The 800 VDC Gambit: Wolfspeed, LITEON, and the Hidden Spec War Inside AI Data Centers
On paper, the pairing is elegant. Wolfspeed owns the most advanced silicon carbide substrate line in the Western world. LITEON ships power supplies that keep hyperscale racks alive. Together they are promoting 800 VDC as the next high-voltage direct-current bus for AI data centers. The blockchain remembers; the architect forgets. The world's first fully 8-inch SiC line, Wolfspeed's Mohawk Valley fab, was still running at an estimated 20–40% utilization in late 2024. Fiscal 2024 gross margin was negative. This is not a victory lap. It is a rescue mission dressed as a technology roadmap.
The AI rack power curve has moved from 30 kW toward 120–140 kW per cabinet with NVIDIA's GB200 NVL72 generation. At those levels, the traditional 48 V DC bus produces excessive copper losses and conversion overhead. Higher DC buses reduce I²R losses, and 800 VDC is the next step after 400 VDC. The catch is that 800 VDC is not a ratified standard; it is a pre-standard architectural bet. Wolfspeed supplies SiC MOSFETs capable of 1200–1700 V, while LITEON handles system integration and bus converters. The source material frames this as a component partnership. It is not. It is a movement to lock specification early.
Core analysis: Let me start where the gloss ends. Wolfspeed's real advantage is its vertical IDM structure: SiC substrate, epitaxy, device design, fabrication, and module packaging. A 6-inch to 8-inch substrate migration is the critical cost lever. Substrate cost accounts for 40–50% of total SiC device cost; moving to 8-inch reduces unit substrate cost by 40–45%. But that reduction only materializes if yield and utilization are healthy. The source report estimates Mohawk Valley yield climbed from 50–60% in early 2023 to roughly 85–90% by late 2024. Industry-wise, good yield means nothing at 20–40% utilization. Depreciation is fixed. With utilization below 60–70%, the fab cannot cover its own depreciation. That is why Wolfspeed's gross margin was negative in fiscal 2024, while STMicroelectronics and Infineon sit at 40–45%.
The financial picture is not a minor background detail. Wolfspeed's capital expenditures ran at 50–80% of revenue, against a typical 20–30% for the sector. Operating cash flow was consistently negative, around -300 to -400 million in fiscal 2024; free cash flow was even worse. This is a structure financed by debt and equity dilution. From my audit experience, I would not approve this as a going concern without a clear path to 60% utilization on the 8-inch line.
Now add the demand side. The AI data center segment is still less than 10% of Wolfspeed's revenue, but it is growing at triple-digit annual rates. LITEON's own AI power revenue doubled from 2023 to 2024. The opportunity is real. But it is not stable. Wolfspeed's revenue remains concentrated in industrial/energy and EV, which have been weak. LITEON's top five customers account for maybe 50–60% of revenue, concentrated in hyperscale cloud vendors. That concentration cuts both ways; a hyperscaler's design win can be huge, but a pause in cloud capex hits the PSU vendor first.
Competition is the other neglected variable. In SiC MOSFETs, ST holds roughly 25–30% share, with Infineon and Rohm close behind. Wolfspeed leads in substrate with 25–30%, but the device lead is narrower. Chinese suppliers like Sanan, SICC, and TankeBlue are ramping 8-inch SiC and cutting prices. The source report notes that SiC MOSFET prices fell 15–20% in 2024. That favors 800 VDC's eventual adoption, but it squeezes Wolfspeed's margins in the interim.
Supply-chain risk is fixed-asset risk. Wolfspeed's furnaces and high-temperature implant tools depend on Japanese and German suppliers. Chinese equipment makers are closing the gap for crystal growth, but 8-inch substrate equipment localization remains below 20 percent. That means the bottleneck is equipment availability, service contracts, and spare parts. An export-control expansion could slow capacity expansion even if it does not stop output.
The deeper insight is spec-in. When a power architecture is not yet standardized, the vendor that enters the reference design for Nvidia, AMD, or a cloud provider can set the physical and electrical interface. That means every subsequent power module must match that form factor and bus voltage. This is analogous to a settlement layer: first mover defines compatibility. LITEON is a tier-two power ODM relative to Delta's 40–50% share in AI server PSUs. Delta would be the natural incumbent for a next-generation option. If LITEON and Wolfspeed can get their 800 VDC design into one major reference platform, they will not simply sell components; they will install a toll booth.
But the bull case is not wrong. The physics is unforgiving. I have stress-tested power-stage designs for years, and the advantages of SiC over silicon IGBTs at 800 V are measurable: lower switching losses, higher operating temperature, and better thermal stability. GaN is strong below 650 V but weak at 800 V. There is no commercial alternative to SiC for this voltage class today. If AI rack power reaches 200 kW and beyond, 48 V buses become physically implausible. The transition to 800 VDC, or something close to it, is not a luxury; it is an engineering necessity.
The contrarian angle is about timing, not direction. Wolfspeed may not have enough runway to reach the standard. Its financial fragility—negative gross margin, negative operating cash flow, high leverage—makes it a weak partner in a zero-revenue standard war. LITEON is healthier, with 18–22% gross margins and positive free cash flow, but its revenue is tied to cloud capex cycles. If AI capex pauses for two quarters, the value chain that needs 800 VDC becomes a lab project. Meanwhile, Nvidia has every incentive to integrate power conversion into its own rack design, turning Wolfspeed and LITEON from partners into suppliers. A spec-in advantage can be expropriated by a platform owner.
There is another possibility that the bullish reading misses. The fact that LITEON, a conservative ODM, co-develops an unproven 800 VDC bus suggests that at least one hyperscaler has privately signaled demand. That signal is enough to justify a hedge, but not enough to justify Wolfspeed's current valuation. The market is pricing Wolfspeed's resurrection narrative, not its current earnings. That is apparent from a price-to-sales ratio of 1.5–2.5x while negative earnings make price-to-earnings meaningless.
The geopolitical layer adds noise but not a signal. Wolfspeed is not on the U.S. Entity List, and SiC devices are far less restricted than advanced logic chips. China's gallium and germanium controls do not directly affect SiC. However, an 800 VDC standard led by a U.S.-Taiwan pair would compete with China's 240/336 V HVDC standards used in telecom and some data centers. This is not a gunfight; it is a standards land grab.
Takeaway: The blockchain remembers; the architect forgets. In 2027, you will either be reading about 800 VDC as the data center settlement layer—or about a once-promising SiC pioneer that bonded itself to a voltage standard that never arrived. The signal to watch is not Nvidia's GB200; it is the power architecture of the next-generation Rubin platform. If that platform moves to 800 VDC, Wolfspeed and LITEON have a genuine moat. If it stays on 48 V or shifts to a different high-voltage bus, this collaboration will be remembered as an expensive architecture bet. I have seen this pattern too often before in many audits: the architecture is critical, the timing is fatal.