Power, Not Silicon: The Real AI Bottleneck Is the Grid
On 12 June 2026, at the All-In Summit's Best Ideas Pitch Competition, four fund managers walked onto a stage in front of a live audience and a panel of judges to pitch their highest-conviction longs. The judges picked Talen Energy and MGM Resorts; the audience voted MGM first. The most interesting trade of the night was the one that didn't win.
The thesis from Daniel Draffen, the manager of the Talen Energy pitch, is uncomfortable for a market that has spent two years treating artificial-intelligence capex as the story. The binding constraint on the AI buildout, he argued, is no longer compute silicon. It is electrons, transmission lines, and the decade-long permitting cycle that decides whether a megawatt ever reaches a substation. The same evening, Oleg Nodelman pitched Akari Therapeutics as a biotech with a China-proof moat. Aaron Cowen pitched MGM Resorts against a $48 floor set by Barry Diller. Kyle Samani pitched the Geodnet token. But the power pitch was the one that explains why the rest of the trade is mispriced.
The grid as the binding constraint
Draffen's argument runs in two moves. First, US power markets are structurally tight for reasons that have nothing to do with AI. Coal retirements have outran gas and renewables buildouts. PJM, the Pennsylvania–Jersey–Maryland grid operator that runs the largest competitive wholesale market in the country, has said it needs 106 GW of new capacity over the next ten years — roughly the size of Japan's total current electricity consumption. "We do not need AI demand to keep the power markets incredibly tight for the next 20 years," Draffen said. "AI demand just turbocharges. That's all it does."
That is a different framing from the one equity investors have been running with. For most of 2024 and 2025, the AI trade was a chip trade: Nvidia, TSMC's Arizona build, the HBM memory complex. The grid was treated as a slow-moving utility sector that might benefit at the margin. Draffen's pitch turns that hierarchy on its head. If the grid cannot add 106 GW in a decade — and there is no realistic construction schedule that gets there — then every existing nuclear and gas baseload asset in PJM is a de facto monopoly on the marginal megawatt, and the price of that megawatt gets set by the next data-center hyperscaler that signs a long-dated contract, not by the spot market.
A repricing of politically toxic assets
The marquee example is Three Mile Island. Microsoft signed a 20-year power-purchase agreement with Constellation Energy at roughly $100 per megawatt-hour to restart the Unit 1 reactor, against a spot price that has hovered around $50/MWh. "Three Mile Island, brought to you by Microsoft Azure," Draffen quipped. The deal is doing three things at once: repricing a politically radioactive nuclear asset for the AI era, locking in a multi-decade fixed cost for a hyperscaler that has decided forward power prices are uninsurable, and setting a precedent for behind-the-meter colocation that other data-center operators will be forced to match or be priced out of.
Jensen Huang, Nvidia's chief executive, has said publicly that the company needs 1,000 times more power than is currently available to it. Even if that number is loose — and it is loose; it conflates nameplate with delivered, training with inference, and present with steady-state — the directional signal is clear. Draffen's read is that the US needs every available power source: nuclear, gas, solar, and eventually orbital solar, plus an acute buildout of critical-mineral refining. He compares it to China's twenty-year capacity expansion, noting that China now generates roughly three times the electricity the United States does. "Make America great again on power" is, in this framing, an industrial-policy problem disguised as a permitting problem.
The Zell playbook applied to electrons
Talen Energy is the equity expression of that view. The company is sitting on 2 GW of nuclear and 6 GW of natural-gas baseload, trades at a roughly $25 billion enterprise value against an estimated $45 billion replacement cost, and has an Amazon data-center contract already rolling into the run rate. Draffen presented a three-scenario free-cash-flow model: about $50 a share on current contracted volumes, $70 a share if more power-purchase agreements sign, and over $100 a share if Talen builds into the 4+ GW of incremental capacity PJM will need. The stock sits in the high $300s. He framed the trade in the language of the late Sam Zell: "If you can buy a hard asset at below replacement cost for an asset that's going to be needed in the future, you buy it at the discount to replacement cost, you hold it, and you sell it at a big premium to replacement cost when the market wakes up."
The bet is that the market has not yet repriced existing baseload for an era in which data centers are baseload customers. A data center, as Draffen put it, "is the exact same thing as a refinery. You put electricity in, and on the other end instead of gasoline or jet fuel out comes photons or tokens or intelligence." That metaphor is doing work. Refineries, when they were scarce, traded at multiples of replacement cost for the better part of a decade. There is no structural reason a 2 GW nuclear plant in a constrained PJM zone cannot do the same.
What the other pitches said
The other three pitches were sharp in their own lanes but tell a different story. Cowen's MGM pitch was an activist-arbitrage case. Barry Diller owns 26% of the company — 80% of his net asset value — and has bid $48, up from $37 two weeks before the pitch, providing a hard floor. Above that floor, Cowen argued, the market is mispricing three distinct assets: a Las Vegas core worth roughly $60 a share, an Osaka, Japan casino complex worth another $50 (MGM owns 40% of the property, in a country with a $40 billion gambling market anchored by pachinko and horse racing), and a Dubai option worth $40 to $50. The audience voted MGM first; the judges placed it second behind Talen. David Friedberg, citing a contact in Las Vegas, noted that the Sphere at the Venetian is producing about a million dollars a day in incremental EBITDA on show days — a useful data point on how entertainment draws amplify gaming revenue.
Nodelman's Akari Therapeutics pitch was about a biotech moat that most generalist investors miss. The company uses actinium-225, a radioisotope derived from legacy US nuclear-weapons-program radium-223, as the payload for its lead radiopharmaceutical. That supply chain is unavailable to China because Beijing's nuclear fuel cycle is structured differently. In a sector where Amgen-versus-Sandoz-style biosimilar erosion has become the default bear case for any biologic, a structurally non-replicable isotope is a real moat. Nodelman also noted that $15 billion in M&A and dealmaking has flowed into radiopharmaceuticals over the past few years, with Bristol, Novartis, Bayer, and Eli Lilly all building radiotherapy pipelines. Lilly, notably, backstopped Akari's $300 million IPO with a $100 million order. The radiopharmaceuticals space, in other words, is the second- or third-best expression of the same "bottleneck physical asset" theme that the power pitch made primary.
Samani's Geodnet pitch is the smallest in market cap but the most network-effect-laden. The network runs 22,000 RTK base stations across 150 countries — roughly twice the combined footprint of Trimble, Hexagon, and Topcon, which between them have spent twenty to thirty years building out their own RTK networks and have about 12,000 stations between them. The economic model is a token-incentivised flywheel: individuals buy roughly $300 base stations, mount them on rooftops, and earn Geod tokens. Eighty percent of network revenue buys tokens back on the open market — about $8.8 million a year, visible on the Solana blockchain, against a roughly $1 million annualised revenue run rate that is growing at about 3x year-over-year. Average customer spend, Samani said, grows from about $60,000 in year one to $170,000 in year two as users expand use cases. Flagship customers include John Deere's GUS spraying system, DJI drones, and TomTom's autonomous-vehicle map updates. "There's no chance they can compete on cost. Geodnet base stations are a few hundred bucks. You're just not going to compete on cost with Geodnet [from space]," Samani said, dismissing the orbital RTK threat.
The counter-narrative
The contrarian read on the power pitch is straightforward. PJM's 106 GW number is a forecast, not a procurement schedule. Interconnection queues, state-level siting fights, and the political economy of rate increases for residential consumers all push against the buildout. The 20-year Microsoft–Constellation deal at $100/MWh is a single data point, not a curve — and if regulators force a higher pass-through to consumers, the political backlash could slow every subsequent hyperscaler deal. The Zell-style "buy below replacement" framework also assumes that replacement cost itself is a stable reference, which is less true in a sector where the cost of a new combined-cycle gas plant has risen 40 to 60 percent in three years. None of that kills the thesis, but it narrows the addressable mispricing.
The contrarian read on the MGM pitch is that Barry Diller's 26% stake is not a permanent feature of the cap table; activist blocks tend to monetise, and a successful Osaka opening could mark the top of the trade rather than the bottom. The contrarian read on Akari is that a single-asset radiopharmaceutical with a $500 million enterprise value, no approved product, and a moat that depends on isotope supply discipline is exactly the kind of name that gets marked down 50% on a single failed Phase 2 readout. The contrarian read on Geodnet is that token-incentivised hardware networks have a poor historical record of surviving the moment real competition or real regulation shows up.
The structural read
The through-line across all four pitches is that physical, regulated, hard-to-replicate assets are the places where the next decade of bottleneck rents will accrue. Silicon is a manufacturing problem; manufacturing problems get solved. Permitting, transmission siting, nuclear restart, isotope refining, and rooftop-mounted RTK networks are coordination problems; coordination problems persist. The market is starting to price that distinction. The PJM forward capacity curve, the Constellation–Microsoft PPA terms, and the radiopharmaceutical M&A wave are the leading indicators. The judges and the audience on 12 June were not wrong to pick the casino and the power plant as the two best ideas of the night. But the more durable lesson is the one that ran through all four pitches: in an economy that is repricing for an AI-eats-everything world, the trade is not the chip. The trade is what powers the chip.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- https://www.youtube.com/watch?v=fO5sC7qS04E