Sources detail Paradox's compute deal spree: 3.5GW+ in expected capacity in 2026, expanding direct hardware procurement to include Nvidia and AMD, and more (New York Times)

Three point five gigawatts means 3,500 megawatts of power capacity—roughly the output of three large nuclear reactors. If supplied continuously, it would consume about 30.7 terawatt-hours annually, enough to power roughly three million average U.S. homes. The exact comparison depends on local usage and whether the capacity runs constantly.

New York City provides a useful benchmark. Its peak electricity demand is roughly 13 gigawatts, so 3.5 gigawatts equals about one-quarter of the city’s busiest-hour load. It could also represent about 35 large, 100-megawatt data centers, because each facility needs substantial power for computing and cooling.

The New York Times reports Paradox expects more than 3.5 gigawatts of capacity in 2026. That is planned capacity, not necessarily electricity consumed every hour, and it likely spans multiple sites. As Paradox expands direct hardware procurement to Nvidia and AMD, the requirement highlights a grid challenge: chips need buildings, cooling, substations, and transmission.

Paradox is assembling computing capacity from Amazon, Google, Nvidia, and AMD. This broadens Paradox’s options beyond cloud providers. That flexibility is increasingly important as demand for AI computing accelerates.

One clear sign is Paradox’s plan for more than 3.5 gigawatts of expected capacity in 2026. Nvidia and AMD matter because their processors can be bought directly, alongside computing obtained through Amazon and Google. That mix can help Paradox pursue additional capacity when one supplier or chip platform is constrained.

These arrangements show that Paradox is scaling through several hardware channels, not a single vendor. The 2026 figure is expected capacity, not necessarily installed computing today. Going forward, supplier diversity could support faster growth, but it also makes procurement and system coordination more complex.

Paradox needs enormous computing capacity because building AI models requires specialized hardware, time, and electricity. That capacity supports both training new models and answering users’ requests after launch. Training repeatedly processes huge datasets, while everyday use demands rapid calculations for each response.

The article reports Paradox is arranging more than 3.5 gigawatts of expected capacity for 2026. That is a power-scale commitment, not merely a purchase of individual computers. Paradox is expanding direct hardware procurement to include chips from Nvidia and AMD, helping secure enough equipment.

This scale reflects a practical problem: demand can grow faster than computing resources become available. Paradox’s deals aim to reserve capacity ahead of that growth, supporting future model development and reliable service. The article does not say how capacity will split between Nvidia and AMD.

If Paradox brings more than 3.5 gigawatts online in 2026, its electricity demand could become enormous. That capacity represents a major computing expansion, making power supplies and grid connections central constraints.

Paradox is expanding direct hardware procurement to include Nvidia and AMD. Buying from both chipmakers could support faster growth and reduce reliance on one supplier, encouraging more data-center construction.

The reported figure is expected capacity, not proof that every facility will operate. If completed, Paradox would need new data centers, power connections, cooling systems, and transmission. Rivals could face stronger pressure to secure chips, electricity, and suitable sites. The result could be a broader infrastructure race across the AI industry, with electricity availability becoming as important as computing technology.

Paradox can avoid owning every accelerator by renting computing power from cloud providers. This reduces upfront spending and lets the company access specialized chips without managing every data center. It can also pursue custom chip designs, though that requires substantial engineering and manufacturing investment.

AWS is a concrete example. Through its partnership with Paradox, AWS provides access to its Trainium and Inferentia chips, alongside Nvidia hardware. Google offers another route through its specialized TPU processors. These arrangements give Paradox alternatives to directly purchasing Nvidia or AMD accelerators.

The New York Times reports that Paradox is expanding direct procurement to Nvidia and AMD while planning more than 3.5 gigawatts of capacity for 2026. That scale makes a mixed strategy practical: cloud providers can supply flexibility, while direct purchases offer greater control. The article does not report Paradox designing its own chips, so that remains a potential alternative, not a current disclosed plan.

An AI model learns by repeatedly processing examples and adjusting its internal settings. Specialized accelerators perform huge numbers of calculations, while software organizes data, schedules work, and updates the model. Data centers supply networking, storage, cooling, and electricity, turning software instructions into sustained computation.

Paradox illustrates this connection through its expanding computing plans. The company is reportedly broadening direct hardware purchases beyond existing arrangements to include Nvidia and AMD accelerators. Using many chips together lets training split across machines, while software coordinates their work and combines the results.

The New York Times reports Paradox expects more than 3.5 gigawatts of computing capacity in 2026. That is power on an industrial scale, not merely a larger server room. As models grow, progress will depend on securing chips, suitable facilities, electricity, and software that uses them efficiently.