Wall St's AI Reload: Hyperscaler Capex Hits $1T, but Memory and Drives May Be the Cleaner Play

Hyperscaler spending is still rising, but the market wants a less crowded AI exposure

This is a rotation call, not an AI-end call. Bank of America's July survey found 82% viewed semiconductors as the market's most crowded trade, after chip-focused funds drew record $10 billion in net inflows through May. That is the warning sign. When the broad semiconductor trade gets this packed, the more interesting move is not abandoning AI altogether. It is shifting toward tighter bottlenecks where pricing power and demand can still outrun the crowd.

The bill is still getting bigger

The spend story remains intact. Five hyperscalers guided to $750 billion for 2026, that pace is set to cross $1 trillion next year, and Moody's estimated more than $3 trillion in five-year capital investment across the same group. That is not a wind-down in the AI buildout. It is a larger bill, and a bigger bill changes where investors want exposure: away from the most crowded names and toward the suppliers that can absorb more spending more directly.

Why the market is drifting physical

That is why memory and storage look cleaner now. Reuters tied next year's $1.1 trillion in Big Tech AI spending directly to demand for memory chips, while broader infrastructure commentary places storage alongside power, cooling, and storage required to keep racks humming. If you want AI-infrastructure exposure outside the most crowded trades, the physical layer looks increasingly important.

Memory and storage are becoming central to AI infrastructure

The rotation argument only works if you understand the bottleneck. AI storage is no longer a side category. Data centers now consume over 50% of DRAM and NAND bit TAM for the first time, which changes the economics for memory and storage suppliers in a direct way.

Why storage is the choke point

More compute does not help if the system starves for data movement and data persistence. Training and inference need fast memory, but they also need storage that can handle models, checkpoints, and datasets at scale. That is why the current AI-driven storage demand is being framed as a supercycle rather than a normal cycle. The pattern looks more like a step-change tied to AI workloads than gradual consumer or enterprise upgrade demand.

This matters because capex is still moving in the right direction. Even before the latest round of guidance, outside analysts were modeling about 75% of hyperscaler capex in 2026 going to AI infrastructure, representing roughly $450 billion in AI-related spending. If that spending continues to run into memory and storage constraints, the suppliers sitting between compute and capacity are well placed to benefit.

Storage is tiered, so the opportunity is not just "buy memory"

Bulls do not need every byte of AI data to live in flash. They need the tiering model to hold. And the evidence supports that. HDDs still provide cost-efficient capacity for warm and cold workloads, while SSDs are better suited to latency-sensitive access. AI systems therefore do not need one replacement technology; they need a balanced storage stack.

That is why the cleaner exposure is not just "buy memory." It is owning the companies that monetize different layers of the stack.

Micron, Seagate, Western Digital, and SanDisk each solve a different part of the bottleneck

Micron gives investors the fast-memory side of the bottleneck. The source material highlights a shortage of memory and storage chips from hyperscalers, which supports the view that AI memory demand is translating into stronger pricing conditions.

Seagate, Western Digital, and SanDisk offer the capacity and flash angles. Seagate and Western Digital have been among the strongest performers in this group this year, with Seagate up 65% year-to-date, Western Digital up 77%, and SanDisk more than tripling. That outperformance suggests the market is treating these names as part of the core AI storage trade, not as an afterthought.

The economics matter more than the headline

The key number is a 5x–10x $/TB premium for SSDs versus HDDs. That premium helps explain why storage is not one trade. It supports a tiered model in which HDDs remain important for scalable capacity, while SSDs capture more value in performance-sensitive layers. The setup, then, is straightforward: Micron leans into the high-speed memory squeeze, while Seagate, Western Digital, and SanDisk offer exposure to capacity, tiering, and flash growth. The main risk is that supply additions cool pricing faster than expected.

The trade works if spending stays directional

The bear case is slowing growth, not dead AI demand

The bear case is credible. UBS estimates hyperscaler capex growth slowing to 25% next year and 6% in 2028. If that deceleration arrives before memory and storage suppliers can improve mix or raise prices, the trade starts to look late-cycle rather than supercycle. The issue is not AI demand disappearing. It is marginal spending weakening enough to pressure pricing power.

Why bulls still have the better current evidence

In Deloitte's 10-day update after hyperscaler guidance updates, top hyperscaler capex plans rose to roughly $670 billion, implying about $960 billion in total AI data-center spend. That does not look like a theme that investors are abandoning.

The important distinction is slope versus direction. Hyperscaler AI capex is still rising and still expected to cross $1 trillion next year. If the bill keeps getting bigger, memory and storage can still rerate even if headline multiple expansion cools. What would actually break the trade is not slower growth by itself. It would be weaker supplier pricing or hyperscalers showing they can expand AI capacity without absorbing proportional memory and storage volumes.

What to watch next

Watch the filings and supplier signals. If the biggest spenders keep putting capital behind AI buildouts, the storage lane remains a live trend trade. If not, it starts to look more like a timing trade driven by commodity cycles.