What does analog do that digital can't?
Texas Instruments, Analog Devices, NXP, Infineon, and a handful of others run a parallel semiconductor industry that nobody talks about because it is unfashionable, durable, and embarrassingly profitable. Every digital system depends on it. AI systems depend on it more than most.
Analog chips do the work that the physical world demands and that digital cannot do cleanly — converting between continuous signals and discrete bits, regulating power, driving motors, conditioning sensors. They are made on trailing nodes, are not commoditized, and the gross margins are higher than logic. Most investors do not pay attention to this industry. They should.
Where analog actually lives
Digital electronics moves bits — discrete ones and zeros — and everything inside a microprocessor or memory chip is doing that. But the real world is not digital. The output of a microphone is a continuous voltage. The reading from a pressure sensor in a tire is a continuous voltage. The current draw of a motor is continuous. The voltage on a power rail has to be regulated to a continuous value within microvolts. None of these signals can be processed by a digital chip until they have been converted, and the chips that handle the conversion are the analog industry.
The interface device that turns a continuous voltage into a stream of bits is an analog-to-digital converter — an ADC. The reverse direction — turning bits back into a controlled voltage or current to drive a speaker or a motor or a laser — is a DAC, a digital-to-analog converter. Every smartphone has dozens of ADCs and DACs. Every modern car has hundreds. Every server has at least one ADC per voltage rail being monitored. The performance characteristics of these converters — sampling rate, bit depth, noise floor, linearity — directly limit what the digital chip can do, because the digital chip cannot see the world more clearly than the ADC that fed it.
Beyond converters, analog encompasses power-management ICs (PMICs that regulate voltage rails), motor drivers, RF transceivers, sensor signal-conditioning amplifiers, audio codecs, op-amps, voltage references, and a hundred more sub-categories. Each category has a few specialist suppliers, deep moats around long-cycle customer relationships, and prices that do not erode the way digital prices do.
Why trailing nodes are fine
Digital chips have spent sixty years racing to smaller process nodes because each node shrink improves transistor density and gives a generational performance gain. Analog chips do not get the same benefit. The physical phenomena that analog circuits depend on — accurate voltage matching between transistors, controlled noise, large stable capacitors, well-characterized resistors — get worse at smaller nodes, not better. At 5nm or 3nm, transistor mismatch is higher, voltage headroom is lower, and the design effort required to deliver a clean analog signal is enormous.
As a result, most analog chips ship on what the industry calls trailing nodes. A new automotive sensor chip might be designed at 180nm or 130nm — process generations that were leading-edge in 2002 and 2005 respectively. A power-management chip for a phone might be at 65nm or 40nm. Some analog parts are still made at 350nm. The fabs that run these nodes have been depreciated for years. The equipment is paid off. The processes are mature and predictable. Gross margins are routinely 55-65%, which is higher than logic and dramatically higher than memory.
The cost structure flips the usual narrative. Where logic chips need leading-edge fabs (and the resulting $30B capex bill), analog companies can run profitable businesses on mid-2000s equipment in mid-2000s fabs that they often own outright. The economics resemble specialty chemicals or industrial machinery more than they resemble Silicon Valley. The customer relationships are sticky — once you have qualified an analog part in an automotive or medical or industrial design, swapping suppliers is a six-to-twelve-month requalification — and the average product stays in the catalog for ten to twenty years.
The industry structure
The analog industry is concentrated but not as concentrated as logic. Texas Instruments is the largest pure analog supplier; it makes power management, signal chain, ADCs and DACs, and embedded processors for industrial and automotive customers. Analog Devices is the second-largest, with strong positions in high-performance converters, RF, and instrumentation. NXP (a 2006 carve-out of Philips Semiconductors) is dominant in automotive microcontrollers and identification chips. Infineon is the European leader in automotive, industrial, and power semiconductors. STMicroelectronics serves a mixed automotive and IoT customer base. Renesas (a Hitachi-Mitsubishi-NEC microcontroller merger) is the Japanese leader.
Together these six companies account for roughly half of the global analog and mixed-signal market. A long tail of specialists — Microchip, Maxim (now owned by ADI), Skyworks, Qorvo, Cirrus Logic, On Semiconductor, Diodes Incorporated — fills the rest. Concentration is rising slowly through acquisition (ADI bought Linear Tech in 2017 and Maxim in 2021; Microchip bought Microsemi in 2018), but the industry remains far more horizontal than the logic side.
The Chinese analog industry is the one to watch. Until 2018, Chinese chipmakers were close to absent from the analog category. Since then, companies like SG Micro, 3PEAK, and Will Semiconductor have built credible product lines, often by copying or partially licensing reference designs from US incumbents. Analog is a category where Chinese self-sufficiency may arrive faster than in logic, because the fab requirements are modest and the design IP, while difficult, is not gated by ASML's export controls.
Where analog shows up in an AI rack
It is everywhere. Every voltage rail on a GPU board is regulated by a PMIC stack — typically a multi-phase buck converter stepping 48V down to 0.8V, with each phase delivering 50-100A. A typical H100 board has more than 30 analog ICs just for power conversion. The signals between GPUs over NVLink and between racks over PCIe and InfiniBand are driven by SerDes (serializer-deserializer) circuits that are partly analog at their core. The board management controller (BMC) on every server is reading dozens of temperature, current, and voltage sensors — each one an ADC and a sensor-signal-conditioning circuit. The optical transceivers that connect rack to rack are full of analog circuits for laser drive, photodetector readout, and clock recovery.
In a $200,000-$400,000 HGX H100 8-GPU server, the analog content might be $8,000-15,000. As a fraction it is small. As a supplier dependency it is large, because if any of the analog parts go on allocation, the whole server cannot ship. During the 2021-22 supply shock, server builds were delayed by analog parts more often than by GPUs themselves. The phrase that circulated was that a server is gated by 'the slowest fastener,' and analog parts are often that fastener.
Strategic read
If you are tracking AI infrastructure, the analog suppliers are a less-exciting but more-reliable exposure to the same buildout. The volume of analog content per AI server is not collapsing. The price points are not collapsing. The customer base is not consolidating away from the existing suppliers. And the gross margins are protected by the very things that make digital margins fragile — long product lives, qualified-in customer relationships, fragmented end markets.
The interesting question is whether AI accelerator-side innovation creates room for new analog suppliers. Power delivery for a chip pulling 1000+ amps at sub-1V is an unsolved problem; current approaches stack dozens of phases and lose efficiency. Vertical power delivery (where the regulator sits directly on top of the chip rather than next to it) is one architectural direction; integrated voltage regulators that move some of the regulation onto the compute die itself are another. Either could create $5-10B of new analog content per year at scale, and the suppliers that capture it may or may not be today's incumbents.