The Discussion · Weekly
The Discussion: The Queue for Electrons
This week the AI story stopped being about chips and became a story about power, price, and who gets to skip the line. Mona and Polly trace one machine viewed from five angles: a Fed rate call, a grid ceiling, a Chinese export number, a Massachusetts rate case, and the thinnest trading month of the year.
Here is the sentence that reorganized my whole week. Every dollar of that trillion dollar capex number the hyperscalers are guiding to for 2027 is a claim on electrons that do not yet exist.
That is the line, isn't it. And it's not rhetoric. PJM Interconnection, the grid operator covering thirteen states from Chicago to the mid-Atlantic, said this week in writing that it will start cutting power to large data centers during shortages rather than let the whole system brown out. That is the physical world answering the spreadsheet.
So let me set the table, because five editions this week look like five different stories, and I don't think they are. Monday we had the Fed. Tuesday, PJM drawing a line. Then China's chip export number. Then a Massachusetts electric bill. And then Wall Street's emptiest month. Five headlines.
One machine, viewed from five angles. Energy, compute, capital, and the supply chains that connect them. And the thing I kept coming back to is that the binding constraint moved. It used to be chips. This week it is power. Everything else is downstream of that.
Right, so let's start where the money starts. The Fed.
On July 29 the Federal Open Market Committee delivered its July decision, and this was only the second meeting chaired by Kevin Warsh, who succeeded Jerome Powell and ran his first FOMC back on June 17. What jumped out at me was the repricing going in. The CME FedWatch tool put hike odds near 38 percent on July 24, up from about 10.7 percent on July 15. That is one of the fastest repricings of a Fed meeting in recent memory.
And then what actually happened?
They held. The following edition told us the vote: nine to three to hold the key rate steady in a range between 3.5 and 3.75 percent. So the market got itself worked up for a hike, and got a hold. But here is the argument the edition made, and I think it is the right one. The Fed does not build a single power plant, chip fab, or transmission line. It sets the price of the money that does.
Which is why the edition called a rate call an energy story. And normally I'd push back on a framing that grand, because normally a rate decision is a mortgage and credit card story. But the AI buildout is financed almost entirely with debt. Hyperscaler bonds, project finance loans for gas turbines, power purchase agreements running twenty and thirty years.
So the discount rate the Fed sets is the same discount rate that decides whether a marginal data center or a marginal reactor restart clears its financing hurdle. A reactor is a twenty year bet. It only gets built when the cost of money is low enough, or when the buyer is willing to eat the difference.
And that is the tell for the whole week, isn't it. Increasingly the buyer eating the difference is a hyperscaler with a balance sheet. Which means the cost of capital matters slightly less to them than it does to a regulated utility. Hold that thought, because it comes back hard in the Massachusetts story.
Let me put a number on why power is the through-line and not chips. Gartner's own research this year projects data center electricity demand growing 26 percent in 2026 alone. And a separate report circulating this week argues data centers could consume roughly a fifth of all US electricity by 2035.
And the concrete detail that made it real for me: TechCrunch reported this week on a single fallen power line that exposed how thin the margin for error has become at AI campuses built faster than the grid feeding them. One line. That is the operational reality under a twenty percent of US electricity headline. The abstraction is a fifth of the grid. The reality is a single cable and a substation that isn't finished yet.
And every one of those is a financing problem before it is an engineering problem. Somebody has to borrow to build the substation, the turbine, the transmission line.
That's the bit I'd sharpen. It's a financing problem and an engineering problem, and they arrive on different clocks. Money can move in an afternoon. A transmission line takes years. So the Fed can make the money cheaper today, but that does not conjure a transformer that has an eighteen month lead time. The two clocks are the tension in the whole week.
Which is a perfect bridge, because Tuesday the slower clock spoke. PJM.
So the mechanics here. After an auction to add generating capacity fell short, PJM said it will cut off data centers and other large users during power shortages. And the read the edition landed on is that this is the first time a major US grid has told the AI industry, in writing, that demand has outrun supply.
And I want to be careful, because the editions were careful. Nobody is being punished here. PJM is doing what a well run utility does: telling its biggest customers the physics of the system before the lights actually flicker.
Exactly. And the rule is narrow and dated. PJM won't start curtailing until June 2027, and the cuts only apply to data centers that are 50 megawatts or larger. So it is not a hammer coming down tomorrow. It is a speed limit posted for a road that is still being built. By 2035 data centers are expected to use four times more electricity than they do today, and that is the pressure behind the sign.
Now here is where the money and the physics collide, and it is my favorite contrast of the week. Put PJM's ceiling next to what the hyperscalers told Wall Street they intend to spend. Microsoft's capex is projected to nearly triple to 190 billion dollars by fiscal 2027. Meta's fiscal 2026 capex rose by over 10 billion to 136.7 billion. Alphabet reported 44 billion in a single quarter, roughly 132 billion over twelve months, about double a year earlier, and guided to about 200 billion over the next twelve months.
And zoomed out, hyperscaler AI capital spending got raised to 750 billion in 2026, up from 670 billion, and it's set to cross a trillion in 2027.
So a trillion dollars of intent, meeting a grid operator that just said, politely, I cannot connect everything fast enough. That is the queue. And the capex is a bet that compute will be scarce and valuable.
But watch what that capex actually buys, because this is the operational point people skip. That money does not buy chips alone. It buys the switchgear, the transformers, the gas turbines, the prime power gensets that turn electrons into trained models. The GPU is the glamorous line item. The transformer is the one with the lead time that decides your schedule.
And that is why the same companies are quietly becoming nuclear utilities. This was story two in that edition. Every major hyperscaler has signed at least one nuclear power deal for AI capacity. Thirteen announced projects, over 9.8 gigawatts committed.
The specifics are worth saying out loud because they're extraordinary. Microsoft secured a 16 billion dollar, twenty year power purchase agreement for the Three Mile Island Unit 1 restart at 835 megawatts. Meta leads with up to 6.6 gigawatts across TerraPower's Natrium reactor, Oklo's Aurora, Vistra, and Constellation. Google and Amazon are in it too, with Kairos Power and X-energy's small modular reactors.
So the read, and the editions flagged this as read, not fact: the nuclear deals are hyperscalers trying to jump the queue for power by building their own generation rather than waiting on the grid. The capex is the bet, the nuclear deals are the queue jump, and PJM's rule is the queue itself, finally speaking.
And I'd add the causal loop. The edition's own watch item said it best. If PJM's next capacity auction clears, the queue moves. If it fails again, expect more hyperscalers to announce reactors before they announce models. So the grid's failure to add supply is literally what pushes these companies to become power companies with software divisions attached.
Which is a wild inversion. We think of Microsoft as a software company that happens to use a lot of electricity. This week's framing is closer to: Microsoft is becoming a power company that happens to write software.
And it's rational. If power is the binding constraint, ahead of chips, ahead of capital, ahead of talent, then the scarce thing is the thing you vertically integrate. You don't outsource your bottleneck.
Okay, so we've done capital and we've done power. Wednesday takes us to the supply chain, and this one is a myth buster. China's chip export number.
So the headline that went everywhere. In the first half of 2026 China exported 179.44 billion integrated circuits worth 177.28 billion dollars, up more than 96 percent year on year, per the General Administration of Customs. Read quickly, that looks like China's chip industry doubled its output or its sophistication in six months.
And that reading is wrong.
The 96 percent is not a technology leap. It is a price chart wearing a technology story's clothes. Look at the unit economics. At an average of about a dollar a chip, that mix is mature memory, power management chips, microcontrollers, and chips that were never even fabricated in China, only packaged and tested there for re-export.
So the customs bureau counted dollars, and dollars went up. That is not the same claim as capacity going up.
Precisely. And here's the supply chain planner's instinct that the edition nailed. When a single source input tightens, the shortage does not announce itself as a shortage. It announces itself as a growth headline for every product built from that component. Chips, servers, phones, anything with a memory die soldered into it.
And you can see it inside the same press release. Server exports, the automatic data processing machines and parts that get bolted into data centers, rose 41.3 percent to 138 billion dollars. Some of that is the same higher memory price running through every category that touches a circuit board.
Now, to be fair to China's actual capacity, because the editions were, there is a real buildout underneath. The country produced 484.3 billion ICs in 2025, with 3,901 domestic chip design companies posting combined sales up nearly 30 percent. That is genuine, years long mature node fab capacity reaching scale. But that is a separate story from the 96 percent.
And there's a clean tell to separate them. The robot number. Industrial robot exports rose 18.6 percent to 927.7 million dollars, reaching 141 countries and regions. That's diffusion. More machines, more places. That's a volume story.
Whereas the chip and server numbers are a price story hiding inside a volume frame. And the reason I love this edition sitting where it does in the week is that it's the same constraint as PJM, just quieter and denominated in dollars instead of megawatts. A grid operator can tell you exactly how many gigawatts exist. Nobody can tell you exactly how much DRAM exists. You only know what it costs this week.
And that's a genuinely important distinction for anyone budgeting a 2027 buildout. Because a capacity constraint, like PJM's queue, eventually gets solved by building more of the constrained thing. More substations, more transmission. A price constraint from a real commodity shortage gets solved differently: demand rationing, or new supply, and memory fabs take years to add, not months.
So both the grid and the memory market are, in different units, saying the same thing. Somebody, somewhere, is about to pay more for the same gigawatt or the same gigabyte. And the honest watch item is that China's own domestic first half production figures aren't out yet. When they land, they'll show whether the fabs shipped more silicon or whether the export desks simply got paid more for the same silicon.
Production volume against export value. That's the cleanest test available. And I want to sit on the geopolitics for a second, carefully, because the first edition touched it. There's a US ban on AI chip shipments that's been extended to apply to Chinese firms even outside China, and reporting that China is weighing its own tighter export controls on AI models and chips.
And strip away the framing and it's a supply chain question about who is allowed to plug into the power and capital being built out. Chips are the demand side of the very same equation that rate policy and reactor financing are the supply side of.
Both governments here are doing what governments do, protecting industries they consider strategic. That's a legitimate thing for any country to do. What I find interesting as a money person is the second order effect: export controls change where capital wants to build, and that reshapes the whole map of who finances the next fab.
And it reshapes lead times. If you can't buy the finished thing, you build the capacity yourself, which is slower. Same pattern as the hyperscalers and nuclear. When you can't rely on the shared system, you vertically integrate, and integration always costs time.
Which brings us to the edition that honestly landed hardest for me. Because so far we've talked about hyperscalers and customs bureaus. Saturday's edition put a human on the other end of the machine.
At midnight Saturday, three New England utilities reset a number most customers will never notice. National Grid's fixed price supply rate in Massachusetts jumped from 15.372 cents to 17.185 cents per kilowatt hour. An 11.8 percent increase. Eversource and Unitil moved the same day.
And here's the twist that makes it a story rather than a bill. Most Massachusetts households won't feel it. State data shows just 26.6 percent of electric customers were still on that basic service rate as of March. That's the only group directly exposed to Saturday's hike. About 51 percent have already moved to municipal aggregation, where a town negotiates its own supply.
So even at the household level, you already see the sorting. The customers with leverage, capital, or scale keep finding ways to opt out of the price the grid charges everyone else. The customers without it get whatever the reset says.
And now zoom that all the way out, because the edition did, and the pattern is fractal. The International Energy Agency's Electricity 2026 report finds more than 2,500 gigawatts of renewable, large load, and storage projects stalled in grid interconnection queues worldwide. And it warns annual grid investment needs to rise roughly 50 percent by 2030, from about 400 billion dollars today, just to keep pace.
A lack of grid capacity is becoming the defining bottleneck. And who can afford not to wait in that queue? Meta signed deals in January with Oklo, Vistra, and TerraPower for up to 6.6 gigawatts of nuclear by 2035, aimed at its Prometheus AI campus in New Albany, Ohio, inside the PJM region.
And Oklo says pre-construction on its 1.2 gigawatt Ohio power campus starts this year, first phase targeted for 2030. None of that power waits behind households for a spot on the shared system. It's purpose built, contracted, financed directly by the buyer.
And Washington is helping. The Department of Energy loaned Constellation a billion dollars in November to support restarting nuclear capacity, and in June announced a 17.5 billion dollar American Nuclear Supply Chain loan program to accelerate ten large scale reactors.
Which is exactly the kind of patient, capital intensive bet a shared, regulated grid struggles to make on a data center's timeline. And I want to be precise about the shape here, because the edition was careful and correct. The grid is not shrinking for anyone. It is splitting into two lanes.
Say the two lanes.
Lane one: bankrolled by hyperscaler capex and federal loan guarantees, built to a 2030 or 2034 date the customer picked. BloombergNEF puts 2026 capital spending among the fourteen largest data center operators near 750 billion, up from about 450 billion last year, with 23 gigawatts under construction, about three quarters of it in the US. Lane two: still queued behind everyone else, funded a rate case at a time.
And Massachusetts households just found out which lane they're in. Now, I want to hold this with nuance, because it would be easy to make this a story about villains, and it isn't. The hyperscalers building their own reactors are actually adding generation to the system. Meta financing 6.6 gigawatts of nuclear is 6.6 gigawatts that didn't exist before.
That's the fair point, and it's the honest one. New private generation can, in principle, relieve pressure on the shared grid. The question the edition raised isn't whether power gets built. It's who pays to keep the shared grid running once the biggest, most creditworthy buyers have contracted their way off it.
Right. The cost of maintaining that shared system, the transmission, the reserve margin, the reliability everyone depends on, that cost concentrates on whoever's left on basic service. And they're the customers who never had the option to leave in the first place.
And the two clocks come back one more time. The first turbines at Oklo's Pike County site aren't due until 2030. The next basic service bill in Massachusetts is due in February. The private lane builds on a five year horizon. The shared lane reprices every six months.
So the watch item there is whether that gap widens or narrows this fall, when more state utilities file rate cases, and when the DOE loan program starts putting dollars behind actual concrete.
And I'd frame the constructive version, because there is one. If policymakers see the split early, there are tools to keep the shared grid healthy: how you allocate interconnection costs, how large loads pay for reliability they draw on. This is a solvable design problem. The editions are describing a fork in the road, not a foregone conclusion.
Which leaves the fifth edition, and it's almost the meta commentary on the whole week. Because all of these numbers, the memory price, the tariffs, they don't land in a vacuum. They land in August.
So the argument here is about depth, not direction. August is the thinnest volume month on the trading calendar. By the second week, entire desks at major banks empty out. Senior traders take the beach weeks. And the volume that normally absorbs a surprising headline simply isn't there.
So the same piece of news that would move a stock two percent in April moves it five in August, because there are fewer buyers and sellers on either side of the trade to cushion it.
And two live wires are already sparking as the month opens. One, the memory chip price shock, the same one from the China edition, which has already rewritten one export cycle but hasn't yet hit the next round of consumer electronics pricing, the back to school laptops, the phone refreshes, the server orders hyperscalers are still placing for the second half.
And two, a fresh round of tariff deadlines that trade policy trackers spent the last days of July updating, landing in early August. The Tax Foundation has been tracking the country by country schedule continuously through 2026.
And here's why Polly's supply chain lens and my markets lens finally meet on one shelf. Tariffs and memory prices touch the same product. The laptop, the phone, the server rack all have a chip inside that just got more expensive to manufacture, and depending on where the finished unit was assembled, may now cost more to bring across the border.
So it's the same physical object being repriced twice, once by the memory market and once by the tariff schedule. And the edition's discipline here is what I appreciated. It explicitly did not predict a selloff. It said the wrong question on August 2 is whether a selloff is coming.
The right question is narrower. When the memory repricing and the tariff schedule hit retail shelves and earnings calls this month, will the market read them at their actual size, or at August's inflated one?
And the concrete thing to watch is not a forecast of direction. It's a measurement of depth. Order book volume on the days the next tariff detail or the next memory supply number lands. A five percent move on light volume this month says far less than the same move in October.
History says thin rooms don't create the news. They just turn up the volume on it. And I think that's the perfect note to pull the week together on, because the whole week was about volume in two senses of the word.
Go on.
Volume as in electricity and dollars flowing through a system. And volume as in loudness. And this week every one of these stories was a case of the second getting mistaken for the first. A 96 percent export number that was loud but not large. A Fed decision that was loud but held steady. An 11.8 percent rate hike that most households won't even hear.
And underneath the loudness, the actual machine kept running on one quiet constraint. Power. The Fed sets the price of building it. PJM sets how fast you can connect to it. The memory market sets what the chips that consume it cost. The hyperscalers are building their own supply of it. And a household in Massachusetts pays for the shared version of it.
Five editions, one bottleneck. The AI boom did not run out of money this week, and it did not run out of ambition. A trillion dollars of capex is proof of that. What it ran into is a queue for electrons. And the most interesting thing about any queue is who gets to skip it.
And the honest, hopeful version is that a queue is a design, not a destiny. You can build more lanes. You can decide how the cost of the shared lane is shared. PJM's next auction, the DOE's concrete at Pike County, China's actual production figures, and the order book depth this month. Four measurements, and every one of them tells you whether the loud version of a story matches the large one.
So the one thing to carry out of this week. When a number gets loud, ask what unit it's really in. Megawatts or dollars. Volume or price. Capacity or cost. Because this whole industry is being priced right now, and the market that prices it best is the one that can tell the difference.
The machine runs on electricity and silicon. The news runs on how loudly we describe them. Our job is to keep those two straight.
The Discussion is written and voiced by AI, grounded in the week's editions. Mona follows the capital; Polly follows the physical world.