OCP Conference Focus: Manufacturing and packaging have significantly expanded production, AI chip bottlenecks have shifted downstream, including memory, racks, power, etc

Author: Bao Yilong

Source: Hard AI

The AI semiconductor industry will usher in another strong growth year in 2026, but the investment logic in the AI hardware field is undergoing profound changes.

On October 20, Morgan Stanley's research report pointed out that for the past two years, the market focus has been on upstream capacity bottlenecks such as TSMC's CoWoS packaging and advanced processes.

However, according to the latest statements from NVIDIA and TSMC, as well as signals released at the 2025 OCP conference, this situation has changed. The core contradiction restricting AI development is no longer in chip manufacturing and packaging, which have been significantly expanded through large-scale capacity increases.

Morgan Stanley emphasized that the real bottleneck is shifting 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 investment opportunities are spreading from upstream wafer foundry and packaging to a 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 only bottleneck; chip manufacturing and packaging have significantly expanded

Once upon a time, 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 dynamics show that this situation has greatly improved.

TSMC revealed in its recent earnings call that "AI demand is stronger than we imagined three months ago," and the company is working to "narrow the supply-demand gap." Importantly, TSMC stated that the lead time for expanding CoWoS capacity is only six months, which provides great flexibility for the supply side.

Although the front-end capacity of advanced nodes such as 4nm and 3nm remains tight, AI semiconductors clearly have a higher priority than cryptocurrency ASICs or Android smartphone SoCs.

NVIDIA CEO Jensen Huang also recently stated that semiconductor capacity is no longer a limiting factor as it once was. After experiencing a surge in demand a few years ago, the entire supply chain's manufacturing and packaging segments have "achieved significant expansion," and the company is confident in meeting customer demand.

Overall, while total demand continues to grow rapidly, the report predicts that global CoWoS total demand will reach 1.154 million wafers in 2026, a year-on-year increase of 70%, but the rapid responsiveness of the supply side has significantly improved.

Bottlenecks shift; downstream infrastructure becomes a new challenge

When chip supply is no longer the biggest challenge, bottlenecks naturally shift downstream.

NVIDIA pointed out that the current greater constraints come from the availability of data center space, power, and supporting infrastructure, and the construction cycles in these areas are much longer than chip manufacturing.

The content of the OCP conference also confirms this trend. As AI cluster scales move towards "hundred-thousand-level GPUs," the entire design concept of data centers is being reshaped:

Power and Cooling:

• The deployment of large-scale GPU clusters means huge power consumption and cooling challenges. At the OCP conference, liquid cooling has become the default configuration for new AI racks, and the demand for power supply solutions such as high-voltage direct current (HVDC 800V) is also increasing
• This benefits companies like Aspeed, whose BMC (Baseboard Management Controller) is not only used for servers but has also expanded to various devices including cooling systems.

Storage and Memory:

• AI workloads place extreme demands on data storage and access speeds. Meta has clearly stated that, for cost reasons, its data centers will prioritize the use of QLC NAND flash memory. Meanwhile, Seagate mentioned that HDDs (hard disk drives) will still maintain 95% of their capacity online to meet the needs of large and remote data centers.
• More critically, the demand for HBM (High Bandwidth Memory) is experiencing explosive growth, with reports predicting that global HBM consumption could reach 26 billion GB by 2026, of which NVIDIA alone will consume 54%. This highly concentrated strong demand makes HBM supply a key variable affecting AI server shipments.

Racks and Networking:

• To achieve ultra-large-scale deployment, OCP has introduced standardized blueprints such as "AI Open Data Center" and "AI Open Cluster Design," covering racks, liquid cooling, power interfaces, and more.
• In terms of networking, Alibaba stated that pluggable optics remain the preferred choice due to their total cost of ownership and flexibility, while technologies like LPO (Linear Drive Pluggable Optics) are also gaining attention.
• CPO/NPO (Co-Packaged/Near-Packaged Optics) is expected to be realized by 2028 as manufacturing processes mature.

Demand Forecast Indicates Explosive Growth for Downstream Components

The importance of downstream infrastructure can be validated by demand forecast data.

Morgan Stanley analysts expect global cloud service capital expenditures to grow by 31% year-on-year in 2026, reaching $582 billion, far exceeding the market's general expectation of 16%.

Additionally, if we assume that the proportion of AI servers in capital expenditures increases, this means that capital expenditures for AI servers could achieve approximately 70% year-on-year growth in 2026.

From the demand side, major AI giants are still frantically "stockpiling." The report details the AI chip demand for 2026:

• CoWoS Capacity Consumption: NVIDIA is expected to account for 59% of the share, followed by Broadcom (18%), AMD (9%), and AWS (6%).
• AI Computing Wafer Consumption: NVIDIA leads with a 55% share, followed by Google (22%), AMD (6%), and AWS (6%).

In summary, the signals conveyed by the OCP conference, corroborated by industry chain data, clearly point to a new direction for AI hardware investment. As capacity bottlenecks in chip manufacturing and packaging gradually open up, the market's focus will inevitably shift to the infrastructure that supports ultra-large-scale AI computing.

The research report suggests that for investors, this means expanding their focus from individual chip companies to the entire data center ecosystem, seeking "shovel sellers" with core competitiveness in downstream areas such as power, heat dissipation, storage, memory, and networking.