A silent change is taking place somewhere between Santa Clara and the dusty outskirts of Phoenix, and most Americans won’t be aware of it until their electricity bills alert them. This fall, Nvidia’s Vera Rubin platform—named for the astronomer who discovered dark matter—will go into full production. The racks themselves, which weigh almost two tons apiece and are assembled from about 1.3 million parts, will begin to arrive at data centers around the nation. It’s the kind of moment that seems important in a way that can only be verified in retrospect.
The way the industry has been discussing Vera Rubin gives the impression that this isn’t just another cycle of upgrades. When the previous generation, Grace Blackwell, shipped in 2024, it was already deemed overwhelming. Nvidia is now claiming ten times the performance per watt, which sounds dubious until you consider that the company has essentially fulfilled its previous promises. At CES in January, Jensen Huang, wearing a leather jacket, told reporters that a single prompt can now initiate “a thousand-step journey.” These machines are designed to handle that kind of workload.
It’s not just the silicon that makes Vera Rubin intriguing in a strictly American setting. It’s the surrounding buildout. Through 2029, Nvidia has pledged to produce up to $500 billion worth of AI infrastructure in the US, including Blackwell GPUs at TSMC’s new Arizona facilities. To ship racks north, Foxconn is building a plant in Mexico. Vera Rubin systems will be housed inside Meta’s data centers by 2027, the company said last week. The list of early clients, which includes OpenAI, Anthropic, Amazon, Google, Microsoft, and CoreWeave, is a veritable who’s who of businesses that have already committed to tens of billions of dollars in capital expenditures just this year.
It is worthwhile to consider the physical reality of these systems. The weight of each rack is almost 2,000 kg. In contrast to Blackwell, where superchips were soldered to boards and a single failure could result in a lengthy, costly repair, the compute trays slide out in a matter of seconds. Additionally, Vera Rubin is Nvidia’s first fully liquid-cooled platform, which is more significant than it might seem because conventional evaporative cooling uses water in amounts that have already put a strain on local governments in areas like Northern Virginia and central Texas. This is significantly reduced by liquid cooling, which will be quietly appreciated by county boards and utilities.
Nevertheless, the system will use about twice as much power as its predecessor. The next stage of building data centers in the United States will be defined by this trade-off: more efficient per token, but heavier overall. Operators are already redesigning facilities with denser power feeds and reinforced floors in Loudoun County, Phoenix, Iowa, and the corridors outside of Atlanta. With its server aisles and raised-floor cooling, the older boxy data center might serve as a visual reminder of a transitional period, similar to how mainframe rooms now resemble museum pieces.
Although Nvidia’s price is unknown, Futurum Group projects that each Vera Rubin rack will cost between $3.5 million and $4 million, which is about 25% more than Grace Blackwell. The McKinsey estimate of nearly $7 trillion in global data center investment by 2030 ceases to seem like a wild figure when you multiply that across the hyperscaler order books. It begins to resemble an accounting line.
But beneath it all lies a fragility. Supply chains have been shaken by shortages of memory chips. The competition from Google’s custom silicon, Broadcom, and AMD is more intense than it was a year ago. Furthermore, there is still disagreement over whether the demand for AI justifies the amount of money being invested in it. As this develops, it’s difficult to avoid wondering if late 2026 will be remembered as the time when American infrastructure discreetly reorganized around the roadmap of a single company or as the time when someone, somewhere, finally questioned whether the math holds.
