Where will the main line of AI be next year? Morgan Stanley: The "de-bottlenecking" theme will replace chips, optimistic about energy infrastructure

After NVIDIA has undoubtedly dominated the computing power market, Wall Street's smart money is beginning to think about the next opportunity.

According to the Wind Trading Desk, Morgan Stanley pointed out in its latest report that with the non-linear growth of AI computing power demand, the market logic is undergoing a fundamental change: investors need to shift their focus from "chips" to "de-bottlenecking" by 2026.

Morgan Stanley has once again raised its power demand forecast for data centers, based on the latest growth expectations for NVIDIA's sales. The cumulative power gap for U.S. data centers from 2025 to 2028 is expected to reach 47 gigawatts (GW), up from the previous forecast of 44 gigawatts, which is equivalent to the total electricity consumption of 9 Miami areas or 15 Philadelphia areas.

The report further indicates that after accounting for various "fast power" solutions, U.S. data centers will still face a power shortage of 10-20%, roughly equivalent to a gap of 6-16 gigawatts, which will be most severe in 2027.

Surge in Power Gap: From Chip Crisis to Grid Crisis

Based on the latest chip sales data, Morgan Stanley has once again raised its power demand forecast, signaling a more severe energy crunch. The report shows that by 2028, the power gap faced by the U.S. has been revised from the previous 44 gigawatts (GW) to 47 gigawatts. This supply-demand imbalance indicates that relying solely on the existing grid is no longer realistic.

"Given the non-linear improvements in AI and the diffusion of use cases, we believe investors will shift their focus in 2026 to alleviating the 'intelligent bottleneck.' Key 'intelligent bottlenecks' include: power, political support, labor, and various data center equipment."

Pragmatic Solutions and Ongoing Gaps

In the face of the long wait for grid interconnection, "Time-to-Power" has become the core metric for measuring asset value. Morgan Stanley listed four solutions to bypass grid congestion: gas turbines, Bloom Energy (BE) fuel cells, site-specific nuclear power plants, and the transformation of Bitcoin mining sites. Despite these workaround solutions, analysts remain cautious, believing that the gap cannot be fully filled.

"Even accounting for various 'Time-to-Power' solutions, the net power gap in the U.S. will still reach 10-20% of the construction volume required for data centers (approximately 6-16 gigawatts). We believe this gap will be most severe in 2027, when chip demand grows rapidly, while most turbine solutions have yet to be operational."

The "Arbitrage" Moment for Bitcoin Miners

Among all energy solutions, cryptocurrency mining companies, holding ready access to power permits, are evolving into the "fast lane" of AI infrastructure.

Morgan Stanley specifically mentioned two types of models: "New Neocloud" providers like IREN that directly lease GPUs, and the "REIT endgame" model like APLD, which involves building shells leased to hyperscale enterprises. This transformation of asset attributes is re-evaluating the value of mining companies

"We continue to believe that Bitcoin sites provide AI participants with the fastest power online time and the lowest execution risk... Given the recent weakness in many AI infrastructure stocks, we recommend focusing on those most promising 'de-bottlenecking' participants."

Non-linear Growth of AI Capabilities Drives Surge in Demand

Supporting this massive energy consumption is the exponential leap in AI capabilities. The report points out that the continuous non-linear improvement in AI capabilities and the emergence of more compute-intensive application scenarios are the fundamental drivers behind the ongoing upward adjustment of data center power demand. Morgan Stanley defines this trend as 'the diffusion of intelligence.'

Specific evidence includes: Morgan Stanley analyst Brian Nowak predicts that by 2030, Agentic Commerce will achieve a GMV scale of $190 billion to $385 billion, accounting for 10-20% of U.S. e-commerce. Currently, 45% of U.S. respondents use ChatGPT, 32% use Gemini, and 36% of ChatGPT users made purchases through the platform in the past month.

In terms of enterprise AI applications, by the third quarter of 2025, 24% of AI-adopting enterprises reported quantifiable benefits, up from 15% in the third quarter of 2024. Morgan Stanley expects that AI-driven efficiency improvements will contribute an incremental 30 and 50 basis points to the net profit margins of S&P 500 constituents in 2026 and 2027, respectively.

More importantly, in the most challenging general intelligence test, ARC-AGI-2, the latest cutting-edge large model, Gemini 3 Deep Think, has scored approximately 45%, whereas a few months ago, such models scored only 10-20%. Considering that the average human score is 60%, and that cutting-edge model training will use about 10 times the computational resources by 2026, the industry expects AI capabilities to potentially surpass human levels in multiple complex reasoning tests.