Jensen Huang sees a trillion-dollar boom in AI factories

If you still think building a data center is just about housing servers and running databases, Jensen Huang wants you to drop the term entirely. They are no longer data centers. They are AI factories: you feed them electricity, and they spit out tokens.

Kicking off COMPUTEX 2025 at a packed Taipei Music Center with over 4,000 attendees, the NVIDIA founder laid out an audacious vision: AI has become foundational infrastructure, rivaling electricity and the internet, destined to build a multi-trillion-dollar industry.

Deskside supercomputers and semi-custom silicon

To power these factories, NVIDIA is spreading its Blackwell architecture across every tier of hardware.

For individual developers, NVIDIA showed off DGX Spark, a personal AI supercomputer now in full production and shipping in a few weeks through partners like ASUS, Dell, Gigabyte, Lenovo, and MSI. Right above it sits the new DGX Station, which plugs straight into a standard wall socket, delivers up to 20 petaflops of compute, and can run a 1-trillion parameter model locally.

For enterprise customers deploying agentic AI, RTX PRO Servers and a reference architecture for the NVIDIA AI Data Platform are now rolling off assembly lines.

Perhaps the most strategic announcement was NVLink Fusion. NVIDIA is allowing hyperscalers to design semi-custom systems using its high-speed NVLink interconnect, opening up what used to be a rigid proprietary stack so cloud giants can optimize hardware around their own specific workloads.

On the regional front, NVIDIA announced a deepened partnership with Foxconn to build an AI factory supercomputer for Taiwan's research community and companies like TSMC, alongside unveiling NVIDIA Constellation, a major new office in Taiwan.

Disney physics and five-trillion-dollar factories

Beyond digital agents, Huang argued that the next big frontiers are physical AI and general robotics.

To train robots before they step into the messy real world, NVIDIA partnered with Google DeepMind and Disney to build Newton, an advanced physics simulation engine. It also released the Isaac GR00T-Dreams blueprint to generate synthetic training data, alongside the Isaac GR00T N1.5 foundation model for humanoid robots.

Industrial manufacturing is the immediate proving ground. With roughly $5 trillion worth of factories being built globally, companies like TSMC, Foxconn, and Pegatron are leaning on NVIDIA Omniverse to spin up digital twins, optimizing assembly lines before laying a single physical brick.

Why it matters

NVIDIA's rhetoric shift is brilliant positioning. Calling facilities "factories" instead of data centers reframes GPU purchases from an IT expense into manufacturing capital expenditure. In Huang's framing, you are not burning money on infrastructure; you are investing in a factory that prints valuable tokens.

For developers, hardware like DGX Spark and DGX Station matters immediately. Being able to run trillion-parameter reasoning locally from a wall plug removes cloud latency, data privacy friction, and monthly rental fees for teams prototyping complex agent systems.

NVLink Fusion is an equally shrewd defensive maneuver. Major hyperscalers want custom silicon to reduce their reliance on NVIDIA. By unbundling NVLink for semi-custom designs, NVIDIA essentially says: build whatever compute blocks you want, as long as the nervous system connecting them remains ours.

The real risk, in our view, is downstream demand. Trillions of dollars in physical and digital infrastructure require end-user software that generates enough revenue to justify the electricity bill. If autonomous agents and humanoid robots take longer to commercialize than hardware deployments suggest, the industry could face a hangover of underutilized capacity.