OCP Conference Focus: Manufacturing and packaging have already expanded significantly, and the bottleneck for AI chips has shifted downstream, including memory, racks, power supply, etc.
```
Author: Bao Yilong
Source: Hard AI
The AI semiconductor industry is expected to witness another strong year of growth in 2026, but the investment logic in the AI hardware sector is undergoing a profound change.
On October 20, a Morgan Stanley research report pointed out that over the past two years, the market has been focused on upstream capacity bottlenecks such as TSMC's CoWoS packaging and advanced manufacturing processes.
However, according to the latest statements from NVIDIA and TSMC, as well as signals from the 2025 OCP conference, this situation has changed. The bottlenecks in chip manufacturing and packaging have been addressed through large-scale capacity expansions, and are no longer the core constraints for AI development.
Morgan Stanley emphasizes that the real bottleneck is shifting to downstream, focusing on supporting infrastructure such as data center space, power supply, liquid cooling, high-bandwidth memory (HBM), server racks, and optical modules.
The report believes this means that investment opportunities are spreading from upstream wafer fabrication and packaging to the broader downstream supply chain. In the future, data centers that cannot secure sufficient power and physical space will fall behind in the AI computing power race.
Upstream capacity is no longer the sole bottleneck; chip manufacturing and packaging have greatly expanded
There was a time when the market was filled with doubts about whether AI chips could be supplied in sufficient quantities, and TSMC's CoWoS advanced packaging capability was seen as a key bottleneck. However, the latest industry developments show that this situation has improved significantly.
TSMC revealed at its recent earnings call that "AI demand is even stronger than we imagined three months ago," and the company is striving to "narrow the supply-demand gap." Importantly, TSMC stated that the lead time for expanding CoWoS capacity is only six months, which provides tremendous flexibility for the supply side.
Although its advanced node wafer front-end capacity, such as 4nm and 3nm, remains tight, AI semiconductors clearly have higher priority than crypto ASICs or Android smartphone SoCs.
NVIDIA CEO Jensen Huang also made it clear in a recent conversation that semiconductor capacity is no longer the limiting factor it once was. After the demand surge experienced a few years ago, the entire supply chain's manufacturing and packaging segments have "greatly expanded," and the company is confident in meeting client demand.
Overall, although total demand continues to grow rapidly, the report predicts that global CoWoS demand will reach 1.154 million wafers in 2026, a year-on-year increase of 70%, yet the supply side's response capability has improved significantly.
Bottlenecks shift, downstream infrastructure becomes the new challenge
When chip supply is no longer the biggest issue, the bottleneck naturally shifts downstream.
NVIDIA pointed out that the current greater constraints come from the availability of data center space, power, and supporting infrastructure, which have much longer build cycles than chip manufacturing.
The contents of the OCP conference also confirm this trend. As AI cluster sizes move toward "hundred-thousand-scale GPUs," the entire concept of data center design is being reshaped:
Power and Cooling:The deployment of large-scale GPU clusters means massive electricity consumption and cooling challenges. At the OCP conference, liquid cooling has become the default configuration for new AI racks, and demand for power supply solutions such as HVDC 800V is also growing.Companies like Aspeed thus benefit, as their BMCs (baseboard management controllers) are used not only in servers but in various devices including those for cooling.
Storage and Memory:AI workloads place extreme demands on data storage capacity and access speed. Meta has made it clear that for cost reasons, its data centers will prioritize using QLC NAND flash. Meanwhile, Seagate mentioned that HDDs (mechanical hard drives) will still provide 95% of online capacity to serve large and remote data centers.More critically, demand for HBM (high-bandwidth memory) is exploding. The report forecasts that by 2026, global HBM consumption will reach 26 billion GB, with NVIDIA alone consuming 54% of that. Such highly concentrated and strong demand makes HBM supply a key variable influencing AI server shipments.
Racks and Networking:To enable ultra-large-scale deployment, OCP has released standards such as "AI Open Data Center" and "AI Open Cluster Design," covering racks, liquid cooling, power interfaces, and more.In networking, Alibaba said that pluggable optics remain the first choice because of their total cost of ownership and flexibility, while LPO (linear-drive pluggable optics) technology is also drawing attention.CPO/NPO (co-packaged/near-packaged optics) is expected to be implemented around 2028 as manufacturing processes mature.
Demand forecasts point to explosive growth for downstream components
The importance of downstream infrastructure can be verified using demand projection data.
According to Morgan Stanley analysts, global cloud service capital expenditures are expected to increase 31% year-on-year in 2026 to $582 billion, far exceeding the market's general forecast of 16%.
Furthermore, if the share of AI servers in capital expenditures increases, this means that capital expenditures for AI servers in 2026 may see year-on-year growth of about 70%.
On the demand side, major AI giants are still aggressively "stockpiling." The report breaks down AI chip demand in 2026 as follows:
CoWoS capacity consumption: NVIDIA is expected to occupy 59% of the share, followed by Broadcom (18%), AMD (9%), and AWS (6%).AI compute wafer consumption: NVIDIA leads far ahead with 55% of the share, followed by Google (22%), AMD (6%), and AWS (6%).
In summary, the signals from the OCP conference, cross-verified with industry chain data, clearly indicate a new direction for AI hardware investment. As capacity bottlenecks in chip manufacturing and packaging progressively ease, the market’s focus will inevitably shift to the infrastructure supporting ultra-large-scale AI computing.
The report suggests that for investors, this means looking beyond single chip companies, and expanding vision to the entire data center ecosystem, seeking out the "picks and shovels" with core competitiveness in downstream segments such as power, cooling, storage, memory, and networking.
This article comes from WeChat public account “Hard AI”. For more AI frontier information, click here
Risk Warning and DisclaimerThe market involves risks, and investments should be made with caution. This article does not constitute personal investment advice, nor does it take into account the special investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their particular situation. Investments based on this article are at your own risk. ```