The AI Data Center Energy Problem: Colossus Power Consumption and the 267% Electric Bill

The strongest ten minutes of Marc Chaikin’s “The End of Elon?” presentation have nothing to do with stocks. They are a case that the current way of building AI, in ever-larger clusters drawing ever-more power, has hit a wall. Host Molly Hendrickson sums it up: “we’re almost at a crisis point with data centers in America right now.” Chaikin uses the argument to set up his pitch for small, precise national lab machines. Whether or not you buy the pitch, the numbers deserve a look.

The figures the promo cites

Household equivalents. “A typical AI data center consumes as much electricity as get this, 100,000 households – not people, households. One data center the size of Elon’s Colossus can consume 2 million households’ worth of electricity.” A 100-megawatt facility does roughly match 100,000 average U.S. homes, and Colossus at its planned 2 gigawatts scales to about 2 million. These are fair ballpark comparisons.

The IEA projection. “According to the International Energy Agency, data centers will double their electricity use between now and 2030… to around 945 terawatts. That’s enough to power the entire nation of Japan.” The IEA’s April 2025 report does project roughly 945 terawatt-hours of global data center consumption by 2030, up from around 415 TWh in 2024, and does compare it to Japan’s total consumption. The promo says “terawatts” where it means terawatt-hours, a common slip.

U.S. capacity. “By 2028, data centers in America will consume an estimated 150 gigawatts of energy… the equivalent of 150 nuclear reactors.” Estimates for 2028 U.S. data center load vary widely, from about 75 to over 130 gigawatts depending on the source. 150 is at the high end but within the range of published forecasts. The “150 reactors” comparison assumes one gigawatt per reactor, which is roughly right.

The electric bill. “By an average of 267% in areas with high data-center concentrations, according to a Bloomberg study. Imagine your electric bill going from $500 to $1,800 a month.” This is the headline-grabbing number, and it needs context. Bloomberg’s 2025 analysis found that wholesale electricity prices at grid nodes near heavy data center concentrations had risen as much as 267% over five years. That is a wholesale spot-price figure at specific locations, not an average residential bill. Retail bills in affected regions such as the PJM territory have risen materially, but in the range of 10 to 30 percent, not 267. The direction of the claim is right; the magnitude as applied to a household bill is not.

Colossus by the numbers

Chaikin spends real time on the physical scale of xAI’s Memphis facility, which the headline calls “SpaceX’s Colossus” (it is xAI’s, though Musk’s companies have since consolidated). “When completed, the world’s most powerful AI data center campus will cover 1,156 acres.”

Then the football field exercise. “Now multiple that by 100 – 100 football fields… Now multiple those 100 football fields by 10… so about 1,049 football fields. Elon’s Colossus facility is THAT big, Molly.” An NFL field including end zones is about 1.32 acres, so 1,156 acres is roughly 875 fields; using the playing field alone (about 1.1 acres) gets you to Chaikin’s 1,049. Either way, the campus is enormous.

Power: “Colossus will consume about 2 gigawatts of energy at full capacity.” xAI has stated ambitions in that range for the expanded Colossus 2 site, and the company has been importing gas turbines and a Tennessee Valley Authority connection to get there. The turbines have been the source of local air quality complaints, which the promo alludes to with “They’re pumping noise into the air and disrupting lives.”

Chaikin’s contrast: Lux, the new AMD-powered machine at Oak Ridge, “will consume as little as 24 megawatts of electricity, according to our estimates… That represents a savings of about 99%.” That is Chaikin’s estimate, not a published figure, but tens of megawatts is the right order of magnitude for a machine of Lux’s size. We cover Lux in detail in Oak Ridge and Frontier Supercomputer Stocks.

The other costs

The promo does not stop at electricity. “They’re gobbling up our land and crowding out much-needed housing.” “More than that, they’re now draining our water supplies.” Both have become live local political issues. Data centers use water for evaporative cooling, and in drought-prone regions like Arizona and parts of Texas, permitting fights over water allocation have delayed projects. Land use fights are common in Northern Virginia, the densest data center market in the world.

Chaikin’s conclusion: “And that’s coming with major costs to society right now. Something has to give.” That is the setup for what he calls the “AI White Swan of 2027,” a disruption he argues is predictable because “We simply can’t continue on this path.” We unpack that framing in The AI White Swan of 2027.

Where the argument is strong

The energy constraint is the most credible part of the American Atlas thesis. Utilities are the binding constraint on AI buildout in 2026. Interconnection queues stretch for years. Hyperscalers are signing nuclear power purchase agreements and building gas plants behind the meter. The political backlash is real and growing, and it is bipartisan in a way few issues are. Any technology that delivers scientific compute at one percent of the power draw has a genuine advantage in that environment.

Where it is stretched

The promo’s leap is from “big clusters have an energy problem” to “small precise machines make big clusters obsolete.” Those are different jobs. Colossus trains large language models, which requires massive parallel low-precision compute. Lux runs AI on scientific data at FP64. A 24-megawatt science machine does not replace a 2-gigawatt training cluster any more than a surgical laser replaces a bulldozer.

Chaikin knows this, and concedes it late in the presentation: “if you’re already positioned in today’s top AI plays, that’s a good thing. What I’m showing you today is a different layer of the same opportunity.” The more accurate version of his thesis is that the energy constraint creates a market for efficiency, and precision compute at national labs is one slice of that market.

What it means for investors

The energy problem is investable from several directions, and the promo only takes one. The Homestacks and Musk Master Key promos we have covered take the supply side: buy the power producers. Chaikin’s American Atlas takes the efficiency side: buy the chipmaker whose FP64 hardware does more science per watt. The free pick is AMD.

Both angles can be right at once. Power demand grows, and the premium on efficiency grows with it. For our full assessment of how Chaikin turns the energy argument into a stock pick, and whether the Power Gauge Report is the right vehicle, read the End of Elon / American Atlas review.

NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.