On a day when AI-related shares were sliding for a second straight session, Bank of America’s analysts were moving their numbers in the other direction. On September 15, the bank raised its forecast for US semiconductor industry growth, lifting its expected compound annual growth rate for 2026 through 2030 from 14 percent to 18 percent. The note landed in a market that could not decide whether the AI trade was slowing down or still speeding up.
The timing is striking. The ten-year Treasury yield touched 5 percent for the first time since 2007, and the Nasdaq and the S&P 500 drifted lower ahead of a Federal Reserve meeting. The selloff in AI names has been framed in some quarters as a reckoning with the pace of capital spending. BofA’s forecast is a direct counter to that framing, delivered in numbers rather than argument.
The bank also pushed its estimate for the total addressable market higher, from $2.7 trillion to $3.2 trillion by 2030. The revision rests on brighter expectations for memory chips and server-related components, the parts of the supply chain that have become the tightest bottlenecks in the data center buildout. BofA said the industry took roughly fifty years to reach $1 trillion in annual sales, and that a $1.7 trillion market could now double within four years.
The equipment numbers are where the forecast gets concrete. The bank expects wafer fabrication equipment spending of $156 billion in 2026, up 33 percent from the prior year, rising to $210 billion in 2027, another 34 percent gain. Within that, DRAM equipment is projected to reach $61 billion in 2026 and $85 billion in 2027. Those figures describe an industry still buying the machines to make the chips, which is a bet that the chips themselves will keep selling.
The memory component of the forecast is driven by the same dynamics reshaping the whole industry. High-bandwidth memory has become a prerequisite for the accelerators that power AI training and inference, and its makers have been expanding capacity as fast as their suppliers will allow. BofA’s equipment numbers capture that expansion, since every new memory line requires a matching investment in fabrication tools.
The revision matters because it comes from a bank whose semiconductor research is followed across Wall Street. When BofA raises its growth outlook in the middle of a two-day selloff, it is effectively arguing that the market’s anxiety about spending is a repricing of valuations, not a signal that orders are drying up. Analysts at other firms have made a similar case this week, even as memory and equipment makers led the declines.
The disconnect between stock prices and order books is the story of the moment. Memory makers and the companies that build their equipment remain sold out for quarters ahead, according to people familiar with the industry. No major customer has publicly cut an order. What has changed is the mood, as a debate about the pace of frontier AI development has spilled into the prices of the companies that supply it.
The bank offered one data point that cuts both ways. Nvidia’s B200 accelerator rents for $5.72 per hour, according to BofA, down less than 10 percent from its March peak. The fact that the price has held so close to its high suggests demand for compute has not collapsed. The fact that it has fallen at all is the kind of detail the bears are now pointing to.
The forecast raises the question of who pays for all of it. The AI buildout is financed by a small number of companies whose willingness to keep spending depends on their own revenue trajectories. If those companies slow their purchases, the entire chain from fabs to equipment makers feels it. BofA’s answer, in effect, is that the spending has not slowed and shows no sign of slowing soon.
Analysts cautioned that forecasts this large carry their own risk. A forecast of $3.2 trillion by 2030 assumes current customers keep spending, and that the next generation of applications arrives to justify it. The industry has grown by betting ahead of demand before, and it has been wrong before. The difference now is the size of the numbers involved.
What the note does not say is when the spending becomes too much. The debate that rattled the market this week is not about whether AI demand is real. It is about the pace, and about whether the labs and cloud providers racing to build should slow down before the bill comes due. BofA’s forecast is one answer to that question, delivered in the form of a spreadsheet rather than an essay.
For now, the two stories run in parallel. The stocks are repricing, and the order books are full. Whether the repricing eventually reaches the order books is the question that will decide which story turns out to be right. The bank has placed its bet on the order books.
