AI infrastructure is expanding, and overbuild fears are rising
Investment in AI data centers keeps growing, and the market is starting to worry that GPUs, power and compute supply could end up repeating the kind of overbuild seen during the internet bubble.
Paul Bonnet, an investor at 20VC, offered a different scenario: if AI does not end in a bubble burst but instead develops successfully, the result may be lower AI token prices and, at the same time, a market that grows exponentially.
He said falling prices and industry growth are not contradictions. What matters is whether every step down in cost creates a new application market that is larger than the last one.
History suggests shortages often turn into surpluses
Bonnet said the concern around an AI bubble has historical precedent. Most supply shortages eventually turn into oversupply within about 3 to 5 years, destroying large amounts of capital along the way.
He pointed to the fiber buildout around 2000, which later left behind large amounts of dark fiber. Shares of related companies such as Corning, Lucent and Nortel were also hit hard. Earlier still, the railroad investment boom of the 1890s in the U.S. pushed roughly one quarter of rail mileage into receivership.
By that logic, today’s rush to build AI data centers, buy GPUs and develop power infrastructure could end in a similar way.
Aluminum fell more than 99.9% over a century, yet the market grew
Bonnet said aluminum offers a useful exception.
Aluminum is not rare at all; it is one of the most abundant metallic elements in Earth’s crust. The problem was that early humans could not separate and refine it cheaply. According to the data he cited, aluminum traded as high as $1,200 per kilogram in 1852, while gold was about $600 per kilogram at the same time.
Legend has it that Napoleon III served his most honored guests with aluminum tableware, and the roughly 100 ounces of aluminum on top of the Washington Monument were once among the most expensive metal products in the U.S.
The real turning point came in 1886, when Charles Martin Hall and Paul Héroult separately discovered electrolysis methods for smelting aluminum. In 1889, the Bayer Process solved ore refining. By 1954, U.S. aluminum output had reached about 1.3 million tons, while prices had fallen to about $0.48 per kilogram, a decline of more than 99.9% over a century.
Yet the industry did not disappear. Bonnet said the market value tied to aluminum rose from about $500,000 in 1895 to about $640 million a year by 1954.
Lower prices created new markets, not just more of the old one
For Bonnet, this is where aluminum most closely resembles AI tokens. Each collapse in price did not simply mean selling more of the old product. It opened entirely new markets.
At $1,200 per kilogram in 1852, aluminum could only be used for jewelry and ceremonial objects. When the price fell to about $37 in 1859, it entered scientific instruments. In the early 1890s, at around $4.40, it began appearing in cookware, competing with copper and cast iron. Around 1900, when it fell to about $0.72, aluminum moved into overhead transmission lines.
Even after nominal prices stopped falling sharply, performance kept improving. Duralumin appeared in 1909, making aluminum a true structural material for the first time and helping drive all-metal aircraft and aviation, including the DC-3, which entered the market in 1935. From the 1950s onward, aluminum was also used on a large scale in foil, packaging and beverage cans.
AI tokens may be following the same curve, only much faster
Bonnet said AI tokens may be moving through a similar process, only at a much faster pace than aluminum ever did. Aluminum took about 100 years to achieve a cost decline of more than 1,000 times. AI tokens could see a comparable drop in less than 10 years.
The real question, he said, is not what happens when tokens become 10 times, 100 times or 1,000 times cheaper. It is what kinds of applications, if any, will exist that do not exist today.
ChatGPT, Claude and other chatbots are still capped by human demand. Even with more than 1 billion people using similar AI platforms each month, human time for reading, typing and work remains limited.
Reasoning models and AI agents may change that constraint. Reasoning tokens do not need to be read word for word by a human. Future agents may also call other agents, and even build, manage and execute additional agents. If that happens, the final consumer of AI tokens will not necessarily be a human.
Bonnet said this could shift the demand ceiling for AI from “global population × each person’s time” to two variables: cost and utility. Both are improving right now, as token costs keep falling while model capability keeps rising.