Fiber lasts decades, GPUs last three years: the AI boom isn't the dot-com boom

Executive summary

The presenter argues that AI is genuinely useful, but parts of the AI investment cycle resemble the blockchain and dot-com booms: a real technology is wrapped in claims of inevitability, aggressive valuations, self-reported benchmarks, and pressure on executives to adopt it before its business value is clear.

The central business distinction from the dot-com comparison is the durability of the infrastructure. Fiber-optic cables could remain useful for decades and gain capacity as equipment improved. GPUs, by contrast, may lose economic value within a few years as newer generations arrive. If AI investment unwinds, buildings and power infrastructure may remain valuable, while older chips and debt tied to them could be harder to recover.

Business strategies, frameworks, and playbooks

• Match technology to the task. The featured startup’s product is described as a fast classifier that returns a label and confidence score for tasks such as support-ticket triage or fraud signals. The presenter argues that many such jobs can be handled with smaller, cheaper models—or established methods such as logistic regression, gradient boosting, or BERT—instead of a frontier LLM.
• Evaluate business value, not adoption activity. Pilots, integrations, and an “AI strategy” are not proof of measurable impact. Track whether deployments produce revenue, reduce costs, improve service, or otherwise show a clear operational benefit.
• Validate vendor claims independently. Treat speed, cost, and performance figures from a company’s own benchmark as unverified until tested against your workloads and alternatives. Confirm that a product’s advantage holds under realistic conditions.
• Assess vendor and product continuity. Before placing a critical workflow on a young AI vendor, consider what happens if it is acquired, pivots, or fails. Evaluate portability, replacement options, and the operational consequences of losing the service.
• Examine asset economics and financing. For AI infrastructure projects, assess hardware resale value, useful life, depreciation assumptions, debt exposure, and whether projected revenues can support operating and capital costs.
• Account for opportunity cost. AI spending also consumes engineering time, research talent, electricity, and capital that might otherwise go to other products or technologies.
• Separate durable infrastructure from fast-depreciating assets. Data-center buildings, grid connections, cooling systems, and power contracts may last much longer than GPUs. The presenter recommends not assuming that the entire AI buildout will retain value equally.
• Watch the “efficiency” rebrand. Smaller, cheaper, specialized models may be a sound technical direction. But a new label or acronym does not by itself justify a high valuation; evaluate the business fundamentals behind it.

Metrics and claims highlighted

Figures below are claims or estimates cited by the presenter. The video notes that some company benchmarks are self-reported and that external estimates may be debated.

• TypeSafe AI and its classifier: The company reportedly raised a $40 million seed round at about a $200 million valuation, then announced an $870 million Series A at a $7.5 billion valuation just 24 days later.
• The classifier reportedly responds in 70–500 milliseconds, compared with 3 seconds to several minutes for frontier LLMs on comparable structured tasks. Its stated price is $0.042 per million input tokens, with output described as effectively free. The presenter says several platforms integrated it within days and that it attracted hundreds of thousands of developer sign-ups.
• Benchmark caution: The video cites company-run benchmark claims of roughly 193 times faster and 447–467 times cheaper, while also mentioning a claim of up to 200 times faster. The presenter questions the precision and independence of these figures.
• Enterprise AI returns: An MIT NANDA report is described as finding that about 95% of the organizations studied had no measurable P&L impact from generative-AI initiatives. The presenter also cites $30–40 billion in enterprise spending and says roughly one in 20 pilots produced real revenue.
• GPU depreciation: Microsoft and Alphabet are said to use six-year useful-life assumptions for some equipment, while Meta uses five and a half years. The presenter says Meta’s extended assumption reduced depreciation expense by about $2.3 billion in the first nine months of 2025. Amazon reportedly shortened the life of some servers from six to five years in early 2025.
• Michael Burry is cited as estimating that depreciation may be understated by about $176 billion across large technology companies from 2026 to 2028.
• GPU rental prices: A leading Nvidia GPU reportedly fell from about $8 an hour in early 2024 to around $2 an hour by late 2024.
• AI infrastructure spending: Estimates cited include AI capital spending growing 46% faster than AI revenue growth, hyperscaler capital expenditure consuming 90–100% of operating cash flow in 2026, and about $300 billion in annual AI revenue being needed to cover depreciation and operating costs.
• Debt and financing: CoreWeave is described as having borrowed $2.3 billion against GPUs; its debt reportedly grew from under $8 billion in 2024 to over $20 billion. Meta’s Louisiana data-center financing is described as involving about $27 billion of off-balance-sheet debt.
• Market exposure: The presenter says the 20 largest AI and chip stocks reportedly lost more than $1 trillion in a week amid financing concerns.

Examples and case studies

• Long Island Iced Tea / Long Blockchain: In 2017, the beverage company changed its name to Long Blockchain. Its stock reportedly rose more than 200% at the open and ended the day up about 180%, despite the company saying it would continue its existing business.
• KodakCoin: Kodak’s 2018 cryptocurrency announcement reportedly more than doubled its stock in two days, but the product did not become a lasting business.
• ICO market: The presenter cites a Satis Group assessment that about 81% of ICOs were scams, and an Ernst & Young analysis finding that a basket of 2017 ICOs would have lost about two-thirds of its value within a year. 71% reportedly had no product on the market.
• Australian Securities Exchange (ASX): An independent Accenture review reportedly found that its blockchain settlement-system project was 63% complete and excessively complex before it was scrapped in 2022, with about A$250 million written off.
• TradeLens and Libra/Diem: The blockchain shipping platform TradeLens shut down in 2022. Facebook’s Libra project was renamed Diem, with the remaining assets sold in early 2022.
• Builder.ai: The London startup, valued at about $1.5 billion, collapsed into insolvency in 2025. Reports cited by the presenter alleged that much of the work marketed as AI was done manually by engineers, alongside allegations of inflated revenue.
• Humane AI Pin: The device was marketed as a smartphone replacement and reportedly sold for about $700. HP later bought the remaining assets and team for around $16 million, after which the service was switched off.
• Circular AI financing: The presenter describes Nvidia investing up to $100 billion in OpenAI while OpenAI buys Nvidia chips; Oracle agreeing to provide around $300 billion of computing capacity over five years, requiring Nvidia purchases; and an OpenAI–AMD arrangement tied to chip purchases. The presenter characterizes the resulting commitments as approaching $1 trillion, far beyond current revenues.
• Dot-com fiber buildout: The video says about $30 billion was spent laying roughly 19 million miles of U.S. fiber, with more than 90% of long-haul fiber estimated to be “dark” by 2002. More than 60 telecom companies reportedly failed between 2001 and 2003. Fiber later became valuable because it could last for decades and gain capacity through wavelength-division multiplexing without replacing the cable.
• AI hardware contrast: GPUs are described as having a much shorter economic life than fiber. The presenter argues that newer chip generations can sharply reduce the rental value of older hardware, while buildings, grid connections, cooling, and power infrastructure may remain useful.

Practical implications for managers

• Require a clear business case for each AI deployment: the task, baseline cost and performance, expected gains, and a way to measure realized impact.
• Use the simplest model that meets the quality, latency, privacy, and cost requirements; do not default to a frontier LLM for routine classification or extraction.
• Test vendor claims independently, and plan for service interruption or supplier failure before making a vendor mission-critical.
• Review GPU useful-life and depreciation assumptions against actual utilization, resale value, and the pace of model and hardware changes.
• Scrutinize debt-backed infrastructure and financing structures, particularly when equipment serves as collateral and its market value can fall quickly.
• Distinguish a durable asset from a temporary boom in demand: useful AI applications and power infrastructure may persist even if speculative companies, products, or valuations do not.

Presenters and sources

• Presenter: The video’s host is not named in the subtitles.
• Sources and organizations cited: TypeSafe AI; Vercel; Cloudflare; LangChain; Satis Group; Ernst & Young; Australian Securities Exchange; Accenture; TradeLens; IBM; Facebook/Meta; Builder.ai; Bloomberg; Humane; HP; Nvidia; OpenAI; Oracle; AMD; Microsoft; Alphabet/Google; Amazon; CoreWeave; Blue Owl Capital; PIMCO; BlackRock; MIT NANDA; Andrew Odlyzko; Ali/Alliance Research (name unclear in captions); JPMorgan; Goldman Sachs; David Cahn of Sequoia; Michael Burry; Bank of England; the IMF; and The Verge.