The Appetite: How a $150 Million Monthly Bill Became a Power Plant Decision

Somewhere south of downtown Memphis, in a neighborhood called Boxtown that has spent a century absorbing other people's industrial exhaust, dozens of gas turbines hum around the clock next to a data center the size of several football fields. On July 1, 2026, the meter on that facility effectively started running a little faster. An open-source AI lab called Reflection AI began paying SpaceX — yes, the rocket company — $150 million a month for the privilege of renting Nvidia's most powerful chips inside that building. If the contract runs its full course to 2029, the total comes to roughly $6.3 billion. That is not a typo, and it is not really a story about rockets or even, primarily, about one AI startup. It's a story about what happens when computation itself becomes an industrial input voracious enough to bend electricity markets, corporate emissions targets, and the geography of American power grids — all within about eighteen months of the phrase "AI compute shortage" becoming a boardroom cliché.

The deal, in plain terms

Reflection AI is not a household name, and that is rather the point of it. Founded in 2024 by Misha Laskin and Ioannis Antonoglou, two researchers who cut their teeth at Google DeepMind, the company has positioned itself as a kind of Western answer to DeepSeek — an open-weight model developer aiming to give governments and enterprises an alternative to both Chinese open models and the closed, proprietary systems sold by OpenAI, Anthropic, and Google. It has barely shipped a public frontier model yet. What it has done is raise money at a pace that makes even 2026's frothy AI market blink: from a $545 million valuation to roughly $25 billion in about a year, with Nvidia alone pouring in close to $800 million across funding rounds — an investment that conveniently also primes the pump for Nvidia's own chip sales, since Reflection's biggest expense is Nvidia hardware.

That hardware sits inside Colossus 2, the newer of two enormous data center campuses SpaceX operates near Memphis, built originally to serve Elon Musk's other company, xAI, and its Grok chatbot. Under the new agreement, Reflection gets dedicated access to racks of Nvidia's GB300 chips — the current top tier of AI training and inference silicon — starting July 1 and running through 2029, with a flexible exit clause that lets either side walk away with 90 days' notice after an initial three-month lock-in. It is less a purchase than a very expensive, very serious lease, priced like a mortgage on a small city's worth of computing power.

A rocket company becomes a power company's landlord

Here is the detail that ought to stop you mid-scroll: SpaceX, a company whose entire founding mission is getting things off the planet, has quietly become one of the more consequential landlords of AI computing power on it. Reflection is not even SpaceX's biggest tenant. Anthropic is reportedly paying on the order of $15 billion a year to lease the entire capacity of Colossus 1, giving it access to roughly 220,000 Nvidia GPUs. Google is paying SpaceX around $920 million a month for compute using about 110,000 GPUs housed in SpaceX facilities. Toss in a reported acquisition-in-progress of the coding-assistant company Cursor, and SpaceX's cloud-and-compute business alone is now humming along at a run rate north of $80 billion in committed revenue through 2029 — a bigger number, by some estimates, than the entire annual revenue of the rocket-launch business that made SpaceX famous in the first place.

The strategic logic is straightforward once you see it: building a gigawatt-scale data center campus and filling it with the newest, scarcest chips is enormously expensive, but the demand for that capacity is currently so intense that whoever has spare rack space can practically name their price. SpaceX had the land, the capital, and — crucially — the willingness to build power generation on-site rather than wait years for a utility interconnection. That last part is where this story stops being a curiosity about Elon Musk's business empire and starts being a story about the electric grid itself.

Why Memphis, and why that matters

Data centers this large don't simply plug into the grid the way a house does. They need firm, uninterrupted power at a scale most regional utilities can't provision quickly, and interconnection queues around the country now stretch years, not months. SpaceX's solution at Colossus — first for Colossus 1, and again for Colossus 2 — was to bring in dozens of gas turbines and generate power on-site, bypassing the wait. That decision has made Memphis one of the most closely watched flashpoints in the entire AI-energy story, and not for flattering reasons.

Community groups, the NAACP, and legal organizations including the Southern Environmental Law Center and Earthjustice have accused the facility of running turbines without the air permits normally required under the Clean Air Act — turbines capable of emitting more than 1,700 tons of nitrogen oxides and dozens of tons of fine particulate matter and formaldehyde each year, in a metro area that already earns failing grades for ozone pollution and has been nicknamed an "asthma capital." Boxtown, the neighborhood closest to the plant, reportedly faces a cancer risk roughly four times the national average, and residents live within a mile of both the turbines and an elementary school. Lawsuits are ongoing. Whatever their outcome, the underlying pattern is now familiar across the industry: when a data center's timeline collides with the grid's timeline, the grid usually loses, and the improvisation that fills the gap — diesel or gas generation stood up fast, sometimes ahead of permits — lands hardest on whoever already lives closest to the industrial edge of town.

The bigger current this deal is riding

Zoom out from Memphis and the pattern repeats at a scale that dwarfs any single contract. According to the International Energy Agency's 2026 analysis, global electricity demand from data centers rose about 17% in 2025, with AI-focused facilities growing even faster — on the order of 50% — against a backdrop of roughly 3% growth in overall global electricity demand. The IEA now projects that data center electricity use will roughly double by 2030, climbing from about 485 terawatt-hours in 2025 toward 950 terawatt-hours, with the AI-specific slice of that consumption tripling. For a sense of scale, that eventual figure approaches the entire current electricity consumption of a country like Japan.

The uncomfortable part is not just how much new demand is arriving — it's what's being built to meet it. Analysis from the American Action Forum found that natural gas' share of newly planned U.S. power generation capacity rose from roughly 11.1% in 2024 to about 18.1% in 2026, even as renewable capacity growth nearly flatlined, inching up only about 2% over the same period. Separate industry tracking cited by outlets covering the sector describes planned non-renewable capacity additions jumping by as much as 71% between 2025 and 2026. Put simply: faced with a wall of AI demand that can't wait years for a new wind farm's interconnection study, much of the power industry's answer has been to reach for the fuel that can be built fastest — gas — rather than the fuel that's cleanest.

Big Tech's own numbers give up the game

If the aggregate statistics feel abstract, the individual corporate disclosures are not. On July 1, 2026 — the same day Reflection's contract with SpaceX took effect — Amazon and Google both published sustainability figures showing their own emissions climbing, explicitly because of AI infrastructure. Amazon's greenhouse-gas emissions rose about 16% in 2025 to roughly 81 million metric tons of carbon dioxide equivalent, with emissions from purchased electricity alone up 34%, a jump the company tied directly to data center expansion — it added more data center capacity globally in 2025 than any other company, including more than 1.2 gigawatts in the fourth quarter alone. Google's emissions climbed even more sharply, by around 18%, the largest single-year increase the company has ever reported. Both companies have public net-zero commitments — Google by 2030, Amazon by 2040 — and both are now openly acknowledging, in their own official filings, that the AI buildout is the reason those targets are getting harder to hit, not easier.

That is the real significance of the SpaceX-Reflection deal: it is not an outlier. It is a single, unusually well-documented data point inside a systemic shift that is already visible in the sustainability reports of the largest companies on Earth. When a $150-million-a-month contract for chip access is treated as unremarkable business news, it's worth remembering that money has to convert into megawatts somewhere, and increasingly, those megawatts are coming from turbines built faster than anyone can permit them properly.

What happens next

None of this means the compute build-out stops, or should. Reflection AI's ambitions — an open-weight alternative to closed frontier labs, with backing from national-security-adjacent customers like the Department of Energy's Genesis Mission — are a legitimate and, in some tellings, strategically important goal. SpaceX turning a data center into a second business line is a rational response to genuine scarcity in advanced chips. The problem was never that AI needs power. The problem is the timeline mismatch: AI labs are signing three-year, multibillion-dollar contracts on a planning horizon measured in months, while transmission lines, permitting processes, and utility-scale renewable projects operate on a horizon measured in years. Something has to fill that gap in the meantime, and right now, gas turbines — built fast, sometimes ahead of the paperwork — are filling it.

Watch three things over the next eighteen months. First, whether regulators in Tennessee, Mississippi, and beyond start treating data center power generation as its own permitting category rather than an afterthought, given how many similar disputes are surfacing in Georgia, Virginia, and Texas. Second, whether the "flexible exit clause" language now standard in compute contracts — including Reflection's 90-day out — becomes a genuine check on overbuilding, or just fine print that never gets used while demand keeps climbing. And third, whether any of the hyperscalers' next annual sustainability reports show the emissions curve bending back down, or whether 2025's numbers turn out to be the first of several years in which the industry quietly redefines what "net zero by 2030" is allowed to mean. The appetite for compute is not going away. What remains genuinely up for grabs is what we're willing to build, and burn, to feed it.