Vicor (VICR) - the AI-power name that isn't a utility, and the 48V bottleneck it owns

The standard AI-power trade is about generation. Where do the electrons come from: gas turbines, nuclear restarts, grid interconnects, the merchant power names like $VST and $CEG. That trade is real, and it is crowded.

$VICR sits one layer the story skips: the last inch of power delivery, from the 48-volt rack bus down to under one volt at hundreds of amps, at the die edge of the GPU. An NVIDIA NVL72 rack pulls roughly 120kW today; the Rubin-class roadmap points toward ~250kW. At those densities the legacy 12V distribution bus stops working, and how you get current the final centimeters to the silicon becomes a distinct engineering problem with a distinct set of winners. Vicor is one of the few production architectures that solves it, and it holds patents on the way it does. This piece walks through what Vicor actually makes, how it earns, where it sits in the stack, and the risks that make it a genuinely two-sided name.

What Vicor actually makes

When traders hear "AI power" they picture substations and cooling towers. Vicor is upstream of the socket and downstream of the rack: it makes the converter modules that turn a rack-level DC bus into the exact voltage and current a processor needs, as close to the processor as physically possible.

Its core idea is Factorized Power Architecture. Instead of one converter doing everything, Vicor splits the job: a regulation module (PRM) handles the control loop, and a current-multiplier transformation module (VTM) does the final step-down right at the load. The VTM can deliver 1500 amps or more at under one volt at the die edge, which is the regime modern accelerators actually run in. Packaging this as ChiP-scale modules (Converter housed in Package) lets a board designer drop power conversion into the tight real estate beside a GPU rather than routing high current across a board and losing it to resistance.

That is the whole pitch in one line: at 120kW and climbing per rack, the loss and the space you spend moving current the last few centimeters starts to dominate, and vertical 48V-to-point-of-load conversion is now the default GPU power topology rather than an exotic choice.

How they make money

Vicor sells components, priced per module, into boards that ship in volume once a design is qualified. Revenue is roughly $474M over the trailing twelve months (data as of 2026-09-11), which for a $9.1B company tells you the market is paying for the ramp ahead, not the trailing print.

The structure worth understanding:

• Design wins that ramp with each accelerator cycle. A power module qualified onto a GPU board is sticky, often sole-sourced, and rides that platform's volume. In 2024, 48V-direct-to-PoL ChiP modules were qualified into next-generation AI accelerator boards (the customer is unnamed, widely read as inside the $NVDA ecosystem).
• An in-house fab as a margin lever. Vicor brought its Andover, Massachusetts ChiP fab online across 2023-2024, its first high-volume in-house module manufacturing. Utilization is the swing factor: underused it weighs on gross margin, fully loaded it lifts the ceiling.
• A non-cyclical defense keel. Long-standing MIL-spec and VITA-62 power modules sit in radar, avionics, and satellite buses, revenue that does not move with the AI capex cycle.
• An optional IP-royalty leg. In Q4 2024 Vicor filed an ITC Section 337 complaint against Delta Electronics, Foxconn, and others, alleging infringement of its factorized-power patents, and announced a royalty-bearing licensing program. If that litigation lands, running royalties on rival 48V modules become a second income stream. It is binary and slow, so treat it as an option, not a base case.

Net margin runs around 31%, which for a component maker reflects the IP-heavy, module-level pricing rather than commodity silicon.

Where it sits: the last inch of AI power

On the site, Vicor's structural cluster is the Semi Equipment / Litho bubble (an extended-tier fit, not a pure play), and it now carries the AI Power theme and AI Hardware theme alongside Power Semis. That combination is the point: it is a semiconductor-adjacent hardware name whose demand driver is AI rack power, which is why it does not trade like the merchant-generation utilities it gets lumped with.

Its measured co-movement leans toward the semicap and precision-hardware complex rather than the power utilities: names like $LRCX, $TER, $AMKR, and $LAZR show up as its closest correlates, not $VST or the gas names. For the generation side of the AI-power question, see the data center power bubble; Vicor is the delivery side of the same electricity story.

The competitive frame matters. Merchant power-management peers MPWR, ADI, and TXN all sell 48V multi-phase voltage regulators and compete for the same board slots. Vicor's differentiation is the patented current-multiplier topology and the vertical-power packaging, not a price advantage.

The numbers

Two things stand out. First, the valuation is a growth multiple: at ~19x sales the market has priced a meaningful ramp, so the risk is as much multiple compression as business execution. Second, the return profile is violent in both directions. Up nearly 4x on a one-year view, then down a third off the recent high. That is a small-float name (~34M shares) levered to a design-win narrative, and it trades like one.

The bull case

• 48V-to-PoL ChiP design wins ramp with the next accelerator cycle, and each qualified board is sticky, sole-source volume.
• The ITC complaints against Delta and Foxconn-tier module makers could convert into running royalties, a second income leg on top of product.
• The Andover ChiP fab lifts the gross-margin ceiling as utilization climbs into the ramp.
• The defense and aerospace backlog is non-cyclical ballast under the AI-cyclical top line.
• Once a Vicor module is qualified into a GPU board, requalifying a competitor is costly and slow, which protects the socket.

The bear case

• The largest buyers are integrating power delivery in-house. Hyperscalers like $GOOGL and $META designing their own accelerators can design their own power stacks too.
• $MPWR, $ADI, and $TXN compete aggressively on 48V multi-phase VRMs and have deeper customer relationships in some sockets.
• Revenue is lumpy on single-customer concentration; one platform's timing can swing a quarter.
• Andover fab underutilization weighs on margins until the volume actually arrives.
• The IP-licensing thesis is binary and slow: an ITC outcome could take years and may not convert to material royalties.

How to access

$VICR is a direct NASDAQ listing. For US-retail access to the name, see /stack/ibkr.

For a diversified wrapper, it currently appears in the VistaShares Artificial Intelligence Supercycle ETF (AIS) at roughly 2.8% weight (as of 2026-08-05), so buying that fund gives fractional exposure alongside the broader AI-infrastructure basket. It is a modest position, not a proxy for the stock.

Bubble shifts, ETF-weight changes, and rule-based alerts on $VICR are part of /pro.

What to watch

• The ITC Section 337 ruling against Delta and Foxconn: any decision converts the IP leg from option to fact, in either direction.
• Andover fab utilization in the quarterly prints: the margin story is a utilization story.
• Design-win disclosures on the next accelerator platform, and any sign a large customer has moved power delivery in-house.
• The correlation cluster: Vicor trades with the semicap complex ($LRCX, $TER) today. If it starts trading with the merchant-power names instead, the market is re-reading what kind of company it is.
• On the tape, the site's auto-computed levels bracket the recent price around the $175 and $215 zones; those are the nearest observable structural levels, not targets.

Live data on this ticker: /stocks/vicr - price, ETF holdings, bubble correlation, bot positions.

Bubble context: /bubbles/semi-equipment - the cluster this name belongs to and how it's moving.

QuantAbundance is educational research. Nothing here is investment advice. See /disclosures.