AI infrastructure: rethinking risk, reward and reality

Why The AI Pullback Isn’t the End of the Story

Just a year ago, the AI investment narrative felt unstoppable. Today, some are wondering if the cycle has already peaked. But let’s be clear: we are still in the early stages of a generational technology shift. The recent market volatility doesn’t mark the end of the AI opportunity—it’s more likely just a short-term reset in a much longer structural cycle.

Pella continues to believe that AI represents a revolutionary computing platform, comparable to the rise of the internet or the mobile era. We maintain a significant allocation to what we define as “AI Enablers”—companies that provide the physical and digital infrastructure needed to build, run, and scale AI models. The universe of AI Enablers includes companies like NVIDIA, TSMC, ASML, Broadcom, Cadence Design Systems, Vertiv, and Schneider Electric.

What’s Driving the Weak Performance?

The first quarter of 2025 was difficult for this group. Most AI Enablers posted negative returns, with only SK hynix delivering positive performance. Several others saw share price declines of over 20%, including Marvell Technology and Synopsys. Yet, when we examined analyst expectations and earnings revisions, the picture was surprisingly stable. Outside of AMD and Generac, there were no material downgrades to revenue or earnings forecasts.

That led us to explore three broader explanations for the drawdown:

1. Lower Long-Term Growth Expectations: Using our valuation framework, we back-solved for the long-term growth implied by market prices. In nearly every case, the market’s current growth assumptions (for years 4–10) are now below each company’s historical growth rate—despite the arrival of what could be the most transformative technology trend in decades. That suggests markets may be underestimating the structural growth potential of these companies.
2. Higher Discount Rates: Despite a decline in risk-free rates, the discount rates implied by market prices have actually increased—by 0.50% to 0.75% on average. This points to an increase in perceived risk or equity risk premia. In other words, the market is demanding a higher return to own these stocks.
3. Were They Overvalued? We also reviewed valuations using our FCF Yield-to-Growth framework. As of 19 January—right before the selloff—most AI Enablers were still trading above our valuation hurdle of 8.5% to 9.5%, depending on their risk level. That suggests they were not expensive relative to expected cash flows and growth. Therefore, overvaluation doesn’t appear to be the cause either.

What Changed the Narrative?

The shift in sentiment seems to have been triggered by three events in quick succession:

· 20 January: DeepSeek unveiled its R1 model, which matched performance benchmarks of Google and OpenAI’s models, but with significantly lower compute requirements. Investors interpreted this as a signal that infrastructure needs might be lower than previously assumed.

· 24 February: TD Cowen reported that Microsoft cancelled several U.S. data centre leases, raising concerns of overcapacity and excess supply.

· 25 March: Alibaba’s Chairman, Joseph Tsai, warned that AI data centre investment may be forming a bubble, driven by speculative capital.

While impactful in the short term, we believe these stories are being misinterpreted. None of them materially alter the structural demand outlook for AI infrastructure.

The Bigger Picture: Capex Keeps Rising

Data tells a more optimistic story. Hyperscalers—Amazon, Microsoft, Alphabet, and Meta—continue to increase their capex budgets. In fact, consensus forecasts for 2025–2027 capex have been revised up by 14% since DeepSeek’s R1 release, from US$847bn to US$962bn. This underscores the continued momentum behind AI infrastructure spend.

And it’s not just the tech giants. New players are entering the arena. Project Stargate—backed by OpenAI, SoftBank, Oracle, and MGX—plans to invest US$500bn in U.S. AI infrastructure by 2029, starting with a US$100bn commitment. Meanwhile, AI Infrastructure Partners, including BlackRock and NVIDIA, have already allocated US$30bn, with an ambition to reach US$100bn. These initiatives are incremental to the Hyperscalers’ capex and broaden the base of long-term investment.

What Are the Companies Saying?

Recent commentary from management teams reflects continued strength in AI demand:

Amazon: “AI represents, for sure, the biggest opportunity since cloud and probably the biggest technology shift and opportunity in business since the Internet.”

·AMD: “We remain confident in our AI strategy and believe we are still in the very early innings of a multi-year adoption cycle.”

Eaton: “Data centre construction build rate doubled between 2023 and 2024. At 2024 build rates, it would take seven years to consume the current backlog.”

Schneider Electric: “We’ve seen an acceleration of that market… that translates into really double-digit growth for the future.”

Some companies have flagged a likely moderation in the growth rate starting second half of the year—hardly surprising after a period of triple-digit growth. Importantly, our investment cases are not predicated on continued hyper-growth. In most cases, mid-single-digit long-term revenue growth is enough to justify our valuations.

Where We Stand

Pella’s AI Enabler positions currently offer an implied long-term return of ~12% p.a., significantly above our 8.5% to 9.5% return hurdle. These companies are not priced for perfection. If anything, they are priced for mediocrity—and we believe they will deliver far more than that.

It’s also worth noting the phase shift now underway. The AI cycle is transitioning from the Training phase—capital-heavy and concentrated—to the Inference phase, which will be longer in duration, involve more players, and be characterised by iterative optimisation. While the growth in data centre capex may become more uneven, it will also become more diversified and durable.

Final Thoughts

The recent pullback in AI Enablers looks less like the end of a boom and more like a market that has momentarily lost perspective. The fundamentals remain strong. Capex is rising. Valuations are compelling. We see short-term sentiment shifts as an opportunity, not a warning sign.

Structural technology shifts like AI don’t resolve in a year or two. They play out over decades. The companies building the backbone of this transition are likely to be long-term beneficiaries—and we remain confident in our positioning.