AI Server Power Semiconductors in 2026: SiC, GaN, DrMOS and PMIC Under Supply Pressure

Why power semiconductors became the AI compute bottleneck #

For two decades, the story of the data center was told in processors: more cores, more memory bandwidth, more interconnect. In 2026 the bottleneck moved to the part of the system most teams used to take for granted — the path that gets electricity from the rack busbar down to the sub‑volt rail feeding a GPU.

The trigger is simple arithmetic. A conventional general‑purpose server draws roughly 500 W and carries around $100 of power semiconductors. A current‑generation AI server pulls 3,000 W or more, and the power‑semiconductor content per box lands in the $400–$500 range — a three‑to‑four‑fold jump (industry estimates cited across 2026 semiconductor coverage). The jump is not linear across the bill of materials: it concentrates in the high‑current, high‑voltage devices that legacy server designs never stressed.

For buyers, the consequence is concrete. Power‑stage and PMIC lead times, which looked benign a year ago, are now a sourcing risk on their own — independent of the GPU allocation everyone is already chasing.

What actually grew inside the box #

Three device classes carry most of the增加 (increase):

• DrMOS / Smart Power Stage (SPS). A flagship AI GPU needs 40–80 phases of local regulation. That pushes DrMOS count per server from the low tens (traditional x86) to several hundred, and per‑box value from a few hundred dollars to a few thousand. Renesas' ISL99390FRZ‑TR5935 (90 A SPS) and Infineon's TDA21590 / TDA22590 (90 A OptiMOS integrated power stages) sit squarely in this demand pocket.
• Multi‑phase controllers and PMICs. These set the precise voltage for GPU, CPU and HBM. High‑current, wide‑VIN controllers are capacity‑constrained; MPS has long led the high end, with domestic suppliers (Novosense, Xinpeng) accelerating substitution.
• High‑voltage SiC and GaN. These serve the 400 V / 800 V front‑end and the intermediate bus — the part of the system legacy silicon struggles to scale efficiently.

The notable shift versus the 2021 auto‑driven supercycle: the incremental demand is now led by AI data centers, not vehicles. One widely cited 2026 estimate puts 37%–62% of new power‑semiconductor market growth through 2030 as coming from AI compute and data centers, ahead of automotive and industrial.

The 2026 price cycle: what the signals say #

Several independent 2026 reports describe the same structure — a supply‑rigidity and AI‑demand resonance, not a one‑off spike:

• Second round of price increases. CITIC Securities flagged a second power‑device price‑up cycle starting July 2026, expected to run into 2027. Multiple suppliers — Infineon and Texas Instruments among them — announced second‑round increases effective 1 July 2026 in the 10%–20% range, according to Chinese financial coverage.
• Differentiated increases by tier. AI‑dedicated power ICs and high‑voltage analog rose ~15%–25%; industrial/energy‑storage isolation parts ~10%–15%; low‑end consumer parts only mildly. Scarce, high‑barrier AI parts saw the steepest moves, with some NVIDIA‑qualified exclusive codes reported up over 80% in a single adjustment.
• Lead times stretched. Mainstream power devices reportedly moved to 30+ weeks; some high‑voltage parts longer. A Chuzhou‑based power‑semiconductor maker cited in 36Kr coverage moved to two shifts with order backlog extending four to five months.

Two structural facts explain why this is a cycle and not a blip. On the demand side, AI‑server shipments are forecast to reach ~3.7 million units in 2026, up ~51% year on year in one estimate. On the supply side, TrendForce expects global 200 mm (8‑inch) capacity to enter negative growth in 2026 as TSMC and Samsung shift capacity toward advanced nodes; utilization at the top‑ten foundries' 200 mm lines had already returned to ~90%.

SiC and GaN: the two growth lines under high‑voltage architecture #

The architecture move reshaping the device mix is the shift from 48 V to 800 V HVDC distribution, promoted most visibly by NVIDIA's proposed 800 V rack. The logic is physical: at 800 V DC, the same copper carries 150%+ more power, cutting busbar weight and loss.

How the wide‑bandgap split is expected to fall:

Growth figures cited across 2026 research: Yole‑type estimates put SiC and GaN data‑center CAGRs at ~29.5% and ~46.3% respectively. J.P. Morgan‑cited forecasts (reported by 36Kr) size the AI power‑semiconductor market at ~$2.7 B in 2025 rising toward ~$16 B by 2028 (≈82% CAGR), with silicon ~$11.2 B, SiC ~$3.1 B and GaN ~$1.7 B by 2028. Mordor‑type data places the PMIC market around $44.7 B in 2026, growing to ~$63.7 B by 2031 (~7.3% CAGR).

What 800V HVDC rewrites on the BOM #

The 800 V move is not just a voltage change — it re‑orders which parts sit where:

• Input side: AC distribution, UPS and PDU give way to HVDC busbars, solid‑state transformers (SST) and DC protection. SiC modules, isolated drivers and protection ICs gain.
• Rack side: 48 V/54 V busbars become 400 V / 800 V HVDC. High‑voltage switches, hot‑swap and protection ICs enter.
• Board side: the 800 V → 50 V / 12 V / 6 V conversion becomes a high‑ratio DC‑DC problem — GaN's sweet spot — alongside the DrMOS / power‑stage final stage.

Infineon has guided AI‑data‑center revenue above €2.5 B for fiscal 2027 (roughly 10× versus 2024), and ST has raised its expectation above $2 B for 2027 — both official investor‑communication figures, not third‑party forecasts. That alignment between supplier guidance and architecture roadmaps is why most buyers now treat the wide‑bandgap shift as a planning assumption rather than a maybe.

What buyers should check before RFQ #

The risk is not "power ICs exist" — it is matching the exact ordering code, confirming availability against a real allocation, and not assuming a same‑current device is a drop‑in substitute.

• Confirm the full ordering code. TDA21590 (PG‑IQFN‑39, 90 A OptiMOS stage) and TDA22590 (PG‑UFLGA‑34, 90 A integrated stage) share package intent but are not interchangeable without layout and thermal review. Match suffix, package and carrier.
• Confirm date code, lot and packing. For SPS devices like ISL99390FRZ‑TR5935, reel label, MSL handling and lot traceability matter as much as price.
• Confirm current lead time before release. With 30+ week signals on mainstream parts, a quoted price is meaningless without a committed delivery date.
• Treat "equivalent current" with caution. A similar‑current SPS is not an automatic substitute without controller, pinout, thermal and layout review.

Conclusion #

The 2026 pressure on AI‑server power semiconductors is a structural supply story, not a sentiment swing: AI‑server shipments and 800 V HVDC adoption are pulling DrMOS, PMIC, SiC and GaN demand ahead of 200 mm capacity that is, if anything, shrinking. Buyers who plan around the exact ordering code, verify allocation and lead time up front, and track the SiC‑gray / GaN‑white split as a real BOM change will source through the cycle more reliably than those chasing spot price alone.

For exact stock, date code and lead‑time confirmation on Infineon TDA21590 / TDA22590 or Renesas ISL99390FRZ‑TR5935, an RFQ with target quantity and date‑code requirement is the fastest way to get a committed answer.

Sources:

• 36Kr / 格隆汇 coverage of AI data‑center power‑semiconductor surge and 2026 price‑up cycle (July 2026)
• J.P. Morgan AI power‑semiconductor market forecast (2025–2028), cited via 36Kr
• Yole / Mordor Intelligence power‑semiconductor and PMIC market estimates (2026)
• CITIC Securities note confirming second power‑device price‑up cycle (July 2026)
• TrendForce 200 mm capacity and utilization commentary (2026)
• Infineon and STMicroelectronics AI‑data‑center revenue guidance (FY2027), investor communications