Last week, a crypto outlet published a short item telling its readers that ON Semiconductor โ onsemi โ is repositioning itself to capture a $213 billion AI power market.
I stopped at the number.
Global power semiconductor revenue, the actual silicon that switches and converts electricity, runs somewhere between $50 billion and $60 billion a year. The $213 billion figure is not wrong by a rounding error. It is wrong by a factor of four. The most plausible reading is that it counts power supply units, rack-level power shelves, busways, switchgear, transformers and grid interconnection โ an ecosystem-level total addressable market, or a multi-year cumulative one, packaged as a single-year opportunity and handed to a reporter who had no reason to question it.
I have spent years inside these decks. In 2022 I modeled oracle latency during the Terra collapse and calculated that a 15% price-feed deviation combined with lighthouse-node delay could have liquidated $2 billion in positions. That paper was cited by three security firms. The lesson was not that markets are irrational. The lesson was that the number everyone quotes is almost never the number anyone meant.
The same discipline applies here, and it applies with unusual force, because the crypto industry is now physically, legally and financially entangled with the AI power buildout. We are financing it, hosting it, and increasingly tokenizing it. And we are repeating the same error we made with every previous TAM slide: accepting the headline as the ground truth.
The anomaly is not the $213 billion. The anomaly is how few people flinched.
The physical stack under every GPU cluster
To see why the number is wrong โ and why it still matters โ you have to look at what sits beneath an accelerator.
A modern AI datacenter converts electricity five or six times between the substation and the transistor. Medium-voltage AC arrives at the site. A transformer steps it down. UPS and switchgear stages condition it. Inside the rack, a power shelf converts AC to 48V DC, or in newer designs the site runs 800V DC all the way to the row. From 48V, a multi-phase controller and a bank of smart power stages step down to roughly 0.8 to 1.0 volts at the die edge, at hundreds of amps per device.
Every stage loses energy as heat. Every stage has a bill of materials and a vendor. ON Semiconductor sells into a narrow slice of that chain: the multi-phase controller, the smart power stage, and the SiC and IGBT devices used further upstream in conversion and site electrical infrastructure.
That is a real business. It is also a business where onsemi is a challenger, not an incumbent. Monolithic Power Systems holds the dominant position in the GPU voltage regulator module market. Vicor leads in vertical power delivery, the architecture that pushes current up through the board instead of across it. When I led a comparative benchmark of Optimistic and ZK rollups in 2023 โ 10,000 simulated transactions across Arbitrum and StarkNet โ the useful finding was not which system won on average throughput. It was identifying where the bottleneck actually sat, which was almost never where the marketing claimed the innovation was. Same instinct here. The bottleneck in AI power is not the controller silicon. It is upstream of it, in copper and permitting.
Crypto touches this stack in three ways, and each carries a different truth content. Mining fleets are converting to AI hosting. Decentralized compute networks are brokering GPU capacity on-chain. Tokenized energy and power purchase agreements are being sold as real-world assets.
All three are priced against one underlying assumption: that power is scarce, that it stays scarce, and that whoever holds the interconnection holds the margin. Two of those three are defensible. The third deserves an audit.
Modeling the conversion chain as a settlement layer
Treat the power delivery path as a protocol. Each conversion stage is a block producer. End-to-end efficiency is the product of the stages, not the minimum and not the average. This is a critical distinction that most infrastructure commentary gets wrong.
Run realistic numbers on a 100 kW rack in an 800V DC architecture. Site-level transformer and UPS conditioning at 95%. The 800V-to-48V conversion stage at 98%. The 48V-to-core VRM at 90%, which is generous for a 0.8V rail at 800 amps. Multiply: 0.95 ร 0.98 ร 0.90 = 0.838.
At a 100 kW rack, that means 16 kW is dissipated as heat before a single instruction retires. Sixteen percent of the power bill is thermal overhead. Sixteen percent of the thermal load must be removed by cooling that itself consumes power. The efficiency chain compounds exactly like validator uptime across a sharded network: one weak stage drags the whole path, and the loss is multiplicative, not additive.
Now the arbitrage that is actually driving capital.
Post-halving, with hashprice hovering in the mid-forties per petahash per day, one 100 kW rack of S19-class machines running at roughly 30 joules per terahash produces about 3.3 PH/s. That is roughly $150 a day in gross revenue. Power at $0.045/kWh costs $108 a day. Call it $40 of daily margin before opex, before difficulty repricing, before the next generation of ASICs resets the denominator.
Run the same 100 kW into 120 to 140 high-end accelerators at 700 to 800 watts each. At $2 per GPU-hour and 70% utilization, that rack grosses somewhere near $4,300 a day. Even at 30% utilization the comparison is not close.
The incentive gradient is not a trend line. It is a cliff. Cliffs produce migration events, not gradual transitions. Every publicly listed miner with a substation and a signed interconnection agreement has already run this arithmetic, which is why the announced hashrate-to-AI conversions keep arriving in clusters rather than drips.
And this is where the analysis stops being about capital and starts being about latency.
The latency cost of electrons
In 2024 I published a critique of modular data availability titled "The Latency Cost of Modularity." I had measured blob submission delays of roughly 12 seconds during peak block production on Celestia and argued that this undermined real-time settlement guarantees for rollups that depended on it.
The electrical equivalent is worse, and it does not degrade gracefully.
Interconnection queues in most US ISOs run three to five years. Large power transformers carry lead times of two to three years and are still working through the aftermath of the 2021 grain-oriented electrical steel shortage. Gas turbine slots at the major OEMs are effectively sold out into 2027 and 2028. A 200 MW campus does not get built because a token launched.
A modular blockchain under congestion can absorb latency and keep producing blocks. A datacenter that does not receive its contracted power by its energization date does not degrade. It fails to exist. The capital is stranded, the depreciation schedule keeps running, and the revenue line stays at zero.
This is the structural asymmetry that the AI-power narrative systematically hides. Compute is elastic. Electrons are not.
What on-chain compute actually proves
In 2025 I designed a verification protocol for AI inference using zero-knowledge proofs, cutting verification overhead by about 30% versus the methods then in production. I presented it in Tel Aviv and it drew venture interest, which was flattering and slightly beside the point.
Here is the part that matters. Verifying that a model ran is one problem. Verifying that the electricity existed is another problem entirely.
Code does not lie, but it often omits the truth. An on-chain receipt can bind a weights hash, an input, an output and a timestamp. It cannot prove which physical hardware executed the job, which jurisdiction the hardware sat in, what the marginal electricity cost was, or what the carbon intensity of that electron was. Every one of those claims is currently an attestation. Attestations are signed by exactly the centralized operators that decentralized compute networks claim to route around.
The same gap runs through tokenized energy. Most of what is marketed as power-backed real-world assets terminates in a legal wrapper around an off-chain cash flow, with the on-chain leg handling settlement and distribution rather than verification. As a product, that is fine. As a claim about trustlessness, it is a category error, and it should be labeled as one in every token's documentation rather than in the footnotes of a blog post.
Reading onsemi as a mirror
The onsemi case is instructive precisely because it is the mirror image of a crypto infrastructure token.
Strip the narrative and you find a cyclical power and analog business with roughly 30% of revenue historically exposed to Chinese substitution from Silan, CR Micro, StarPower and Times Electric. Utilization near 70%, below the 85% to 90% band that keeps depreciation from eating gross margin. Gross margin drifted from roughly 48% to roughly 45% across two years. Automotive EV demand is decelerating. Industrial is flat.
The response is not a new product cycle in the classical sense. The response is a new TAM slide.
I have seen this pattern. It is the 2021 institutional DeFi total addressable market deck, reskinned with silicon. The underlying strategic move is genuine โ SiC substrate self-supply acquired through GTAT is a real structural advantage, and the Fab-lite restructuring is real capex discipline rather than accounting theater. But the headline number is a narrative hedge against a maturing core, and it should be read exactly the way an analyst reads a token with a shrinking fee base and an expanding roadmap.
In 2020, while finishing my undergraduate work on elliptic curve pairing efficiency, I independently audited the Zcash Sapling upgrade and found a side-channel in the Merkle tree implementation that could leak privacy under load. I spent 120 hours on the report and filed a pull request. The finding never appeared in the release notes. It appeared in the diff.
That habit is the only durable edge in a market where narrative velocity exceeds verification velocity by an order of magnitude. Read the PR, then read the pull request.
The blind spot nobody is pricing
Everyone is treating AI power as a demand story. It is a supply story with a single point of failure.
The chain is only as strong as its weakest node, and in this chain the weakest node is not silicon, not GPUs, and not capital. It is the substation and the interconnection queue โ a centralized, rate-regulated, permitting-bound chokepoint that no amount of on-chain coordination can parallelize.
Notice the structural echo. Layer2 sequencers are single centralized nodes, and decentralized sequencing has been a roadmap slide for two years. Decentralized compute networks replicate that architecture one layer down: a permissionless market presented on top of a handful of hyperscale datacenter operators and a grid that allocates capacity by queue position. The interface is permissionless. The substrate is not.
The second blind spot is duration mismatch. A token can be deployed in a week. A 200 MW interconnection agreement takes three to five years and can be voided by a single utility commission ruling. Any AI-power thesis that requires capacity which has not yet been energized is, by construction, pricing a narrative with a multi-year settlement delay โ and repricing that narrative is a function of policy calendars, not of protocol upgrades.
Takeaway
The forward-looking question is not whether AI needs power. It does, and the demand is real. The question is which layer absorbs the failure when the interconnection does not arrive on schedule, and whether the crypto instruments layered on top of that capacity will have disclosed that dependency before or after the drawdown.
Watch three signals: large power transformer lead times, whether the semiconductor vendors begin breaking out datacenter power revenue in quarterly filings rather than in keynote slides, and whether proof-of-inference systems mature from proving FLOPs to proving joules.
Scalability is a trilemma, not a promise. When the grid is the sequencer, who audits the block?