TSMC is pushing more of its advanced chip packaging to outside contractors, according to people familiar with the matter, as orders for Nvidia’s AI processors strain the company’s production lines to the limit.
The world’s largest contract chipmaker has decided to outsource more of the chip-on-wafer, or CoW, step of its CoWoS packaging process to ASE Technology and other packaging and testing houses, the people said. CoWoS is the 2.5D packaging technology TSMC uses for AI chips: the processor sits in the center, high-bandwidth memory modules ring it, and an interposer connects them into a single package. The technique lets chip designers combine a powerful accelerator with the memory it needs to feed, and it has become the indispensable manufacturing step of the AI boom.
Demand is the problem. Nvidia’s GPU orders have pushed TSMC’s packaging lines to their limits, and the company has been expanding its own CoWoS capacity for more than a year, building new plants in Taiwan and reconfiguring older fabs. But the packaging step remains the tightest constraint in the AI supply chain, and every accelerator Nvidia sells needs one of these packages. Outsourcing the CoW step, which stacks the chips on the interposer, lets TSMC focus its own capacity on the more valuable parts of the process while pulling in outside capacity to keep the lines moving.
The move is a sign of how far the bottleneck has extended. Two years ago, CoWoS capacity was the subject of worried supply-chain reports; today, TSMC’s own expansion has eased the worst of it, but the sheer volume of Nvidia’s orders, combined with growing demand from AMD, Broadcom and the cloud providers’ custom chips, has pushed the system to its limit again. ASE, the world’s largest packaging and testing company, has been adding capacity in anticipation, and the outsourcing deal gives it a role in the most important manufacturing process in the industry.
The packaging step is also where the AI supply chain meets the memory makers. The HBM modules that surround the processor come from SK Hynix, Samsung Electronics and Micron Technology, and the interposer, a thin slab of silicon that carries signals between them, is made by TSMC and a few other suppliers. Any delay in any one of those steps holds up the finished package, which is why TSMC’s decision to share the CoW work is being watched so closely: it effectively adds a fourth partner to a process that was running at full speed.
Analysts said the outsourcing reflects a pragmatic calculation. The CoW step, while technically demanding, is less valuable per unit than the wafer processing TSMC performs in its fabs, and handing part of it to ASE preserves TSMC’s own capacity for the steps where its technology advantage matters most. It also hedges against the risk that AI demand, after two years of explosive growth, takes a pause: the company can pull the outsourced work back in if the lines empty.
For Nvidia and its customers, the practical effect is shorter lead times and, eventually, more GPUs in the market. For the industry, the deal is another sign that the AI boom has outgrown even the capabilities of the company that builds its engines, and that the surrounding ecosystem, from packaging houses to memory makers to interposer suppliers, will carry an increasing share of the load. The next constraint, analysts said, is likely to be HBM supply, and the companies that solve that problem will define the next phase of the AI cycle.
The decision also has implications for the pecking order of the semiconductor supply chain. For years, TSMC has been the indispensable company, the one supplier no one could do without, and its pricing power has reflected that. Sharing the packaging work with ASE is a recognition that the AI boom has created volumes that no single company can handle alone, and that the ecosystem around TSMC, from ASE to the memory makers to the interposer suppliers, will share in the growth. ASE’s shares rose on the report, a sign that investors read the news as a transfer of business to the packaging giant.
The question for the industry is how long the constraint lasts. TSMC’s own expansion, combined with the outsourced CoW work, should ease the packaging bottleneck over the next year, and the AI chipmakers are already designing around the limits they know. The next constraint will be HBM, where supply is tight and the memory makers are spending heavily to expand. If the pattern of the past two years holds, the bottleneck will simply move down the chain, and the companies that own the next scarce step will reap the rewards. For now, the immediate effect is more GPUs reaching customers sooner, which is good for everyone selling AI, from Nvidia to the cloud providers to the startups that buy their capacity.
