A GaNsta's Paradise

Scavenging Through the Valley of Death

TL;DR

• onsemi acquired a shuttered vertical Gallium Nitride facility for $20 million and inherited NexGen's full patent estate out of its bankruptcy in late 2024. This transaction represented a discount of more than 90% on the total private and public capital previously invested in the technology lineage. That acquisition serves as a practical case study for making bets on stranded hard-tech capabilities.

• These bets operate at the intersection of technology commercialization and options-pricing logic. They treat stalled startups as deeply discounted call options and use the commercialization hurdles as a search screen rather than a verdict.

• The four commercialization hurdles ask whether the technology works, whether a market exists, whether it offers a value proposition, and whether the surrounding ecosystem can support it. A stall on any one strands a technology in the so-called valley of death. For vertical GaN, three of the four hurdles are no longer purely speculative. The remaining hurdle is ecosystem support, and that now turns on whether native GaN substrate cost, quality, and scale can catch up.

• The underlying commercial market for vertical GaN was enabled by the migration of AI data centers to native 800V power architectures. This architectural shift created an opportunity for vertical devices to deliver a sharp value proposition within high-voltage intermediate bus converters.

• onsemi restarted the facility and began sampling vertical devices to early-access customers in October 2025. However, wide-scale commercial viability will depend on the surrounding ecosystem reducing native bulk substrate costs.

• Corporate leaders can replicate this approach in other areas by establishing a dedicated M&A function to patrol the hard-tech valley of death. This systematic process allows organizations to buy the right to participate in future markets by targeting advanced capabilities that failed because the market, value proposition, or ecosystem had not yet caught up. Success depends far more on the current discount rather than the ability to forecast an often non-linear external environment.

In 2024, a Gallium Nitride (GaN) fabrication plant in DeWitt, New York, sat dark. The cleanroom air filtration systems were silent. Sixty-six thousand square feet of advanced manufacturing floor space lay idle. Built in part with public funds under a New York State economic development initiative, the facility embodied fourteen years of vertical GaN power-device engineering.

The lineage began in San Jose in 2010 with Avogy, a venture-backed startup that consumed roughly $60 million in private capital before the runway ended. Out of the liquidation, the patent estate sold for $200,000 and became the basis for a second startup, NexGen. NexGen then raised $106 million in private capital and moved into a $90 million chip fab New York had already built for a previous tenant that walked away. By the time NexGen shuttered the DeWitt facility in late 2023, total industry-wide spending to commercialize vertical GaN power devices had reached about $350 million across these and other efforts.

In December 2024, with New York State likely eager to find a buyer, onsemi wrote a $20 million check. For that, it inherited the advanced manufacturing footprint and the full NexGen vertical GaN power JFET patent estate. The price was a discount of more than 90% on the approximate quarter-billion dollars in private and public capital that had been poured into the lineage.

NexGen was not the only one. An hour away, in Cornell University's technology park, Odyssey Semiconductor had been pursuing vertical GaN on a separate track. It exhausted its capital on the same timeline and sold its assets to Power Integrations in 2024.

Deep hard-tech assets are essentially worthless until they are not. Their value depends on downstream market events and an ecosystem that may take years to arrive. The chain reaction is fragile and hard to forecast.

Venture funding and industrial-policy subsidies are poorly matched to that timeline. A hardware tech pioneer routinely runs out of money before the market it was built for shows up. Bankruptcy here is not necessarily a verdict on the technology. More often, it is a transfer mechanism. The value does not evaporate when the company does. It strands, waiting for someone to claim it.

The challenge is distinguishing durable stranded capability from entrepreneurial experimentation. The acquirers, if there are any, are making a particular kind of bet, one that sits at the intersection of basic technology commercialization frameworks (link, link) and options-pricing logic (link, link).

Technology commercialization needs to clear four basic hurdles. Does it work? Is there a market? Is there a value proposition? Can the surrounding ecosystem support it? Miss any one and commercialization stalls in the infamous "valley of death."

Options-pricing logic asks a different question. If the upfront cost is small and the downside is capped, uncertainty becomes an asset rather than a liability.

Strategic scavengers treat a stalled startup as a deeply discounted, long-dated call option. The bet is that the technology will clear all four hurdles before an expiration date set by the organizational hurdle rate.

There are many examples across the technology hardware industry, but vertical GaN power devices make a clean and timely case study. The technology languished, waiting for the market and a value proposition that would pull it into commercialization. Then the market arrived from a direction no one had foreseen. The migration of AI data centers to high-voltage power is just beginning, and it hands vertical GaN a chance to clear hurdles its pioneering startups never had. With the market suddenly real and a value proposition that can be measured, the remaining question lies in the surrounding ecosystem: the cost of the starting substrate.

The Market Arrives: Native 800V AI Server Motherboards

The business case for migrating data center power to high-voltage DC is already established. The definitive frameworks live in the primary-source materials from Delta, Nvidia, and Google presentations to the Open Compute Project. Whether the bus is ±400V split or native 800V, the topology converges on one principle. Energy should be delivered at the highest voltage for the longest distance, converting as few times and as close to the silicon load as possible.

The roadmap is the industry walking toward that end state one phase at a time. In the first phase, facilities retrofit separate sidecar racks that convert AC into 800V DC and cable it to top-of-rack DC/DC converters. Within a product cycle, those top-of-rack DC/DC converters are engineered out, replaced by on-board intermediate bus converters (IBCs) that step 800V down directly on the AI server motherboard. Eventually, centralized solid-state transformers (SST) absorb the sidecar racks too, and the distributed conversion layer disappears.

The market for any single conversion phase over time spikes and then collapses. A box appears, ships in volume, and is engineered out one or two cycles later. The market looks painfully discontinuous because it is. Stack the phases on top of one another and the same picture resolves into a clean growth curve.

For SiC and GaN, the 800V transition is a windfall. Across most of the power chain, the two materials stay in their lanes. SiC power devices block higher voltages and pull heat out of the package more easily, which relaxes the thermal design around it. GaN has lower losses and switches faster, which shrinks the inductors and capacitors around it. As a result, SiC will own the multi-kilowatt facility entrance, and GaN will own the low-voltage point-of-load near (or on) the processor.

The battleground sits squarely in the middle node, where the IBC steps 800V down on the AI server motherboard. That motherboard is the most expensive real estate per square centimeter the electronics industry has ever built. The power stage must convert six to ten kilowatts. Minimizing this footprint is a straight commercial imperative. Every square millimeter reclaimed from a smaller IBC can be handed back to high-margin AI compute.

HVDC IBC Prototypes at Computex 2026

SiC clears the 800V blocking requirement without trouble, but its lower frequency ceiling keeps the surrounding passives bulky and expensive. GaN can reach the megahertz speeds the application wants, but today’s commercial devices are still constrained at the high-voltage input. The transition to native 800V motherboards, colliding with the limits of today’s mainstream GaN, is the catalyst for the chain reaction that will drive demand for vertical GaN.

The Value Proposition for Vertical GaN

Conventional GaN power devices are lateral. Current runs sideways across a thin sheet at the surface, in a two-dimensional electron gas that forms in the GaN layer. At low voltage, that geometry works. It switches fast and conducts with very low loss.

At higher voltage, the geometry turns on itself. To block more voltage, a lateral device has to stretch the distance between source and drain. More distance means a longer travel path for the electrons across that thin sheet, which means higher on-resistance, higher leakage, and higher loss. The die area spent to survive the voltage erases the conduction advantages that made GaN worth using.

To get around the ceiling, GaN vendors have resorted to stacking two or three lower-voltage power stages in their reference designs, splitting the 800V bus across them. The same ceiling is a quiet argument behind the Open Compute Project's ±400V split-bus architecture, which halves the voltage any single switch has to stand off.

But stacking devices adds complexity to an already delicate powertrain, and power electronics designers are paid to optimize for efficiency, cost, reliability, and risk. At least for now, the first wave of native 800V IBC prototypes suggests many are not buying the workaround. Contrary to the reference designs from GaN device vendors, the primary-side switching socket in these first prototypes has gone to SiC, not GaN.

Vertical GaN makes the workaround unnecessary. Rotate the current path ninety degrees so it flows down through the chip instead of across it, and voltage blocking becomes a function of how thick the material is rather than how wide it is. Current is no longer confined to a thin electron gas layer on the surface. It can now spread across the whole cross-sectional area of the device, erasing the lateral penalty at high voltage. The device holds off the full bus and still switches at megahertz speed, with no stacking and no split-bus gymnastics.

The Remaining Ecosystem Challenge

Vertical power switches are already standard in silicon and SiC, and the basic device architecture for vertical GaN had been demonstrated years ago. After acquiring NexGen, onsemi restarted the DeWitt facility and in October 2025 began sampling 700V and 1200V vertical GaN devices to early-access customers. That moves the case beyond laboratory possibility. It suggests that the device works well enough to sample, that customers are willing to evaluate it, and that the value proposition is no longer purely theoretical. But sampling only proves the device is available and there is interest. The remaining hurdle is cost, and that problem begins one layer back, with the native GaN starting substrate.

Lateral GaN power devices are fabricated from epitaxial layers grown on silicon. The mismatched lattices require a highly resistive buffer layer to manage strain. They also leave a minefield of defects at the epitaxial boundary. A lateral device tiptoes around it, hugging the clean top layer where the current already wants to run.

Vertical GaN power devices operate by sending current straight down through the epitaxial stack. On a silicon substrate, that current would run into the buffer layer and the defect structure created by the lattice mismatch, ruining device quality. As a result, vertical GaN power devices must be fabricated on high-quality native GaN substrates. GaN epitaxy grown on native GaN eliminates the silicon mismatch and the need for a strain buffer layer, reducing the defects that compromise device performance.

But native GaN wafers are expensive, and that is the roadblock that may yet stall commercialization. It is also the part of the value chain that can potentially unlock onsemi’s acquired assets.

The economics of the AI motherboard socket set the ceiling for native GaN starting substrate cost. A back-of-the-envelope calculation is enough. It shows how much value the IBC socket can support, and therefore how much room vertical GaN has to absorb substrate cost. We use the more conservative, lower-cost, two-stage lateral GaN case as the baseline because, despite the first-generation SiC prototypes, future iterations are likely to migrate toward GaN.

A 10-kilowatt 800V IBC built on two stacked 650V lateral GaN input stages needs four discrete power devices, two devices per stage. The baseline reference device content is about $25 per IBC at volume. Scaled to the equivalent single-stage vertical GaN device area of 0.12 cm², or 0.06 cm² per device, that translates to a baseline reference value of about $200/cm² of yielded vertical GaN die.

Vertical GaN adds value beyond direct substitution. Because it can block the full 800V bus, it requires only one input stage, simplifying the whole power architecture. It reduces overhead and complexity. We estimate vertical GaN could eliminate about $10 of component overhead per IBC, or $80/cm².

A strict accounting should discount value for customer adoption risk. But on an ultra-premium AI motherboard, the physical real estate reclaimed by a single-stage design offsets some or all of that penalty. For this back-of-the-envelope model, we treat the two effects as offsetting, leaving a total economic value for vertical GaN of about $280/cm².

At low volumes (there is no high volume application today), a 100mm hydride vapor phase epitaxy (HVPE) GaN wafer commands roughly $5,000 to $8,000 ($64 to $102/cm²). The primary bottleneck is substrate defect density, which depresses processing yields and elevates reverse leakage current. Yet early 1200V work suggests HVPE material quality is no longer an impossible barrier.

If the supply chain can successfully scale, starting substrate costs sit right in the ballpark of economic viability. Volume deployment will then trigger a virtuous cycle, driving both material purity improvements and aggressive cost deflation down the learning curve.

It is just like SiC. The key to unlocking value begins at the substrate. We made that case before in SiC and EVs [link, link]. The conclusion was that the SiC supply chain could not profitably serve the EV opportunity at its existing cost structure, and that vertical integration would be the only durable path. Two years later, that assessment has aged well. BYD proved the captive model. Bosch moved upstream. Wolfspeed filed for bankruptcy under the old one.

But different forces are shaping vertical GaN. The pull from AI is sharper than the EV pull that shaped SiC, and the value created at the motherboard is high enough to support premium pricing. The supply chain is also far narrower. High-quality native GaN substrate supply is still concentrated among a small group of entrenched suppliers, including Sumitomo Electric, Mitsubishi Chemical, and SCIOCS. The first device maker to win the socket may capture the early market.

The opportunity is straightforward to estimate. At $3.5/kW across 46 GW of phase 2 and phase 3 AI compute deployed during 2030, the vertical GaN opportunity in the IBC socket is about $160 million. Small, but not insignificant, particularly if one supplier enters with a 90% discount on the investment required to build the capability.

The opportunity expands when 800V conversion moves closer to the AI processor. As the IBC closes that distance, more of the energy storage requirement shifts from the low-voltage side to the high-voltage side, pushing the bandwidth requirement of the IBC higher. Vertical GaN, switching higher voltages at higher frequencies, is the most direct device-level solution. The shift happens when the die area needed to carry the current exceeds the safety clearance needed to route the high voltage across the motherboard. As each generation of AI processors draws more power than the last, that tipping point sits not far past 2030.

Scavenging as Strategy

The onsemi deal happened at the device technology level, and that is where the attention is focused. Three of the four hurdles are no longer purely speculative. The device has been demonstrated, the 800V AI server market is in focus, and the IBC value proposition is sharply defined.

The remaining hurdle lies where it has always been for these WBG materials: one level back in the supply chain ecosystem, at the substrate. To walk through the valley of death, onsemi will most likely have to partner with one of the major GaN substrate makers and make a companion bet on scaling substrate supply. That is not a bad prospect for a company that paid almost nothing for assets that took fourteen years and a quarter-billion dollars to build. The same barrier that delayed commercialization may ultimately become the barrier that protects it.

Technology commercialization frameworks are usually applied as a filter, asking whether an innovation will succeed. They are equally useful as a search, a way to discover stranded, bargain assets. Searching this way requires broad cross-domain exposure and deep in-domain expertise. It requires making bets with incomplete information on where demand will migrate, which value propositions will become measurable, and which parts of the ecosystem will become strategic chokepoints. The discipline is pricing the asset against what is possible for it, not against a forecast of when the world will arrive. Organizations built on standard operating procedures rarely do this well.

The pattern is not unique to vertical GaN. When Beacon Power's flywheel business went through Chapter 11 in 2011, Rockland Capital bought the plant and most of the company for a fraction of its build cost, months before a FERC rule change began paying fast-response resources what they were worth. When Northvolt collapsed in 2025 after raising some $15 billion, Lyten acquired the gigafactories, the laboratories, and the remaining intellectual property at a small fraction of the capital that built them, along with the engineers who knew how to run the lines. There are others. The common thread is timing. Scavenging works when a buyer can see which stranded capabilities may be revalued as the market, value proposition, or ecosystem finally catches up.

The lesson for corporate leaders is to build a disciplined M&A function. Patrol the valley of death for technologies that failed from bad timing, not bad technology. Ask what kind of market, value proposition, and ecosystem support it will take to justify the purchase. The job is not to know the future. It is to buy the right to participate in it before the price reflects it.

If you find these posts insightful, subscribe above to receive them by email. If you would like to learn more about our consulting practice and how we assist organizations in the Tech hardware supply chain, please get in touch with us at info@rcdadvisors.com.