TSMC won’t rush new factories despite Nvidia’s push for more supply

TSMC is holding back on rapid expansion plans, even as technology companies push for faster production to meet soaring artificial intelligence demands.

TSMC, the Taiwan-based factory that manufactures nearly every high-performance AI processor, faces mounting pressure from its biggest clients.

Nvidia’s chief, Jensen Huang, recently disclosed his firm needs to fulfill orders worth $500 billion over the coming two years. OpenAI has separately arranged multi-year agreements with Advanced Micro Devices for roughly 3 million to 6 million chips generating six gigawatts of power, plus another deal with Broadcom for ten gigawatts more.

This rush for AI-powered processors has sparked worries about whether enough can be made. The Taiwanese manufacturer handles production for almost all premium AI chips currently available.

Huang traveled to Taiwan personally, pushing the company to increase Nvidia chip output by 100%. Elon Musk has considered whether Tesla might construct its own massive semiconductor plant to support the carmaker’s AI and robotics projects. Sam Altman, who leads OpenAI, has publicly asked TSMC to “just build more capacity.”

TSMC’s spending increases but not fast enough

The manufacturer is responding, though not at the pace customers want. This year’s planned spending on new equipment and facilities stands at approximately $41 billion, with over two-thirds directed toward the sophisticated chips AI companies need. That represents growth from $30 billion spent last year. Industry analysts predict spending will climb to $52 billion by 2027 as reported by Economic Times.

However, when measured against company income, TSMC’s infrastructure investments have actually decreased and are projected to keep falling in coming years, despite louder complaints from customers.

Chip designers have limited options beyond TSMC. Intel, once a manufacturing leader, has fallen behind in producing cutting-edge processors. Samsung remains the only other serious choice, but faces its own troubles. Tesla signed a $16.5 billion agreement with the South Korean electronics company last July to start making AI chips in Texas beginning next year, though the venture reportedly lags behind schedule.

Taiwan’s manufacturing powerhouse now finds itself buried in orders. Warning signs have emerged. Last month, C.C. Wei, the company’s top executive, acknowledged that capacity for AI-related products “is very tight” and the firm is working “very hard” to address the shortfall. Meanwhile, orders continue growing. Investment bank Jefferies forecasts Nvidia will receive 6.3 million AI chips next year, representing a 23% jump from this year’s volume.

Broadcom, which creates chips for cloud computing companies like Google and Meta, is expected to require 5.4 million units, a 29% increase. JPMorgan Chase analysts believe TSMC’s two most advanced chip varieties will operate at maximum capacity for years ahead.

Boosting production, though, takes significant money and time. A state-of-the-art manufacturing plant costs around $20 billion and requires three to four years to finish. Construction in America increases expenses further. TSMC has committed to building six new advanced factories in the United States, with combined costs near $165 billion. America’s troubled construction sector means these plants cost substantially more than comparable facilities in Taiwan, while also taking longer to complete.

Customers want more but won’t pay upfront

Expanding capacity also carries risks for TSMC, which explains its careful approach. The semiconductor business has historically experienced painful cycles where tight supply gives way to excess capacity. Manufacturers frequently expand aggressively during good times, only to end up with unused factories when orders eventually drop.

Daniel Wu from hedge fund Bristlemoon Capital points out that TSMC increased investment during the pandemic to address shortages of simpler chips.

Today, its capacity for making those processors sits underutilized. The company likely wants to avoid repeating that error. It may also worry that the AI chip market could flip into oversupply if Intel and Samsung finally fix their production issues.

Companies demanding more capacity don’t pay anything until TSMC actually builds it. But Wu notes that if TSMC commits $80 billion to $100 billion over two years only to see the AI boom fade, it would face years of idle factories. Its caution may frustrate customers. But they shouldn’t expect TSMC to change course.