ω below 1.
This is the academic version of what practitioners are observing in the field. SMRs are real but arrive in 2032-2035. Nuclear PPAs help but don't solve grid connection queues. Efficiency improvements are erased by the Jevons rebound effect — as compute becomes more efficient, demand for more compute increases proportionally.
The one intervention that does reduce substrate pressure to near zero in the model: making all substrate layers simultaneously unlimited. That is, obviously, not achievable. But its collapse to Ω → 0 confirms the model is falsifiable — it is a science, not a belief system.
THE GRIDREADINESS CONNECTION
GridReadiness was built on an intuition that the paper now formalises: AI deployment is blocked by the physical layer, not the cognitive one. The firm power constraint is real. The grid capacity constraint is real. The transformer constraint is real.
France brownfield sites address two of the six substrate layers simultaneously — firm power (nuclear baseload available) and grid capacity (existing HV connections, 18-36 months vs 7-10 years). They do not solve cooling, water, materials or capital. But they compress the critical path on the two layers that are currently the dominant bottleneck.
That is precisely what the SLIH minimum operator predicts: compress the slowest layer, and the system becomes faster — until the next slowest layer becomes the constraint. Brownfield sites buy the time to solve the downstream layers before they become the binding constraint.
HOW TO CITE THIS IN YOUR ANALYSIS
For developers and investors evaluating European AI infrastructure, SLIH provides a useful analytical frame for due diligence:
- Has the project addressed P_firm (firm power source confirmed)?
- Has the project addressed G_grid (grid connection timeline confirmed)?
- Has the project addressed C_cool (cooling architecture confirmed for target MW density)?
- Has the project addressed M_materials (transformer procurement, copper, GOES steel)?
- What is the current weakest layer — and what is the plan for when the bottleneck migrates?
A Grid Deployment Risk Audit from GridReadiness maps these layers for a specific site and commissioning target. It does not replace engineering due diligence — it compresses the timeline for identifying which layer requires immediate action.
Source: Shaub, J. (2026). Artificial Intelligence or Substrate Drift? A Falsifiable Model of Physical Limits to Intelligence Scaling. Independent Research. ORCID: 0009-0002-9256-0285. June 1, 2026. GridReadiness independently verified the alignment between SLIH findings and field data from EU transformer procurement and RTE grid connection timelines.