by Dr. Chris Kacher
**Energy is the #1 bottleneck for AI right now — but it’s not the only one.**
There’s a second major constraint that’s getting less attention: severe shortages of critical electrical infrastructure — especially power transformers, switchgear, and batteries — combined with continued heavy dependence on China for these components. According to recent reporting, nearly half of all planned U.S. data center builds for 2026 have already been delayed or canceled because of these shortages.
Even when power capacity exists on paper, actually delivering and connecting it to new AI facilities is proving extremely difficult.
### How AI Recursivity Changes the Picture
AI is not just consuming power — it is actively working to solve the problem. Recursivity means AI is speeding up the design of new transformers and grid equipment, discovering better materials, optimizing factory layouts, and improving supply chain logistics. This self-accelerating loop is real and will shorten the painful period.
However, recursivity cannot magically overcome physical and regulatory delays. Building new factories, installing transformers, and getting approvals still takes real time.
### Revised Timeline (factoring in recursivity)
**2026 – early 2027**: Noticeable slowdown.
Power and infrastructure bottlenecks are hitting hard. Many large training runs are being delayed or scaled back. Progress shifts toward efficiency, agentic systems, and smarter architectures rather than raw scaling.
**Mid-2027 – 2028**: The crunch begins to ease.
AI-driven improvements in design, materials, and grid optimization start compounding. Recursivity helps shorten the worst of the bottleneck, so the slowdown is less severe and shorter than a pure physical view would suggest.
**2029 onward**: Recovery and new acceleration.
AI has had enough time to meaningfully improve the energy supply chain. Cheaper and more abundant power becomes available, unlocking the next leg of capability growth.
**Bottom line**:
Energy and physical infrastructure are real, binding constraints today. AI recursivity is a powerful force that will help shorten the difficult period, but it cannot instantly eliminate the physical world’s limits. We are still looking at a real slowdown through 2026 and into 2027, followed by a stronger recovery starting in 2029 as the feedback loop gains momentum.
Market headwinds
-Along with the bottlenecks cited above, AI capex overspend issue which created a sideways to downtrending NASDAQ from Nov-Mar. AI‑infra/AI‑chip names are vulnerable to any disappointment on capex ROI, grid delays, or regulatory friction; we’ve already seen some cases where higher capex guidance hurts the stock.
-Iran issue which started in 1979 so it likely to persist. Any escalation/de‑escalation of Iran/Middle East tensions will meaningfully shifts oil prices.
-The pace of global liquidity continues to slow as central banks look to keep rates steady.
Market mixed bag
The median S&P 500 company is growing earnings per share in the double digits right now — the fastest pace since 2021. FactSet is tracking 17% EPS growth for this year, with acceleration expected through the next three quarters. Barclays just raised its 2026 S&P 500 EPS target to $321, which would be roughly 16% year-over-year growth.
This is the exact opposite of what happened before the big oil shocks of the past — 1973, 1979, and 1990. In those cycles, earnings were already rolling over or falling sharply when oil prices spiked.
Earnings are accelerating from already high levels. That’s why the S&P 500’s 18% drop in forward P/E since last October as a classic **bull market correction** — valuations compressing while profits keep improving — not the beginning of a bear market. In bear markets, both earnings and multiples collapse together. In corrections, multiples fall while earnings hold up or get stronger.
CAVEAT: That said, valuation support exists vs. late‑2024 extremes, but it’s not a cheap market. If AI monetization or macro slows more than expected, there’s a lot of air under those EPS forecasts. The current setup leaves plenty of room for downgrades if 2026–27 turns out more “grindy bottleneck” than “smooth AI super‑cycle.”
Oil prices are also nowhere near as extreme as they were in previous shocks once you adjust for inflation. And American consumers have a bigger cushion this time around, with tax refunds running more than 10% higher than last year.
Stanford University just published the Stanford 2026 AI Index Report. It is a bullish report illustrating how AI usage continues to accelerate. It is excellent at showing **what has happened** but weaker at acknowledging **what is now constraining future progress**. The acceleration narrative is becoming outdated as we move from the “easy” scaling phase into the harder infrastructure-constrained phase.
- The Index acknowledges rising compute demand but does not sufficiently highlight that power and electrical infrastructure (transformers, switchgear, grid capacity) have become the #1 limiter. Nearly half of planned 2026 U.S. data center projects have already been delayed or canceled. This is not a minor hiccup — it is a structural constraint that is already slowing raw scaling.
- Recursivity is real but not instantaneous. The report correctly notes AI’s self-improving nature (recursivity), but it implies this will smoothly overcome bottlenecks. In practice, AI can speed up design and material discovery, but it cannot magically build new factories, install physical transformers, or change regulatory timelines. The physical world still moves slower than software.
Bottom line: Markets trade ahead of the headlines, but given the powerful arguments, markets could remain range bound until the next news-driven catalyst hits.