Boom, bubble, bust, boom. Why should AI be different?

Perceived Value vs. Present Capabilities

• Many see modern AI’s potential value as extremely high, especially when including the political power of controlling it, but note that concrete, high-margin applications are still thin.
• Current mainstream uses are described as: text summarization, style transfer, code generation under close supervision, image generation that still needs human cleanup, speech-to-text, homework/test cheating, and a bit of customer-service triage.
• Several commenters argue this is far from the promised “revolutionize the workplace” narrative and not what trillions in capex were ostensibly funding.

Bubble Dynamics and Historical Parallels

• Repeated comparison to the dot‑com era: a technology can be transformative and still have a spectacular financial bubble.
• Some think the AI bubble is more speculative than 1999 because valuations assume a Hollywood-style AGI that doesn’t exist yet.
• Others argue key AI players today (big cloud/consumer tech firms) already have large revenues and earnings, unlike many 1999-era dot‑coms, so fallout may hurt investors but not destroy the incumbents.

Business Models and Where Value Accrues

• Strong skepticism that token/API sales or low-price subscriptions can ever justify multi‑trillion data center buildouts; margins are thin and users resist high pricing.
• Concern that open and local models will commoditize general LLM capabilities, undermining moats and long-term pricing power.
• Many see real value in “boring” narrow ML (NLP, medical, scientific, weather models), but note those were underfunded pre‑hype and may get starved when the bubble bursts.

Infrastructure, Energy, and Hardware

• Worry that huge GPU/data-center capex, fast-wearing hardware, and massive projected power consumption are economically unsustainable if monetization lags.
• Some argue much of the spend is effectively an arms race among giants, not grounded in proven ROI.

Jobs, Productivity, and Labor Markets

• Debate over whether LLMs truly replace junior developers or designers; many say supervision overhead makes them worse than training humans.
• A recurring view: AI may make office workers modestly more productive (e.g., 10–15%), but capturing that productivity as AI revenue is hard; companies will only pay a small fraction of saved salary.
• Customer service automation is cited as a target job, but current systems are seen as unreliable and often disliked by customers.

Data, IP, and Power Concentration

• Dispute over whether training on scraped content is “theft” versus legitimate use, but broad agreement that power and benefits are concentrating in a few massive players.
• Some see AI’s clearest, least-ambiguous applications in surveillance, military, and population control, with consumer benefits lagging.