Bank of America Resets Micron Stock Forecast as AI 'Memory Tax' Rises

The analyst sat in a room with a spreadsheet and a problem. The problem had a name. Vivek Arya at Bank of America called it the "memory tax."

Here is what that means in plain language. Every time a company builds an AI data center, a chunk of the budget goes to memory chips. Not the processors. Not the networking gear. The memory. High-bandwidth memory, the stuff that feeds the accelerators, the stuff that stores the weights and the context windows and the endless streams of tokens. That chunk now takes up roughly 35% of all AI capital spending.

Thirty-five percent. For memory.

A decade ago, memory was the commodity you bought at the bottom of the cycle and forgot about. Now it is a line item that makes chief financial officers wince.

Bank of America looked at this and did something counterintuitive. The bank raised its forecasts. Not lowered them. Raised them. Fiscal 2027 revenue estimate: $275.4 billion, up from $230.3 billion. Fiscal 2028: $317 billion, up from $244.1 billion. Earnings per share for 2027 jumped to $171.78 from $140.24. For 2028, $197.90 from $145.45.

The price target stayed at $1,550.

What the memory tax actually is

You need to understand the physics before you understand the finance.

High-bandwidth memory requires more silicon wafers to produce the same capacity as conventional DRAM. That is not a design flaw. That is the technology. HBM stacks memory dies vertically and connects them with thousands of tiny wires, and the manufacturing tolerances are brutal.

Micron, Samsung, and SK Hynix, the three companies that make the vast majority of the world's memory, looked at their fab capacity and made a decision. They would rather build HBM than the DDR5 that goes into your laptop. The margins are better. The demand is louder. The customers, Nvidia, the hyperscalers, the AI labs, sign contracts that stretch years into the future.

So the wafers went to AI. And the prices for everything else went up.

Conventional DRAM contract prices climbed 93% to 98% in the first quarter of 2026. Then another 58% to 63% in the second quarter. NAND flash prices rose 70% to 75%. A mainstream PC memory kit that cost $75 a year ago now sells for as much as $460.

That is the memory tax. You pay it at Best Buy. The data center pays it in the capital budget. Micron collects it.

The tax is not a fee anyone voted on. It is an economic consequence. AI memory generates higher returns per wafer. Manufacturers follow the margins. No coordination required.

Why Bank of America raised its numbers anyway

Arya and his team did not raise their forecasts because memory prices are high. High prices invite competition. High prices invite capacity. High prices historically destroy themselves.

They raised the forecasts because Micron changed the structure of its business.

The company now has 26 strategic customer agreements. These are take-or-pay deals. The customer either takes the memory or pays for it anyway. The contracts include price floors and price ceilings. They run from 2026 through 2030. They cover an estimated 35% of Micron's revenue. More than 75% of fiscal 2027 output is already committed.

This is not how memory companies used to operate. The old model was spot pricing and quarterly negotiations and wild swings from shortage to glut to shortage again. The new model looks, and Arya does not use this word, but the numbers imply it, boring. Predictable. The kind of business a pension fund can own without checking the quote every morning.

Bank of America's data shows Micron's AI chip shipments for 2026 have already more than doubled since 2024.

The cash flow nobody expected from a memory company

Micron reported $54.2 billion in revenue for fiscal Q4. Gross margin: 87%. Operating cash flow: $44 billion. Free cash flow: $33.2 billion. The company ended the quarter with $73.5 billion in cash and investments and a net cash position of $68.3 billion.

Read those numbers again. This is a memory company. Memory companies are supposed to be capital-intensive, cyclical, and perpetually one bad quarter away from an earnings miss. Micron just printed a quarter that would make a software company jealous.

The cash opens a door. CHIPS Act restrictions on buybacks expire around December 2026. Bank of America expects that expiration to give Micron flexibility for shareholder returns. The current buyback authorization is $2.2 billion. Analysts expect a larger one. UBS estimated Micron could repurchase roughly 30% of the company over two years.

Micron's CEO says there is no line of sight

Sanjay Mehrotra did not hedge. The CEO said Micron has "no line of sight" to when supply and demand return to balance. He expects memory conditions in 2027 and 2028 to be tighter than in 2026.

The reasons are technical. Larger AI models. Longer context windows. Greater concurrency. Every one of those trends consumes more memory. Every one of them pushes demand further out on the curve.

Then there are the robots.

Micron now believes each humanoid robot will require around 200 GB of DRAM and multiple terabytes of NAND. Morgan Stanley estimates humanoid robot shipments could reach 1.2 million units per year by 2030 and 930 million units per year by around 2050. At the high end of the 2030 range, that is 240 million GB of DRAM demand from robots alone. And that is before the robots start making more robots.

Even if the AI bubble pops tomorrow, Micron has attached itself to the next demand curve.

What could go wrong

Bank of America flagged three risks in its own note.

The first is HBM de-spec. If AI chip designers decide they do not need the most advanced memory for every workload, the pricing premium for HBM could compress. Micron's margins would feel it.

The second is Chinese memory suppliers. If CXMT or YMTC scale up and start shipping competitive DRAM, the supply picture changes. The memory tax becomes a memory discount. Micron's take-or-pay agreements protect revenue, but they do not protect market share.

The third risk is the one every cyclical investor knows by heart. Memory prices and margins will normalize. Bank of America's own analyst acknowledged this. New capacity arrives from mid-2027 through 2028. The question is not whether normalization happens. The question is how much of the current pricing power is structural and how much is a cycle wearing a nice suit.

Arya's answer: the take-or-pay agreements, the price floors, the five-year duration, those things are structural. The spot pricing spikes are cyclical. Micron now has more of the former and less exposure to the latter.

The stock market decides

Micron shares fell roughly 4% after earnings before reversing to finish about 3% higher. The market is still arguing with itself about what this company is now. A cyclical memory maker trading on commodity multiples? Or a critical AI infrastructure supplier with contracted revenue and expanding margins?

Bank of America says the latter. The bank called Micron one of its top AI picks. The average analyst price target sits around $1,521. Rosenblatt raised its target to $1,900. D.A. Davidson went to $2,100.

The memory tax is rising. Bank of America is not running from it. The bank is pointing at it and saying: that is the business now. Get used to it.