Fiber optic cable laid during the dot-com expansion is still in use today. The chips that power artificial intelligence will not last nearly that long. Eric Hippeau estimated that AI chips have a useful life of roughly five to seven years, speaking on CNBC on October 16, 2026. CNBC introduced him as managing partner at the venture firm Lerer Hippeau. His overall view is that the AI boom shares real features with the dot-com era, with one difference that changes the economics.
Spending Is Running Ahead of Revenue, Just Like 1999
In Hippeau’s telling, the money going into AI infrastructure is running well ahead of the revenue needed to justify it. He also sees financing becoming more leveraged and more creative, which mirrors the late-1990s telecom expansion, when carriers borrowed heavily to lay networks before demand showed up.
Lenders are feeling it too. Axios reported Friday that banks are pulling back as AI’s massive debt pile disconnects from economic reality.
AI Chips Run on a Much Faster Replacement Clock
Hippeau said today’s AI chips “are going to have to be replaced by better, hopefully cheaper, faster chips.” Dot-com fiber, by contrast, is still carrying traffic. Once it was paid for, it kept earning a return.
That contrast works in two directions for investors. For chip suppliers, a replacement cycle means repeat business: every accelerator installed today is a likely future order. For the large cloud companies and other buyers, the same cycle means spending that keeps coming back. Buyers spread a chip’s cost over its expected life through depreciation, an annual expense that reduces reported profit.
A shorter useful life makes that yearly charge bigger. This is our reading of his point. He did not spell out these accounting effects.
Falling Token Prices Make AI Cheap to Use
A token is a small chunk of text, often part of a word, that AI models process and that providers use to price usage. Hippeau said the price of AI usage measured in tokens is falling dramatically. He credited that mostly to open-weight models, which are models whose underlying parameters are released publicly so anyone can run them, and he said there are now dozens of them. In his characterization, you can now run AI for almost nothing. Cheaper usage helps adoption. It also makes it harder for buyers who paid top dollar for compute to charge premium prices for it.
Power, Water and Neighbors: Limits the 1990s Never Faced
Hippeau framed the scale this way: “We’re talking about spending in one year what it took ten years to build the internet infrastructure.” He added that the dot-com expansion never ran into today’s water and power constraints, or into opposition from local residents and municipalities.
Government research points the same way. A Government Accountability Office report notes that the Department of Energy projects data centers will account for up to 12% of U.S. electrical demand by 2028. A Colorado legislative review found that data center load growth is the primary reason for recent capacity market conditions in the PJM grid region, and it lists a growing number of state bills on large-load tariffs and permitting.
Agents Could Take Over the Jobs Apps Do Now
Hippeau also described where he expects AI to end up: “You’re not going to have to go to websites, you’re not going to have to go to apps. You’re just going to be asking your agents to do your bidding, your ordering.” An agent is software that carries out tasks for you, such as booking, buying or scheduling, without you clicking through each step.
His comments came during a rough week for chip stocks. A widely followed semiconductor exchange-traded fund was down 9.49% over the past week as of noon Friday. Hippeau did not comment on this week’s moves. The selloff does show why the question of how long these chips last matters right now.
What to Watch as AI Spending Grows
The most useful signal will be in the accounting. If buyers of AI chips start reporting shorter useful lives for their servers, it will confirm Hippeau’s timeline, raise depreciation costs and test whether AI revenue can keep up. The second signal is physical. If power shortages, water limits and local permitting fights start delaying new data center projects, the spending curve will level off no matter how much financing is available. Both will show up in company filings and local zoning decisions well before they show up in quarterly guidance.