The most common question I’ve received over the past two months has been: “Is AI a bubble?” Hindsight will eventually provide full clarity, but like everyone else, I have an opinion. All the classic signs are screaming yes (i.e., valuations), which means the next logical question to ask is: What kind of bubble is this—a good, rational one or a bad, irrational one? I do believe it is important to distinguish the difference, as that guides investment decisions into the future.
First, I’ll state the obvious: When a bubble pops, stock prices fall regardless of whether it is a good or bad bubble. The difference between the two, however, is that a good bubble leaves lasting value. The infamous “bad” bubble was the 1637 tulip mania, where prices surged as everyone had to have a tulip, then prices collapsed as cultivation made flowers abundant—leaving society with… well, the same old tulips. Arguably, the most famous “good” bubble was the 1990s dot-com rush. Yes, shell companies vanished, but it birthed the internet, connection to global markets, and e-commerce infrastructure we rely on today (to name a few).
The dot-com era thrived on fixed upfront costs, winner-take-all network effects, low energy intensity, equity-market funding, and—crucially—valuation rationality (perpetual call options on explosive growth). Today’s AI bubble is vastly different: Business models demand ongoing massive capex and electricity bills; network effects are weak to moderate (anyone can access the same base models); energy intensity is enormous; funding mixes drying private equity (now chasing retail investors) with drained Mag-7 cash piles ($330–405 billion annually in 2025, per Reuters/BofA estimates); and valuations feel irrational amid unclear monopolies. In other words, today’s AI bubble lacks the 1990s low marginal cost structure, potentially labeling this as a bad, irrational bubble.
| Factor | Dot-Com Era | AI Bubble Today |
| Cost Structure | Fixed upfront, low marginal | Ongoing capex + energy bills |
| Network Effects | Winner-take-all | Weak-moderate (open models) |
| Funding | Equity boom | PE drying + Mag-7 cash |
| Energy Intensity | Low | Enormous |
| Valuations | Rational growth options | Irrational, unclear moats |
But bubbles mature slowly, and much can change over the coming months and years that push this into the “good” territory.
The Two Metrics That Matter: Inference Costs and Electricity
Inference costs are the ongoing expenses of using a pretrained AI model for predictions or outputs to the user. Inference costs are sky-high today but have already plunged 50–100x in the last 24 months (Stanford AI Index 2025: up to 280x for GPT-3.5-level performance; Epoch AI: up to 40x/year for science tasks). Projections show another 10x drop every 18 months, driven by hardware (Nvidia Blackwell’s up to 4x efficiency) and techniques like quantization (4–8x savings). This trajectory could rationalize valuations soon, flipping the script on marginal costs.
If so, the Achilles’ heel becomes electricity generation. (And I don’t use the term “Achilles’” lightly since I tore my Achilles years ago, ouch!) US generation has stagnated for 25 years (~4,100 TWh annually, per EIA), a shock given tech’s boom. China, our AI rival, has surged 5–8x (to ~10,000 TWh in 2024), dominating coal, nuclear, wind, and solar (see chart below). They import ~40% of LNG but secure cheap Russian supply via Power of Siberia (38 bcm/year). The results are clear: Chinese electricity prices are at ~$0.07/kWh—half the US average rate—fueling a 30–50% opex edge for AI inference.
Energy: The Foundation of Every Economy
As the saying goes, economies run on applied energy, directed by human knowledge. For the US, AI’s power hunger (data centers to account for nearly half of electricity demand growth by 2030, per EIA) outstrips supply. Microsoft CEO Satya Nadella confirmed this on the October 2025 BG2 podcast: “You may actually have a bunch of chips sitting in inventory that I can’t plug in” due to power shortages. This isn’t hype—it’s the “power bottleneck,” where hyperscalers hoard GPUs awaiting grids.
The Trump administration’s $80 billion nuclear push (October 2025 Westinghouse/Japan deal: Stake in 8+ reactors) is the right long-term play for sustainable baseload. But completion takes ~4 years—too slow for today’s needs. I expect in the short-term the US will ramp up coal and natural gas.
Energy is the backbone of any economy, but in particular, the AI bubble. If inference costs fall as projected, the remaining hurdle is energy. This could help AI morph into a “good” bubble. And what is a “good” outcome for a rational bubble? Cheaper intelligence + abundant power.
