Why One Top Fund Manager Is Betting on TSMC — Not NVIDIA — in the AI Era

Jonathan Cofsky of Janus Henderson says the real money in AI isn't in chips, but in the factory that makes them.

The Contrarian Top Holding

When Cofsky told MarketWatch that his $9.35B Janus Henderson Global Technology & Innovation Fund had made TSMC its largest AI position — bigger than NVIDIA — it raised eyebrows. After all, NVIDIA has been AI's poster child for two straight years. But Cofsky's thesis is simple: ask how much computing the world will build, and whoever wins, TSMC wins — because every major chip designer, NVIDIA included, manufactures at TSMC.

His fund has beaten the S&P 500 on both three-year and ten-year annualized returns.

Physical Constraints Set This Cycle Apart

Cofsky draws a sharp distinction between today's AI buildout and the dot-com bubble: real-world constraints matter now. Advanced fabs and power infrastructure can't be conjured from software alone. These physical bottlenecks naturally enforce capital discipline in a way 1999 never saw.

TSMC: The One Door Everyone Walks Through

TSMC is the world's only pure-play contract chip manufacturer — it builds everyone's chips without competing with them. That neutrality is the entire point. Whether AI's future belongs to generative models, agents, or physical robotics, the silicon has to come from somewhere. Currently, that's almost exclusively TSMC. Cofsky calls it the "only answer" in the computing power race.

The Arm Bet: AI Agents Need CPUs Too

Cofsky has also been adding Arm Holdings. The logic: agentic AI workflows depend heavily on CPU coordination — scheduling, memory management, parallel task orchestration — rather than pure GPU compute. Arm's architecture dominates mobile and embedded markets with what Cofsky describes as "the broadest architectural penetration." He's also watching Arm's transition from IP licensing into chip manufacturing, with Meta as its first customer.

Software: Only 10–20% Will Survive

On software stocks, Cofsky is cautious. AI threatens to automate labor at a scale that will devastate most incumbents. Historical precedent suggests it takes 4–6 years for markets to identify true survivors, and typically only 10–20% make it through. His focus: companies with hard-to-replicate advantages that can attach AI as incremental revenue — Datadog, Snowflake, Cadence.

Contrarian Pick: DoorDash Mispriced

Market concerns about DoorDash's investment cycle and consumer macro sensitivity have weighed on the stock. Cofsky sees those fears as overdone: founder-led strategy, international grocery expansion, and dual受益 from generative AI (logistics) and physical AI (autonomous delivery) make it a potentially significant value play.

Key Takeaways

Source: Macrostream — "Wall Street Fund Manager: The Most Important Company in the AI Era Is Not NVIDIA, but TSMC!"