The AI gold rush examined¶

The Patrician has observed many enthusiasms sweep through Ankh-Morpork over the decades. There was the year everyone became convinced that swamp dragons were the future of personal transportation, which ended predictably with scorched eyebrows and renewed appreciation for horses. There was the brief period when pyramid schemes were considered innovative financial instruments, which concluded with several entrepreneurs acquiring unexpected familiarity with the city’s correctional facilities. And there was the infamous tulip incident, which we don’t discuss in polite company.

The current AI enthusiasm bears certain resemblances to these previous episodes. Everyone is convinced that fortunes will be made, that the current trajectory cannot possibly be wrong, and that anyone not participating urgently will be left behind permanently. Vast sums are being deployed based on projections ranging from “optimistic” through “wildly speculative” to “have you actually met reality?” The people deploying these sums are not fools, which makes the situation either brilliantly prescient or expensively mistaken depending on outcomes we won’t know for years.

The Patrician’s approach to evaluating such matters is to examine what is actually occurring versus what people claim is occurring, to identify who benefits from particular narratives regardless of their accuracy, and to remember that new technology typically takes longer to arrive and matters less than enthusiasts predict while eventually mattering more than sceptics expect, though not in the ways anyone anticipated.

What’s actually happening beneath the hype¶

Large language models can generate remarkably coherent text, write code that frequently works, and answer questions with confidence that occasionally correlates with correctness. These are genuine capabilities enabling real applications. They’re also considerably narrower than the hype suggests because the models work through pattern matching on enormous training datasets rather than understanding in any meaningful sense.

Training frontier models costs tens to hundreds of millions of euros. GPT-4 allegedly cost over €100 million to train. ChatGPT’s subscription revenue likely doesn’t cover the infrastructure costs of serving hundreds of millions of users. This engagement-without-profitability problem is familiar from previous technology waves. Enterprise applications have better economics when businesses pay for measurable productivity improvements, but consumer AI remains more impressive than profitable.

Cloud providers are collectively investing over €100 billion annually in AI infrastructure based on assumptions about demand that may or may not be accurate. Whether utilisation justifies these investments is carefully not being answered publicly. The equipment suppliers, particularly Nvidia, are benefiting enormously by selling GPUs as fast as they can manufacture them at premium prices. The suppliers have less risk because they’re selling capital equipment rather than renting capacity that must generate returns over years.

AI expertise commands seven-figure compensation packages while the definition of “AI expertise” has expanded dramatically. Someone who took an online course can market themselves as an AI engineer through credential inflation that makes assessing actual capabilities increasingly difficult. The academic pipeline produces limited numbers of genuine experts annually because training requires years of specialised education that cannot be rushed to meet demand.

Actual AI deployments are concentrated in narrow applications that work reliably enough for production use. Customer service chatbots, code completion, and document summarisation are seeing genuine adoption. These applications work well enough often enough that failure cases are manageable. They’re also considerably less transformative than the rhetoric suggests. The enterprise sales process reveals that despite the hype, businesses are adopting cautiously through pilots and incremental deployments rather than wholesale transformation.

AI startup valuations require extraordinary outcomes to justify current prices. Companies with €10 million in revenue raising at €1 billion valuations must achieve market dominance while maintaining high margins. These outcomes are possible but statistically rare. The FOMO dynamics driving valuations higher benefit entrepreneurs but create situations where investors are paying prices that require nearly perfect execution to generate returns.

The bubble indicators and the base case¶

Several characteristics suggest bubble dynamics. Prices disconnected from fundamentals, investment driven by fear of missing out, infrastructure spending racing ahead of clear demand, and the pivot where every company suddenly emphasises AI capabilities all resemble historical patterns. These don’t prove a bubble exists but suggest caution is warranted.

Despite these indicators, the case for genuine transformation has merit. Productivity improvements from AI tools in software development and content creation are measurable and meaningful. Developers using AI-assisted coding produce code faster. These gains are modest rather than revolutionary but real and valuable across large populations. Research progress suggests continued improvement is plausible even if not guaranteed. The integration into workflows is proceeding gradually but directionally across industries.

The geopolitical competition between the United States and China over AI leadership ensures sustained funding regardless of commercial returns. Governments view AI as strategically important, which provides a floor under development that pure commercial dynamics wouldn’t necessarily provide. The Patrician notes that genuine transformation and bubble dynamics can coexist because important technologies often experience investment booms that exceed what current capabilities justify while still ultimately proving transformative. The internet was absolutely a bubble in the late 1990s and also absolutely transformed society.

The Patrician’s assessment¶

Looking at the available evidence with appropriate scepticism about motivated reasoning from all parties, The Patrician concludes that AI is genuine technology with real capabilities that will have meaningful impacts, that current investment levels are probably excessive relative to near-term commercial opportunities, and that the technology will likely matter more over the long term than current sceptics predict while mattering less than current enthusiasts claim.

The next few years will likely see disappointing returns for many AI investments as the gap between valuations and business fundamentals becomes apparent. Some companies will succeed spectacularly, many will fail quietly, and the aggregate returns for investors will probably be modest despite occasional spectacular successes. This is the normal pattern for investment in emerging technology and should surprise nobody familiar with technology history.

The infrastructure being built will find uses though perhaps not the uses currently imagined. The data centres will run workloads, the GPUs will process applications, and the investment will enable capabilities, though the specific applications and business models might differ substantially from current plans. Infrastructure tends to be more durable than the businesses that initially justify it.

The capabilities will improve gradually with occasional breakthroughs that are impossible to predict. The trajectory will likely be incremental progress punctuated by occasional jumps when architectural innovations produce step changes in capabilities. This will continue until fundamental limitations are encountered or until returns from additional investment diminish to the point where other technologies become more attractive research targets.

The economic value will eventually emerge through productivity improvements, new applications, and efficiency gains that are individually modest but cumulatively substantial. AI will likely become valuable infrastructure that enables other innovations rather than a standalone industry generating the returns that current valuations imply. This is less exciting than revolutionary transformation but probably closer to how the actual value will materialise.

The sensible response for observers is maintaining appropriate scepticism about extraordinary claims while remaining open to genuine progress, distinguishing between what AI can do now versus what it might do eventually, and remembering that technology transformation typically takes longer and proceeds differently than enthusiasts predict while eventually mattering more than sceptics expect.

The Patrician has seen many enthusiasms come and go. Some proved to be genuine innovations that transformed how things were done. Some proved to be expensive mistakes that people prefer not to discuss. Most proved to be somewhere in between, where the technology was real but the initial expectations were misaligned with what the technology actually delivered. AI will likely follow this pattern, being neither the complete transformation that enthusiasts promise nor the hollow bubble that sceptics claim, but rather a genuinely useful technology that takes years to find its proper applications and whose ultimate impact is determined by factors that cannot be predicted from current vantage points.

In the meantime, the wise approach is watching where the serious money goes rather than listening to what people say, remembering that all parties have incentives to shade their descriptions toward their interests, and maintaining the scepticism that comes from having watched previous enthusiasms play out in ways that consistently surprised everyone involved. The AI gold rush is real, but whether the participants strike gold or merely acquire expensive experience in prospecting will be determined by events yet to unfold.