New cracks in the AI storage boom: Customers are looking for ways to get rid of memory dependency

Memory chips are experiencing the strongest boom cycle in history, but the same force driving the industry's record-breaking profits is also nurturing the next wave of technological change.

Micron Technology's latest earnings report shows that memory chip supply has tightened further compared to three months ago. The company expects its operating profit for the current quarter to exceed the highest revenue of any full fiscal year in its history, and it has signed long-term supply agreements with 15 new customers, extending the expected duration of supply tightness beyond 2027.

However, the continued rise in memory prices is reshaping the incentive structure of the AI supply chain. As HBM becomes an increasingly costly component in AI systems, major customers like Qualcomm, Nvidia, and Cerebras have begun exploring new architectures and algorithms to reduce reliance on HBM.

This means that the high prices underpinning the current memory boom could, in the long run, become a catalyst for innovation on the demand side.

The more expensive memory becomes, the greater the incentive for customers to use less of it

The supply-demand imbalance is continuously strengthening memory manufacturers' pricing power, while forcing downstream customers to accept higher costs.

Apple on Thursday unexpectedly raised prices on several Mac and iPad models between product cycles, explicitly attributing the move to rising memory chip costs.

Micron, meanwhile, stated that most of the newly signed long-term agreements span five years and include price floors, ensuring that even if the industry enters a downturn, selling prices will remain significantly above previous troughs. Analyst Rolf Bulk of Futurum noted that even at the contract's minimum prices, profitability would still exceed previous peaks in the industry cycle.

Although historically, long-term chip supply agreements have not always been strictly enforced, the market currently believes that supply will remain difficult to improve significantly over the next two to three years, leaving customers with little room for price negotiation in the short term.

However, it is precisely because supply remains tight and prices keep rising that more tech companies are beginning to consider an alternative solution—not securing more HBM, but minimizing its use.

Tech giants begin exploring "memory reduction" strategies

Qualcomm highlighted its "High-Bandwidth Compute" architecture at this week's investor conference, aiming to reduce reliance on HBM in AI computation through new system designs. There have also been reports that Nvidia is adjusting some designs for its next-generation Vera Rubin platform to lower overall memory requirements.

AI chip company Cerebras, meanwhile, has made "not using HBM" a core selling point. In its first post-IPO earnings report this week, the company stated that its wafer-scale chip is completely free from HBM constraints. CEO Andrew Feldman bluntly said: "HBM supply is tight and expensive—and we don't use it at all."

At the same time, algorithmic optimization is advancing. In March this year, Google released TurboQuant research, which uses new model compression methods to significantly reduce AI model memory usage with almost no performance impact. At that time, Micron's stock price plummeted by nearly one-third, though it later recovered more than twofold as the market reassessed. However, this event showed that the market is highly sensitive to any technological breakthrough that could weaken HBM demand.

Prosperity is breeding the next round of competition

In the short to medium term, HBM remains an irreplaceable key component for AI computing, and the supply tightness that Micron sees has solid fundamental support.

But over a longer industrial cycle, price is often the most powerful driver of innovation.

When the world's largest AI customers are willing to sign five-year long-term supply agreements—unprecedented in the past—it also means they have a sufficiently long time window and strong economic incentives to invest in new chip architectures, software algorithms, and system designs to accomplish more computing with less memory.

For the memory industry, today's strongest competitive advantages—tight supply, high prices, and unprecedented pricing power—both solidify this super-cycle and may become the biggest force driving customers to reduce reliance on HBM. Historically, every sustained rise in core hardware prices has spurred new technological paths; this time, the memory industry's own boom may be the starting point for the next round of competition.

Risk Warning and Disclaimer

        Markets are risky; investment requires caution. This article does not constitute personal investment advice, nor does it consider the specific investment objectives, financial situations, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this content is at your own risk.