NVIDIA CEO Jensen Huang: Data Centers Should Be Referred to as Super Intelligent Factories
NVIDIA CEO Jensen Huang stated at the America.gov launch event in Washington that he dislikes the term "data center," and that these facilities should actually be called "SI factories" (Super Intelligent Factories). Elon Musk, who shared the stage, proposed a rule of thumb: every additional 5 gigawatts of power corresponds to a 1% increase in U.S. GDP.
The two estimated the annual output per gigawatt to be around $60 billion. Huang mentioned that NVIDIA is skilled at calculating these figures. He estimated that building 10 to 20 gigawatts annually could create about 1 million jobs in construction, plumbing, electricity, and cooling. As for Musk, SpaceX had approximately 1.4 gigawatts of data center capacity as of June. The event was hosted by Gavin Baker from Atreides Management, focusing on how power determines the super intelligence race.
Huang connected this statement to the White House's promotion of the "Super Intelligence / SI" designation. He had previously referred to the facilities as "AI factories": using electricity to produce tokens, which then transform into code, answers, designs, and actions. At GTC, he stated that tokens are the new commodity, with gigawatt-level factories potentially processing tokens from about 2 million per second to around 700 million; he estimated the construction cost per gigawatt to be between $50 billion and $60 billion, possibly rising to $80 billion to $100 billion. NVIDIA is using the DSX blueprint to design chips, racks, networks, power, and cooling according to factory production lines.
He also discussed the Open Agent Safety Platform: OpenShell delineates boundaries for agents, while Sentry monitors on an independent BlueField-4, with the ability to isolate breaches. At a White House luncheon, several tech leaders signed a voluntary "White House Super Intelligence Accord," committing to internal controls on risks, with signatories including Trump, Huang, Musk, and others.
In market mechanisms, this naming rights service serves the narrative of capital expenditure. Buyers are cloud vendors and model companies looking to secure power, land, and shells, while sellers include NVIDIA, which sells GPUs, networks, and factory blueprints, as well as utilities and contractors selling electricity and civil construction. Funds are flowing from model training and inference budgets to gigawatt-level civil construction. Beneficiaries are chipmakers who hold the definition rights of "factories" and power projects that can quote prices per gigawatt; those under pressure include grid capacity, transformers, and cooling supply chains, as well as data center owners who still depreciate according to traditional server rooms. Renaming the server rooms as production lines for manufacturing intelligence facilitates framing electricity and chip orders as industrial policy.
Source: Public Information
ABAB AI Insight
Huang has made two significant name changes: first from data centers to AI factories, and then to super intelligent factories following Trump's push for "SI." The vocabulary follows policy, while the capacity formula remains unchanged—electricity in, tokens out. He and Musk, sharing the same stage, equated 5 gigawatts to 1% GDP and 1 gigawatt to about $60 billion, directly linking grid increments to national accounts. NVIDIA is simultaneously securing land, power, and facilities for projects like Ohio's PORTS-Pike, extending supply discipline from chips to shells. The factory narrative serves to make capital expenditures appear as building steel mills rather than renting cabinets.
The capital pathway is "computation equals revenue." DSX designs the entire facility as a production line for computable output, with customers ordering by gigawatt, and NVIDIA recouping $150 billion to $200 billion in equipment revenue per generation of systems. Commitments from clients like OpenAI for ten gigawatts pre-sell future wafers and HBM. Money flows from cloud vendors and sovereign computing budgets to GPUs, optical interconnects, liquid cooling, and substations. Musk supplements the other end of the same chain with Starship frequencies and self-built power. The motivation is not merely a rebranding but to turn power bottlenecks into sellable factory metrics.
In contrast to the first time "computer rooms" were called "cloud" and "search rooms" were termed "advertising factories": whoever first defines the output unit also defines depreciation and financing language. Jensen Huang has defined the output unit as tokens, elevating it to super intelligence, while still positioning himself as a seller of picks—only now the pick has transformed into an entire factory. TSMC, Foxconn, and domestic packaging plants are upstream, while utilities represent parallel bottlenecks. The industry phase is transitioning from training clusters to continuous inference production lines, with construction speed dependent on electricity rather than the next product launch.
Structurally, this belongs to the reconstruction of the industrial chain. The mechanism is: model companies need tokens, chip companies convert tokens into factory capacity, the government writes factories as super intelligent infrastructure, and the grid becomes a true production throttle. Pricing power shifts from individual card quotes to how much computing power can be installed per gigawatt and how many tokens can be produced. The name change merely aligns this chain with White House terminology, allowing land, subsidies, and grid connection applications to proceed as factory projects rather than as internet server rooms.
ABAB News · Cognitive Laws
- Those who change names first are usually already selling capacity based on the new unit.
- The limits of intelligence are first written on the electricity meter, then on the model card.
- Referring to server rooms as factories is intended to allow the grid to approve electricity based on industrial projects.