The Power Play: Resistance Is Futile

What You Need to Know:

• As AI infrastructure demand inflects, energy is increasingly a bottleneck
• New architectures are needed to support power-hungry processors running complex and dynamic workloads.
• As datacenter power climbs, demand from and for power semis is growing, presenting the next growth vector for the cyclical broad-based semis group.
• We dissect the shift towards an 800 VDC architecture and the key sockets and technology needed to support the industry.

The TD Cowen Insight

With Artificial Intelligence (AI) datacenter infrastructure now measured in gigawatts, power is the highest operating cost, and energy capacity is often gating compute capacity. High voltage architectures are being evangelized to more efficiently deliver power to more dense server racks. In our full report, we model a US$13 billion greenfield Total Addressable Market (TAM) (approximately 60% 5-year Compound Annual Growth Rate (CAGR), outline key sockets and highlight companies best positioned to benefit.

Majority of AI Datacenter Power TAM driven by Stage 2 Power and IBCs

Our Thesis

As AI demand appears to continue unabated, increasing rack density of power-hungry AI accelerators is driving the industry towards a new 800 volts direct current (VDC) architecture. This is critical as investors struggle with the datacenters increasing from approximately 4.5% of U.S. electricity consumption in 2024 to 7-12% by 2028, per Lawrence Berkeley National Lab. As AI continues to proliferate, increasing rack density of power-hungry AI accelerators is driving the industry towards a new 800 VDC architecture that can drive 92-95% energy efficiency, versus just 75-85% efficiency in traditional alternating current (AC) datacenters.

Power semiconductors are the building blocks of this transition. The demand for and the demands from power semis are increasing and changing as processor power requirements grow and become more diverse and more dynamic. Net, we see AI Power becoming the next meaningful growth vector for broad-based semis, including compound semiconductors.

Beyond just the AI narrative halo for stocks, we see AI datacenter growing from approximately 2% in 2025 to more than 10% of the US$108 billion 2030 power semis TAM. We see the most meaningful growth in the components closest to the processor — the low and medium-volt stage 2 power and Intermediate Bus Convertors (IBCs) — lifted by both the demanding standards and amount of accelerator power draws. Value will be accrued by both Gallium Nitride (GaN) and Silicon Carbide (SiC) providers, but we see more materiality to GaN due to a lower base and its high frequency switching properties relevant in low- and medium-voltage applications.

As the AI ecosystem quickly ramps up, we've seen concentrated capacity constraints in a few areas including power-related components. This is partially a result of structural underinvestment in the 2010s, but we think today's shortages are more reflective of "growing pains" as the supply chain works to catch up to demand. AI power demands could drive incremental capacity investment and conversion, particularly given reluctance of foundries to invest in mature-node output. That said, we do not believe AI datacenter power alone will drive the industry into widespread shortages given its relative size to the broader power semis market.

What To Watch For

• Thermal Design Power (TDP) levels from successive generations of AI accelerators and racks and the pace of adoption.
• Availability of power and analog semiconductor supply, which could have further pricing implications.
• Ongoing Capital Expenditure (CapEx) investment in AI by large hyperscalers, frontier model makers, enterprises and sovereigns.
• The pace of datacenter and energy infrastructure buildouts.
• Potential consolidation, particularly in the fragmented GaN space;
• Government regulation and focus on AI power implications.

Key Takeaways

Investors are struggling to differentiate how companies are addressing the AI power market. We believe the AI datacenter power market will provide accretive growth to other companies in our coverage.

Further, we see a shift towards behind-the-meter (BTM) power generation and energy storage, with fuel cells leading to strong growth opportunities for companies where their native 800 VDC architecture aligns well with next-generation datacenter architecture. We also see a switch to solid-state transformers as this migration happens.

We come away from this work broadly constructive on the AI datacenter power opportunity for analog and power semis companies. We think it has the potential to become a meaningful part of company revenues — even for some large cap participants — adding another leg of secular growth to a historically cyclical group. While it is important to consider a company's key exposures to different parts of the AI datacenter power market, we recommend investors to also consider the materiality of the TAM to each company's revenue.