Khan Capital | July 2024
Key Takeaways
- The AI debate is a false binary: history shows that genuine technological revolutions (railroads, internet) routinely produce speculative bubbles, and the crash of the bubble does not invalidate the technology.
- The bull case is supported by real and growing revenue: Nvidia’s data centre business has grown from $15 billion to over $90 billion in two years, and enterprise AI adoption is generating measurable productivity gains.
- The bear case is supported by valuation extremes: the Shiller CAPE approaching 35, the capex-to-revenue gap widening, and market concentration reaching levels that create structural fragility in the benchmark index.
- The most useful framework accepts that AI is both a genuine revolution and potentially overpriced simultaneously, and positions portfolios to participate in the long-term thesis while surviving the corrections that every technology cycle produces.
- Investors should own AI exposure but manage sizing, diversify beyond the Magnificent Seven into less crowded AI beneficiaries, and maintain cash reserves to deploy during the inevitable moments of panic.
Nvidia’s market capitalisation briefly surpassed $3 trillion in June, making it the most valuable company in the world. The stock has gained over 200% in the past 12 months. The Magnificent Seven have accounted for the majority of the S&P 500’s year-to-date gains. AI-related capital expenditure from the hyperscalers is on track to exceed $200 billion in 2024. And the question that is now dividing Wall Street, academic economists, and technology strategists with increasing urgency is: are we witnessing a genuine technological revolution or the inflation of the next great financial bubble?
The honest answer is that we may be witnessing both simultaneously, and understanding why requires a more nuanced framework than the binary “bubble or revolution” framing that dominates the debate.
The Bull Case: This Time Is Different (In Some Ways)
The case for AI as a genuine revolution rests on three pillars that distinguish it from previous technology hype cycles.
The revenue is real. Unlike the dot-com era, where speculative valuations were attached to companies with no revenue and no business model, the AI boom is being driven by companies with enormous, growing, and profitable revenue streams. Nvidia’s data centre revenue has grown from $15 billion to over $90 billion in two years. Microsoft, Google, Amazon, and Meta are reporting accelerating cloud and AI services revenue. Enterprise adoption of AI tools (coding assistants, customer service automation, data analysis) is generating measurable productivity gains and demonstrable return on investment. This is not pets.com; these are the most profitable companies in history investing their own cash flows in a technology they are already monetising.
The use cases are proliferating. In the 18 months since ChatGPT’s launch in November 2022, AI has moved from a parlour trick to a productivity tool deployed across industries. Legal firms use AI for document review. Healthcare systems use it for diagnostic assistance. Financial institutions use it for risk modelling and trading. Software developers use it for code generation. The breadth of adoption, while still in early stages, suggests a general-purpose technology with applications across virtually every sector of the economy.
The competitive moat is deep. Building and training frontier AI models requires billions of dollars in compute, access to proprietary data, and engineering talent that is concentrated in a small number of companies. This creates natural barriers to entry that protect the leading firms’ margins in a way that the low-barrier internet business models of the late 1990s did not. The hyperscalers’ scale advantages in AI infrastructure, data, and talent are widening, not narrowing.
The Bear Case: Echoes of 1999
The case for bubble is equally grounded in observable data and historical pattern recognition.
The valuation extremes are unprecedented. Nvidia trades at approximately 40 times forward earnings after its extraordinary run. The Magnificent Seven collectively trade at valuations that price in years of continued above-trend earnings growth. The S&P 500’s forward P/E has expanded to levels that, historically, have been followed by below-average returns over the subsequent five to ten years. The Shiller CAPE ratio is approaching 35, its highest level since the dot-com bubble.
The capex-to-revenue gap is widening. The hyperscalers are investing over $200 billion annually in AI infrastructure, but the incremental revenue generated from AI-specific applications is a fraction of that amount. The bulls argue that the revenue will come; the bears argue that the investment cycle is running ahead of the monetisation cycle by a margin that will eventually trigger write-downs, guidance cuts, and the repricing of expectations. The 1990s telecom analogy is instructive: the thesis that internet traffic would grow exponentially was correct, but the companies that built the infrastructure went bankrupt before the revenue arrived.
Market concentration is a structural fragility. The Magnificent Seven account for approximately 30% of the S&P 500’s market capitalisation. When a third of the index is driven by a single investment thesis (AI), the index’s diversification benefit is illusory. Any negative development in the AI narrative, whether a competitive disruption, a regulatory intervention, or simply a quarter of disappointing revenue growth, would have an outsized impact on the benchmark that most institutional portfolios are measured against.
Every bubble is built on a true story. The dot-com bubble was built on the true insight that the internet would transform commerce. The housing bubble was built on the true insight that homeownership was an aspiration for millions of Americans. The AI bubble risk is not that AI is fake (it is clearly not) but that the market has extrapolated the technology’s long-term potential into near-term valuations that require flawless execution over an implausibly long horizon.
The Framework: Revolution AND Bubble
The most useful analytical framework rejects the binary choice between revolution and bubble and embraces the possibility that both are true simultaneously. History provides numerous examples of genuine technological revolutions (railroads, electricity, automobiles, the internet) that were accompanied by speculative excesses and subsequent crashes, without the technology itself being invalidated. The crash destroyed investors’ capital; it did not destroy the technology’s value. The companies that emerged from the wreckage often went on to become the dominant enterprises of their era.
The question for investors is not whether AI is real (it is) but whether the current prices adequately compensate for the execution risk, the competitive risk, and the timing risk that separates the technology’s promise from its full economic impact. At current valuations, the market is pricing a scenario in which virtually everything goes right: AI capex continues to accelerate, monetisation keeps pace with investment, competitive moats hold, and no regulatory intervention constrains the dominant platforms. Any deviation from this scenario, even a modest one, would trigger a repricing.
Implications for Investors
Own the theme but manage the sizing. AI is a multi-decade investment theme that belongs in portfolios. But the sizing should reflect the valuation risk: a 10-15% allocation to AI-exposed equities is a strategic position; a 30-40% allocation (which is what the S&P 500 cap-weighted index effectively delivers) is a concentrated bet that assumes perfection.
Look beyond the Magnificent Seven. The AI beneficiary map is broader than the mega-caps: data storage companies, power and utility providers, networking equipment makers, and industrial companies supplying data centre infrastructure offer AI exposure at more reasonable valuations and with less crowded positioning.
The picks-and-shovels approach carries its own risks. Nvidia’s dominance in AI chips is frequently compared to the “picks and shovels” strategy of the Gold Rush. But even picks-and-shovels companies can be overvalued, and the emergence of custom silicon (from Google, Amazon, and others), open-source model efficiency (which reduces compute requirements), and potential antitrust scrutiny all represent risks to Nvidia’s margin structure that current valuations do not fully account for.
Maintain dry powder for the inevitable correction. Every major technology cycle has included corrections of 30-50% that, in hindsight, were buying opportunities. The AI trade will produce similar moments of panic, and investors who have maintained cash reserves will be positioned to exploit them. The best returns in technology investing often accrue to those who buy during the corrections, not those who chase the rallies.
Conclusion
The AI bubble debate is ultimately a question of timing, not substance. The technology is real. The use cases are genuine. The earnings growth is demonstrable. But the market has a well-documented tendency to front-load decades of value creation into a few years of price appreciation, and to punish the resulting overvaluation with corrections that are painful even for investors who are correct about the long-term thesis. The prudent approach is not to avoid AI (which would be a strategic error) nor to maximise AI exposure at any price (which would be a valuation error) but to own the theme at a sizing that allows the portfolio to survive the inevitable correction without being forced to sell. The revolution will continue. The question is whether your portfolio is structured to participate in it.
Related Reading
The bubble debate intensified after the events covered in Nvidia’s AI Supercycle. The question was partly answered by DeepSeek Shock: China’s AI Challenger and the subsequent Tech Volatility in 2026.
Leave a Reply