The real chip war is packaging

Everyone talks about the chip war as a fabrication story. Who can print the smallest transistors? Which country controls the most advanced fabs? But there's a quieter, more consequential battle playing out right now, and it has nothing to do with how chips are designed or etched. It's about how they're assembled. Advanced packaging, the process of connecting multiple chip dies into a single functional unit, has become the true bottleneck in AI hardware. You can design the most brilliant silicon in the world, but if you can't package it, it doesn't ship.

What advanced packaging actually is

Traditional chips are monolithic: a single slab of silicon that does everything. But as chips have grown more complex, the industry has shifted toward a "chiplet" model, where multiple smaller dies are manufactured separately and then connected together into one package. This is where technologies like TSMC's CoWoS (Chip on Wafer on Substrate) come in. CoWoS places chip dies and high-bandwidth memory (HBM) side by side on a silicon interposer, then mounts the whole assembly onto a substrate. It's what makes modern AI GPUs like Nvidia's H100, B100, and GB200 possible. Without it, these chips simply can't reach the performance levels that AI training and inference demand. Other approaches exist too. TSMC's 3DFabric platform includes SoIC for vertical 3D stacking and InFO for fan-out packaging. Intel has its own alternatives: EMIB (Embedded Multi-die Interconnect Bridge) uses tiny silicon bridges embedded in the substrate to connect chiplets, while Foveros enables full 3D die stacking. These aren't theoretical, they're shipping in production hardware today.

The bottleneck nobody saw coming

For years, the semiconductor conversation revolved around transistor density and process nodes. Could TSMC hit 3nm? Would Intel catch up at 18A? But while the industry obsessed over fabrication, a different constraint quietly tightened. TSMC's CoWoS capacity couldn't keep up with demand. In late 2024, the company was producing roughly 35,000 CoWoS wafers per month. That sounds like a lot until you consider that Nvidia alone has reserved the majority of that capacity, and every major AI chip customer, from Google's TPU team to Amazon's Trainium group to Broadcom's custom ASIC platform, is competing for whatever remains. The result is a packaging bottleneck that has constrained GPU shipments, delayed product launches, and forced hyperscalers to scramble for alternatives. The most advanced chips in the world were sitting in queues, waiting to be assembled.

The scramble to scale

TSMC has responded with an aggressive expansion plan. The company is targeting approximately 130,000 CoWoS wafers per month by late 2026, nearly quadrupling its output in just two years. New advanced packaging facilities are going up across Taiwan, in Zhunan, Chiayi, Taichung, and Tainan, with existing fabs being converted from other processes to CoWoS production lines. TSMC is also outsourcing a portion of its CoWoS workload. Estimates suggest 240,000 to 270,000 wafers annually will be handled by OSAT (Outsourced Semiconductor Assembly and Test) partners, primarily Amkor and SPIL. This lets TSMC focus its internal capacity on the highest-margin, most complex work, like silicon interposer fabrication and front-end chip-on-wafer assembly, while partners handle substrate-level steps. But even with all this expansion, demand continues to outpace supply. Analysts predict CoWoS capacity will remain tight through at least 2027, particularly as next-generation HBM4 stacks move to 12 and 16 layers with hybrid bonding, adding yet more complexity to an already strained process.

Intel enters the ring

Here's where the story gets interesting. Intel, long struggling to find its footing in the foundry business, is positioning its advanced packaging capabilities as a genuine alternative to TSMC. In early April 2026, reports emerged that Intel is in advanced talks with Google and Amazon to provide packaging services for their custom AI chips using EMIB and the newer EMIB-T technology. Intel's stock jumped over 23% on the news. The company's packaging facility in Rio Rancho, New Mexico is being prepared for production, and its operations in Malaysia are expanding. Intel's pitch is compelling. EMIB-T, announced in mid-2025, promises improved power efficiency and signal integrity between chiplets. And for U.S.-based customers concerned about supply chain concentration, Intel offers something TSMC currently can't: domestic packaging at scale. With potential gross margins around 40%, packaging could become Intel's bridge to foundry profitability well before its leading-edge wafer manufacturing catches up. Elon Musk's SpaceX and Tesla have also reportedly tapped Intel for custom chip packaging in Texas, further validating the demand for alternatives to TSMC's dominance.

The US angle

Advanced packaging is where the CHIPS Act money arguably matters most. While headlines focus on new leading-edge fabs, the quieter investments in packaging infrastructure may have a bigger practical impact. In January 2025, the U.S. Department of Commerce finalized $1.4 billion in awards through the National Advanced Packaging Manufacturing Program (NAPMP), specifically aimed at building domestic advanced packaging capacity. The program's goal is straightforward: ensure that chips fabricated in the U.S. can also be packaged in the U.S., rather than being shipped to Asia for assembly and then shipped back. TSMC is building its first U.S. advanced packaging facilities in Arizona this year, alongside its wafer fabs. This is significant. Right now, even chips manufactured at TSMC's Arizona fabs have to take a round trip to Taiwan for packaging before they can be used. Closing that loop domestically is essential for the CHIPS Act's vision of a self-sustaining American semiconductor supply chain.

Why this is the real chip war

The geopolitical dimension is hard to overstate. Taiwan currently handles the vast majority of the world's advanced chip packaging. That's a single point of failure for the entire AI industry, and by extension, for every company and government betting on AI as a strategic capability. Fabrication gets the attention because it's dramatic: extreme ultraviolet lithography, sub-3nm transistors, billion-dollar cleanrooms. But fabrication without packaging is like having an engine factory with no assembly line. The parts exist, but the product doesn't. This is fundamentally an atoms problem, not a bits problem. AI progress, for all its software sophistication, is bottlenecked by physical manufacturing. The algorithms are ready. The chip designs are ready. The constraint is in the physical act of connecting tiny pieces of silicon together with sufficient precision, bandwidth, and thermal management to make them work as a system.

What this means for builders

If you're building anything that depends on GPU compute, the packaging bottleneck affects you directly. A few practical takeaways:

The chip war everyone talks about is fabrication. The chip war that actually determines who gets AI hardware, and when, is packaging.

References