TSMC Sees AI Accelerator Sales Soaring by 50% Annually

TSMC Sees AI Accelerator Sales Soaring by 50% Annually

TSMC projects that revenue from AI accelerators will increase by 50% yearly over the next five years after doubling this year.

On Thursday, amid ongoing geopolitical and macroeconomic uncertainty, Taiwan Semiconductor Manufacturing reduced its full-year outlook for the semiconductor industry.

Also read: Limitless AI Pendant Builds Upon the Mistakes of the Human AI Pin

Even though demand will be uneven overall, TSMC anticipates producing a significantly higher volume of AI accelerators this year than it did in 2023. By 2024, the company expects that revenue from AI accelerators will have more than doubled and account for low teens of total sales. TSMC projects revenue from AI accelerators to increase at a 50% annual rate over the next five years.

Boosting productivity

TSMC’s 3-nanometer process node has encountered a few problems. Apple employs the best process available for its customized chips, but AI accelerators are limited to 5nm technology. Although it was widely speculated that Nvidia’s recently announced Blackwell data center GPUs would be built on TSMC’s 3nm node, they will employ an altered version of the same methodology as their predecessors.

TSMC’s proactive measures to increase its 3nm capacity for AI chips, such as converting some of its 5nm tools, demonstrate its confidence in its ability to address challenges and meet future demands.

The move’s drawback will be a reduction in profitability. In the second half of this year, TSMC anticipates a 1 to 2 percentage point decline in its gross margin. Although the company hasn’t modified its long-term gross margin projection, it will look for cost-cutting measures to counteract the impact on gross margin.

Sadly, this technique can only generate so much additional capacity for more recent nodes. During the earnings call, an analyst asked TSMC CEO C.C. if it could increase 5nm capacity by transferring some resources from the presently underutilized 7nm node.

Wei clarified that other nodes would not be able to make a similar shift and that the move was made possible by the physical proximity of the 5nm and 3nm equipment.

TSMC anticipates a more seamless launch of its upcoming 2nm node. In the first two years of operation, Wei observed that the 2nm node has more tape-outs than the 3nm or 5nm nodes, primarily due to AI demand. In 2025, the 2nm node will start producing in large quantities.

The wild card in Intel

The market for AI chips is expanding, but starting next year, TSMC and Intel (INTC -2.40%) will compete to control the semiconductor foundry. By 2030, Intel hopes to become the second-largest foundry in the world. This expansion will be driven by new process nodes that have the potential to surpass TSMC in terms of technology.

While Intel is equally confident that its 18A process, which is set to go into production early next year, will be the best, TSMC is optimistic about its 2nm process. One significant advantage that Intel has is that the Intel 18A will use backside power delivery technology, which can significantly increase efficiency and performance. 

Demand for AI chips will support TSMC’s future growth, but the market won’t be all its own. TSMC’s best process nodes may face competition from Intel and Samsung, whose capacity could put pressure on the company’s market share and profitability.

Image credits: Shutterstock, CC images, Midjourney, Unsplash.