Today, NVIDIA and Doosan Group announced a collaboration to advance physical AI and AI factory infrastructure.
Our companies will explore robotics, AI factory power solutions and advanced electronics materials for next-generation data center systems.
Read more:
The Terafab does not swallow one terawatt of electricity.
People smash the numbers together. Current output of AI compute is roughly 20 gigawatts a year. That is chip output, not the factory power bill.
The 1 TW target is hardware you ship in a year. The 10 GW by 2027 is the power you stand up so the line can actually run.
20 GW is the planet now. 1 TW is the target. 10 GW is what you build so it turns on.
Own the photons. Then own the watts.
AI server maker Quanta (2382 TT) says its 2027 expansion will focus on electrical infrastructure upgrades. Its SMT equipment is largely ready, but newer AI racks consume more power and take longer to test.
We think that investors tracking $NVDA's manufacturing partners need to look beyond assembly-line capacity. The ability to power and test completed racks also affects how quickly hardware can ship.
At its Nov. 11 earnings briefing, CFO Tim Yang said the company needs to keep increasing factory power capacity to support those requirements.
SMT lines mount components onto circuit boards. Completed racks still need powered testing before shipment. Longer tests occupy testing capacity for more time, while higher rack power raises the electrical capacity needed to run them.
Quanta expects total 2027 capex of NT$40bn–45bn.