BlockBeats News, June 26th. The Wall Street attention to the AI semiconductor chain is transitioning from NVIDIA-led GPU transactions to a broader infrastructure bottleneck—storage chips, server CPUs, advanced processes, packaging testing, and device materials.

Over the past two days, several broker reports have shown that AI capital expenditure has not slowed down but is instead spreading deeper into the supply chain. Nomura is bullish on Samsung Electronics and SK Hynix's storage cycle, Goldman Sachs stated that Micron's long-term customer agreements increase profit visibility, Bernstein believes Qualcomm's data center story is beginning to take shape, and Bank of America sees the 'agentic AI'-driven server CPU demand as a new source of tension in the Asia-Pacific semiconductor supply chain.

In a June 24th report titled "Global Memory," Nomura stated that the rising storage prices are expanding from HBM to a broader range of DRAM and NAND products. The bank expects that by the third quarter of 2026, commodity DRAM prices will rise by 24% QoQ, far higher than the previous expectation of about 5%; NAND prices are expected to increase by 25% QoQ. Nomura maintains a buy rating for Samsung and SK Hynix, with a target price of KRW 670,000 for Samsung and an upward revision of the target price for Hynix from KRW 4 million to KRW 4.7 million.

The key of this report is not just short-term price hikes but long-term agreements. Nomura stated that major memory manufacturers are negotiating long-term supply agreements with cloud service providers and customers like NVIDIA, with each supplier negotiating with roughly two to four customers. Agreement terms may include pricing, prepayments, and supply commitments, which may determine profit differentiation among different manufacturers in the future.

Goldman Sachs' latest report on Micron also points to the same trend. The bank stated that Micron's quarter performance was strong, with revenue, gross margin, and EPS significantly higher than market expectations; more importantly, the company disclosed that it has signed 16 strategic customer agreements covering data centers, consumer, automotive, and other fields. These agreements currently cover about 20% of the expected DRAM shipments and 30% of the expected NAND shipments, with the long-term goal to cover around 50% of revenue. At the base price, the agreements represent approximately $100 billion in committed revenue over the next five years, with gross margins higher than the previous cycle peak.

However, Goldman Sachs still maintains a Neutral rating on Micron with a target price of $1,100. Their reservation is that in 2027 and 2028, HBM supply may significantly increase, thereby dampening price momentum. This also reflects the current pricing logic of the AI chip chain: the demand remains strong, but the market is becoming more concerned about supply expansion and cycle tops.

Meanwhile, AI transactions are now moving towards CPUs. Bernstein, following Qualcomm's analyst day, raised its target price from $140 to $235 but maintained a Market-Perform rating. Qualcomm's revenue framework for the 2029 fiscal year is approximately $65 billion, with a target of $15 billion in data center revenue. The company plans to enter the AI infrastructure market with data center CPUs, ASICs, accelerators, and connectivity products.

Bernstein is not so bullish on this. The report states that Qualcomm's data center narrative is "promising," but its mobile business still faces pressure, including Apple's exit from revenue contribution, storage price hikes impacting Android phone demand and margin, as well as the cost increase from the company's continued investment in new businesses. In other words, Qualcomm has secured a ticket to the AI data center, but still needs to prove itself with orders and profits.

Bank of America's Asia-Pacific semiconductor report takes this line further. The bank believes that as AI shifts from training to inference, enterprise deployment, and agentic workflow, the role of CPUs will rise again. Server CPUs will no longer just schedule tasks for GPUs but will take on higher value in data preparation, memory access, model loading, agentic workflow, and system coordination.

Bank of America expects the global server CPU market size to grow from around $35 billion in 2025 to over $170 billion in 2030, with a compound annual growth rate of about 37%; by 2030, AI server CPUs may account for over 80% of the overall server CPU market. With AMD, NVIDIA, Qualcomm, and cloud vendors expanding their in-house ARM CPUs, more capacity will shift to outsourced manufacturing, further boosting demand for TSMC, advanced packaging, testing equipment, and material suppliers.

The bank has raised the target prices of several Asia-Pacific supply chain companies, including TSMC, ASE, Kinik, Chroma, Hon Precision, and GPTC. Bank of America states that the demand overlap for server CPUs, GPUs, and ASICs will require CoWoS suppliers to expand at a compound annual growth rate of around 52% from 2026 to 2028.

These reports collectively outline a new stage in AI semiconductor transactions: the first stage was dominated by GPU supply shortages, while the second stage involves the entire infrastructure chain being repriced. Storage manufacturers lock in cash flow through long-term agreements, CPU vendors compete for inference and agentic AI workloads, and TSMC and packaging and testing facilities become bottlenecks for advanced processes and backend capacity.

Risks are also accumulating. Pitfalls mentioned in brokerage reports include HBM supply expansion, AI capital expenditure slowing down in a high-interest-rate environment, data center construction constrained by power, mobile demand suppressed by storage price hikes, and disruptions in the commodity market from Chinese storage factory expansions.

But at least from these research reports, the current mainstream on Wall Street is clear: the AI transaction is not over; it has just moved from the "buy GPU leader" phase to the "find the next batch of supply chain bottlenecks" stage. Storage, server CPUs, advanced foundries, packaging testing, and equipment materials are becoming the core of a new round of AI capital expenditure narratives.