Key Takeaways

  • Samsung’s forecast reflects strong demand and limited supply for AI-related memory, including high-bandwidth memory used with Nvidia accelerators.
  • Investors remain concerned that hyperscalers may struggle to sustain infrastructure spending or earn adequate returns from AI services.
  • Memory scarcity benefits Samsung and Micron through stronger pricing, but it could increase server costs and delay enterprise deployments.

Samsung has forecast third-quarter 2026 operating profit of 107.4 trillion won, nearly nine times its year-earlier result, as demand for memory used in artificial intelligence infrastructure continues to outpace supply.

The scale of that forecast offers a striking view of the AI investment cycle from inside the semiconductor supply chain. Nvidia accelerators receive much of the attention, but those processors depend on high-bandwidth memory, storage and other components to operate inside AI servers. Samsung’s position in HBM and conventional memory puts it near the center of the resulting capacity crunch.

Yet financial markets are not treating soaring chip earnings as proof that the boom can continue indefinitely. Samsung shares have fallen more than 25% from their June peak, according to RTÉ, as investors weigh record earnings against questions about hyperscaler spending, future supply and the commercial returns generated by AI services.

Samsung benefits from current scarcity, while its customers face a tougher economic equation regarding future AI infrastructure returns.

Strong momentum emerged early in the year as Samsung’s semiconductor operating profit reached 53.7 trillion won in the first quarter, compared with 1.1 trillion won a year earlier. The company also indicated that customer demand pointed to a widening supply gap into 2027.

Memory is no longer a relatively interchangeable line item in a server budget. HBM sits next to accelerators in advanced AI systems and determines how quickly data can move through a workload. Limited availability can therefore constrain an entire deployment, even when customers have secured access to Nvidia processors.

Pricing power cuts both ways. Higher memory prices support Samsung’s earnings and can also benefit Micron, which competes in the expanding market for AI-server memory. For cloud providers and enterprises, however, the same conditions raise the cost of each cluster and complicate procurement schedules.

The broader spending figures help explain why supply remains tight. IDC estimates that global AI-infrastructure spending rose from $153 billion in 2024 to $318 billion in 2025 and forecasts $487 billion in 2026 (source). IDC has also warned that memory and storage scarcity could increase server costs and delay purchases.

This is not a small, isolated hardware cycle. Gartner forecasts worldwide AI spending of $2.67 trillion in 2026, an increase of 49.5% year over year (source). AI infrastructure alone is projected to account for $1.48 trillion, illustrating just how much of the market remains concentrated in capital-intensive computing capacity.

Still, can hyperscalers convert that infrastructure into revenue quickly enough to justify another round of spending? That is the issue hanging over Samsung’s share price. Training models, operating inference services and building data centers consume large amounts of capital, power and technical labor. Even strong demand for AI applications does not automatically produce attractive margins for every provider.

The uncertainty does not imply that infrastructure spending is about to stop. Large cloud operators often plan capacity over several years, while enterprises are still moving AI projects from experiments into production. Supply commitments can also reinforce the cycle because customers facing shortages may place orders earlier or reserve more capacity than they would in a balanced market.

For enterprise technology leaders, the practical response is disciplined procurement rather than a simple bet on whether an AI bubble exists. Organizations can separate model experimentation from production infrastructure, measure utilization and business outcomes, and avoid assuming that accelerator availability is the only constraint. Memory capacity, storage throughput, networking and power can each alter the economics of a deployment.

Samsung’s forecast shows that the AI supercycle remains powerful at the component level. The next test is less about demand in isolation and more about durability: whether customers can keep financing today’s buildout, put expensive systems to productive use and support high memory prices into 2027.