Tomasz Tunguz, a venture investor who writes a blog, has put a date on a worry that runs through the industry: the year the United States starts buying AI output from abroad because its own data centers cannot keep up. His basic model, published on his blog, tomtunguz.com, on 4 November, lands between 2028 and 2030. It is one investor’s estimate, built on assumptions he lists himself, and the spread between his scenarios is wide.
First, the unit. A token is the small chunk of text, often a piece of a word, that an AI model reads or writes, and it is how the industry counts and bills its work. Every chatbot answer and every line of code from an AI assistant is a pile of tokens. Tunguz argues that tokens behave like any other commodity, so a country that cannot make enough of them will import them, much as the US already imports rare earth metals.
The model has three inputs. For demand, he starts from a Goldman Sachs Commodities Research estimate that US data center power draw climbs from 31 gigawatts in 2025 to 41 in 2026 and 66 in 2027. For efficiency, he assumes chips deliver about 2.9 times more tokens for each watt every year, a blend of Nvidia’s Vera Rubin, Google’s Ironwood, and Amazon’s Trainium3 chips. For supply, he assumes that just 55 percent of scheduled US capacity arrives on time.
From those, he concludes that today’s systems plus new data centers can absorb token demand growing 2.5 times a year. Past that pace, better chips can no longer cover the gap, and the country needs more power than it has. Cancellations decide the date. If cancellations or delays hit 45 percent of the projects in the pipeline, imports begin by 2028. At 20 percent, the date slips to 2029. Even with every proposed facility built on schedule, the gap opens by 2030.
Tunguz is open about the largest hole in his own model. It holds the price of computing constant, and real prices rise when supply tightens. So far, he writes, demand has ignored that: short-term GPU rental prices jumped from about $5 to more than $13 an hour over two quarters, and Anthropic signed billions of dollars of short-notice computing at premium rates, yet buyers kept paying. He reads those figures from Anthropic’s amended S-1 filing.
He grants that some price must exist at which people stop buying, but argues that cheaper production would push demand up by orders of magnitude, so rationing through high prices only exports the business. If the US does not build the power and facilities, tokens “will simply be produced elsewhere & imported,” in his words.
A footnote adds the sharpest detail. Power is only 5 percent of the cost of an hour of GPU time, while paying back the capital is 64 percent, so a swing in the cost of borrowing matters more than a swing in electricity prices. Latency does not protect domestic builders either: a round trip from New York to Frankfurt takes 88 milliseconds. Where the US does build, he says, the limit becomes gas turbines, which are sold out into 2030, citing SemiAnalysis and GE Vernova.
He expects the coming limits on the number, size, power use, and location of data centers to reshape what he calls “Startupland.”
That is a forecast from an investor who funds software companies, with no independent check, and the 55 percent on-time figure does most of the work in it. Still, the price evidence is already on the table. A team budgeting AI spend for 2027 should price in rental costs that have more than doubled in two quarters, because his own best case only delays the squeeze until 2030.
Reported by Tomasz Tunguz on his blog (tomtunguz.com) on 4 November 2026.
