The Add Every Memory Bull Needs

We acknowledged memory as the bottleneck first, in July (The Bottleneck of All Bottlenecks). If you have any exposure to memory stocks, this is for you.

If HBM stays scarce, whatever chips do ship still need somewhere to plug in. If it eases, more chips ship, and every one of them still needs somewhere to plug in. Both paths end at the same wall: someone has to generate the electricity.

That wall already has a price tag.

PJM runs the power market across thirteen states, Virginia, Ohio, Pennsylvania, Illinois and Maryland among them. Its most recent capacity auction cleared at $269.92 per megawatt-day. A year earlier it was $28.92. An 833% increase in twelve months, and PJM's own market monitor attributes 63% of it directly to data center demand. The next auction already cleared higher still, $329.17, a FERC-mandated ceiling. The one after that is set at $333.

Dominion customers in Virginia and ComEd customers in Illinois are already paying more because of it.

The other side of the wall is turbines. GE Vernova's gas turbine backlog sits at 116 gigawatts, up from 100 gigawatts in the first quarter of 2026. The CEO said on the earnings call they're now taking reservations for 2031 delivery. Three companies make these turbines, GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries, and all three are booked years out with pricing still climbing, heavy-duty turbines running around $790 per kilowatt, HA-class combined-cycle turbines near $950.

The auction price already cleared. The turbine wait is already five years. Neither one is a forecast.

Most pair trades work through tension. You're long one side and short the other, or long two things expected to move in opposite directions, and you're only right if your view on that relationship holds. Get the relationship wrong and both legs can hurt you at once.

Memory and power work through synergy instead. They're the same underlying bet, AI deployment keeps growing, expressed through two different mechanisms.

If HBM stays scarce, the memory trade performs as designed. SK Hynix keeps its pricing power, and that leg carries the position on its own. If HBM eases, the outcome that normally undercuts a memory thesis, more supply, falling prices, stops mattering here. Easing memory expands deployment. More chips ship, faster, and every one of them still needs somewhere to plug in.

Whichever leg loses ground, the other one picks it up. The position is built around the one variable that was never actually in question, whether the buildout continues.

The energy leg of this position has a specific vehicle right now, and it just explained why in its own earnings report.

Vistra reported on August 7. Adjusted EBITDA jumped 30% year over year. The stock fell anyway, the second time this pattern has shown up in three weeks. Alphabet did the same thing in July, best quarter in company history, stock down 7% on the print (we covered it here). That drop was wrong, and two weeks later the stock was up 21% from the low.

Guidance held steady instead of rising, and the second half needs a 20.8% sequential jump in EBITDA to hit the number already promised, real risk worth naming. That's what the market reacted to. One paragraph down in the same filing sat the actual news.

Vistra is forming Helix Digital Infrastructure alongside KKR, the Kuwait Investment Authority, and Nvidia, committing up to $1 billion as a founding investor and taking the role of preferred power partner. Vistra supplies the electricity, from its existing fleet or new contracted builds, while the partners fund the land and infrastructure around it.

Nvidia showing up as a strategic investor in a power joint venture, not a chip customer, is new. The logic is simple. A chip with nowhere to plug in sells nothing. Nvidia investing in the electricity that runs its own chips is Nvidia protecting its own growth.

Talen and Meta are locking in the same kind of position, at even longer duration. Talen signed an $18 billion, 17-year power purchase agreement with Amazon, up to 1,920 megawatts of nuclear output from the Susquehanna plant through 2042. Vistra signed its own 20-year deal with Meta in January, 2,609 megawatts of nuclear supply. Contracts that long compound cash flow for longer than most fund managers stay at the same firm, which is part of why the market has trouble pricing them correctly while they're still building.

Not everyone buys this. Jim Chanos said in June that power is only 5 to 7% of data center revenue, and pointed out that alt-energy stocks were trading at 50 to 70 times earnings for that sliver of exposure. Fair, for the highest-flying names in this trade, fuel cell companies and pure-play alt-energy stocks priced for a perfect outcome. Vistra trades around 27 times earnings, in a range one recent valuation model called roughly fair rather than cheap. Chanos is right about the expensive end of this trade. He's describing a different stock than this one.

The more serious pushback comes from Epoch AI, an independent AI research group with no position to defend either way. Their research argues America's power stagnation reflects weak demand rather than an inability to build, and that power remains a small slice of total data center cost next to chips. Both of those things can be true and the position still works, because Epoch AI is answering a question about total spending and this piece is answering a question about marginal scarcity, whether enough power exists right now at a price buyers will pay. The PJM data already answers that in real numbers. An 833% price increase happened, in an actual auction, over the past year.

Then there's the political risk, and it's real. Ohio's utility regulator now forces new loads over 25 megawatts to pay a minimum of 85% of their subscribed capacity for up to 12 years. Virginia created a new rate class for large customers starting in 2027, with 14-year minimum contracts. Senator Mark Warner introduced a bill in June aimed at making data center operators cover more of their own cost. All three rules do the same thing underneath the politics: they lock large power customers into long, contracted commitments with the exact utilities providing their power. The mechanics point toward more locked-in, multi-year contracts, the structure this position already depends on.

Vistra traded as low as $134.75 in the days after it reported, nearly 39% below its September 2025 high of $219.82, before closing back at $140.59. The post-earnings analyst consensus target is $228, roughly 62% above that close, with 17 of 21 analysts rating it a buy. Short interest fell 20.8% into the print, from 15.92 million shares to 12.60 million, so the bears weren't piling on ahead of earnings the way they would if this were a broadly hated stock.

A level that's held for most of a year breaking down would send one signal. Bouncing off it again sends a different one, and this dip did the second thing.

None of the upside above is in the current guidance. Vistra flagged roughly $700 million of potential 2027 EBITDA from the pending Cogentrix acquisition and the Meta nuclear deal, neither baked into the numbers the company has actually given investors. Helix adds a business Vistra is now getting paid to help build, not just sell power into, and no model has that in it either.

We called memory the bottleneck first. This is the other half of that position, the leg that grows no matter which way the memory story resolves. Own SK Hynix for the scarcity case. Own Vistra for the deployment case. Whichever one turns out to be right this year.

The whole position rests on one assumption: that easing memory expands deployment instead of just making it cheaper. If HBM supply improves and chip deployment doesn't accelerate to match, sitting flat instead of scaling with the cheaper input, that's the scenario that actually breaks this. It would mean the AI buildout has a spending ceiling independent of any physical bottleneck, chips, memory, or power, and every falsifiable trigger in the memory piece and this one would need rewriting.

Shorter-term, there's plenty that can go wrong with the vehicle without breaking the thesis. Vistra could miss its own second half guidance. PJM's next auction could clear flat. GE Vernova's backlog could stall if hyperscaler capex actually slows. A regulatory ruling could cap what utilities charge instead of shifting cost to data center operators. Any of those would hurt the specific trade. None of them would mean chips stopped needing somewhere to plug in.

If you got something out of this, the most useful thing you can do is send it to one person who'd get something out of it too.

Share neym research

{{rp_personalized_text}}

Or copy and paste this link to others: {{rp_refer_url_no_params}}

This is not financial advice. neym is an independent research newsletter. The author may hold positions in securities mentioned. Do your own research.