Key takeaways
- The AI boom can be both a transformative structural shift and a bubble - these aren’t mutually exclusive.
- What’s unusual today is how widely and early the “bubble call” is being made; historically, bubbles are named after they burst.
- Mega-cap leaders are driving this cycle, while many of the most attractive exposures for mid-cap investors sit in “picks and shovels” niches (cooling, HVAC, utilities, power).
- Valuations in large-cap tech are elevated but anchored by profitable, cash-generative businesses; the notable outlier is OpenAI’s combination of sky-high valuation and extraordinary capital needs.
- If an AI-led correction comes, it may start at the top and ripple down, with mid-caps potentially holding up relatively better than in typical recessions.
It’s peculiar to live through the first bubble that everyone seems comfortable calling in real time. Historically, bubbles are christened in hindsight, after the collapse, with the benefit of post-mortem data and moral clarity. Tulip mania, the roaring ’20s, the railroads, dot-com: their names stuck long after the narrative had run its course. Today’s AI cycle is different. The “bubble” label is being pinned on it while capital is still rushing in and the headlines remain euphoric. That’s abnormal, and it reflects a more informed market, faster media loops, and the collective memory of prior manias, yet it doesn’t automatically make the label wrong. It simply raises the odds that we’re watching two overlapping movies: a genuine, world-changing technology wave and a sentiment-driven overshoot.
I don’t treat “bubbles” and “structural growth story” as opposing camps. The coexistence matters because it shapes strategy.
What a bubble looks like when narrative outruns reality
Bubbles, in my experience, reveal themselves when narrative takes the wheel and fundamentals slide into the back seat. Prices detach from what businesses can plausibly earn today, and valuation becomes a story about a future that may or may not arrive. Fear of missing out spreads. Duration stretches, investors look further and further into the horizon to justify today’s multiples. We’ve seen this movie before, and it tends to run longer than people expect. Railroads didn’t boom for a quarter or two; they boomed for a decade-plus. The Internet didn’t peak and vanish; it rewired commerce while investors overpaid, under-earned, and then eventually right-sized expectations.
We’re seeing familiar markers in AI: excitement and hype, a surge of capex, and returns that, so far, are nascent relative to the spend. It doesn’t mean the returns won’t come. It means that, for now, the scoreboard is heavy with promised points, light on realised ones.
Structural change and speculative excess are not mutually exclusive
I don’t treat “bubbles” and “structural growth story” as opposing camps. Historically, the biggest bubbles have clustered around technologies that genuinely changed the world: rail, radio, the Internet. AI will join that list. The coexistence matters because it shapes strategy. If you acknowledge both forces, you avoid the false comfort of binary thinking. You accept that some assets can be priced for perfection while the technology itself creates enduring winners. You look for ways to participate that don’t require perfection to be profitable.
One important way today’s cycle diverges from the dot-com era is the nature of the players. Outside of OpenAI, the names at the centre - Microsoft, Amazon, Alphabet, Meta, Oracle - are not speculative startups with thin business models. They are massive, cash-generative machines, profitable across multiple lines, trading at valuations that, while full, are not the 60–80x earnings (or-no-earnings) of 1999. That difference matters. It tethers the top of the market to real cash flows, even as the narrative inflates around AI.
OpenAI is the exception that proves the rule. The company’s valuation and capex aspirations dwarf its current revenue base, which makes it an emblem of both promise and imbalance. It could grow into that valuation; it could also become a high-profile case study in how ambitious roadmaps collide with power, supply, and monetisation constraints. Either way, it reinforces the need to distinguish durable cash engines from story stocks, even when both are riding the same technological wave.
How we play AI from the mid-cap seat
As co-portfolio manager of a small- and mid-cap strategy, we don’t have the luxury of simply owning the hyperscalers. My job is to find the enabling ecosystem—“picks and shovels” businesses whose fortunes rise with the buildout, not just the hype. AI isn’t only algorithms; it’s heat, power, cooling, airflow, zoning, permits, and grid capacity. Data centres are thermodynamic realities. They throw off heat and devour electricity. That makes cooling specialists and end-to-end HVAC designers critical. Energy availability is, right now, a bigger bottleneck than chips in several regions. I’ve heard of hyperscalers sitting on silicon they can’t power (e.g. Microsoft). That’s not a theoretical bottleneck; it’s an operational constraint that reshapes timelines and spend.
Utilities and grid infrastructure, greenfield and brownfield generation, refurbishment to meet modern standards, all of these threads will take years to resolve. In the meantime, the work is tangible. The backlog is visible. The hyper-scalers have publicly signalled that they’re not done spending. For mid-cap investors, that’s the sweet spot: companies whose revenue lines are tied to the buildout’s physics rather than the eventual monetisation curve of AI services.
We still respect the dependency chain. If returns disappoint, capex will eventually moderate, and the picks and shovels will feel it. But today, with 12–24 months of pipeline visible across parts of the ecosystem, we can participate with an eye on risk and valuation, not just narrative.
Why this correction (if it comes) might be led from the top
Market sell-offs are typically small and mid-cap-led on the way down; cyclicality and beta bite first. This cycle is unusual. The boom has been mega-cap-led from the start. If the shock is specifically AI scepticism - questions about monetisation, unit economics, or power constraints - the first domino may be upmarket, not down. Concentration at the top of the indexes magnifies the impact. When 40% of a benchmark clusters in a handful of AI-heavy names, an AI-specific derating is a benchmark event, not a sector scuffle.
Contagion would follow, as it did in 2000, though the mechanics differ. Back then, the NASDAQ rolled over months before the S&P 500. Today’s S&P is more tech-heavy, so it may rhyme more with the NASDAQ than with its former self. Mid-caps would still fall in absolute terms, but there’s a case they could hold up relatively better. Valuations in many mid-cap niches remain more in line with long-term history, while the mega-cap decile sits near the top of its range. When perfection gets repriced, the largest weights bear the brunt.
What we’re hunting now
Given where we are in the cycle, we’re intentionally turning over rocks where AI exposure exists but hasn’t yet been fully priced by the market. I’m particularly interested in companies with a division or capability tied to AI infrastructure that can grow without relying on hero assumptions, and where the rest of the business anchors valuation discipline. We don’t want to pay an “AI premium” for work that looks like industrial execution. We want the opposite: industrial execution with an AI tailwind the market hasn’t fully credited.
Bottom-up remains our compass. Company by company, contract by contract, backlog by backlog. In an environment this noisy, the temptation is to trade narratives. Our discipline is to buy cash flows and optionality at reasonable prices, and to let others chase adjectives.
The long arc of an abnormal bubble
Calling a bubble before it bursts is abnormal, but it may be the right call. The trick is to avoid letting the label flatten your thinking. AI can be a bubble and a secular shift. The mania can run long and still resolve into enduring value. Our job is to own the parts of the buildout that earn through the cycle, to be positioned for corrections that start at the top, and to keep our eyes on the physics (heat, power, time) while everyone else argues about the poetry.
If there’s one lesson from past manias, it’s this: the narrative eventually meets gravity. In AI, gravity looks a lot like megawatts, coolant, permits, and cash returns. That’s where we’re standing.
