Idea → Industry
From a hunch to where to dig
Where does the AI buildout create durable new business for analog and power chip makers, and who ends up on the wrong side?
Technology-adoption theme. This is a technology being adopted, not a physical event or a market wave, so I follow the generic chain: who is forced to spend, who supplies them, and which suppliers profit durably. Claims tagged (sources/) come from filings in your folders; (web) means outside sources/.
How the idea fans out — why each industry is on the list
Link 1 — the cause: AI is built out at gigawatt scale
Who is forced to spend: hyperscalers, AI rack builders, robot makers
NVIDIA told investors the top five hyperscalers are expected to spend nearly $800 billion of capex in 2026 and $1.3 trillion in 2027 (NVDA Q2 FY2027 call, 26 Aug 2026) (sources/). Every GPU rack they buy needs power converted many times and data moved at very high speed, and both jobs are analog work. Robot makers are a smaller, later buyer of the same kinds of chips.
Link 2 — what analog chip makers sell them (the picks and shovels)
800 V power conversion: SiC and GaN chips, solid-state transformers
NVIDIA is moving AI data centers from 54 V to 800 V DC power, with the shift starting in 2027. Grid power gets converted once at the building edge, then stepped down near the chips. That needs high-voltage silicon carbide (SiC) and gallium nitride (GaN) transistors, which today's server racks barely use. NVIDIA names Infineon, STMicro, onsemi, Texas Instruments, Navitas, Monolithic Power, Rohm, Renesas, Power Integrations and ADI among its chip partners. Infineon targets about €1.5bn of AI-related revenue in FY2026 and €2.5bn by 2027, and is supplying SiC to Eaton for a solid-state transformer (web, outside sources/).
Power next to the GPU: voltage regulators and power stages
A modern GPU pulls well over 1,000 amps at under one volt, so the last few centimetres of power delivery are a specialist market. Monolithic Power (MPS), Infineon, Renesas, TI and Vicor fight for these sockets, and Vicor also licenses its vertical-power patents (licensing run-rate of about $90m, which the CEO wants to grow to $200m) (web, outside sources/).
Fast analog for moving data: optical and copper links, plus the foundries that make them
Linking GPUs needs high-speed mixed-signal chips (drivers, amplifiers, SerDes, retimers) and silicon photonics. Tower Semiconductor, a specialty analog foundry, reported silicon photonics revenue up more than 270% year over year to an annualised $680m+ in Q2 2026, targeting a $1bn run-rate by Q4 2026, with about $1.3bn of customer contracts for 2027 (TSEM Q2 FY2026 call, 4 Aug 2026) (sources/). Credo, Astera Labs, Marvell and Broadcom sit in the same branch and already have folders here.
The broad-line analog catalog in every rack: sensing, protection, monitoring
Each AI rack also carries hundreds of ordinary analog parts: hot-swap controllers, current monitors, isolators, clocks. Texas Instruments' data center revenue reached $1.5bn in 2025 (about 9% of revenue) and doubled year over year in Q2 2026; ADI's communications segment, driven by data centers, grew 63% to 84% in recent quarters (web, outside sources/).
Physical AI: robots and machines that act in the world
Humanoid and industrial robots are full of motor drives, position sensors, radar and power chips. TI and NVIDIA announced a humanoid-robotics collaboration in March 2026; Infineon, ST, NXP, Renesas, Allegro and Melexis also compete for those sockets (web, outside sources/). Volumes are still tiny.
Link 3 — who profits durably
The suppliers whose chips are hard to qualify and sit where the architecture is changing
Content that is new to the rack (800 V conversion, photonics) is where share is still being decided. Catalog parts that the rack already used grow with volume but are a small slice of a big company's revenue (inferred).
The wrong side — who the same shift works against
48 V bus converters and old-style server power shelves
800 V-to-6 V boards are designed to remove the 48 V intermediate stage altogether (Navitas demo) (web, outside sources/). Whoever lives on that stage loses the socket.
Pluggable optical module assembly
NVIDIA calls Spectrum-6 the first 800G co-packaged optics switch and says copper "has no chance" at scale-out (GTC Taipei analyst Q&A, 2 Jun 2026) (sources/). Co-packaging moves optics onto the switch package and away from separately assembled modules.
Analog makers tied to phones, PCs and consumer gadgets
AI is soaking up memory and leading-edge supply, which raises device costs and can cut unit volumes, so consumer-heavy analog names miss the AI pull and may feel the squeeze (inferred).
The shortlist, ranked — click a heading to sort
| Industry | Link | Durable / passing | Where the money pools | Crowded / overlooked | The catch |
|---|---|---|---|---|---|
| 1. 800 V power conversion (SiC / GaN, solid-state transformers) | Direct | Durable | Qualified 1200 V SiC and high-voltage GaN device makers with volume capacity; Infineon has the clearest disclosed AI revenue (web, outside sources/) | Overlooked | SiC was overbuilt for EVs, so pricing may stay weak; Wolfspeed runs at a negative gross margin (web, outside sources/) |
| 2. Specialty analog foundry for silicon photonics and SiGe | Direct | Durable | Scarce qualified SiPho capacity; Tower is spending $920m mostly on SiPho and SiGe capacity (FY2025 20-F) (sources/) | Overlooked | Tower's own 20-F warns AI wafer prices and demand may be volatile, and capacity races may follow (sources/) |
| 3. GPU-adjacent voltage regulators and power stages | Direct | Durable | Highest-current-density point-of-load parts plus patents (Vicor royalties) (web, outside sources/) | Crowded | Sockets are re-won every GPU generation; Vicor stock rose 54% in five days in Sept 2026 (web, outside sources/) |
| 4. High-speed mixed-signal connectivity (SerDes, retimers, optical drivers) | Direct | Durable | Mixed-signal IP owners (Credo, Astera, Marvell, Broadcom) (inferred) | Crowded | Co-packaged optics could move value from today's link chips to the switch vendor (inferred) |
| 5. Broad-line analog catalog in the rack (TI, ADI) | Indirect | Durable | Spread thin: data center was about 9% of TI revenue in 2025 (web, outside sources/) | Crowded | TI and ADI trade at around 45 to 50 times earnings on this story (web, outside sources/) |
| 6. Physical AI: robot motor control and sensing | Indirect | Passing (for now) | Motor drives, radar and position sensors; no supplier dominates (web, outside sources/) | Overlooked | Humanoid volumes are tiny and the ramp date is unknown (web, outside sources/) |
Parked — weaker links: analog in-memory computing (a stretch; research stage), AI-assisted analog design tools at Cadence and Synopsys (helps analog makers' costs, not their demand), semiconductor test for analog parts such as Teradyne (indirect and cyclical).
The wrong side of the idea
The same force that lifts the industries above pushes against these. Two uses: they're shorts / avoids in their own right, and they're a confirmation test — if the industries that should be bleeding aren't under real pressure, treat the whole theme as weaker or slower than the bull story claims.
| Industry | Why it's hurt (mechanism) | Exposure |
|---|---|---|
| 48 V intermediate bus converters and 54 V power shelves | 800 V-to-6 V conversion is designed to delete the 48 V stage in AI racks from 2027 (web, outside sources/) | Direct hit |
| Low-voltage silicon MOSFETs in server power supplies | GaN and SiC take over the high-voltage stages as racks pass 1 MW (inferred) | Partial |
| Pluggable optical module assemblers | Co-packaged optics on NVIDIA's Spectrum-6 puts optics on the switch package (sources/); pluggables keep scale-up and legacy builds for now | Partial |
| Copper busbar and power cabling inside racks | Over 200 kg of copper per 1 MW rack at 54 V is the problem 800 V is meant to solve (web, outside sources/) | Partial |
| Consumer-heavy analog makers (phones, PCs, gadgets) | AI pulls memory and wafer supply away, raising device costs and squeezing unit volumes (inferred) | Partial |
Study this first
The one slice worth a deep dive
800 V DC power conversion chips for AI data centers: SiC and GaN devices from the grid to the rack
It is new analog content that today's racks barely contain, it locks in for a whole facility design rather than one GPU generation, the profit sits with the few makers of qualified high-voltage SiC and GaN, and the market still prices it behind the crowded GPU-adjacent regulator trade.
The one thing to verify: Whether 800 V DC actually ships in volume on NVIDIA's 2027 racks and gets adopted by hyperscalers for their own custom-chip racks. If it slips or stays NVIDIA-only, the slice shrinks back to a niche.
run learn-any-industry on 800 V DC power conversion semiconductors (SiC and GaN) for AI data centers
New to this space? Start with the line above. Already fluent in it? Skip straight to the analytical primer:
run industry-primer on 800 V DC power conversion semiconductors (SiC and GaN) for AI data centers
The catch hiding in the idea
For the big analog names, AI is real but small. Texas Instruments' data center business was about 9% of revenue in 2025, while TI and ADI already trade near 45 to 50 times earnings on the AI story (web, outside sources/). So most of the price already assumes AI wins, and the bulk of their revenue still rides the industrial and auto cycle. Second, every AI dollar here flows from hyperscaler capex that NVIDIA expects to reach $1.3 trillion in 2027 (sources/). Analog makers have a long history of inventory booms and busts, and a capex pause would hit the newest AI lines hardest. Third, the losers are not yet bleeding: 800 V only starts in 2027 and pluggable optics are still growing, so the architecture shift that creates the overlooked winners is still a forecast, not a fact. Treat the theme as early.