There is a growing disconnect at the heart of the AI investment narrative, and UBS Chief Investment Officer Ulrike Hoffmann-Burchardi is naming it directly. In the latest edition of Signal over Noise ([#14, September 23, 2026](https://theideafarm.com/wp-content/uploads/2026/09/UBS-Signal-over-Noise-14-AI-Superexponential-Speed-Limits.pdf)) 1, Hoffmann-Burchardi and her co-authors lay out a rigorous and sobering framework: AI capabilities are accelerating faster than exponential curves can capture, and yet the commercial value those capabilities generate is accumulating along a far more modest, far more constrained, linear path. The gap between these two trajectories is where investors need to focus.
Intelligence Without a Ceiling
The UBS team introduced the term "AI super-exponential" earlier this year to describe AI capability growth whose rate of improvement itself keeps improving. The METR metric, which tracks the longest human task an AI can complete with at least 50% reliability, shows new model releases consistently placing above the exponential trendline. Recent state-of-the-art models have begun resolving aspects of the Navier-Stokes equations, placing AI at the frontier of problems once considered beyond machine reach. As Hoffmann-Burchardi observes, this "strongly signals that bottlenecks in the value chain increasingly lie outside the intelligence layer." The question of value capture, in her framing, has migrated away from what AI can do and toward the physical and organizational world's ability to absorb it.
The Binding Constraint Is Not the Longest Lead Time
The analytical core of this report is the distinction between lead times and substitutability. Long lead times can be managed, Hoffmann-Burchardi argues, where buyers can switch supplier, technology, or geography. The more dangerous constraint is one that "cannot easily be switched around," capable of holding up every subsequent layer of the value chain while simultaneously supporting pricing and margins in the scarce input. This principle has already played out in semiconductors, where early AI buildout pressure on accelerator compute migrated upstream into high-bandwidth memory, advanced packaging, and leading-edge foundry capacity. The most durable scarcity premiums, the UBS team concludes, lie in inputs "where switching suppliers, technologies, or production routes is particularly difficult."
Power Is the New Upstream Constraint
The same logic now applies to the physical infrastructure stack. Power should not be treated as a single bottleneck, the report is careful to note. Aggregate generation capacity does not guarantee dependable power at the required site, on a specific timetable, and under a financeable commercial structure. Interconnection queue data illustrate the severity: median lead times from connection request to commercial operation increased by more than 60% for projects built between 2015 and 2025, compared to the prior decade. Microsoft's CEO stated the predicament plainly in late 2025, noting "it's not a supply issue of chips. It's actually the fact that I don't have warm shells to plug into." Hoffmann-Burchardi cites this directly, because it places the constraint in unmistakable terms. Even with capital deployed and chips on hand, idle compute cannot generate returns without somewhere to plug in.
Battery storage, life extensions for existing plants, and behind-the-meter generation via natural-gas fuel cells and aeroderivative turbines are all being deployed as workarounds. But as the report notes, these solutions "shift rather than remove the critical path." Equipment delivery, permitting, fuel infrastructure, and system integration become the new constraints as soon as grid interconnection is bypassed. Very few behind-the-meter providers can meet all the requirements developers need, contributing to complex contracts and extended negotiations.
Permitting and Organizational Inertia Close the Gap
Political and institutional friction compounds the physical constraints. The Bipartisan Policy Center documented 54 local moratoriums on data center construction nationwide, with proposals for state-level moratoriums in 15 states. Data Center Watch identified 75 blocked or delayed projects representing approximately USD 130 billion in planned investment. While these figures do not equate to lost near-term capacity in every case, political acceptance has become a genuine site-selection and execution risk.
Inside enterprises, organizational inertia presents a parallel drag. AI adoption is expanding, but measurable economic benefit remains uneven across occupations. Better models alone, Hoffmann-Burchardi is clear to point out, "do not remove these barriers." Enterprises need to adapt processes, integrate data, and restructure how people work, a transformation that moves on its own timetable regardless of what AI can do.
Three Responses, No Single Solution
The UBS team identifies three possible paths to narrowing the gap: shifting more compute from training toward revenue-generating inference; scaling behind-the-meter and off-grid power supply; and targeted policy intervention on permitting, interconnection reform, and grid modernization. None of these alone will resolve the growth-rate mismatch. In combination, alongside more energy-efficient hardware and better load management, they could bring infrastructure deployment closer to pace with AI development.
For investors, the UBS team's focus remains on the bottlenecks themselves: suppliers of scarce, hard-to-substitute inputs across the AI and power value chains; application-layer use cases with the highest enterprise return on investment; and sectors such as pharma, biotechnology, and health care services where AI's long-term benefits are not yet priced.
5 Key Takeaways for Advisors and Investors
- Technical progress is not commercial progress. AI capability is accelerating super-exponentially, but value capture is constrained by physical and organizational limits. Pricing a similarly rapid rise in earnings into valuations is a mistake.
- The binding constraint is the hardest input to replace. In both the semiconductor and power stacks, durable scarcity premiums accumulate not in the highest-profile bottleneck, but in the input with the longest resolution time and the lowest substitutability.
- Power is a multi-layered constraint, not a single one. Grid access, transmission capacity, substations, transformers, and permitting each represent distinct friction points. Behind-the-meter workarounds shift, rather than remove, the critical path.
- Organizational inertia inside enterprises is a second major drag. Better models do not automatically translate into faster adoption or measurable economic benefit. Enterprise workflow adaptation and process integration are rate-limiting factors.
- Investment focus belongs at the bottlenecks. UBS favors exposure to scarce, hard-to-substitute inputs across the AI and power value chains, high-ROI application-layer use cases, and sectors such as pharma and biotech where long-term AI benefits remain underappreciated by markets.
Footnote:
1 Hoffmann-Burchardi, Ulrike, Nikolaos Fostieris, and James Dobson. "Signal over Noise #14: The AI 'Super-Exponential' Meets Its Speed Limits." UBS Chief Investment Office GWM Investment Research, 23 Sept. 2026, https://theideafarm.com/wp-content/uploads/2026/09/UBS-Signal-over-Noise-14-AI-Superexponential-Speed-Limits.pdf.
