Back Research Notes The Lights Out Regime: AI, Covariance, and the Compression of Certainty Published on February 20, 2026 By Jordi Visser Over the past few weeks, one question has surfaced in nearly every conversation I’ve had with institutional allocators: When does this stop? What I initially believed would be a difficult but manageable rotation over the course of 2026, from software dominance to hardware and infrastructure leadership, has moved faster than expected. It has triggered a massive dispersion under the hood of what appears, on the surface, to be a quiet market. So far through February 19th, there are a total of 185 stocks in the SPX with greater than ±15% absolute performance. Last year on this day, there were only 81. All this while the SPX is effectively flat at +24 bps YTD. The Dual Engines of Dispersion This instability is being driven by AI on both sides of the ledger. We are witnessing a historic divergence between the “Atoms” and the “Bits”: The Winning Side (The Capex Beneficiaries): These are the names benefiting from the massive, nearly $650 billion infrastructure spend by hyperscalers. Investors are rewarding the “Atoms,” the compute, power, and hardware, where demand is tangible and the “blank check” is being written. The Losing Side (The Multiple Compression Victims): These are the names being hurt by a collapse in forward visibility. The speed of AI progress is vaporizing the “long-duration” assumptions that software multiples rely upon. When investors can no longer credibly model what a business looks like three years from now, we are no longer dealing with a sector rotation. We are dealing with a structurally driven regime shift. The Turbulence Amplifier A covariance matrix built in calm water becomes a turbulence amplifier when the storm arrives. AI is that storm. It does not simply increase volatility. It alters the fundamental structure upon which volatility models were built. Traditional risk models assume that relationships between assets remain relatively stable and that diversification works because correlations are bounded. Yet history shows that in structural shocks, volatility rises, correlations shift abruptly, and what appeared diversified reveals itself as one concentrated bet. In this case, this is actually the unwind of a 15-year concentrated software and technology bet. AI represents this kind of structural shock. By lowering barriers to entry and accelerating obsolescence, it synchronizes capital flows. When the economic structure shifts, a matrix calibrated to a “low-AI” regime can amplify stress rather than dampen it. The Compression of MICAP This matters profoundly for long-duration assets. Michael Mauboussin’s Competitive Advantage Period (CAP) framework reminds us that valuation depends not just on growth, but on the duration of excess returns. CAP is determined by the rate of industry change and barriers to entry. AI acts as a “CAP Crusher.” It lowers switching costs and replication costs while increasing the speed of innovation. This creates a double-edged sword: even if current cash flows remain intact, the Market-Implied CAP (MICAP) contracts. The result is aggressive multiple compression, not because earnings have imploded today, but because the “duration assumption” embedded in the valuation has become suspect. In a world of Vibe Coding, monthly model releases, China open-source competition, and agentic automation, a five-year moat can be drained in a weekend. Leverage: The Enemy of Resilience Leverage amplifies this fragility. As Nassim Taleb famously noted, “Leverage is the opposite of resilience. It makes you a prisoner of the tail.” Leverage is a bet on the future being sufficiently predictable, a bet on stable correlations, bounded volatility, and slow obsolescence. When uncertainty rises and investors cannot confidently model cash flows even 36 months out, the “tail” begins to wag the dog. Spreads widen, liquidity drops, and equity multiples compress to reflect the new “physics” of the market. What once looked like a conservative capital structure quickly looks fragile. Leverage comes down not because growth has disappeared, but because predictability has. The Emerging Markets Parallel In my emerging markets days, I saw how quickly this dynamic unfolds. A market can move from a euphoric bull run to a violent liquidation in six months, not because earnings imploded overnight, but because confidence in the future did. When the certainty of forward cash flows evaporates, capital retreats first and asks questions later. AI is producing a similar “liquidity of confidence” crisis today. It is increasing structural uncertainty about dominance and competitive advantage. In this environment, turbulence models become essential. We must recognize that regime instability is fundamentally different from cyclical risk. So When Does This Stop? The question deserves a real answer, not a framework. In most historical episodes of structural uncertainty, the turbulence eventually found a floor. The dot-com unwind, the GFC, the taper tantrum each resolved not because the underlying disruption stopped, but because a stabilizing event emerged: a Fed pivot, a regulatory framework, an earnings season that reset expectations to a survivable level. Markets are remarkably good at finding a new covariance regime once they have enough data to rebuild the matrix. This episode will likely end the same way. Some combination of a high-profile software casualty that clears the air, a hyperscaler earnings cycle that resets capex expectations in either direction, or a model release that either accelerates the fear or proves the ceiling. Historically, the market needs an event to crystallize the new order. We will get one. Capital will stabilize, spreads will tighten, and a new consensus will form around the survivors. But here is where this cycle diverges from every prior one. In past regime shifts, the stabilizing event restored a degree of forward visibility that lasted. Investors could rebuild their models. The CAP could be re-estimated with some confidence. What makes the AI regime fundamentally different is that the source of the uncertainty, the pace of model improvement, the emergence of open-source competition, the cost collapse of intelligence itself, is not episodic. It is structural and accelerating at a speed markets and investors have never experienced. Playing traditional chess is now speed chess. Every stabilizing event will be temporary, because the next disruption is already in training. The traditional institutional response to regime uncertainty is to wait for clarity and then re-risk. In 2026, that may be the most dangerous strategy available. Clarity will arrive in windows, not in a sustained regime. The year ahead is less likely to look like prior periods of significant dispersion, when volatility faded and multiples expanded on durable new assumptions, and more likely to look like a series of compressions and reprieves, each one resetting the baseline lower for long-duration software assumptions and shorter for the assumptions behind them. The Path Forward: Designing for Episodic Clarity This is not simply a portfolio construction problem. It is a leverage philosophy problem. What I expect to follow is not simply a period of elevated volatility that eventually mean-reverts. I expect this environment to serve as a preview of what the AI world actually looks like for investors, and that preview is deeply unfamiliar to most of the people managing institutional capital today. Consider who has been running the dominant strategies for the past seventeen years. A generation of growth investors was trained in an era of compounding, stable, non-cyclical cash flows. They learned to identify moats, underwrite duration, and hold with conviction. That playbook was not just rewarded. It was the entire game. The businesses that won were predictable engines, and the edge came from correctly identifying which engi