Intel chief executive Lip-Bu Tan warned that the memory shortage battering PCs and smartphones could get even worse in 2027.
Speaking at the AI Infrastructure Summit in Santa Clara on 15 September, Tan said the industry had underestimated how serious the memory bottleneck would become.
“Early last year, I said that ‘memory would become a bottleneck.’ Many people did not realise it back then, but as a result, I was right,” Tan said.
According to DigiTimes Tan expects conditions to deteriorate further next year as AI infrastructure continues swallowing increasingly large quantities of memory. Capacity has failed to keep pace with demand, particularly as suppliers chase the lucrative server and accelerator markets.
Tan said memory production capacity remained severely constrained while prices had risen rapidly. “As you know, they have gone up five, six, or seven times,” Tan said.
That is starting to hurt manufacturers well beyond the giant AI server builders. Low-end smartphones and notebooks are particularly exposed because memory makes up a much larger share of their total component costs.
Tan said memory could now account for 70 to 80 per cent of the cost of some cheaper phones and laptops. At those levels, manufacturers have little room to absorb further increases without raising prices, changing specifications or abandoning products.
The squeeze is already forcing smaller device makers to rethink designs and hunt for whatever components they can secure. Some are placing orders far further in advance while others are testing parts more carefully as scarce supply creates an inviting market for counterfeit chips.
SK Hynix has been similarly gloomy about what comes next. Its chief executive warned in July that 2027 could be “the worst year in the industry’s history from the supply perspective,” with demand potentially exceeding available capacity beyond 2030.
AI has shifted the economics of the memory business because manufacturers can make considerably more money supplying high-bandwidth memory, server DDR5 and other data-centre products. Consumer hardware makers are left fighting for capacity after the more profitable customers have had their fill.
Tan said demand for CPUs used in AI inference systems has surged as processors are still needed to manage GPUs, applications and general-purpose workloads around increasingly large accelerator installations.
The problem extends beyond processors and memory. Tan pointed to electricity and cooling as further infrastructure constraints, mentioning air, liquid and microfluidic cooling as areas requiring investment as rack power densities continue climbing.
Chipzilla is consequently looking more closely at complete rack-level systems rather than individual processors. “It is important to look at the system rack, resolve the load, adjust to customer requirements and work with them,” Tan said.
