Bank of America: Chip stocks pull back as AI spending forecasts keep rising: BofA sees room left in the cycle

What happened to customer spending as chip stocks pulled back?

The semiconductor index rose 100% in the first half of 2026, then retreated about 20%–25% from its peak; meanwhile, BofA continued to raise its spending forecasts for major hyperscalers and emerging cloud providers.
BofA expects these customers to spend about 1 trillion dollars in 2026, nearly double the previous year, and 1.35 trillion to 1.4 trillion dollars in 2027, up about 35%–40% year over year. Its forecasts for both years are about 20% higher than three months ago.
BofA believes AI is helping major cloud providers attract customers and launch new products, while growth in core businesses such as search and e-commerce is also accelerating at some providers, strengthening their ability and willingness to keep investing.

Emerging cloud providers are taking on some spillover demand from major cloud providers, which need computing capacity but may lack readily available land, power and facilities or be unable to bring capacity online quickly enough.
But these providers must both generate revenue from the computing capacity they have built and raise funds for the next round of expansion. Some emerging cloud providers' bonds already yield around 15%; as the cost of capital rises, the tension between the pace of monetization and the pace of expansion will also increase.

Memory chip prices and margins have historically been sensitive to additions of capacity. BofA notes that new supply from China and manufacturers' current capital spending could create supply pressure when demand slows.
BofA believes the difference this time is the high-bandwidth memory needed for AI: producing the same amount of capacity requires about 3–4 times as much production capacity, so the supply-demand gap could persist for some time even if demand growth slows or supply increases.

Prices for AI models or tokens are falling. BofA notes that a new generation of GPUs can produce 5–10 times as many tokens as the previous generation using the same amount of power; greater computing efficiency can partly account for lower prices. More advanced models can also be offered to enterprise customers willing to pay.
Open-weight models also require computing infrastructure. For on-device AI, BofA expects simple tasks to be handled on devices such as phones, while complex inference will still depend on the cloud; more local processing will also consume battery life.

BofA believes the semiconductor industry is closer to the middle of the cycle, while some valuations already reflect concerns about its later stages. The bank points to the buildout of 3G and 4G infrastructure: the cycle lasted about 10 years, but most of the gains in related stocks were concentrated in the first five years.
The bank believes the growth drivers for 2027 are relatively clear, while the spending outlook for 2028 is the bigger question. If cloud spending and infrastructure utilization continue to develop favorably, the cycle could last at least another two years.