Empower Semiconductor: What ADI Bought for $1.5 Billion

Figure E.1

Where accelerator power delivery physically sits, and where it is moving. My description, not an industry-standard taxonomy.

01The deal

Analog Devices announced on May 19, 2026 that it would acquire Empower Semiconductor for $1.5 billion in cash. The deal closed on July 7, 2026. ADI's August 19, 2026 Form 10-Q confirms both the close date and the approximately $1.5 billion purchase consideration. Public disclosures do not indicate material antitrust or CFIUS complications during the seven-week gap between announcement and close.

Empower had raised approximately $215 million across its life per Tracxn and Crunchbase records. The most recent round, a $140 million-plus Series D in September 2025, was led by Fidelity Management and Research, with participation from Maverick Silicon, CapitalG, Atreides Management, Socratic Partners, Walden Catalyst Ventures, Knollwood, and an Abu Dhabi Investment Authority subsidiary. Barclays served as placement agent. The purchase price is roughly 7x Empower's cumulative external funding, which is a useful transaction reference multiple but is not the same as a 7x return for any specific shareholder. Actual investor returns depend on entry price, liquidation preferences, dilution across rounds, secondaries, and employee equity treatment, none of which are publicly disclosed.

The core investment thesis is architectural, not incremental. Multi-kilowatt AI processors are forcing more of the power-delivery chain toward the compute package. The companies that own the integrated combination of voltage regulation, magnetics, silicon capacitance, and advanced packaging at that migration point capture the position for the next architectural cycle. ADI paid $1.5 billion for the most commercially advanced entrant at that position, plus optionality on a broader power-delivery franchise if the migration hardens across the AI accelerator market.

A rough forward-revenue frame. If Empower's Crescendo platform captures $50 to $200 of content per next-generation accelerator (illustrative range derived from module count, integrated capacitance, and comparable content-per-part economics at nearby layers), and if next-generation accelerator volumes reach three to five million units per year by 2028, the Crescendo revenue could imply an annual run-rate of roughly $150 million to $1 billion. At $1.5 billion, ADI is paying 1.5x to 10x that projected run-rate range, depending on which endpoint holds. The wide multiple reflects genuine uncertainty about content capture and volume rather than a precise price signal. The revenue base from ECAP silicon capacitors (already in mass production per Empower's February 2026 announcement) sits on top of that as a separate commercial line.

02The Empower stack

Empower Semiconductor was founded in 2014 by Tim Phillips, Gene Sheridan, and David Lidsky. Phillips remained CEO through the acquisition. The company is headquartered in Milpitas, California. Its product line groups three closely-related product families under a common FinFast platform architecture.

FinFast platform. The underlying integrated voltage regulator technology. FinFast combines power switches, control loop, and drivers on a single die with dozens of components that would normally sit as discrete parts on a printed circuit board. Empower describes FinFast as a co-designed stack across power architecture, FinFET power device design, advanced power packaging, integrated magnetics, and silicon capacitors. The claimed footprint reduction is around 10x versus a comparable discrete design. Quad-output, low-voltage (1.8V-class), and higher-current variants have shipped over the last four years.

ECAP silicon capacitors. Wide-bandwidth, low-impedance decoupling capacitors fabricated in silicon rather than as ceramic or tantalum discrete parts. ECAPs can be substrate-embedded, die-side mounted, or land-side mounted, with sub-5-picohenry equivalent series inductance and bandwidth extending into the multi-gigahertz range. Empower announced on February 10, 2026 that the ECAP product family was in mass production, with catalogue parts including the EC2005P (9.34µF), EC2025P (18.68µF), and EC2006P (36.8µF). The ECAP line is Empower's most commercially mature product and the closest thing to a proven revenue base within the acquired asset.

Crescendo. The kilowatt-class vertical power delivery platform announced in the run-up to the Series D and the specific platform at the centre of the strategic thesis. Empower's May 2026 Advanced Power Modules brochure (available at empowersemi.com/wp-content/uploads/2026/05/Empower_APM-Brochure_May2026_spreads_digital-opt.pdf) describes Crescendo as a system combining multiple FinFast IVR dies with ECAP silicon capacitance in a single package, mounted on the compute substrate beside the accelerator die. Vendor-published claims from that brochure include 2,600+ amperes peak current, 65 amperes per power leaf, greater than 90% efficiency, approximately 1 degree Celsius per watt thermal performance, 2.2 millimetres ultra-low package profile, backside-cooling compatibility, PMBus and AVSBus digital control, and a published 500-ampere load-step transient example resolving in 450 nanoseconds. Empower has revised these specs across dated releases (an October 2024 launch cited "3,000A+ power domains" and "10% loss reduction"; a September 2025 OCP announcement led with "20x faster response"; the May 2026 brochure is the current set). These are vendor-published performance claims, not independent measurements.

The acquisition thesis needs to hold all three product families at once. ECAP is generating revenue today from customers that use silicon capacitors independent of Empower's IVR modules. Crescendo carries most of the strategic optionality on where kilowatt-class AI power delivery ends up, but its production ramp is largely ahead. FinFast underneath both is the co-designed platform IP that makes the rest of the stack difficult for a competitor to reproduce.

03Why on-package power delivery is the architectural question

Two physical constraints push the architectural progression: aggregate current at the die and transient response speed. Both worsen with each generation of compute silicon.

Current at the die. My modeled estimate for a Blackwell B200-class GPU is roughly one thousand amperes at the die at core voltages below one volt. Next-generation accelerators are projected at fifteen hundred to twenty-five hundred amperes at similar voltages. As aggregate accelerator current moves into the one-to-two-kilo-ampere range at sub-volt rails, board-level distribution becomes increasingly difficult to manage within practical thermal, area, and loss budgets. The exact threshold at which board delivery becomes infeasible depends on the specific architecture, because different topologies see different currents at different points in the distribution chain. Factorised approaches transform voltage upstream so the current seen on any single distribution segment is smaller than the current delivered at the load. Non-factorised approaches carry the full processor current on the last-hop copper. What is true across all topologies is that the case for moving conversion and current multiplication closer to the package strengthens as processor power draws into the multi-kilowatt regime.

Transient response. Modern accelerators change power draw at microsecond timescales as workloads switch across parts of the die. A regulator centimetres away cannot respond to that transient without bulk capacitance sized to absorb the excursion, and the bulk capacitance itself takes board area. A regulator millimetres away with integrated silicon capacitance responds faster and needs less bulk capacitance.

The Marvell collaboration announced in June 2025 targets integrated power delivery for four-kilowatt-plus processors. That number matters. Four kilowatts is where the architectural pressure to move voltage regulation, magnetics, capacitance, and current multiplication materially closer to the processor becomes difficult to ignore. It is also where the acquisition price for a commercially advanced on-package player becomes economically sensible: the buyer is paying for an architectural option on on-package power at the moment on-package is becoming decisive.

04The competitive field on-package

Four commercially relevant approaches are the focus of this analysis at the on-package position. Three are integrated-voltage-regulator startups. The fourth is Vicor with a factorised current-multiplier topology that mounts at the same physical location but requires an upstream regulator. Two additional Marvell PIVR ecosystem partners (Endura Technologies and Photeon Technologies) target the same on-package position with less-detailed public technical positioning. What separates these approaches is topology and the degree of silicon and packaging integration, more than physical placement.

Empower (now ADI) uses a system-in-package integrated voltage regulator architecture in a low-profile power-delivery package positioned vertically underneath or immediately adjacent to the processor, depending on implementation. Crescendo integrates regulators, magnetics and capacitors in a vertically coupled package and scales through multiple power leaves. It appears to be the most commercially advanced of the three IVR startups by funding base ($215M cumulative) and by disclosed customer engagement (the June 2025 Marvell collaboration is the only publicly named strategic partnership among the three). Architecturally the SiP approach is the most conservative among the three IVR options: highest die count per module, largest physical footprint, but the shortest path to production because system integration is well-understood.

Ferric plays chip-scale IVR. The Fe1766, launched in August 2025, delivers 160 amperes from 35.5 square millimetres of silicon with an integrated inductor on the same die. That is roughly 4.5 amperes per square millimetre of current density with the inductor integrated. Ferric says the device delivers 3x the power density of competitive solutions. Ferric's cumulative funding is approximately $41 million per PitchBook records, including the $32 million Series C in December 2024. The company is smaller and newer to commercial scale than Empower, and its architecture is more integrated on a per-die basis.

PowerLattice plays a highly integrated power-delivery chiplet. It emerged from stealth on November 17, 2025 with a $25 million Series A led jointly by Playground Global and Celesta Capital, bringing total funding to $31 million. Its Rainier architecture is a single silicon die integrating on-die magnetic inductors, power transistors, control circuits, and a programmable software layer, designed for integration into the processor package via substrate embedding, interposer, or land-side mounting. PowerLattice claims its architecture reduces total compute power needs by more than 50% and effectively doubles processor performance. Per press coverage of the Series A (including TechCrunch and semiconductor press), the company has silicon in hand and engineering samples in progress at TSMC for kilowatt-class GPUs, CPUs, and accelerators. No shipping product or publicly named production customers.

Vicor plays a different topology in the same last-hop physical real estate. Vicor's Vertical Power Delivery uses its Geared Current Multiplier (GTM) module mounted directly beneath the processor, taking a regulated intermediate voltage (typically 48V) from a Pre-Regulator Module (PRM) placed elsewhere and stepping the voltage down while multiplying current at a fixed transformation ratio (K=48 to produce nominal one-volt output, with current gain greater than 40x). Vicor's second-generation VPD targets approximately 3 A/mm² current density at 1.5mm module thickness, with up to 40x current multiplication. Vicor's approach is factorised: the regulator and the current multiplier are separate physical devices, so system integration requires the upstream PRM plus the under-package VTM/GTM stack.

Endura Technologies and Photeon Technologies round out the Marvell PIVR ecosystem. Endura (CEO Dr. Massih Tayebi) describes its approach as digital power delivery designed for integration into next-generation 2.5D and 3D AI system architectures. Photeon markets an Integrated Core Power (iCP) technology. Neither company has published product-level specs comparable to the four contenders above, and neither has disclosed customer wins outside the Marvell ecosystem announcement. Both are worth tracking as additional on-package participants rather than currently framing as headline competitors.

The Marvell package-integrated-voltage-regulator (PIVR) announcement in June 2025 named six ecosystem partners: Infineon, Monolithic Power Systems, Empower Semiconductor, Endura Technologies, Ferric, and Photeon Technologies. That six-partner list is a useful signal about how customers actually think about this position. Marvell is qualifying multiple package-integrated approaches in parallel so its custom-XPU customers can select the topology best matched to specific design requirements, rather than backing one architecture and betting against the others. The six-partner ecosystem is consistent with a plural on-package market and with a multi-supplier qualification strategy at Marvell. It does not, by itself, establish that multiple architectures will survive at production scale. What it does weaken is the "one architecture wins" narrative that would justify pricing any single vendor as the certain incumbent.

The topological distinction between IVR and factorised approaches matters. An IVR (Empower, Ferric, PowerLattice) integrates the final regulation function on-package: it takes a higher-voltage input and generates the processor rail close to the load. A factorised current multiplier (Vicor) separates regulation from current multiplication: an upstream PRM establishes the intermediate voltage while the VTM/GTM performs the final voltage transformation and current multiplication at the load. The trade-off is between eliminating a separate on-package regulator stage (IVR) and accepting an upstream regulator in exchange for a thinner and denser point-of-load module (factorised). Which topology wins in a given design depends on the specifics of that design. Both topologies are commercially viable, but public evidence of production-scale accelerator deployment remains much stronger for Vicor than for the three IVR startups.

The board-level incumbents (MPS, Infineon, Renesas, Texas Instruments) are increasingly extending their power portfolios toward the on-package position as accelerator currents grow. Their strategic relevance to this analysis includes both potential acquirers of the remaining on-package startups and direct participants at the on-package position through their own product development. MPS's and Infineon's presence in the Marvell PIVR ecosystem is one sign of that direct participation, though the specifics of their internal package-integrated roadmaps are not publicly disclosed.

05Head-to-head specifications

What each contender publicly claims. The metrics are not directly comparable across the four because the products solve different physical scales and use different topologies. What the table is good for is showing where each is willing to be measured.

Metric	Empower Crescendo	Ferric Fe1766	PowerLattice Rainier	Vicor VTM/GTM
Topology	Integrated regulator (SiP)	Integrated regulator (chip-scale)	Highly integrated power-delivery chiplet	Factorised current multiplier
Requires upstream regulator?	No	No	No	Yes (PRM)
Peak current (vendor claim)	2,600A+ (module), 65A per leaf	160A per die	Not publicly disclosed	Up to 40x current multiplication per module
Current density	5x higher than baseline (vendor claim)	~4.5 A/mm²	Not publicly disclosed	>3.0 A/mm²
Package profile	2.2mm, backside-cooling compatible	Fits within processor package	Substrate/interposer/land-side mounting	1.5mm thin, mounts under processor
Efficiency / loss claim	>90% efficient, 20% power-loss savings	Not publicly disclosed	>50% total compute power reduction	Low PDN loss emphasis, not quantified publicly
Transient / bandwidth (claim)	20x higher bandwidth; 500A/450ns example	Order-of-magnitude vs multiphase (vendor claim)	100x faster than DC-DC (vendor claim)	Low output ripple emphasis, not bandwidth-quoted
Digital control	PMBus, AVSBus, telemetry	Not publicly detailed	Programmable software layer	Not publicly quoted
Commercial status	ECAPs in mass production since Feb 2026; Crescendo in final sampling with mass production disclosed for late 2025; Marvell collaboration public	Product launched Aug 2025, Marvell PIVR ecosystem member	Engineering samples at TSMC, pre-revenue	Shipping in disclosed AI accelerator designs
Cumulative funding raised	~$215M	~$41M (PitchBook)	$31M	Public company (NASDAQ: VICR)
Status	Acquired by ADI Jul 2026 at $1.5B	Independent	Independent	Independent public

The specs are not directly comparable across contenders because each solves the same physical constraint through a different topological or integration path. Empower's Crescendo delivers a full accelerator's worth of current in a single system-in-package. Ferric's Fe1766 delivers a fraction of that per die at higher per-die density, deployed in multiples. PowerLattice's monolithic approach targets chiplet-level integration inside or under the substrate. Vicor's factorised approach delivers current multiplication at the load with an upstream regulator handling the regulation function. Multiple designs will use multiple contenders in combination, not one as a universal winner. The Marvell PIVR ecosystem's six-partner list (Infineon, MPS, Empower, Endura, Ferric, Photeon) is the clearest external evidence that customers already treat this as a plural rather than winner-takes-all position.

Figure E.2

FinFast platform: the five-pillar co-design that constitutes Empower's actual moat.

06The moat: architectural co-design, not any single component

Press-cycle presentations frame the moat around individual technical differentiators (FinFET switches, silicon capacitors, integrated magnetics). Reading the underlying platform and customer engagement, the more accurate view is that the moat sits in the co-design across those five pillars rather than in any single one. My assessment:

Silicon capacitor manufacturing capability. ECAP is in mass production per Empower's February 2026 announcement, with a product family covering multiple capacitance values (roughly 9µF to 37µF) and packaging options for substrate embedding, die-side, and land-side mounting. Ferric and PowerLattice do not have equivalent silicon capacitor portfolios. Building an equivalent capability from scratch would likely require multiple product cycles. This is the closest thing to a standalone defensible asset in the acquired stack.

Integrated magnetics IP. The thin-film magnetic inductor design embedded in FinFast is patent-protected. The patent portfolio has some strength, though the space is contested and Ferric has its own integrated inductor patents. This is a defensive asset but not a durable moat against a competitor with equal engineering depth.

Customer qualification and named partnerships. The Marvell collaboration announced in June 2025 targets pre-validated, packaged IVR solutions for Marvell's custom-XPU customers on the four-kilowatt-plus processor class. That is the only publicly named on-package customer partnership I found among the three IVR startups. Empower described its Series D as reflecting "depth of customer adoption," though public disclosures do not identify the specific end customers or programs. If Crescendo is qualified into current-generation programs (a reasonable inference from the Marvell collaboration and Series D language, but not a demonstrated fact), displacing a qualified vendor at mid-program is expensive and rare.

The FinFast co-design across five pillars. This is the strongest structural moat. Ferric's chip-scale approach is architecturally more integrated at the IVR die level but lacks the silicon capacitor line. PowerLattice's highly integrated die-level approach is architecturally deepest of the four but is pre-revenue and has not publicly demonstrated equivalent production history in packaging, magnetics integration, or silicon capacitors. Vicor operates through a different topology entirely. Reproducing Empower's specific five-pillar co-designed platform, at the specific yield and cost points needed for hyperscaler qualification, is a multi-year build for any competitor. This is the moat that justifies the acquisition premium.

What is not defensible in isolation. The SiP architecture itself does not confer durable protection. If Ferric, PowerLattice, or Vicor reach commercial scale with a topologically-cleaner alternative and get qualified into a subsequent silicon revision, Empower's specific architectural position becomes the "old on-package" while the frontier moves elsewhere. The moat protects the platform, not the specific Crescendo product form factor.

07The two acquired businesses: ECAP and Crescendo

The acquired asset splits cleanly into two commercial businesses with very different maturities.

Story one: the ECAP shipping business. Silicon capacitors in mass production since February 2026, with a catalogue of parts across multiple capacitance and packaging variants. This is a real commercial line targeting the power-integrity layer of advanced packaging (substrate embedding, die-side and land-side placement, sub-5-picohenry ESL, multi-gigahertz bandwidth). ECAPs also address a distinct power-integrity problem that any high-performance package needs to solve, and therefore have potential applicability beyond the specific Crescendo integration path. Public disclosures do not provide enough customer detail to quantify how much of that opportunity is independent of the Crescendo roadmap. This is the near-term revenue foundation of the acquisition.

Story two: the Crescendo architectural option. Kilowatt-class vertical power delivery targeting the four-kilowatt-plus processor class, with the Marvell collaboration as the specific customer anchor. This is the strategic bet on where AI accelerator power delivery ends up architecturally. It is early commercial (Series D language references "depth of customer adoption" without disclosing the customers) but the addressable revenue if the architectural migration hardens is the multi-hundred-million-dollar-to-multi-billion range depending on scenario. The Crescendo half of the acquisition is where the leverage is.

ADI's cross-sell opportunity applies to both stories. ECAPs potentially extend into ADI's existing power-management customer base, while Crescendo can layer onto ADI's broader signal-chain, isolated gate driver, and PMIC relationships in AI infrastructure accounts, packaged as a broader "grid-to-core" offering. ADI's acquisition announcement explicitly framed the deal as expanding its strategic grid-to-core position and its AI compute power-delivery TAM through both integrated voltage regulators and silicon capacitors. Whether ADI can actually integrate Empower's product roadmap and engineering culture with its larger-company execution model is a standard M&A execution question. The Maxim and Linear acquisitions provide institutional playbooks but not guarantees.

What ADI did not buy: architectural leadership at the frontier. Ferric's chip-scale approach and PowerLattice's monolithic approach both push per-die integration further than Empower's SiP. Vicor's factorised topology contests the same physical position with different economics. ADI now appears to own the most commercially advanced of the independent on-package IVR platforms. It does not own the next-generation architecture. The follow-on strategic move likely involves acquiring one of the remaining startups or funding internal R&D to hold the frontier through the next silicon revision cycle. Whether ADI can afford or negotiate that follow-on depends on how MPS, Infineon, and Renesas play their own strategic hands over the same window.

08Three scenarios for on-package power delivery through 2030

Three shapes the market could take between now and 2030. These are analytical scenarios, not probability-weighted forecasts, and they are not mutually exclusive.

Base case: Marvell ecosystem model generalises. The on-package position remains a plural market with multiple qualified topologies. Empower/ADI captures the largest single share of on-package IVR revenue via Crescendo and the ECAP power-integrity business. Ferric and MPS split most of the balance across specific programs. Vicor holds its shipped-into-accelerator content and grows it modestly. PowerLattice stays commercially early through the window. Endura and Photeon capture minor content in specific PIVR-qualified sockets. Under this case, ADI's $1.5B looks well-priced because ECAP provides a tangible commercial base against which the more speculative Crescendo option can be valued, and Crescendo captures the largest slice of the growing IVR pool.

Split-architecture case: volume programs use SiP or factorised, leading-edge programs use monolithic. Empower's SiP and Vicor's factorised current multiplier win the volume-oriented programs (mainstream Nvidia SKUs, most hyperscaler custom silicon). PowerLattice or a similar monolithic architecture wins the highest-density, most power-constrained programs (leading-edge Rubin, most-aggressive custom silicon at Microsoft or Google). Ferric occupies a middle position on specific chiplet-level implementations. Under this case, ADI needs to acquire PowerLattice or an equivalent to hold the leading-edge segment, and the $1.5B is a down payment on a larger power-delivery franchise.

Frontier-collapse case: monolithic reaches commercial scale faster than base-case timing. PowerLattice or an equivalent reaches commercial scale by 2027-2028, gets qualified into a next-generation Nvidia or custom silicon program, and demonstrates an architectural advantage that customers cannot ignore. Empower's SiP becomes the "old" on-package solution within 24 to 36 months of the ADI close. ADI would need to acquire PowerLattice or accept that its $1.5B acquisition aged faster than the model assumed. Under this case, the winning play from a public-equity perspective was to own the monolithic startup at Series A pricing, not the SiP incumbent at exit.

09What the reference transaction does to the rest of the sector

The $1.5B ADI-Empower deal is a reference transaction for the on-package competitive landscape. Three specific consequences, each qualified.

Ferric's implied valuation shifts upward, though the transaction establishes a reference point rather than a valuation floor. Empower and Ferric would command comparable strategic premiums only if their commercial and technical assets are comparable, and they are not: Empower has ECAP mass production, a broader integrated platform, and a publicly named Marvell partnership. My illustrative range for a Ferric strategic exit today is $200 million to $500 million, but the specific number depends on customer traction, ECAP-equivalent product capability, current-density differentiation, Marvell ecosystem execution, and revenue that are not publicly disclosed.

PowerLattice at $31M raised has real optionality but is harder to price against Empower's transaction because PowerLattice's commercial position is genuinely different. Pre-revenue, more architecturally integrated, unproven at scale. Playground Global and Celesta Capital at Series A pricing hold optionality on a wide range of outcomes if PowerLattice demonstrates commercial viability in the next 18 to 30 months, but converting that optionality into a specific dollar valuation is not price-discoverable today.

MPS and Infineon just watched ADI pay $1.5 billion for a private power-semiconductor company at their addressable market. Both companies now face a strategic decision: acquire Ferric or PowerLattice (or both) defensively, or continue investing internally to compete with ADI's newly consolidated position. I would expect strategic acquisition interest in at least one of the remaining startups to increase materially following the ADI-Empower transaction. Whether that interest converts into a specific deal, and on what timeframe, is genuinely uncertain. Any follow-on acquirer is more likely to be MPS, Infineon, or Renesas than ADI, which has just completed the $1.5B Empower deal.

10What this changes for the AI Power Chain program

The reframed thesis is architectural rather than incremental. This is a bet that AI's transition to multi-kilowatt processors forces more of the power-delivery chain toward the compute package, and that the companies controlling the critical last-hop architecture (through either tighter integration or factorised conversion) are positioned to capture value as power moves toward the die. The bet is on where power is delivered and which architectural stacks capture that shift, not on any single component being better than an incumbent, and not on AI generally using more watts.

Six things this deal and the reframed thesis change for how I approach on-package power delivery.

The on-package position is becoming a four-way architectural contest, but not necessarily a four-way winner-takes-all market. Vicor's VTM/GTM competes for the same last-hop physical real estate as the three IVR startups, using a different (factorised) topology. Framing this as three IVR contests with Vicor sitting elsewhere undercounts the field. The right investment lens is: which topology (integrated regulator versus factorised current multiplier), and within regulator topologies which integration level (SiP versus chip-scale versus monolithic), wins in which programs. Marvell's PIVR ecosystem announcement is consistent with different accelerator programs selecting different sockets rather than a single architecture consolidating the whole on-package position. That is not the same as proving multiple architectures will survive at production scale, and I would want that tested with real qualification data over the next two silicon revision cycles.

The commercial evidence gap is what an investor should diligence hardest. Empower's ECAP line is in mass production. Crescendo has one publicly named customer partnership (Marvell). Everything else about production volume, program qualification, and content per accelerator sits behind non-disclosure and inference. The gap between excellent technology evidence and demonstrated production-scale evidence is the central diligence question, and my probability weighting reflects that uncertainty explicitly.

The ADI transaction is a reference point, not a valuation floor for Ferric or PowerLattice. The specific factors driving the $1.5B for Empower (ECAP mass production, integrated FinFast platform, Marvell partnership, silicon capacitor manufacturing capability) are not equally present at the other startups, so the two remaining startups will command different pricing at any similar strategic exit trigger.

ADI, MPS, Infineon, and Renesas are the four strategic acquirers to watch. Any could rationally acquire one of the remaining startups within the next couple of years. Timing and price are uncertain and I do not have public evidence for a specific transaction in either direction.

Vicor's competitive position needs careful reassessment. The earlier framing (Vicor at an adjacent 48V position, not disrupted by the Empower deal) was wrong. Vicor competes for the same last-hop physical real estate as Empower, using a topologically different approach. Whether the two are substitutes, complements, or program-dependent depends on the specific design in question. Vicor's disclosed content on current-generation AI accelerator boards is real and grew through 2024-2026. Whether ADI's newly consolidated Empower position expands or compresses Vicor's future content trajectory is a genuine open question.

Power delivery is already moving closer to the processor. Where the migration stops is the interesting question, and PowerLattice's chiplet is the clearest public place to look for the answer. Its architecture pushes integration further than Empower's SiP or Vicor's factorised VPD. If PowerLattice reaches production scale over the next 24 to 36 months, today's on-package modules (including Empower's Crescendo) become an intermediate generation, and ADI's $1.5B ages faster than its model likely assumes. Otherwise, ADI captured the right asset at the right time. Anyone underwriting ADI, MPS, Infineon, or Renesas exposure through 2028 should be diligencing that.