The $213 Billion Mirage: Why onsemi's AI Power Pivot Is a Defensive Trade Dressed as Offense

The number doesn't survive first contact with the math.

$213 billion. That's the market size being floated for AI power in onsemi's strategic narrative. The figure appears in headlines, propagates through syndicated briefs, and gets pasted into investor decks without a single footnote.

Here's the arithmetic problem. The global power semiconductor market — the entire category where onsemi competes — generates roughly $50 to $60 billion in annual revenue. Every MOSFET. Every IGBT. Every SiC module. Every traction inverter in every electric vehicle, every appliance power supply, every industrial motor drive, every data center rectifier on the planet. All of it.

The "AI power opportunity" being attached to onsemi is three to four times larger than the entire category from which it is supposedly carved. You don't need a graduate degree in finance to see that this is either an ecosystem-level TAM — counting power racks, busbars, transformers, cooling infrastructure, the entire electrical spine of every data center ever built — or a multi-year cumulative projection, or both. It is not the addressable market for a company selling discrete power semiconductors.

Tracing the gas leaks before the code compiles, I found the rupture in the first line. The piece that launched this narrative came from Crypto Briefing, a cryptocurrency media outlet. Covering discrete power semiconductors. That tells you everything about the information supply chain: this is narration trailing a stock move, not analysis preceding a position.

What does onsemi actually do?

Formerly ON Semiconductor. Even further back, Motorola's semiconductor division. An IDM — integrated device manufacturer — operating in the power semiconductor lane. They design, fabricate, package, and sell the components that switch and condition electrical energy: trench MOSFETs, IGBTs, SiC MOSFETs, superjunction transistors, gate drivers, power management integrated circuits. Their historical center of gravity is automotive — roughly half of revenue. Industrial adds another quarter. Consumer, communications, and a growing but still small data center sliver fill the remainder.

In power semiconductors, onsemi is unquestionably first tier. The peer set is Infineon, STMicro, Wolfspeed, Mitsubishi Electric. They hold a top-three position in silicon carbide following the 2021 acquisition of GTAT, which bought them control over SiC substrate supply — generally the most important cost variable in SiC device economics. In automotive image sensors, they lead the world with over forty percent share. That is a formidable franchise by any measure.

The story now being told is different. AI data centers. NVIDIA's Blackwell-class GPUs exceed 1,000 watts per chip. Rack-level power is pushing past 100 kilowatts. The power delivery chain — 800-volt HVDC distribution, 48-volt intermediate bus, point-of-load conversion, vertical power delivery — is a genuine bottleneck. Microsoft, Google, Amazon, Meta: all power-constrained, not compute-constrained. That part of the narrative is real.

The strategic question is whether onsemi's pivot into this segment represents a structural growth transformation or a narrative response to structural decay in its core markets.

Watch the timing. Gross margins have compressed from roughly 48 percent in 2023 to around 45 percent currently. Capacity utilization sits near 70 percent — well below the 85 to 90 percent zone that keeps fixed-cost absorption healthy. Automotive demand has cooled. Industrial is deep in a destocking cycle. And Chinese power semiconductor champions — Silan Micro, CR Micro, CRRC Times Electric, Starpower — are systematically eating the mid and low end of the market with state backing and price aggression.

The AI power pivot arrives at a convenient moment. Convenient for the investor relations calendar. Convenient for the narrative. The market should ask whether it's also convenient for the P&L.

A note on scale. The segment onsemi actually competes in — power discrete and module revenue globally — sits in the tens of billions of dollars. AI power demand is a real increment to that. But the difference between incremental demand and an entirely new addressable market is the difference between a cyclical uptick and a new secular engine. The entire re-rating thesis hangs on that distinction.

The reported piece contained five information points. No executive quotes. No financial tables. No sourcing. For context, when I audited the Golem ICO distribution contract back in 2017, I found that the critical vulnerability — an integer overflow in the batch claim function — was only visible after parsing the assembly opcodes directly. The documentation said one thing; the executable said another. That's the discipline required here. Ignore the deck. Read the code. The code is the quarterly financial statement.

Now the core analysis. Seven dimensions. No filler. Let's work through each.

Dimension One: Technology — First Tier on Home Ground, Third Tier Where the Growth Is

Let's be precise about what onsemi is and is not.

onsemi is not a logic company. It doesn't compete with TSMC or Samsung at the leading edge. Its fabs run mature nodes: 0.18-micron to 40/65-nanometer BCD processes. The transistor portfolio is power-specific: trench MOSFETs, IGBTs, SiC MOSFETs, superjunction structures. No GAA. No FinFET. No EUV. Analysts who mechanically apply the "mature node equals weak technology" heuristic miss the point entirely. For power devices, the battle is not in critical dimension. It's in substrate quality, thermal management, parasitic inductance control, packaging, and reliability over a twenty-year operational lifetime. Different arena. Different weapons.

In that arena, onsemi has no generational gap. First tier, full stop. The SiC roadmap is advancing from the M1/M2 generations through M3, with 1,200-volt and 1,700-volt devices moving toward qualification. The 8-inch — 200-millimeter — substrate conversion is the critical cost inflection. Moving from 150mm to 200mm wafers reduces die cost materially. Because onsemi owns substrate production through GTAT, the yield learning curve accrues in-house rather than being paid to Wolfspeed or Coherent as margin. SiC substrate yield industry-wide sits in the 60 to 75 percent range. It is a brutal learning curve. And owning the substrate is a genuine structural moat. I cannot overstate this: in power semiconductors, the substrate is the supply chain.

Now cross into the AI data center power domain. That's a different skill stack entirely. Multi-phase controllers. Smart power stages with integrated current sensing. 48-volt-to-12-volt conversion. Direct 48-volt-to-core conversion. Vertical power delivery — VPD — where the power stage sits directly beneath the processor die to minimize path inductance at sub-milliohm levels. Here the gap shows.

onsemi's share of AI server power is below five percent. It is a challenger. The leaders are MPS — Monolithic Power Systems — which is structurally embedded in NVIDIA's reference designs, and Vicor, which owns the high-density vertical power delivery frontier. onsemi's module packaging for automotive SiC is strong; that's proven. But AI server power requires system-level design: controller, power stages, control algorithms, accurate current and temperature sensing, and integration within a GPU reference architecture. That is a different capability stack than supplying automotive power modules. The overlap is partial at best.

Let me draw on the Golem audit again. Four months parsing assembly opcodes taught me to distinguish between a system that works at the architecture-diagram level and one that survives contact with production. Same discipline applies here. onsemi's AI power roadmap looks credible on paper. Credibility is not adoption. Qualification cycles in hyperscaler power chains run eighteen to thirty-six months. Design wins must be earned board by board, generation by generation, often against incumbents already proven in the field.

My estimate: onsemi trails MPS and Vicor on AI-specific power architecture by one to three years. Not insurmountable. But real. And it means the AI power pivot is a multi-year program, not a near-term revenue event. Any analyst modeling hockey-stick AI revenue in the next two quarters is modeling a fantasy.

Dimension Two: Supply Chain — The Geopolitically Immune IDM

Power semiconductors have a geopolitical superpower: they are not worth sanctioning.

The entire export-control architecture that constrains advanced logic — EUV lithography restrictions, high-bandwidth memory controls, advanced packaging scrutiny — passes over mature-node power devices almost without notice. onsemi doesn't use EUV. Doesn't need it. Doesn't buy from ASML. Its equipment comes from Applied Materials, Lam Research, and Japanese toolmakers with six-to-twelve-month lead times and no license friction. The dependency profile is low. Supply-chain vulnerability rating: minimal. This is a structural dividend that growth investors systematically undervalue.

The GTAT acquisition transformed the upstream picture. Silicon carbide substrates are the critical scarce input in the SiC value chain, historically concentrated at Wolfspeed and Coherent. Substrate costs dominate SiC device economics. By internalizing substrate production, onsemi converted a supplier dependency into an internal capability with yield-learning compounding. China's export controls on gallium and germanium, imposed in 2023 as geopolitical retaliation, add another layer. Those controls constrain supply chains that depend on Chinese-processed materials. onsemi's exposure is modest. To the extent that disruption penalizes competitors, onsemi is a net beneficiary. It's a strange sentence to write about a company with substantial China revenue, but the geopolitical matrix genuinely favors a US-headquartered power IDM.

Downstream, customer concentration runs moderate to high. Top five customers exceed thirty percent of revenue, weighted toward automotive Tier 1 suppliers and OEMs. Automotive buyers hold genuine pricing power; they have spent the last two years pressing SiC prices downward hard. In AI power, hyperscalers and ODMs are also demanding, but the current supply-demand balance for high-density power solutions favors the seller. That advantage is temporary. When power supply chips commoditize — and they will — the leverage shifts back to the buyer. The strategic variable is whether onsemi converts this current window of scarcity into design wins that survive commoditization. Liquidity is just patience with a time limit. The window is open now. It will close. Every quarter of slow qualification is a quarter of the window lost.

Dimension Three: The China Factor Nobody Leads With

Here's what the press release omits.

onsemi's China exposure is roughly thirty percent of revenue. For years, this was the growth engine. Chinese EV manufacturers adopted SiC aggressively. ADAS, automotive image sensors, power modules: China was the profit center. Then the ground shifted.

China's semiconductor industry went all-in on localization. Silan Micro and CR Micro scaled IGBT and power module production. CRRC Times Electric is advancing SiC. Starpower is shipping automotive-class power modules. The first generation of Chinese SiC carried reliability risks that failed automotive qualification. That has been changing, wave by wave. China's Big Fund Phase Three is deploying an estimated $47.5 billion, a meaningful share of it aimed directly at power semiconductors and SiC localization.

The market dynamics follow a textbook commoditization curve. Chinese entrants attacked the low end, matched datasheet promises at a fraction of the price, won design wins in domestic OEMs, then used that scale to fund reliability improvement programs. The data confirms the pattern: Chinese SiC device prices have fallen sharply, and design-win activity at tier-two OEMs has shifted decisively toward domestic suppliers. The trajectory is unambiguous. What was an onsemi-dominated market in 2020 is now a contested market with an accelerating domestic champion presence.

The 2017 Golem audit taught me a related lesson: trust is a matter of code verification, not documentation. Substitute "reliability data" for "code" and the principle applies. Chinese power devices started with documentation that outpaced actual reliability. That gap has been closing. Each successive generation closes it further. At the mid and low end, onsemi is fighting a losing rearguard action.

So the AI power pivot functions as an exit strategy embedded inside a growth narrative. A retreat from a battle where pricing power, political tailwinds, and market share are all moving against the company — toward a segment where American hyperscalers and geopolitical incentive structures favor Western supply chains. That's not intellectual dishonesty. It's capital allocation. A rational company retreats from deteriorating markets and redeploys into growing ones. But investors should understand what the pivot actually represents. It is not the discovery of a new frontier. It is a repositioning born of necessity. The narrative says offense. The balance sheet says defense. Both can be true. Only one is the real reason.

Dimension Four: Capex and Capacity — The Depreciation Game

Semiconductors run on depreciation physics. Any investor who doesn't start there will misread every margin number.

onsemi's capacity utilization is somewhere near 70 percent. The healthy zone for its asset base is 85 to 90 percent. Fabs carry depreciation whether they are full or empty. Underutilization spreads fixed cost across fewer units and compresses gross margin. That is the mechanical driver of onsemi's margin compression from 48 to 45 percent. Management cannot manage this away through narrative. The only cures are higher utilization or a richer product mix.

The AI power pivot targets both. It is a higher-margin product set, and it is a capex reallocation rather than a new greenfield build. onsemi's capital intensity is low by industry standards: capex at roughly 15 to 20 percent of revenue versus 35 to 45 percent at TSMC. The fab-lite trajectory is deliberate. Divestitures of legacy fabs, including the Niigata facility in Japan, reduce fixed cost. The GTAT SiC substrate expansion, the 8-inch conversion, the European power module capacity in the Czech Republic and Slovakia: these are targeted investments aligned with the strategic repositioning.

The timeline creates friction. New SiC capacity takes twelve to twenty-four months from tool move-in to qualified production. The depreciation drag from those new lines will pressure margins throughout the ramp. AI power revenue must scale fast enough to absorb that depreciation. If it doesn't, margins stay suppressed and the narrative strain becomes visible in the quarterly numbers.

I ran a similar exercise in 2022 following the LUNA collapse. Three weeks of back-testing the UST seigniorage mechanism with historical oracle data. The conclusion: economic models fail when they rely on infinite growth assumptions rather than tangible collateral. The parallel here is softer but real. The AI power pivot is capital-efficient — a reallocation rather than a bet-the-company expansion. That's rational. But it also means the revenue inflection will be slow. Anyone expecting AI power to move the revenue needle within two to three quarters should adjust expectations now. This is a multi-year program with meaningful uncertainty at every step.

There's one more angle. If the auto/industrial cycle bottoms in 2025 and 2026 — which channel inventory signals suggest — then onsemi's core business stabilizes just as the AI power narrative needs its proof point. The alignment of the cyclical trough with the narrative inflection is the single most important reason the stock could re-rate. But alignment is not causality. It's a setup. It can break.

Dimension Five: Demand — The Real TAM vs. The Poster TAM

Now the $213 billion question.

Separate the real demand from the poster TAM. Real demand for power semiconductors in AI data centers is strong. Blackwell chips run north of a kilowatt per device. Rubin-class parts will go higher. Rack power at 100-plus kilowatts is being deployed today. The power delivery chain — 800-volt HVDC, 48-volt intermediate bus, multi-phase conversion, vertical power delivery — is a genuine system bottleneck. Data center operators are power-limited before they are compute-limited. That's consensus fact.

It does not follow that the AI power slice is worth $213 billion. The entire global power semiconductor market is $50 to $60 billion. The AI data center slice is single-digit billions today. Even at aggressive growth, it will take years — likely a decade-plus — to reach a size that justifies the $213 billion figure. The number is an ecosystem-level TAM. It includes racks, busbars, transformers, cooling, the full electrical infrastructure of data centers. Or it's a multi-year cumulative projection. Or both. It is not an addressable market for a discrete power semiconductor IDM in any quarter you can plan around.

I've seen this pattern before. In 2021, the market accepted a multi-trillion-dollar metaverse TAM as a planning anchor for companies whose actual revenue was gaming hardware. In 2023, real-world asset tokenization was floated with TAMs larger than the global bond market. In each case, a real underlying trend got inflated into a narrative construct. Then the construct snapped back toward reality.

The false TAM serves a function. It converts a modest strategic story into a headline. It lets a press release compete with attention economics. But it corrupts analysis. Anyone using $213 billion as a planning anchor for onsemi's AI power opportunity is not an analyst. They're an audience member.

Nevertheless, the direction is real. The inventory cycle adds near-term texture. The automotive and industrial destocking that punished onsemi through 2024 and early 2025 appears to be entering its late phase. Channel inventories are approaching historical norms. Semiconductor cycles historically run three to four years from peak to trough. The current downcycle began in 2023, which places normalization somewhere in the 2025-2026 window. If that holds, onsemi's core stabilizes just as the AI narrative needs evidence.

Also worth noting: AI inference demand is the next lever. Training gets the headlines, but inference is what scales with actual usage. Every deployment of an LLM, every edge inference request, draws current through exactly the power conversion components onsemi wants to sell. The inference buildout is a broader, longer-duration demand driver than the training cluster boom. That's a substantive argument in the bull case — even if the $213 billion number is pure marketing.

Dimension Six: Competition — The Elephant Named MPS

Read the official language carefully: onsemi's entry into AI power "may redefine competitive dynamics."

Four words matter. May. Redefine. Competitive. Dynamics. In corporate communications, "may" carries the weight of "we have no idea." "Redefine" means "we're late and need to displace incumbents." This is not offensive language. This is a challenger's statement.

Map the competition. MPS dominates multi-phase controllers and power stages for AI accelerators. MPS is inside NVIDIA's reference designs. That means every hyperscaler building on NVIDIA's platform inherits MPS power silicon by default. MPS's qualification is baked into the ecosystem. Its supply chain is locked. Its system-level integration with GPU reference architectures is deeply embedded. To displace MPS, onsemi must win design migrations board by board, generation by generation. That's a long, grinding, unforgiving process.

Vicor owns the vertical power delivery frontier. VPD is where the next efficiency gains live: placing the power stage directly beneath the processor die to cut path inductance and deliver current at lower voltage with higher efficiency. Vicor's current multiplier technology is widely regarded as best-in-class for the high-density requirements of next-generation accelerators. onsemi's VPD capability is comparatively weak. That's a real technical gap in the most important emerging segment of the power chain.

Infineon brings scale, brand, and SiC strength. TI is present at the system level. The AI power segment is not greenfield. It is a contested arena with entrenched, well-qualified competitors. onsemi's sub-five-percent share in AI server power is not a beachhead. It's a toehold.

There's a subtle point about existing ecosystems. The silence between the blocks tells the real story. Open any next-generation accelerator reference design. Look at the part numbers on the power stages. The BOM is never empty — somebody is already there. In the AI power segment, the names are MPS, Vicor, Infineon, and TI. Changing a power architecture in a production GPU platform is a multi-year project carrying real risk. Incumbents have a structural advantage that new entrants underestimate at their peril.

I built a latency-arbitrage tool in early 2024 around the Bitcoin ETF approvals. Executed over 5,000 micro-trades capturing spreads between the GBTC discount and the new spot ETFs. $42,000 over six weeks. The lesson generalized: incumbents move slower than entrants assume, but they also hold infrastructure advantages that crude speed doesn't overcome. MPS and Vicor have years of qualification baggage, system-level integration, and supplier relationships constructed around their silicon. onsemi has a credible SiC franchise and a powerful balance sheet. Those are assets. They are not a supply-chain position inside NVIDIA's next-generation reference design.

Dimension Seven: Valuation — A Cyclical Stock Telling a Growth Story

Now the pricing question.

The market is being offered a story: onsemi is transitioning from a cyclical semiconductor company tied to automotive and industrial demand into a structural AI infrastructure compounder. The re-rating trade requires buying the stock on a cyclical multiple and watching it expand to a growth multiple as AI revenue materializes.

The current numbers. Gross margin: 45 percent, down from 48. ROIC: 12 to 15 percent against a WACC of 8 to 10 percent. Still value-creative, but the gap is narrowing. ROE: 20 to 25 percent, down from peak. EV/EBITDA: roughly 10 to 12 times versus a peer group around 14. The discount reads either as a bargain or as a risk premium for the exact uncertainty the AI narrative is supposed to eliminate. Or both.

An accounting note. onsemi's R&D treatment is conservative, with most R&D expensed rather than capitalized. No earnings manipulation signals. Operating cash flow sits in the $1.5 to $2.0 billion range. OCF to net income runs above one. Free cash flow is positive but capex-suppressed. The balance sheet is not the issue. The business quality is not the issue. The issue is purely the strategic narrative and its revenue proof.

The re-rating trade hinges on one ratio: AI power as a percentage of total revenue. Below ten percent, this is a cyclical company with a speculative growth option. Above ten percent, with demonstrated hyperscaler design wins, the structural-growth argument becomes credible. Everything before that threshold is narrative.

I've seen this exact structure before. The 2022 LUNA post-mortem. The model was stable above a 60 percent confidence ratio and structurally doomed below it. The model didn't fail because of an external attack; the assumptions did. Its own logic became self-reinforcing in the wrong direction once confidence crossed the threshold. The onsemi story has a comparable falsifiable threshold: the AI revenue mix crossing roughly ten percent of total revenue. Below the threshold, the growth narrative runs on hope. Above it, the narrative runs on revenue. And the market is being asked to pay re-rating premiums before the crossing is demonstrated.

Price action says the market is ambivalent. The discount to peers reflects skepticism. If the AI power story were fully believed, onsemi would trade at a premium, not a discount. The market is telling you it doesn't fully believe. What it's actually pricing is a compound option: the core business doesn't collapse, and AI power grows to a meaningful share. That's a reasonable option to own. But it's an option, not a certainty.

Here's the contrarian framing the narrative avoids.

This entire pivot is an admission. Read the sequence: gross margins in decline. Capacity underutilized. China market eroding segment by segment to state-backed local champions. Automotive demand stalling. Industrial in a destocking cycle. The AI power move is not a proactive discovery of a new frontier. It's a reactive escape from a deteriorating core. That doesn't invalidate the strategy. It reframes the risk. Buying the story at face value means accepting a narrative that omits its own defensive motive.

The comparison to DeFi is instructive. During the DeFi Summer of 2020, I deployed $150,000 of personal capital into Uniswap V2 ETH-USDC liquidity pools to test AMM mechanics against traditional order books. The surface numbers looked like free yield. Everyone piled in. The true economics — impermanent loss during volatility spikes, slippage, gas costs, competing liquidity — told a different story. I ran a rebalancing bot on testnet for weeks, measuring the gap between surface APY and realized returns. The lesson: the story everyone quotes is rarely the math that plays out. The market quotes the TAM. The math lives in design wins, qualification cycles, and margin mix.

Retail sentiment reads the $213 billion headline and sees a growth association on a cyclical stock. The tracker sees a massive TAM and assumes a large addressable piece belongs to onsemi. Smart money sees something different: a sub-five-percent share in a contested segment, facing incumbents with structural advantages, competing against MPS and Vicor at the highest-value layer of the power chain.

There's another blind spot. Cloud providers are developing in-house silicon and power solutions. If the largest AI infrastructure buyers continue verticalizing, the entire independent power semiconductor supply chain faces a structural challenge — not unconquerable, but real. onsemi's AI power entry could find itself caught between entrenched incumbents and the vertical integration wave. That's a pincer movement no press release can escape.

Consider the stablecoin analogy. The real driver of crypto payments adoption in developing countries isn't blockchain ideology. It's local currency inflation forcing people to find survival alternatives. The driver of onsemi's pivot isn't the AI power opportunity. It's the collapsing economics of its core markets forcing a high-margin alternative. The narrative will say "we see opportunity." The balance sheet says "we need cover." Both statements can be true. Only one of them is the real reason.

The MPS problem deserves emphasis. To enter AI power, onsemi must either beat MPS inside NVIDIA's reference designs — a war of attrition with a well-entrenched incumbent — or win external niche designs where MPS's grip is weaker. The second path is realistic. The first path is not. And if the real opportunity is in the niches, the revenue ceiling is lower than the narrative implies.

The market's job is to force a conclusion before the evidence is in. That's the core skill of a trader: filtering signal from noise, recognizing when a narrative is ahead of its proof.

Three signposts define this story.

First, quarterly AI data center revenue disclosure. Until onsemi breaks out the line, there is no evidence — only narrative. When a company believes its strategic story, it presents the numbers. The absence of disclosure is itself a signal.

Second, hyperscaler design wins. If the company announces a supply arrangement with a tier-one cloud provider, that is genuine validation of the entire strategic pivot. If the design wins are limited to secondary ODMs or internal pilots, treat the story as status quo with better marketing. Names matter. A design win with Microsoft, Google, Amazon, or Meta is worth a thousand investor decks.

Third, gross margin inflection. If margins stabilize near 45 percent and begin expanding while AI power revenue grows, the fixed-cost leverage math works. If margins continue sliding through 2025, the depreciation drag is winning and the narrative is carrying the stock. Margins are the honest ledger of whether the strategic pivot is creating value or merely costing money.

The ten percent revenue threshold is the line in the sand. Cross it, with design wins and margin expansion, and the re-rating case is real. Fall short, and this is a cyclical semiconductor stock with a well-produced growth presentation. The direction of travel is genuinely interesting — AI power demand is real, the SiC asset base is deep, the geopolitical positioning is clean. But the honest position right now is: the technology is credible, the numbers are not slogan-deck credible, and the proof is in future earnings.

The noise-to-signal environment here is worse than most. The original article was published by a crypto outlet, optimized for attention, containing five data points and no sourcing. That's not journalism. That's a front-running narrative. The time to investigate the story was before the headline, not after.

Every model has a threshold condition that determines whether its logic compounds or collapses. For LUNA, it was the confidence ratio. For onsemi's AI pivot, it's the ratio of AI power revenue to total revenue. Everything else is commentary.

Buy the P&L, not the deck. The market has a way of confusing the two. The quarter will tell the truth. The deck won't.