Don't believe the narratives that "AI capital expenditure can grow infinitely" and "there is ten-year demand visibility".
Chip analyst P Equity believes that hyperscale cloud vendors are simply unable to accurately predict the situation two years from now, let alone ten years. AI will of course grow in the long run, but any technology will eventually enter a mature stage, and capital expenditure will sooner or later shift from rapid growth to a plateau. Therefore, memory remains a cyclical industry. It's just that this cycle, due to the existence of long-term agreements (LTA), prepayments and price floors, the downward trend may be more moderate than in the past.
But what is truly remarkable in the short and medium term is the memory sector. AI data centers are gradually shifting from training to inference, and inference puts higher requirements on Memory Bandwidth, model weights, KV Cache, and storage for long-term Agent tasks. Different institutions have very different forecasts for the proportion of memory in Hyperscaler Capex, but their common conclusion is: memory has become one of the largest cost items in AI infrastructure. The document even cites UBS estimates that next year, standalone memory spending may reach 900 billion US dollars.
This means that the next stage of the AI industrial chain cannot only focus on HBM. HBM is still one of the core memory products with the highest value content and the highest profit margin, but the popularization of Agent may drive the demand for HBM, DDR5 and NAND at the same time. Especially for long-running Agents, more context, status, data and intermediate results need to be stored, and some data does not need to be kept entirely in the expensive HBM, which can be offloaded to NAND. Therefore, the outbreak of Agent essentially means the rising demand of the entire Memory Hierarchy, rather than a simple increase in HBM demand.
US Stock Investment Network learned that the second very important change is: GPU may no longer be the biggest physical bottleneck. The rental price of old GPUs on the market is still very high, and even H100s have been extended to eight or nine years of use, which shows that the demand for computing power is indeed very strong. However, relevant experts from AMD judge that from the perspective of bare chip supply alone, the shortage may not be as severe as imagined. What really restricts GPU deployment is power, advanced packaging and memory. In other words, to judge AI Capex in the future, we cannot only look at how many GPUs NVDA can produce, but also see whether these GPUs have access to power, memory, packaging and racks to actually run.
The third most noteworthy hidden bottleneck is the ABF substrate. In the BOM of Vera Rubin, several links that have seen triple-digit price increases include ABF substrates, PCBs and memory. Mr. P judges that the ABF shortage may last until after 2028, and some even believe it will last until 2030. Dell, HP, NVIDIA and Broadcom have also mentioned DRAM, NAND, ABF substrates and wafer constraints in recent supply chain discussions. This logic is very important, because the market has fully priced in GPUs and HBM, but ABF may still be a link with relatively low attention, yet it directly determines whether high-end AI chips can complete packaging and shipment.
The fourth long-term trend is "copper retreats and optical advances", but the speed will be slower than the market expects. Optical communication will definitely win in the end, because the scale of AI clusters is getting larger and larger, and the importance of speed, bandwidth and power consumption is getting higher and higher. Copper will face problems such as crosstalk and heat dissipation under high-speed transmission. However, CPO still has problems in cost, yield and heat dissipation at present, so copper and optical will most likely continue to coexist before 2027. NPO may be seen first in 2027, CPO will start to ramp up from 2028 to 2029, and the large-scale dominance of optical technology may not come until after 2030. In other words, copper will not be eliminated immediately, but will be "prolonged in service forcibly".
Why do cloud vendors try their best to delay CPO deployment?
A very key reason is still memory. Memory has become so expensive that it accounts for an increasing proportion of BOM and capital expenditure. Hyperscalers must reduce costs in other links. If copper can still meet the demand, they will try to use it for a few more years instead of fully upgrading to the expensive CPO immediately. This logic shows that not all components in the future AI server industrial chain will see price increases, but a "budget grabbing" will occur within capital expenditure: the more money memory takes up, the more other components need to reduce costs.
The fifth real big bottleneck may be power supply. A data center cannot run immediately after the servers are bought. It requires complete infrastructure such as land, power supply, shell, grid connection, and gas turbines. At present, the backlog of orders for gas turbine suppliers such as Mitsubishi, Siemens, and GE Vernova has been scheduled to after 2030. If you order a gas turbine today, it may not be delivered until 2030. Therefore, the growth rate of AI computing power in the next few years is most likely not determined by GPU production speed, but by "how many GW of power can be actually connected to the grid".
US Stock Investment Network believes that the most valuable sentence for investment is that AI infrastructure is transitioning from the "GPU shortage era" to the "system bottleneck era".
That means in the future, we can no longer simply interpret it as "AI demand growth equals NVDA selling more GPUs", but split it into four main lines: Logic, Memory, Power, and Networking. Mr. P has also clearly categorized AI infrastructure into these four major areas.
The strongest certainty in the short and medium term is still Memory, especially HBM, DRAM, NAND $SNDK, related targets $MU Micron, SK Hynix $SKHY; followed by ABF substrates, PCBs and advanced packaging that are relatively ignored by the market; then comes network upgrades, especially high-speed optical communications, lasers and the final CPO. Although power may be the biggest physical bottleneck, for some assets such as gas turbines, since orders have been scheduled to after 2030, the marginal growth brought by new orders may not be as attractive as it was two years ago. This difference is very important: the most scarce thing in the industry is not necessarily the most worthy of chasing in the stock market right now.
Finally, there is an easily overlooked risk point: LTA has not completely eliminated the memory cycle. Even if the contract is written for five or even ten years, once AI capital expenditure drops significantly, suppliers may not dare to force customers to accept the goods. Otherwise, customers will only hoard DRAM in warehouses and stop purchasing in the next few years. Therefore, LTA is more like "reducing cyclical fluctuations" rather than "eliminating cycles".
According to US stock big data StockWe.com, the biggest Alpha in the next stage of AI may no longer come from the most noticeable GPUs, but from the Memory, ABF substrates, advanced packaging, power and high-speed networks behind GPUs that "cannot boot the device without them".
This article is from WeChat official account "US Stock Investment Network" (ID: tradesmax), author: StockWe.com, published with authorization from 36Kr.