Every 800V rack NVIDIA ships in 2027 is worth roughly 11x more ON silicon than the 54V racks shipping today. Consensus still hasn’t put that math in its model.

The semiconductor cycle is technically still in a trough. ON Semiconductor’s (NASDAQ:ON) revenue is down 15% year-over-year, fab utilization sits around 68%, and the stock just ripped 21% in a single week and sits up 167% from its April 2025 low. Either the market has lost its mind, or it’s figuring out that the “EV recovery story” label everyone has been using for ON is the wrong label.

The setup nobody was pricing

ON makes the kind of chips that sit inside things most people never think about. Power management, sensors, voltage regulation. Historically, the bull case was electric vehicles. Every EV needs enormous amounts of silicon carbide to manage battery power, and ON built one of the largest SiC franchises in the industry. The EV slowdown over the past two years was a direct hit to that thesis, and the stock spent most of 2024 and 2025 bleeding.

Here is what changed. According to a research briefing reviewed by Wolf Financial this week, CEO Hassane El-Khoury laid out a specific framework at the Morgan Stanley TMT Conference in March. Today’s AI data center racks run at roughly 120 kilowatts of power. ON’s content inside one of those racks is approximately $9,500. The next generation of AI racks, which start shipping in the back half of 2027, will operate at close to 1 megawatt using an architecture called 800-volt direct current. ON’s content inside one of those racks jumps to approximately $105,000.

That is 11 times the silicon content, per rack, across a three-generation GPU transition.

Why 800-volt is happening

This is the part where the physics actually matters. NVIDIA’s next flagship rack system, the Rubin Ultra NVL576 (also called the Kyber rack), is expected to run at approximately 600 kilowatts per rack. At that power level, traditional 54-volt distribution stops working. Delivering 600kW at 54V would require roughly 11,000 amps of current moving through copper busbars, which is physically unreasonable.

Switching to 800V direct current reduces the current by about 15 times and cuts copper requirements by roughly 45%. It also removes several conversion stages, which is where ON’s specific chips come in. The research describes a vertical gallium nitride (GaN) process in the 700-to-1200 volt range as a technology no competitor can replicate at scale today. That is the technical moat.

ON formally announced its NVIDIA collaboration for 800VDC power solutions in July 2025. Early adopters are already building infrastructure ahead of Rubin Ultra availability. CoreWeave, Lambda, Nebius, Oracle Cloud Infrastructure, and Foxconn (building a 40-megawatt facility in Kaohsiung, Taiwan) are all referenced as ecosystem participants.

The early supply chain signal

Supply chains leak information before revenue does. Susquehanna analyst Christopher Rolland runs a quarterly silicon carbide channel survey, and the Q1 2026 edition flagged something specific. ON recorded one of the largest lead-time extensions in the entire dataset, more than two weeks quarter-over-quarter. Lead-time extensions on specialty power semiconductors are almost always the earliest fingerprint of real demand pull before it shows up in earnings.

Rolland described the aggregate revenue opportunity in 800VDC AI data center architectures as “multibillions of dollars.”

The trough math

ON did $6 billion in CY25 revenue, down 15% from CY24 and down 27% from its $8.25 billion peak in CY23. Q4 2025 showed real stabilization. Revenue of $1.53 billion landed in line with guidance, EPS of $0.64 beat consensus by two cents, and free cash flow hit a record $1.4 billion (a 24% FCF margin in the trough).

The company also recently exited approximately $900 million of non-core, low-margin revenue. That depresses the headline growth number but improves revenue quality.

Then the margin story. ON’s fabs are running at approximately 68% utilization, which creates roughly 700 basis points of underutilization charges dragging on gross margin. When factory loading recovers toward the low-80% target, those charges unwind. Management’s long-term gross margin target is 53%, about 1,500 basis points above current levels.

The AI revenue ramp

AI data center revenue specifically: $250 million-plus in CY25, doubled year-over-year. Guidance for Q1 2026 is “high teens” sequential growth, implying a run-rate approaching $300-350 million entering CY26. The research briefing lays out an estimated path to $1 billion by CY28 as the 800VDC cycle kicks in.

The mispricing

ON trades at roughly 20x forward earnings and 13.5x EV/EBITDA. Monolithic Power ($MPWR), a peer with meaningful AI data center exposure, trades at 42-52x P/E. Power Integrations ($POWI) trades at 38x. ON is priced at roughly half the multiple of its closest comparables.

Consensus expects approximately 40% annual EPS growth over the next three years, putting the PEG ratio below 1.0. Add a $6 billion share repurchase authorization (~26% of the current $32.6 billion market cap) and a 4.8% buyback yield, and the setup gets harder to ignore.

What to watch

The single biggest calendar event is a Financial Analyst Day in New York City on September 16, 2026. That is where management is expected to formally update long-term financial targets, show the vertical GaN and SiC JFET roadmaps with customer design-win visibility, and attach quantified revenue targets to the 800VDC opportunity. If management delivers credible line-of-sight to $1 billion or more in AI data center revenue by CY28, the re-rating case becomes very hard for the Street to ignore.

Q1 2026 earnings hit May 4.

Thanks for reading! For more updates throughout the week, follow @WOLF_Financial on X.

Benzinga Disclaimer: This article is from an unpaid external contributor. It does not represent Benzinga’s reporting and has not been edited for content or accuracy.