The explosive rise of generative AI has redefined performance benchmarks across industries. From foundational model training to high-powered large language models (LLMs) for consumer applications, large-scale AI deployments are straining global computing infrastructure. Now, mounting supply pressures brought on by rapid adoption of AI technology are creating new challenges for chipmakers and procurement leaders alike.
Three crucial components are in the spotlight: high-bandwidth memory (HBM), high-performance graphics processing units (GPUs), and advanced solid-state drives (SSDs). Once confined to specialized clusters, these parts are now essential building blocks for every hyperscale data center racing to meet the next wave of AI acceleration.
However, the convergence of surging AI demand, fragile semiconductor capacity, and ongoing geopolitical headwinds is putting unprecedented stress on the supply chain. With memory requirements per chip at all-time highs and AI workloads forecasted to grow 25-35% year-over-year (YoY) through 2027, the start of protracted shortages of these sought-after components is actively unfolding. As policymakers in Washington consider sweeping tariffs aimed at reshaping global semiconductor trade flows, shortages will be exacerbated further.
The chip industry anticipated the transition toward AI-centric computing, but few predicted the speed at which demand would accelerate. In 2024, NVIDIA’s H100 and AMD’s MI300 series GPUs saw near-immediate sellouts across all production cycles. A growing segment of AI startups and enterprise AI labs joined major cloud providers in clearing the shelves. As this period of soaring demand shows no signs of slowing, it is reshaping the memory hierarchy.
HBM3, the memory standard used to power data-hungry AI accelerators, is at the center of a new supply bottleneck. SK Hynix, Samsung, and Micron—the three dominant producers of HBM chips—have been operating near full capacity. Yet, the trio is currently reporting six- to twelve-month lead times amid a backlog of orders. In some cases, lead times are even longer when coupled with specialized packaging constraints, particularly for customers seeing TSMC’s CoWoS-style integration.
Analysts have noted that HBM3 pricing has already risen 20-30% year-over-year, a trend expected to persist throughout 2025 as demand continues to outpace capacity expansion.
The GPU market tells a parallel story. Despite aggressive capacity ramp-ups, Nvidia’s next-generation Blackwell GPUs are already backlogged for a year or more. Hyperscalers like Microsoft, Google, and Meta dominate allocation. These supply constraints are unlikely to abate in the near term thanks to buildouts of new AI-dedicated data centers and sovereign AI initiatives.
Meanwhile, the storage segment is beginning to feel this ripple’s effects. Traditional SSDs are proving inadequate for sustained next-gen AI workloads, which require significantly higher endurance and bandwidth. Enterprise-grade NVMe SSDs have thus seen prices climb by 15-20% compared to a year ago. Suppliers are struggling to keep pace with overlapping demand from hyperscale cloud and edge AI deployments.
Thanks to AI, pressure on the semiconductor supply chain is no longer confined to front-end chip fabrication. The industry now finds itself grappling with the complex constraints of packaging and integration needed to bring these high-end components to the market. Many extra wafers produced thanks to upscaled fab capacity are now in line for the advanced packaging steps required to make them usable.
TSMC’s CoWoS packaging, for instance, is critical for stacking HBM alongside AI processors. Demand for this advanced integration technique has exploded in the past year and capacity is already fully booked through the end of 2025. By the end of 2026, the Taiwanese giant’s CoWoS capacity is expected to reach 90,000 wafers per month, but even a CAGR of 50% between 2022 and 2026 isn’t enough to meet demand. Nvidia, AMD, and a growing roster of AI chip startups are booking all available production slots the moment they become available.
Compounding supply issues is the regional concentration of chip production. The Asia-Pacific region, particularly Taiwan and South Korea, dominate advanced packaging and testing infrastructure. That consolidation poses a growing risk to global supply chain continuity. When a 7.4-magnitude earthquake struck Taiwan in April 2024, it temporarily halted output at several key fabs and packaging facilities. Despite production coming back online quickly, the incident underscores the fragility of the system. A single natural disaster or political move could affect deliveries worldwide.
Efforts to expand capacity aren’t simple. New fab costs have jumped to $20-75 billion with long timelines for ROI. Government incentives, while generous in some regions, have struggled to speed up timelines or attract sufficient private capital at the scale required. The capital risk remains significant, especially when demand projections hinge on unpredictable generative AI adoption curves.
Meanwhile, second- and third-tier suppliers—those providing substrates, photomasks, and specialized test equipment—are also stretched thin. In many cases, their production is just as limited as the upstream chipmakers they support.
Taken together, these systemic chokepoints illustrate the unfortunate truth that there is no easy fix. The chip industry cannot simply build its way out of the AI surge. Production and packaging must scale in tandem, and both remain hampered by long investment horizons, limited regional diversity, and a workforce in short supply. Until these factors align, capacity constraints will remain a defining feature of the market.
While supply-side bottlenecks continue to limit component availability, trade policy developments add another layer of complexity. The second Trump administration’s push toward aggressive trade reform has reignited fears of a fragmented semiconductor market, where tariffs and strict compliance requirements matter just as much as fabrication nodes and lead times.
At the heart of Washington’s policy shift is a proposal to levy new tariffs on semiconductor imports. The initiative, positioned as an extension of the administration’s “America First” economic platform, aims to incentivize domestic manufacturing by penalizing imports of critical chip components. While final details remain under negotiation, preliminary estimates suggest tariffs could raise component costs anywhere from 10 to 30%, depending on classification and origin.
For many companies, the cost increase would be unavoidable. The international nature of chip manufacturing means even a single AI chip could involve a web of cross-border steps. An HBM module fabricated in South Korea might be integrated into a GPU die in Taiwan, packaged in Malaysia, and shipped to an assembly facility in Mexico before entering the U.S. market. Under the proposed tariff plan, each of those stages could trigger a separate duty, compounding costs dramatically for such components. For procurement teams already managing tight allocation schedules and rising costs, the additional burden of tariff compliance introduces significant operational risk.
Many vendors are adjusting pricing and delivery terms preemptively. Several chipmakers and component suppliers are incorporating tariff buffers into their quotes, hedging against the possibility of retroactive duties or shifting classifications. Shipments of some components have been delayed by as much as six weeks or rerouted through alternative logistics hubs to mitigate exposure.
Industry leaders have expressed growing concern about these policies. Unlike in other sectors, where sourcing alternatives can be more easily substituted, AI-grade semiconductors depend on a narrow set of vendors operating at the bleeding edge of technological capability. There are no short-term replacements for TSMC’s CoWoS technology or SK Hynix’s HBM3 stacks. Efforts to localize production will take years to materialize—and may never reach the scale or performance of the global incumbents.
For supply chain leaders navigating the web of AI hardware availability, reacting to new variables is no longer sufficient. The intersection of capacity constraints, geopolitical volatility, and escalating demand requires a structural rethink of procurement strategies. Surviving this cycle will depend on the ability to rapidly shift from transactional buying to long-horizon planning.
One of the clearest imperatives is early commitment. Hyperscalers, original equipment manufacturers (OEMs), and even well-capitalized startups are locking in supply through multi-quarter, and in some cases multi-year, purchase agreements. For second-tier buyers, failure to secure allocation windows early is already translating into prolonged delays. Strategic stockpiling, once viewed as inefficient, is becoming a worthwhile hedge against systemic volatility. Though it is pertinent to practice such a method in moderation, as overreliance contributed to the overwhelming amount of excess that plagued the industry in 2023.
Meanwhile, multi-vendor sourcing is more important than ever. While market leaders like Nvidia and AMD remain irreplaceable for top-end AI accelerators, regional silicon players—especially in China, India, and the EU—are a way for procurement teams to diversify exposure. These alternatives may not offer identical performance, but for non-critical inference tasks, or internal model fine-tuning, they provide valuable flexibility. Sourcing across geographies, even at a modest cost premium, can reduce exposure to single-point failures.
What remains absent is regulatory clarity. For companies making multimillion-dollar sourcing decisions, navigating a tariff environment defined by speculation is a minefield. Will component-level tariffs be assessed at the point of fabrication, packaging, or final assembly? Will exemptions apply to countries with strategic partnerships or U.S.-based investment? Without clear answers, it’s imperative that organizations build contingency into every shipment and budget for the impact of changing legal interpretations.
The call from industry stakeholders is straightforward: tariff policy must account for the complexity of the supply chain it seeks to regulate. That includes not only transparency in enforcement mechanisms but also targeted incentives that support domestic investments in high-value bottlenecks.
Much will also depend on the pace of demand acceleration. If AI workload growth remains on its current trajectory of doubling every two to three years, the gap between component demand and available capacity could widen significantly. Tariff enforcement disincentivizing international coordination would expand the gap further. In the most strained scenarios, costs could spiral further, lead times could extend beyond 12 months, and smaller AI firms could be priced out entirely.
Runaway AI demand, limited component capacity, and uncertain trade policy have created a supply chain environment that is volatile by nature. For procurement leaders, the challenge now is less about managing disruption and more about navigating a prolonged structural realignment in how, and where, critical AI components are sourced.
Near-term success depends on strategic foresight. Multi-scenario planning, diversified sourcing, and deeper supplier engagement must become the norm. Procurement strategies that once prioritized lean inventory and just-in-time fulfillment now require buffers, pre-commits, and geopolitical risk modeling.
Still, there is reason for cautious optimism. If policy aligns with operational reality, offering clear tariff enforcement and targeted support for domestic packaging and memory manufacturing, capacity relief could begin to materialize within two to three years.
Until then, the companies that invest in resilience today will be best positioned to lead tomorrow’s AI market. Sourceability’s global experts can help your organization get started.