Are AI stocks in a bubble? 3 reasons we think not

• Tech stocks have risen amid growing excitement over artificial intelligence, but valuations have remained well below dot-com bubble era levels.
• Unlike the speculative excesses of the late 1990s, today’s AI spending is grounded in real demand and strong Q2 earnings reinforced the long-term AI opportunity, with clear evidence that capex is translating to returns.
• We believe investors should stay selective but engaged in the AI theme — focusing on companies with durable cash flow, pricing power, and real user demand.

AI bubble concerns have ebbed and flowed since the release of ChatGPT in late 2022. But AI stocks are not in a bubble, in our view. Gains for AI and related stocks are supported by earnings, AI infrastructure is serving real demand, and we believe the technology’s largest economic applications are still ahead.

Still, comparisons to the late-1990s dot-com era are easy to draw. Stock prices are near all-time highs, driven by AI exuberance and the capital investment fueling the build out of datacenters.

For many, today’s AI economy rhymes with the dot.com era when capital investments were being made to pave the way for the internet. There are also parallels being drawn between the vendor financing arrangements of the dot-com era and so-called “circular financing” today — deals where a supplier helps finance a customer who then spends money back with that supplier (either directly or via a partner) — as the scale of the datacenter build has led to unconventional financial agreements among major companies across the AI stack.

But we see three key differences between today’s environment and the late 1990s:

• AI-related stock valuations are relatively low. Today’s valuations are far below levels reached at the peak of the dot-com bubble when the top four tech leaders of early 2000 (MSFT, CSCO, INTC, ORCL) traded nearly 70 times 2-year forward earnings.¹
• Today, the average 2-year forward Price/Earnings (P/E) is about 18x for the biggest AI datacenter spenders: Microsoft, Alphabet, Amazon and Meta — also known as hyperscalers.²
• The demand for AI far exceeds the supply today. Demand for AI compute is growing exponentially with no observed slowdown.³ AI datacenter capacity is constrained, driving multi-year, prepaid commitments to lock in scarce AI chip and datacenter supply. The demand curve has no clear cap: every leap in capability has been creating appetite for more compute.
• We may be in the very early stages of an industrial revolution-type transformation. AI has the potential to transform almost every sector of the economy, including transportation, manufacturing, and healthcare.

AI-related stock valuations are not egregiously high, in our view. The forward P/E ratio on the PHLX Semiconductor Index (SOX) is about 20x, the same level it was trading in January 2023, less than two months after ChatGPT launched. Since that date (November 30, 2022), semiconductor stocks are up 47% on an annualized basis.

Earnings have pushed semiconductor stocks higher but the market isn’t rewarding the fundamentals, in our view, let alone putting outsized P/E multiples on semi stocks.

A P/E multiple, high or low, still begs the question: a multiple of what earnings? In January 2023, the SOX traded at roughly 17x forward earnings⁵, but consensus was pricing an earnings path that ultimately proved far too conservative, especially beyond the first year.

At the start of 2023, semis traded at 17 times the next 12 months of expected earnings and 13 times what the market expected them to earn this year.

The near-term number wasn’t a major surprise, but the gap between expectations and reality for 2026 earnings is large. Semiconductor earnings in 2026 are now forecast to be 3.4 times what the market expected at the beginning of 2023.

Portfolio risk may rise if a few AI-related names stumble, as detailed in our 2026 Fall Investment Directions, but we believe this isn’t proof of mispricing considering those names are producing the bulk of the earnings growth.

Demand for AI is greater than supply. Not only is demand greater than supply, demand is growing exponentially while supply growth is limited physically. This is in stark contrast with the dot-com era, where years of broadband supply was added before demand existed.

The clearest evidence is revenue. Anthropic is reportedly generating $65 billion in annualized revenue, while OpenAI is now reportedly generating revenue at a $40 billion run rate.⁷ That’s over $100 billion in revenue from products that didn’t exist four years ago. For context, there’s only one software company in the world (Microsoft) that makes over $100 billion a year in revenue.⁸

The largest providers of AI infrastructure are all reporting the same trend: accelerating cloud growth. In just one year, the combined cloud backlog, which represents signed customer commitments, of Amazon, Alphabet, Microsoft, and Oracle nearly tripled, rising from $800 billion to $2.3 trillion.⁹ AI investment is translating into revenue. The hyperscalers also have continued raising capex guidance and describe an environment where customer demand exceeds available infrastructure.

AI compute remains scarce enough that rental prices for NVIDIA H100 GPUs have continued rising¹⁰ despite the chips being roughly three years old. In most technology markets, older hardware becomes cheaper over time. Rising prices for aging AI hardware is clear evidence that demand is growing faster than new capacity can be brought online.

Every incremental unit of AI compute is being rapidly absorbed by revenue-generating workloads. Supply, meanwhile, remains limited by the physical constraint of chips, memory, power and data-center capacity. In the dot-com era, infrastructure waited for demand. Today, demand is waiting for infrastructure.

NVIDIA's latest earnings call reflects this trend as the company gave a revenue growth outlook for next year (fiscal 2028) of “approximately 70%” vs. consensus estimates of 45%. “This is a supply-constrained outlook,” NVDIA's CFO said on a conference call, noting that demand was actually growing much faster than supply.¹¹

We believe we are in the very early stages of an AI era with the potential to transform the global economy at a scale similar to the industrial revolution. As excited as everyone seems to be about AI, it is still mainly limited to printing text and code. In the coming years, the economy may see profound efficiency gains as modern AI moves more deeply into transportation, manufacturing, healthcare and scientific research.

The most advanced frontier models available today were built on infrastructure investments made roughly two years ago. Investment has scaled substantially since then, supporting the development of models that are expected to be far more capable than those we use today. In 2023, Wharton professor Ethan Mollick coined an oft-repeated remark, “Today’s AI is the worst AI you will ever use.”¹² It is still true. We are only scratching the surface.

Adoption is also much earlier than the headlines suggest. According to the 2026 Stanford AI Index, 88% of organizations used AI in some form, but the deployment of AI agents remained in the single digits across nearly every business function.¹³ AI is widely available, but it is not yet deeply embedded in how most companies operate.

Our optimistic thesis could be challenged by any of the following scenarios:

1. Power constraints. If AI builders can’t secure consistent, 24/7 power, then our constructive view may weaken. Interconnection queues and transmission bottlenecks in hubs like Northern Virginia and Texas already suggest power constraints could slow the AI datacenter build-out.
2. Rising idle capacity or sustained utilization slippage. Evidence that new datacenters aren’t filling, or that booked capacity is being canceled, would challenge the “insatiable” view.
3. Public opposition and regulation, which could slow adoption.
4. Circular financing could pull demand forward and concentrate losses if expected revenues do not materialize.
5. Greater reliance on debt financing: repayment obligations are fixed, so even ultimately viable investments can run into liquidity or solvency problems if revenues arrive later or lower than expected.
6. Open-source models could pressure pricing and shift profits across the AI ecosystem. But cheaper models can also make more applications economical and increase total usage.

These risks will create winners and losers, but they do not negate the growth in earnings, revenue and end demand already visible today. Investors risk underestimating the magnitude and durability of AI capital spending.

While vigilance is warranted as the landscape evolves, we feel that current evidence points to a growth story that is fundamentally distinct from past bubbles. Investors willing to remain disciplined and discerning may find the AI theme remains one of the most compelling opportunities of the decade.