The year might be winding down, but pressure on the semiconductor supply chain is only ramping up. We’re entering another period of heightened uncertainty as new operational risks emerge amid growing AI infrastructure demand.  

In Taiwan, Micron’s Taoyuan labor union has secured the legal right to strike after 88.3% of its registered members supported industrial action. Although no strike has been confirmed, potential disruptions at a major DRAM manufacturing facility pose another threat to memory availability at a time when supply is already crunched. As with the strike concerns at Samsung in May, procurement teams face renewed unease around one of the industry’s largest memory manufacturers.  

Meanwhile, the next phase of AI is expected to place even greater pressure on the strained supply chain. According to McKinsey, AI inference will account for approximately 60% of AI compute demand by 2030, shifting industry priorities from training increasingly powerful models toward deploying them efficiently at scale. This transition will create new requirements for memory bandwidth, specialized processors, networking, power management, and advanced packaging.  

Together, these developments demonstrate how short-term manufacturing vulnerabilities and long-term technological changes are converging to reshape semiconductor supply chains.  

Micron labor tensions threaten DRAM production stability

Micron’s labor dispute in Taiwan has taken a significant turn. At the start of October, 1,994 of 2,258 registered members of its Taoyuan union (88.3% of total membership) voted to strike, securing the right to take industrial action. Reuters reports that 99% of the ballots cast were in favor of the strike.  

Micron’s Taichung plant has its own union, which is scheduled for a second mediation session on October 22. Should those negotiations fail, Micron could be facing another strike vote. Together, the two Taiwan unions represent more than 80% of Micron’s workforce in the country, with Taipei authorities describing Taiwan as its largest manufacturing base.

As noted by TrendForce, the Taoyuan facility primarily produces DRAM, exposing the memory supply chain to a major disruption if workers follow through with the strike. A stoppage that pulls engineers away from equipment maintenance and process monitoring could erode yields and operational stability before causing an outright shutdown. However, in a DRAM market with little slack, even a modest yield loss would reach buyers through tighter allocations and longer lead times. TrendForce adds that the consequences for DRAM production largely hinge on the duration of a strike if it occurs, its duration, employee participation, and Micron’s contingency measures.  

The unions are seeking a profit-sharing structure tied to 15% of Micron’s operating profit. Current demands include a one-time FY2026 bonus estimated at about 83 months of pay per Taiwan employee followed by quarterly distributions from FY2027 onward under a redesigned performance program.  

Union chairman Lin Che-jui argues existing performance rewards for employees don’t reflect the company’s financial performance. Moreover, workers have raised concerns about differences in performance bonuses and stock-based awards between corporate-level employees and those on the fab floor, citing limited transparency in how those awards are distributed. The union also points to Samsung and SK Hynix, which it argues already have reward systems similar to those being demanded.  

This development will be important to monitor in the days ahead with DRAM stability hanging in the balance. The Taoyuan union plans a demonstration on October 19 and has said further action could follow. If a surprise strike is set, procurement and engineering teams will have little time for reactionary planning.  

Sourceability helps customers prepare for unexpected supply constraints through global sourcing networks, access to alternative memory suppliers, and market intelligence that tracks product availability, pricing, and early lead-time movements that signal trouble.  

AI inference once again poised to consume more supply

The AI buildout is pivoting from training models to serving them, changing what is required of the hardware propping up data centers around the world. McKinsey projects that inference will account for roughly 60% of AI compute demand by 2030 with the other 40% dedicated to training.  

Under this continued-momentum scenario, total computing infrastructure demand will grow substantially, reaching 278 gigawatts in 2030 compared to approximately 120 gigawatts today. Inference grows about 36% annually in this scenario, outpacing the 30% growth rate for AI overall.  

Every organization that moves a model from training into production adds to the load, which is why experts believe inference will be the main driver of AI-related semiconductor demand through the end of the decade.  

Inference workloads rely on processors that can quickly access and move data, increasing the importance of HBM, SRAM, and other high-speed memory architectures. Different inference applications carry their own requirements for latency, memory capacity, and power, which will force operators to fine-tune their systems to match.  

Within a single AI request, the prompt-processing phase is compute-heavy while the generation phase is bandwidth-heavy. GPU-plus-HBM systems that dominate today will continue to serve a large share of workloads, but custom accelerators, SRAM-based designs, and multi-vendor pairings that split those phases across distinct silicon are gaining ground. So too is demand for cloud infrastructure and hardware tuned for edge environments that will handle a greater portion of AI processing. Software is equally important for lowering server costs on existing infrastructure.

As inference workloads become more complex, experts have warned that interconnect bandwidth will become a more significant factor. McKinsey adds that power delivery and thermal management are increasingly capping rack density well before compute does.  

Improving inference economics will require coordinated progress across the entire stack. McKinsey notes that cost to serve is set by how silicon, memory, networking, software, power, and cooling work together, and that many of these architectural decisions are expensive to reverse once made.  

For manufacturers and procurement teams, the shift toward AI inference means more varied component requirements. Each new architecture brings its own mix of memory, processors, networking, and power components, along with its own qualification process.  

Supplier dependencies that were once tracking a small group of GPU and memory vendors will spread across a broader field, and any one of those points can become a bottleneck. As these trends unfold, procurement teams will need to consider more than current component availability and pricing to ensure supply stability.

Sourceability helps customers prepare for the constraints of an inference-focused AI market through access to our global sourcing network and vetted suppliers across memory and adjacent components. Our Datalynq market intelligence tool gives your team early visibility into component availability, pricing, and lead time changes across the BOM, enabling informed pivots before inventory disappears.