Citi's Semiconductor Bet Shifts to Power and Infrastructure as AI's Structural Bottleneck Emerges

The semiconductor cycle is maturing, and the investment thesis is shifting. The initial phase of the AI trade was a pure compute story, concentrated on a handful of chip designers and hyperscalers. But as systems scale, the limits of existing infrastructure are becoming apparent. The new bottleneck is no longer just processing power; it is memory, storage, and power. This marks a structural pivot from a cyclical tech segment to the foundational layer of a multi-year infrastructure build-out.

The clearest signal is in the memory market. Demand is now being driven by AI servers, where requirements are materially higher than in traditional workloads. This is causing pricing to move independently of consumer device volumes, a sign of a shift toward a more structural, infrastructure-led demand profile. High-bandwidth memory (HBM) sits directly alongside advanced GPUs, scaling with model complexity. Supply is expected to remain tight, and the market is projected to grow at close to an 80% annual rate. At the same time, enterprise SSD demand linked to AI inference is forecast to expand at roughly 50% per year. This infrastructure pressure is already translating into financial results, with DRAM and NAND average selling prices expected to rise by roughly 70–80% into 2026.

This shift creates a powerful tailwind for analog and power management semiconductors. These components are less tied to the volatile consumer cycle and more directly correlated with capital expenditure on physical infrastructure. As data centers demand more electricity and complex system architectures, the need for efficient power delivery and management grows. This moves the sector's growth profile from being driven by discrete product cycles to being more stable and aligned with the ongoing build-out of AI's physical backbone.

This view is reflected in the institutional playbook. Citigroup's sector call to overweight Information Technology is a direct bet on this infrastructure thesis. The sector's stellar 28% year-to-date performance provides evidence of the ongoing capital allocation into this build-out. While the compute layer may see its own cycles, the broader infrastructure story-spanning memory, storage, and power-is establishing a more durable growth trajectory. For portfolio construction, this suggests a move from pure-play AI chips toward a more balanced exposure within IT, capturing the quality factor of sustained CAPEX.

Analyzing the Top Picks: Drivers of Outperformance

Citi's top picks are not a random selection; they are a direct translation of the infrastructure thesis into specific stock picks. Each company's fundamental profile aligns with a different facet of the AI build-out, from the compute core to the power delivery layer.

Broadcom and Nvidia remain the undisputed leaders in the AI compute stack. For both, data-center sales are the core growth engine. Citi notes that demand from the Data Center end market (34% of semi sales) remains solid due to AI, providing a durable revenue base. Their dominance in networking and GPUs, respectively, ensures they capture the highest-value segments of this infrastructure spend. The institutional view here is a conviction buy on quality and scale, betting that their market positions will continue to benefit from sustained capital expenditure.

The real margin upside, however, appears to be in the analog and power management layer. Texas Instruments is highlighted for its gross margin upside as capital spending moderates. This is a nuanced call. As the initial, frenetic phase of AI server deployment settles, the focus shifts from pure volume to efficiency and optimization. TI's shift in R&D toward data-center applications positions it to profit from this transition, capturing value as the infrastructure itself becomes more complex and power-hungry.

Monolithic Power Systems is seen as the purest play on the infrastructure growth story. Citi expects the company to outgrow the industry led by enterprise data sales product growth. This is supported by the broader sector trend: data-center-related sales for analog companies surged 50–70% year over year in 2025. For MPWR, this isn't just a tailwind; it's the primary driver of its outperformance. The company's focus on power management solutions for servers and enterprise systems directly targets the rising power demands of AI workloads.

The bottom line is a portfolio-level allocation. Citi's top picks represent a strategic rotation within semiconductors. It's a move from pure-play compute (where growth is cyclical) toward the foundational layers (where growth is structural). This shift is designed to capture the quality factor of sustained CAPEX while positioning for the next phase of the cycle, where efficiency and power management become as critical as raw processing power.

Portfolio Implications and Risk Assessment

The institutional strategy emerging from this thesis is a classic cyclical barbell. It combines high-conviction, high-beta growth names with more defensive, infrastructure-focused plays. This structure aims to capture the explosive upside of the AI compute layer while hedging against its inherent volatility with the more stable, capital-expenditure-driven demand for foundational components.

On one side of the barbell are the pure-play compute leaders, Broadcom and Nvidia. These are conviction buys on quality and scale, betting that their dominant market positions will continue to benefit from sustained data-center CAPEX. Their performance is the engine of the sector's stellar 28% year-to-date rally. On the other side are the analog and power management specialists, like Texas Instruments and Monolithic Power Systems. These names offer a different growth profile-less tied to discrete product cycles and more directly correlated with the ongoing build-out of AI's physical backbone. This pairing provides a balanced exposure, capturing both the cyclical growth and the structural infrastructure tailwinds.

The primary risk to this thesis is the industry's extreme concentration. The semiconductor sector is navigating a high-stakes paradox: record sales are driven by a single, booming end market. As of mid-2025, generative AI chips were estimated to approach $500 billion in revenue, or roughly half of global chip sales. This creates a vulnerability. If the AI demand cycle were to slow, the entire industry's growth trajectory could be materially impacted. The sector's heavy reliance on this single driver is a key watchpoint for portfolio managers.

Valuation remains a critical factor for selectivity. While the sector as a whole is performing well, not all valuation reads are stretched. This divergence offers an opportunity to construct a more resilient portfolio. Investors can overweight the high-conviction growth names while also finding value in the infrastructure plays, where the growth story is more durable and less prone to the sharp corrections that can hit pure-cycle segments. The barbell approach, therefore, is not just a tactical rotation but a structural defense against the sector's own concentration risk.

Catalysts and What to Watch

The institutional rotation into semiconductor infrastructure is a forward-looking bet. To assess its validity, investors must monitor a set of near-term signals that will confirm the thesis or reveal its vulnerabilities.

The most immediate data point is quarterly earnings. For the analog and power management names at the heart of Citi's thesis, the focus will be on data-center-related sales and margin trends. The call for gross margin upside as capital spending moderates for Texas Instruments hinges on the company's ability to maintain pricing power and operational leverage as the initial deployment frenzy settles. Similarly, the expectation that Monolithic Power Systems will outgrow the industry led by enterprise data sales requires concrete evidence of accelerating server and data-center revenue growth. Any deviation from these trends would challenge the "Phase 2" cycle narrative.

Beyond individual companies, the trajectory of capital expenditure cycles in hyperscalers and telecom providers is the ultimate demand driver. These are the customers for the infrastructure semiconductors. Any signs of a slowdown in their CAPEX plans-whether due to macroeconomic pressures or a perceived saturation of AI deployment-would directly threaten the growth story for analog and power management firms. The sector's heavy reliance on generative AI chips as a primary revenue source makes it sensitive to shifts in this single end market's spending rhythm.

Finally, broader economic indicators will provide context for cyclical support. While the AI infrastructure story is structural, it still interacts with the wider economy. The improving trends in industrial demand alongside stabilizing global manufacturing PMIs are positive for analog companies with diversified exposure. Conversely, a deterioration in these metrics, or a sharper-than-expected decline in China auto production, could pressure the industrial and automotive segments, creating a drag on overall analog sector performance. For names like Texas Instruments, which have significant exposure to these areas, this provides a counterbalance to the AI tailwind.

The bottom line is that the infrastructure thesis is not a static conclusion but a dynamic setup. The rotation into these specific names is a bet on sustained data-center CAPEX and the structural shift in semiconductor demand. The coming quarters will provide the real-time data to validate that bet.