The Rising Demand for Analog Semis in Data Centers
Discover how analog semiconductors are revolutionizing data centers and why investors should take note of this critical shift.
The chips getting all the attention are the wrong ones.
NVIDIA's GPUs dominate headlines. Custom AI accelerators from Google, Amazon, and Microsoft receive breathless coverage with each new generation. But while the industry fixates on compute silicon, a quieter revolution is happening in the power delivery infrastructure underneath it all — and analog semiconductors are at the center of it.
UBS recently flagged rising opportunities for analog semis in AI and data center power, and the observation cuts against the prevailing narrative that this is purely a digital compute story. It isn't. The more AI workloads scale, the more the analog layer becomes the critical constraint — and the critical opportunity.
What Analog Semiconductors Actually Do (and Why That Matters Now)
Analog semiconductors don't process ones and zeros. They manage the real world: voltage, current, temperature, frequency. Where digital chips transform information, analog chips condition power — converting, regulating, protecting, and distributing electrical energy across complex systems.
The distinction matters enormously in a data center context because every digital chip in a rack is completely dependent on analog components to receive clean, stable power at the precise voltages it needs.
A modern AI accelerator like an H100 GPU operates at extremely low core voltages — often under 1V — while drawing hundreds of watts. Getting power from the utility feed at medium voltage down to sub-1V levels across thousands of chips requires a cascade of analog power conversion stages: rectifiers, voltage regulators, power management ICs (PMICs), and precision reference circuits. None of that is digital. All of it is analog.
The irony is structural. As digital chips get more powerful, they become *more* dependent on analog support infrastructure, not less.
AI Is Breaking Data Center Power Architecture
Legacy data center power design was built around a relatively stable, predictable load profile. A server rack draws X watts; you engineer for that, done. AI infrastructure throws that model out entirely.
GPU clusters running inference workloads create massive, instantaneous power fluctuations. A rack that might idle at 5kW can spike to 80kW in milliseconds when a training batch kicks off. That kind of transient behavior is brutal on power distribution systems designed for steady-state operation. Voltage droops, current surges, and thermal spikes all become acute problems at scale.
Analog semiconductors — specifically high-speed voltage regulators and power management ICs — absorb those transients and keep AI accelerators from crashing or throttling under load. The faster and more variable the AI workload, the more precision analog power management becomes a performance variable, not just a reliability variable.
This is the non-obvious insight most coverage misses: analog semis aren't just supporting AI data centers; they're actively enabling higher AI performance by keeping power delivery within the tight tolerances these chips demand.
The Technical Case for Analog in AI Power Solutions
Point-of-load power conversion is where analog semis earn their keep in modern AI infrastructure. Rather than distributing power at standard voltages and converting at the board level, hyperscalers are increasingly moving to 48V direct current distribution architectures — and then converting down to chip-level voltages right at the processor socket using integrated voltage regulators (IVRs) and discrete analog components.
This approach dramatically reduces resistive losses in power delivery (watts wasted as heat in copper traces) and allows for faster, more precise voltage response. Companies like Texas Instruments, Analog Devices, Monolithic Power Systems, and Renesas have been quietly building product lines optimized for exactly this application.
The thermal management layer adds another dimension. AI chips generate extraordinary heat densities — we're talking about power dissipation approaching 1,000W per chip in next-generation designs. Temperature sensing, fan control, liquid cooling management, and thermal protection circuits are all analog functions. A single rack of H100s might contain hundreds of analog temperature sensors and thermal management ICs working in concert to keep the system from destroying itself.
None of this is glamorous. All of it is essential. And the bill of materials for analog components per rack is growing substantially with each generation of AI hardware.
The Investment Picture in Analog Semiconductors
UBS's thesis isn't complicated, but it's well-grounded. The analog semiconductor market for data center applications was relatively modest historically — dominated by server power supplies and networking equipment. The AI buildout is fundamentally different in scale and intensity.
Hyperscalers — Meta, Google, Microsoft, Amazon — are collectively spending hundreds of billions on data center infrastructure through 2026 and beyond. Every watt of AI compute deployed requires a corresponding investment in power delivery, conversion, and management. Analog semis capture a meaningful slice of that spend, and unlike GPU supply, which is bottlenecked by TSMC's most advanced nodes, analog chips are largely manufactured on mature process nodes where capacity is more readily available.
That supply-demand dynamic matters for margins. Analog semiconductor companies don't face the same leading-edge fab constraints as digital chip designers. Their moats are more often circuit design expertise and customer relationships built over decades — harder to replicate than throwing money at a new fab.
The key players worth tracking: Texas Instruments remains the scale leader with the broadest analog portfolio and a strategy of owning its own manufacturing. Monolithic Power Systems has aggressively taken share in high-performance computing power management. Analog Devices brings strength in precision signal chain applications alongside power. Renesas has been expanding its data center power portfolio through both organic development and acquisition.
For investors, the calculus is different from betting on which AI model wins or which GPU generation leads. Analog semiconductor content per data center rack is growing regardless of who wins the AI compute wars. It's a picks-and-shovels play with genuine technical barriers to entry.
Where Data Center Analog Demand Goes From Here
The next wave of pressure on analog infrastructure is already visible in the power specifications being published for next-generation AI chips. NVIDIA's Blackwell architecture and its successors require power delivery systems that previous data center designs simply cannot support. Rack power densities of 120kW, 200kW, and beyond are being seriously planned for, which means power conversion and distribution infrastructure needs to be fundamentally re-engineered.
Liquid cooling integration — whether direct-to-chip or immersion — creates new analog requirements around pump control, flow sensing, leak detection, and thermal regulation. These aren't high-volume commodity applications; they're precision analog design challenges that reward companies with deep application expertise.
There's also the grid connection problem. As data centers scale toward gigawatt-class campuses, managing the interface between utility power and internal DC distribution requires sophisticated power electronics — many of which are, at their core, analog semiconductor applications. Power factor correction, harmonic filtering, and grid synchronization are fundamentally analog problems solved with analog components.
The trajectory is clear: as AI infrastructure scales from kilowatt racks to megawatt pods to gigawatt campuses, the analog semiconductor content required to make that infrastructure function reliably scales with it — and the engineering complexity of those solutions makes them difficult to commoditize quickly.
The market's focus on compute silicon is understandable. But the companies solving the power delivery and management challenges underneath those flashy GPUs are building durable positions in infrastructure that will be required for decades. That's a different kind of opportunity — less volatile, more defensible, and hiding in plain sight while everyone else chases the next training chip announcement.
The smart money is starting to notice. The smarter move is to understand *why* before everyone else does.
[INTERNAL LINK: analog semiconductors]
[INTERNAL LINK: AI infrastructure]
[INTERNAL LINK: data center power management]