Artificial Intelligence & Machine Learning

AI Chip Demand: Semiconductor Industry Faces Pandemic Déjà Vu

Rebounding from the pandemic-induced global chip shortage of 2020-21, the semiconductor industry is bracing for yet another impact - this time due to artificial intelligence. The accelerated adoption of AI across industries has caused an unprecedented increase in demand for purpose-built semiconductors, particularly AI chips, that could upset the intricate demand-supply equilibrium.

A demand increase of 20% or more is enough to disrupt the semiconductor supply chain and cause a chip shortage, according to a report by Bain and Company. AI adoption is driving 31% growth in PC sales and 15% in smartphone sales between 2023 and 2026. To meet this surge, bleeding-edge semiconductor fabs would need to increase output by approximately 25% to 35%, costing an estimated $40 billion to $75 billion.

The supply for advanced chips could "hopefully" balance out demand only in "2025 or 2026," said C.C. Wei, the CEO of TSMC, the global leader in the chip foundry market.

"The demand [for AI] is real," he said.

Unlike traditional processors, AI chips are vital for handling workloads such as machine learning, neural network processing and generative AI tasks. A typical AI chip, such as those used in training large language models or powering advanced computer vision in edge devices, has an architecture with billions of transistors and highly complex circuitry to support parallel processing. This complexity means AI chips not only consume more raw materials but also require advanced fabrication techniques, often restricted to the most advanced semiconductor manufacturing facilities.

AI-enabled PCs are leading the charge. They are expected to make up 43% of all PC shipments by 2025, a jump from 17% in 2024. By 2026, AI-enabled laptops will be the only choice of laptop available to large businesses, up from less than 5% in 2023. (see: 2025 Is the Year of AI PCs; Are Businesses Onboard?)

Far from a simple upgrade, this shift represents a step-change that demands significantly more from semiconductor manufacturers and involves a complex supply chain - from advanced packaging and memory production to constructing data centers and securing reliable power sources. As chipmakers scramble to meet high demands, the semiconductor industry faces a significant challenge: adapt and scale, or risk another pandemic-like global shortage.

But unlike the pandemic that gave little time to the industry to prepare for a boom in remote work and a sharp rise in sales of consumer electronic devices, chipmakers today are anticipating AI overload and future-proofing their operations.

Gartner forecasts the global semiconductor revenue to total $717 billion by next year, combined with a surge in demand for next-generation GPUs and high-bandwidth memory, HBM - a specialized memory essential for high-performance AI servers.

Yet, with only a handful of players capable of producing these advanced components - TSMC, Samsung and Intel among them - the semiconductor industry risks a bottleneck.

TSMC makes all of the world's advanced AI chips. Some of its biggest clients include Apple and NVIDIA. If TSMC were to face a supply crunch, two of the world's largest companies would grind to a halt. "Basically, there is air - and TSMC," said NVIDIA CEO Jensen Huang.

NVIDIA - the leading producer of H100 GPUs predominantly used to train AI models such as ChatGPT - has achieved remarkable success, earning over $30 billion in revenue in second quarter of 2024. The shipments of its H100 processes are projected to reach 2 million in 2024, much higher than its prediction of 500,000 units. But NVIDIA's production remains limited by its reliance on TSMC's 5nm process. Unverified reports say the company has chosen Intel to supply advanced packaging services for a portion of AI chip production, as TSMC struggles to meet the high demand for its own advanced packaging capabilities.

Specialized equipment shortages compound these challenges. Extreme ultraviolet lithography machines, essential for producing chips with dense circuits, are produced exclusively by ASML, creating additional constraints on scaling up production quickly.

But by now, companies have a playbook. The pandemic chip shortage underscored the semiconductor industry's need for resilience. The traditional "just-in-time" inventory model, which kept costs down but left manufacturers exposed to demand surges, proved untenable. In response, chipmakers are now pivoting to "just-in-case" inventory practices, keeping reserves of critical components to guard against future shortages, the Bain report showed. To keep pace with AI's demands, companies are investing in fabrication techniques, such as 3D chip stacking and silicon photonics, which enhance performance without relying on the most advanced lithography nodes.

With major players located in East Asia, the development of regional manufacturing hubs has become a priority for governments. President Joe Biden signed the CHIPS Act, injecting $52 billion into domestic semiconductor production, while Europe's Chips Act targets a 20% global production share by 2030. This localization push aims to build resilience, especially given heightened geopolitical tensions between the U.S. and China, which threaten to disrupt access to critical materials.

Intel's $20 billion Ohio facility is part of its plan to boost output and meet the growing demand for AI chips. TSMC is also expanding its U.S. footprint, with a $40 billion facility in Arizona geared toward manufacturing advanced AI components, while Samsung has committed $17 billion to a Texas-based fab, though these facilities will not reach full capacity before 2025.

The supply-crunch repercussions will be felt directly by enterprises looking to innovate. A chip shortage could slow down their projects in AI-driven analytics, automation and machine learning, potentially disrupting timelines for digital transformation. To mitigate these challenges, enterprises should diversify chip sourcing, prioritize strategic partnerships with key suppliers, adopt adaptive AI architectures that rely on multiple types of processors, and invest in cloud-based AI services to offset hardware limitations. This too, however, is short term.

With the promise of basking in dollars, chipmakers have a lot to gain from AI. But with a delicate supply chain and constantly evolving technologies, they can't afford to cut corners. The next few years will be crucial as chipmakers navigate the complexities of AI acceleration, with stakes higher than ever for an industry foundational to the digital economy.