AI’s Hunger for Memory

The AI infrastructure discussion is often focused around GPUs that power AI model use and training. Nvidia is the market leader in GPUs with demand and market share that dominates other vendors in the market. However, without large amounts of memory to support ever-growing model size and complexity, these GPUs wouldn’t be able to perform at all. To satisfy these needs, companies like Samsung build AI memory chips for major customers, including Tesla and Nvidia.

Memory has become one of the main pressure points for AI’s growth. Memory shortages and tight supply are becoming the hidden tax inside the AI economy. Chipmakers are diverting capacity toward high-bandwidth memory for AI servers, which leaves thinner supply for conventional DRAM used in PCs, phones and other electronics. That shift gives memory producers unusual pricing power, but it squeezes device makers that built their business around steady component deflation.

Gartner has warned that rising memory costs will push buyers toward premium devices and depress lower-end PC and smartphone demand. In a recent report, Gartner said worldwide semiconductor revenue is forecast to exceed $1.3 trillion in 2026, up 64%, with memory revenue expected to triple amid “memflation.” As a result of this intense demand for memory, Gartner estimates that DRAM prices will rise 125% in 2026 and NAND flash prices 234%, with pricing relief not expected until late 2027.

Samsung has warned that the AI boom is worsening chip shortages and creating headwinds for its own smartphone and display units.AI is lifting semiconductor revenue, but it is also redistributing profit toward memory suppliers, raising input costs for consumer-electronics brands and making supply access a competitive factor.

The Critical Role of Labor in AI Chip Supply

Samsung’s labor dispute is happening in a global labor market already short of skilled chip workers. Deloitte has estimated that the semiconductor industry will need more than one million extra skilled workers by 2030. The U.S. faces its own version of the squeeze. The Semiconductor Industry Association has projected that the U.S. semiconductor workforce will grow by nearly 115,000 jobs by 2030, but about 67,000 of those new roles risk going unfilled at current degree completion rates. The gap is concentrated in technicians, engineers and computer scientists.

While AI aims to take a big chunk out of white collar and even many blue collar jobs, AI cannot automate its way out of the chip bottleneck. A chip fabrication plant requires significant amounts of capital and human craft. Operators, equipment engineers, process technicians, materials specialists and yield teams turn billion-dollar plants into reliable output.

Realizing the critical nature of chip fabrication, governments are spending billions to localize chip production. The U.S. CHIPS Act, European subsidies, Japan’s memory push and South Korea’s strategic focus all assume fabs can be staffed with talent even if the money becomes available. A subsidy can support construction and tools, but it cannot instantly produce thousands of experienced engineers.

As a result of this demand and pressure, what’s happening at Samsung is a warning sign for the industry. In a weaker chip cycle when there’s less demand, labor has less leverage. In an AI cycle with scarce supply, tight capacity and impatient customers, skilled workers are a critical resource in the same category as lithography tools, substrates, cleanrooms and power contracts. The human factor is still incredibly important, and perhaps one of the more notable spots where people have real leverage.

As a result, Samsung’s union is pressing for a clearer bonus formula after memory-chip workers were offered far higher bonuses than colleagues in logic and foundry operations. The dispute has also been tied to retention, with discontent feeding concerns that talent may leave for SK Hynix, Samsung’s aggressive memory rival.

Samsung’s looming strike should force customers and investors to ask a new set of questions. Which companies have enough supply to be able to meet continuing growth in demand for increasingly more memory-intense models? Which suppliers are exposed to labor issues that could put a halt to data center growth? Which countries can train enough technicians before subsidy-backed fabs open? Which customers have enough supplier diversity if one supplier can’t meet its obligations? The AI market is moving beyond lab breakthroughs and model benchmarks, and is starting to behave like critical infrastructure, with choke points in labor, memory, power and manufacturing.