According to a thematic research report titled "Decoding the Rhythm of AI Transformation: From Compute Buildout to Industrial Diffusion," Morgan Stanley has put forward a core judgment based on the sixth round of its AI stock mapping database, which covers approximately 3,600 stocks globally: AI investment has moved beyond the stage of simply betting on upstream compute-enabling names and has entered a new era of "barbell-style" dual-track positioning, requiring investors to both capture structural opportunities arising from compute bottlenecks and increase allocation weight to companies that are actually deploying AI.
Morgan Stanley states that its core research theme for 2026 is technology diffusion, and in this industrial cycle that is currently in transition, it expects a barbell-shaped excess return pattern, with core infrastructure enablers and early-stage software vendors and AI adoption companies generating gains in tandem.
The typical pattern of a technology cycle is that semiconductors are the first to deliver excess returns, followed by the infrastructure segment, and then software and services.
The bank judges that the investment logic of the AI industry is entering a new phase: shifting from trades that were highly concentrated in enablers to a broader barbell-shaped opportunity pool that simultaneously covers selected enablers and emerging AI adoption companies.
Since the start of the current AI cycle, the semiconductor and infrastructure segments have recorded 500-700 basis points of excess returns relative to the S&P 500 Index.
In our view, now is the time to begin raising exposure to early-stage software enablers and AI adoption companies.
At the same time, however, the massive wave of AI capital expenditure buildout, combined with key bottlenecks such as power, will lengthen the upside window for some high-quality names.
Therefore, this computer cycle will not replicate the simple, linear rotation of market leadership seen in past cycles.
The bank notes that this AI cycle has unique characteristics, with labor, power, and regulatory policy constituting three core constraints.
The report estimates that global data centers will face a 57GW power gap in 2026-2028, and combined with practical obstacles such as local approvals and labor shortages, the pace of compute supply expansion will be significantly suppressed, with compute scarcity persisting for years and no simple, clear sector handoff rotation as seen in history.
Hyperscaler capital expenditure growth is expected to decline from 93% in 2026 to 14% in 2028, and expectations that the hardware capex boom has peaked are gradually being priced in by the market, though sentiment in certain bottleneck segments will persist.
Data shows that AI enablers, meaning upstream suppliers such as chips, compute, and hardware, have doubled their expected EPS over the past two years and continue to maintain strong earnings momentum.
Meanwhile, AI adopters, meaning companies across industries that apply AI technology to their own businesses to achieve cost reduction and efficiency gains, are expected to see cumulative EPS growth of about 70% over the next 12 months across two years, with an earnings inflection point already visible.
Market consensus expects that adopters with high real AI impact will expand EBIT margins by 460 basis points in 2025-2026, nearly double that of the MSCI ACWI Index, but the market has not yet fully priced in the long-term dividends from AI-driven productivity gains, leaving room for upward revisions to forward earnings forecasts.
On valuation, the forward P/E of high-AI-impact adopters has fallen back to 18 times, compared with 22 times for enablers, and the risk-reward ratio has improved significantly after valuation digestion.
The barbell investment strategy proposed by the bank has two ends.
On one end, it continues to hold bottleneck assets: prioritizing power-related segments, including energized data center service providers, energy storage, grid equipment, and new energy; while also selecting individual names within semiconductors and infrastructure, focusing on second-layer supply chain bottlenecks beyond power.
On the other end, it expands new allocations: increasing exposure to early-stage software enablers, with priority order being infrastructure software, cybersecurity, and then selected application software; and positioning in AI adopters that have already delivered quantifiable business returns, with screening focused on two major indicators: first, the real impact of AI on the company's business, and second, the company's pricing power, since only companies able to retain the cost-reduction benefits of AI can truly convert them into profit realization.
IT services is a late-cycle beneficiary segment and is not yet suitable for major allocation at this stage.
From a global perspective, the path of AI development varies markedly across markets.
In North America, AI dividends are spreading from hardware to software and the real economy; in Asia-Pacific, returns remain concentrated in the upstream supply chain, but more and more companies are disclosing AI-driven revenue growth in their financial reports; in Europe, AI opportunities are more concentrated in application deployment within traditional industries.
The report particularly emphasizes that investors cannot simply chase the "AI concept," and that the real impact of AI on a company's investment logic is the key to success or failure.
Names for which AI is a core part of the logic deliver huge excess returns compared with those where it is only a secondary influence; conversely, companies for which AI poses a core business threat significantly underperform peers that face only moderate disruption.
At the same time, there are reverse risks in the market: if compute demand continues to exceed expectations and labor and power bottlenecks persist, the upcycle in semiconductors and infrastructure could last far longer than the market's baseline expectations, and Wall Street has repeatedly underestimated the development space of the technology industry in the past.
Overall, the AI trade has moved from pure "compute-stacking" thematic speculation toward a new stage of verifying commercial ROI, and balancing upstream bottleneck assets with application names that are delivering results has become the allocation approach better suited to the current environment.
Comments