Nvidia CEO Jensen Huang announced Wednesday that his company will spend $150 billion a year in Taiwan, open a new Taiwan headquarters, and deepen its partnership with TSMC by staying close to advanced packaging technology that does not yet exist at TSMC’s US facilities. The announcement, reported by Ars Technica, lands as the Trump administration continues pushing US firms to onshore AI chip production. Huang’s statement is the clearest signal yet that the physical reality of chip manufacturing and the political ambitions of the White House are not aligned.
The technical argument for Taiwan centers on CoWoS (chip-on-wafer-on-substrate), the advanced packaging process that stacks high-bandwidth memory directly onto Nvidia’s GPU dies. TSMC’s Arizona fabs are not yet equipped to perform this packaging step. Every finished Nvidia AI chip, including those nominally “made in America,” still travels to Taiwan for that process. Huang’s new headquarters and expanded TSMC partnership are bets that this constraint will not resolve quickly enough for the US supply chain to catch up.
Huang put the scale of the shift plainly at the Taiwan ceremony: four or five years ago, Nvidia was spending roughly $10 billion to $15 billion a year in Taiwan; that number has grown to $100 billion and is heading to $150 billion. He also connected the investment to demand for his forthcoming Vera Rubin AI system, which he expects will trigger “the greatest infrastructure buildout in history” while facing supply chain constraints throughout its production life. The Taiwan base also keeps Nvidia close to Foxconn, Wistron, and Quanta Computer, the system integrators who assemble the AI servers that data center operators actually buy.
The $150 billion figure deserves scrutiny. It is Huang’s own characterization of annual spend, not a signed contract with the Taiwanese government, not a capital commitment disclosed in a securities filing, and not an independently verified projection. Nvidia has strong incentive right now to perform Taiwan commitment loudly: it creates counter-pressure against US tariff threats and signals to Taiwan that Nvidia is a reliable long-term partner at a moment when Taiwan’s geopolitical position is contested. The number may be accurate; it may also be rounded up to make a political point. Both can be true simultaneously.
The US policy backdrop makes the timing significant. Semiconductor investigations ordered by the Trump administration are scheduled to conclude in July, and the outcome could include substantial new tariffs on imported chips used in data centers, according to reporting by Supply Chain Dive. US trade representative Jamieson Greer confirmed last week, per Bloomberg, that the administration continues to weigh semiconductor tariffs to encourage domestic manufacturing. Huang is positioning Nvidia’s existing US commitments as sufficient reason to be exempted, while simultaneously announcing that Taiwan is where the serious infrastructure buildout will actually happen.
This dynamic is not isolated to chips. Our coverage on May 28 noted that China is expanding travel restrictions on AI talent and asserting national-strategic control over its AI workforce. Both governments, the US and China, are now treating private AI infrastructure as a sovereignty asset and applying pressure accordingly. Nvidia sits in the middle, generating revenue from both ecosystems while navigating restrictions that limit its access to each.
The physical constraint Huang is describing maps directly onto the infrastructure bottleneck we covered in our Anthropic compute reporting. The limitation is not ambition or capital; it is the inability to produce finished, packaged AI silicon at the rate that demand requires, and the packaging capability is still concentrated in Taiwan. Moving that capability to the US is a multi-year process that requires TSMC to build out new fabs, train specialized workforces, and qualify processes that currently have no domestic equivalent.
Enterprise procurement teams sourcing AI compute through H2 2026 and into 2027 should watch two things: whether the July semiconductor tariff investigation produces new levies on data-center chips (which would raise input costs for US-assembled AI servers), and whether Nvidia’s Vera Rubin supply ramp faces the packaging constraints Huang warned about, which would extend lead times regardless of where a buyer sits in the procurement queue.
Reported by Ars Technica on 2026-05-28.
