JP Morgan (Aug. 5) says AI chips (GPU, CPU, memory, storage devices) will remain
the primary bottleneck for AI compute infrastructure through most of 2026–27,
but semiconductor capacity is accelerating and the industry retains high
expansion flexibility with sufficient lead time. The bank warns that at some
point in the next 18–24 months, delays in datacenter deployment and limits on
power supply (grid and on-site/independent solutions) could supplant chips as
the key constraint on AI compute. Although datacenter rollout has so far
exceeded expectations and not impeded cloud AI deployment or token output, JP
Morgan flags power and infrastructure bottlenecks as the market variables to
watch in H2 2027 and 2028.
