The Semiconductor Crisis Nobody Was Watching For

How the AI data center boom created two parallel semiconductor crises, and why only one was contained.

In late 2024, a well-known concern surfaced in manufacturing boardrooms. Hyperscalers were committing hundreds of billions of dollars to AI infrastructure, and supply chain leaders who remembered the disruption of 2021 were asking the obvious question: are we about to see the same thing again, only bigger?

Nearly two years later, the answer is mixed. Logic chips, including the microcontrollers and mature-node parts whose shortages stalled auto plants in 2021, have largely avoided a repeat. Memory has not. DRAM and HBM prices have roughly doubled in twelve months, suppliers are allocating capacity rather than simply fulfilling purchase orders, and analysts are openly using the term "DRAMpocalypse."

The pressure came from the same source: AI's appetite for compute. Yet the outcomes have been very different. This matters well beyond semiconductors, because the next constrained resource in a supply chain may be just as difficult to anticipate. Five capabilities have helped separate the manufacturers that anticipated the risk from those caught unprepared, whether or not memory is currently part of their bill of materials.

The scale is worth noting. The five largest hyperscalers spent roughly $388 billion on infrastructure capex in 2025 and are guiding toward $630–690 billion in 2026. This is not a short-term increase in demand that businesses can manage with buffer stock. It represents a structural shift in global manufacturing capacity toward one customer segment, with consequences for other industries.

Why logic mostly avoided a repeat

The chips whose shortages disrupted automakers in 2021 are manufactured on older process nodes (28nm and above), in fabs that are structurally separate from the leading-edge lines producing AI accelerators. An airbag controller was never competing with Nvidia for TSMC's advanced packaging capacity. Those two areas of demand were separated by the underlying production technologies, rather than by a strategic supply chain decision.

That doesn't mean the AI supply chain was free of constraints. CoWoS packaging capacity has been a genuine bottleneck, and Nvidia alone is estimated to have secured more than half of it through 2027. But that pressure has remained largely within the AI compute supply chain, rather than spreading across every industry that relies on semiconductors.

The manufacturers that weathered the disruption best had also learned from 2021. In the years since, they had invested in five capabilities without needing an AI-specific justification:

• 1. Mapped supply networks, not just supplier lists. They traced dependencies down to wafer fabs and packaging houses. When HBM and CoWoS constraints began appearing in supplier earnings calls, they could recognize their exposure months before it affected a purchase order.
• 2. Secured capacity, not just inventory. Long-term agreements and reserved capacity reduced the risk of discovering mid-crisis that they were competing with a trillion-dollar capex cycle for the same production lines.
• 3. Treated component risk as an executive issue, rather than a procurement footnote. This helped them make capital and timing decisions earlier, with less internal friction than companies still routing those decisions through quarterly procurement reviews.
• 4. Shared realistic demand signals. Providing demand ranges and actual consumption data, rather than a single inflated forecast, supported better allocation decisions. Suppliers are more likely to prioritize customers whose forecasts they trust.
• 5. Designed for flexibility. Multi-sourcing and platform standardization reduced the number of products dependent on a single part number.

A sixth factor supported all five: restraint. In 2021, the bullwhip effect made shortages worse as buyers doubled their forecasts at the same time. That did not happen on the same scale this time, suggesting a more mature, less reactive approach to managing supply risk.

These capabilities are not unique to semiconductors. They are also helping determine which manufacturers can manage the current memory shortage and which are struggling.

Why memory was different

DRAM and HBM are not separated by the same manufacturing divide that protected mature-node logic. HBM and the DDR5 modules used in enterprise servers are produced on the same fab lines by Samsung, SK Hynix, and Micron. Together, those three companies account for essentially the entire global merchant memory market.

HBM consumes roughly three times the wafer capacity of standard DRAM per gigabyte, so every wafer redirected toward HBM reduces the capacity available for server RAM or memory used in industrial control boards.

That reallocation is already showing up in supply and pricing. SK Hynix has described its entire 2026 output as essentially sold out, DDR5 contract pricing has roughly tripled in a year, and finished-goods inventory at the two largest suppliers has fallen below ten days, a level last seen in the 2018 memory supercycle.

For a manufacturer with substantial DRAM content, the impact is immediate: a bill-of-materials item has doubled in cost in twelve months, with no manufacturing separation from AI demand to protect it.

The situation is familiar from 2021, when a relatively ordinary component could bring an entire production line to a halt. Yet many of the resilience programs developed after that shortage focused on microcontrollers and did not apply the same discipline to memory. They were designed to prevent a repeat of the previous crisis, leaving a different exposure unaddressed.

The problem is also harder to resolve. New logic and packaging capacity has been coming online steadily. Memory has no comparable short-term relief. Adding a gigabyte of HBM output requires roughly three gigabytes of DRAM wafer capacity, and new memory fabs take years to permit and build.

Some of SK Hynix's next-generation capacity is not expected to reach volume output until 2028–2029. This helps explain why several analysts expect the current cycle to last longer than the 2017–2018 DRAM price spike that many procurement teams still use as their benchmark.

What prepared manufacturers did differently

The manufacturers managing this shortage most effectively are not necessarily those holding the largest inventories. They have applied the same five capabilities to memory, ahead of many of their peers.

By mapping memory content down to the module level, they could identify affected product lines within days of SK Hynix disclosing that its 2026 output was fully committed. Multi-year agreements insulated part of their volume from spot-linked pricing. Executive teams had already considered what would happen if DRAM prices tripled, so substitution plans were approved in advance rather than developed during the shortage.

They were also sharing actual consumption data with suppliers instead of relying on single-point forecasts. And they had built in alternative DRAM densities and DDR4-to-DDR5 substitution paths, reducing the likelihood that an unavailable configuration from one supplier would stop production.

Manufacturers without these capabilities are facing input cost increases of more than 100% with no contractual protection, and discovering their memory exposure through supplier price increase letters rather than their own risk models.

The common thread

Across both cases, the main differentiator was not inventory. It was visibility into the extended supply network, contractual access to constrained capacity, executive-level ownership of component risk, realistic two-way demand signals with suppliers, and product designs flexible enough to absorb a supply shock without requiring a redesign.

Manufacturers that developed these capabilities as part of their normal operations, rather than as a one-off response to 2021, could apply them again, more quickly, to a different component.

Their value was never limited to semiconductors. These capabilities strengthen how organizations understand their supplier networks, negotiate access to capacity, and exchange information when supply comes under pressure. They remain useful even as the constrained component changes.

The implications also extend beyond semiconductors. Power and grid equipment — transformers, switchgear, and cooling infrastructure — is already showing similar pressure: rapid demand from a small group of buyers converging on a limited number of manufacturers, with lead times that cannot adjust quickly.

Rare-earth and critical mineral supply faces similar dependencies on individual regions. These constraints will not unfold exactly as DRAM has, but they will require the same underlying capabilities.

What supply chain leaders can do now

For CSCOs and CPOs assessing their own bills of materials, three actions are worth prioritizing this quarter, whether or not memory is currently a concern.

First, identify components by their potential to stop production, rather than simply by spend. Focus on those sourced from fewer than three viable producers, and examine whether their suppliers are reallocating capacity toward AI-adjacent demand. Memory is the clearest current example, but it is unlikely to be the last.

Second, ask semiconductor suppliers what proportion of your current volume is covered by multi-year agreements and how much remains exposed to spot or short-term pricing. If the answer is "mostly spot," that is the exposure to address now, rather than waiting for the next supplier price increase letter.

Third, bring the question of what happens if a critical component doubles or triples in cost or lead time to the executive committee before it becomes urgent. Agree in advance who has the authority to approve substitutions, reallocations, or emergency sourcing above a defined threshold. The governance gaps that slowed many organizations in 2021 remain unresolved for memory today.

The timing matters. Memory's warning signs were public months before price increases reached customers: supplier disclosures of sold-out capacity, inventory days falling below historical lows, and increasingly stark analyst commentary.

Responding to those signals requires effective processes and governance, not just better forecasts. These are areas organizations can address this quarter, regardless of where memory prices go next.

The wider lesson

The AI data center boom created the conditions for a second global semiconductor shock. In the leading-edge logic supply chain, much of that shock was contained, supported by supply network visibility, strategic capacity access, executive-level risk ownership, better demand signals, and more flexible product design.

Crucially, these capabilities had been developed in the years between crises, rather than in response to the latest disruption.

Memory has seen the opposite outcome, and the shortage is still unfolding. Manufacturers managing it more effectively are applying the same approach to a different component category. Others assumed the industry had moved past the chip shortages of 2021, only to find themselves exposed to a new constraint.

The lesson is not that another semiconductor supply shock was avoided. Only part of it was. The ongoing pressure on memory shows why organizations need to identify concentration risks before they become supply disruptions, and why the capabilities developed after one crisis need to extend beyond the components that caused it.

The manufacturers best equipped to handle the next shortage will be those that can recognize these risks early, rather than discovering them through a supplier's price increase letter.

Why Semiconductor Demand Is Outpacing Traditional Planning Models

Global semiconductor demand is accelerating faster than the industry can absorb it. AI workloads, advanced compute, electrification, automotive content growth, and edge devices are driving sustained demand across
both leading and mature nodes.

Download our White Paper to understand how to tackle the challenges facing the Semiconductor industry today.

Download now