Question (Nutty's framing): Which companies make money even if we don't pick the winning power architecture — NVIDIA's full 800V-DC rack vs OCP/Mt. Diablo, SiC vs GaN, single-stage vs multi-stage VRM, lateral vs vertical GaN, paralleled discrete vs co-packaged-with-die?

An automated screen for "power semiconductors AI data center" returned the fab-equipment layer — ASML, LRCX, KLAC, AMKR, TER — as its top picks. That's a true "picks and shovels" answer to a different question (who wins regardless of which fab buys tools). For Nutty's question — who wins regardless of which power-delivery architecture wins inside the rack — the right cohort sits one layer downstream from the equipment makers: the power-semi merchants (MPWR, MPS, VICR, NVTS, POWI, AOSL, IFNNY, ON, STM), the packaging/integration enablers (GLW glass cores, SMTC ALPR, BESIY / AMS:BESI hybrid bonders), and the rack-level integrators that consume all of the above (VRT, ETN). Both layers are in this brief.

Executive summary

The "AI power crisis" thesis as Nutty frames it across his 4-part series + the Next Bloom Energy side story is a parasitic-loss story: somewhere between 25-30% of grid energy is bled off as heat between the substation and the transistor, and every percentage point of efficiency unlocks measurable rack-density. Two architectures are competing — NVIDIA's full 800V-DC bus through the rack, vs the OCP "Mt. Diablo" 48V/intermediate-bus camp — and the fight will likely be settled by whichever one ships at scale through 2027. The picks-and-shovels question is who wins regardless. The answer is the power-semi merchants who sell into both architectures, the packaging suppliers whose technology is required for either, and the rack-level integrators who will sell the cabinet either way. Today's tape confirms it: every name in the cohort has gone parabolic, with MPWR +41% / VICR +70% / NVTS +104% / AOSL +89% / ON +66% / IFNNY +45% over 30 days — RSIs in the 70s-80s, all stretched well above SMA200. The trade is no longer "find the names"; it's "decide whether to chase or wait for the first violent dip." Convergence with today's research-trace (NVTS, ASML, MU, AVGO, GLW, SMTC) and with NuttyCLD's named cohort (MPWR, VICR, NVTS, POWI, AOSL, IFNNY, ON, BESIY / AMS:BESI, AMAT, LRCX, ASML) is unusually tight — eight names appear in two paradigms, three names appear in all three. That's the highest cross-paradigm signal we've seen on a single thesis since the optical supercycle in March.

Top 3 picks-and-shovels names (architecture-agnostic)

1. MPWR (Monolithic Power Systems) — sells into both 800V and 48V

Why architecture-agnostic: MPWR ships full power-stage modules, multi-phase controllers, and PMICs to both the NVIDIA full-800V camp and the OCP intermediate-bus camp. They are the merchant Tier-1 that doesn't have a horse in the architecture race — they sell the bridge components either side will buy. Their products are the most-cited "default" power-stage in NVIDIA reference designs, but they're equally embedded in hyperscaler-custom OCP boards.

Tape: $1,583.48, +41% / 30d, +151% / 12m, RSI 68. Stretched but not yet absurd; this is the one where you can almost defend a chase entry on a 5-7% pullback. Of the cohort, MPWR has the cleanest "merchant power semi to AI" story with no exit overhang from a private-equity sponsor or fab divestiture cloud.

The bear: valuation. MPWR has been trading at sky-high P/S for multiple quarters. A single quarter of decel rerates them violently.

2. VICR (Vicor) — owns the factorized power IP regardless of where it gets deployed

Why architecture-agnostic: Vicor's "factorized power architecture" — pre-regulation followed by current multiplication via their patented multiphase modules — is a topology that maps onto either 800V or 48V intermediate-bus designs. They're licensed/integrated on both NVIDIA's reference platforms and several OCP cabinet vendors. Their patent moat on multi-cell ChiP modules and current multipliers is the rare thing in power semis: a defensible IP position rather than a manufacturing-cost-down race.

Tape: $268.36, +70% / 30d, +553% / 12m, RSI 79. This one is the most parabolic of the picks-list. The good news is the move is grounded in real backlog conversion (rack-mounted current multipliers shipping into hyperscalers); the bad news is RSI 79 and the 12m comparison is a pre-thesis low. Don't chase. Wait for a 15-20% drawdown that would still leave it +400% on the year.

3. ETN (Eaton) — the rack-level integrator who sells the cabinet either way

Why architecture-agnostic: Eaton sits at the layer above the chip — power distribution units, switchgear, busbars, transformers for the data-center floor. Whether the rack inside is full-800V-DC or 48V-intermediate, ETN sells the room that contains it. Pair-trade cousin to VRT (which is more cooling-and-rack and arguably more levered to the cycle but also more crowded). ETN is the slower, less-loved, less-parabolic name; the boring choice; the "infrastructure REIT-grade" cash flow of the cohort.

Tape: $425.55, +16% / 30d, +44% / 12m, RSI 63. By far the least stretched name on the picks list — RSI 63 is borderline still buyable, and 30d/3m moves are tame versus the chip names. The trade-off is obvious: less torque if the thesis goes vertical, but also a much softer landing if the cohort gets violently rerated.

Honorable mention — GLW (Corning): Glass-core substrates for AI accelerator packages came up in both today's research-trace (wave 2027-2028 replacing organic PCB substrates) and Nutty's Part 4 (the "last micrometer"). GLW is RSI 37, down 10% on the week — the only oversold name in the entire cohort. If you believe glass-core is real, this is your contrarian entry. Not in the top 3 because the thesis is one full architecture-shift away from monetizing; the others print revenue today.

Cross-paradigm convergence

Three names appear in all three paradigms: NVTS, ASML, and (by adjacency) MPWR/VICR. ASML is the one we already own conviction in. NVTS is the surprise — small-cap GaN-on-Si pure-play that has been a quadruple in 12 months and is now showing up independently from the article-derived list, the agent trace, and the architecture-agnostic question. NVTS is the cohort's "did three independent paradigms just point at the same name" signal. Coverage gap: NVTS is not yet in ai-power.json despite being in optical-supply-chain.json — should be added.

Sources

• Pairs with today's autonomous-research trace on AI power delivery / parasitic loss
• NuttyCLD parts 1-4 + side story

Action items (for future scan / triage)

• Add NVTS to ai-power.json (currently only in optical-supply-chain.json); it's a 3-of-3 cross-paradigm hit
• Add MPWR, VICR, POWI, AOSL, IFNNY, ON, STM, ETN to ai-power.json or a new power-semis.json watchlist — the "last micrometer" cohort is currently scattered between semis.json, ai-infrastructure.json, and optical-supply-chain.json
• Address coverage gap: AIXG, VEEC, MRAAY, BESIY (SiC/GaN epitaxy + hybrid bonding) need OHLC fetched and a watchlist home
• Consider a dedicated power-delivery perspective with bull (architecture-agnostic merchants compound) / bear (cohort RSI 70s-80s, single decel quarter rerates everything) scenarios

Sources

Price, RSI and trend figures read from the desk's validated daily scan summaries (summaries). No number in this note was computed in prose.