The industry thought it was racing to build the smartest machine. It is discovering that the race is about who can still borrow, still get memory, and still turn the lights on — and all three are tightening at once.
By K. Neelu, ICTpost Global AI Intelligence
Colette Kress had been Nvidia’s chief financial officer for 13 years when, on Nov. 12, she became its chief executive in the space of a morning. Jensen Huang, who had run the company since co-founding it in 1993, resigned effective immediately amid a board investigation into reports that he had been targeted by a foreign intelligence effort. Five days later, Kress walked onto an earnings call that had been on the calendar for months and reported the largest quarter in the company’s history: $115.8 billion in revenue, more than double a year earlier, nearly all of it from data centers.
A founder-CEO would have spent that call talking about demand. Kress, a finance executive by training, spent it talking about limits. Nvidia would be “rigorously evaluating future investments, commitments, and guarantees,” she said, and would stay “disciplined in how we commit capital.” Gross margin would fall three points, because memory suppliers can now charge what they like. Next year’s growth forecast of 75%, she made clear, was a supply number, not a demand number.
Fourteen hours earlier, 2,800 miles east, PJM Interconnection — the grid operator for 67 million Americans from Chicago to the Jersey shore — had filed a 1,067-page proposal asking federal regulators to let it halve the time it takes to connect a data center to the grid. (Bloomberg) PJM is running a supply shortfall of nearly seven gigawatts and has watched its last two capacity auctions clear at the regulatory price ceiling.
Put the two events side by side and you have the whole story of AI in late 2026. The most profitable chip company on earth is constrained by memory it cannot buy, financing it is reconsidering, and power its customers cannot connect to. Capital, chips and electricity were supposed to be three separate problems with three separate owners. They have turned out to be one problem wearing three costumes — and every one of them now lands on Kress’s desk.
Capital: The Month the Spigot Narrowed
For two years the money was the easy part. The hyperscalers paid for data centers out of cash flow, and the argument was about returns, not funding. That ended this autumn.
The four largest cloud companies are on pace to spend roughly three-quarters of a trillion dollars this year, and Goldman Sachs expects the top five to pass $1.2 trillion next year. Operating cash flow is growing at about a quarter of that rate. Oracle already spends more on data centers than its business generates; Alphabet and Meta cross that line in 2027, Microsoft last, in 2028. (Epoch AI analysis, via Better Than Random)
So the industry went to the bond market, and the bond market started to say no — or at least, “not at that price.”
In the last week of October alone, Meta announced plans to raise $95 billion in equity, Microsoft sold $25 billion of debt, SpaceX closed $40 billion in debt financing and Oracle filed to raise $20 billion more. (Fortune) S&P cut Meta’s outlook to negative; Meta told its own shareholders it faced an “increasingly challenging environment for raising capital” and might need to move faster, before other companies absorbed what was left. OpenAI, the demand signal beneath much of this, cut its funding target by a third after a first close that fell short. Oracle bonds issued in February now yield almost two points more than they did then; the cost of insuring Alphabet’s debt has nearly tripled in a year. (Axios)
The people who move capital for a living have stopped being coy. Broadcom’s Hock Tan has said capital may become AI’s biggest constraint. BlackRock’s Larry Fink called the competition for it the fiercest of his career. Apollo’s Torsten Slok put it most plainly: “we have a shortage of capital, and we have an abundance of good ideas.”
This is where the first bottleneck folds into the second, and where Kress re-enters. As customers’ spending outran their cash, Nvidia itself became the lender, helping to arrange more than half a trillion dollars of outside financing and guaranteeing capacity for customers who could not otherwise afford its chips. In August, Kress conceded the obvious objection: “we know some will call this circular financing. We see it differently.” (Nvidia Q2 FY27 call, via Motley Fool) Tuesday’s promise to re-examine those guarantees was not boilerplate. It was the lender announcing it would be a more careful lender — delivered by the person who used to sign the checks and now runs the company.
Chips: The Shortage Behind the Shortage
The chip shortage of 2026 is not the one most people picture. The processors exist. What does not exist in sufficient quantity is the high-bandwidth memory stacked beside each one, and the TSMC packaging that bonds them together. TSMC’s chief executive has said its advanced packaging is sold out, and Nvidia is believed to hold roughly 60% of it.
The memory squeeze is worse, and the best illustration of it is a photograph. In June, at the Computex trade show in Taipei, Huang stopped at SK Hynix’s booth, picked up a marker and wrote “Please Make More” on a display wafer of next-generation memory. (VRLA Tech) It was a joke. It was also the world’s most important customer begging one of only three suppliers on earth for product.
The numbers behind the joke are stark. Nvidia’s long-term supply commitments — mostly memory — more than doubled in a single quarter, to $279 billion. (Nvidia 10-Q, via Capital & Compute) Kress has told investors the constraint will last through the end of fiscal 2028; Intel’s chief executive says there is “no relief until 2028.”
Memory is where chips and capital collide. Nvidia’s margin — the engine that produced its cash, its buybacks and its ability to backstop customers — is falling because memory makers have pricing power for the first time in a decade. The chip bottleneck is, in effect, taxing the capital bottleneck.
It is also rerouting the spending itself. On Oct. 28, Microsoft trimmed its 2026 capital budget by $5 billion — not because demand softened, but because rented GPU capacity it had contracted for slipped into next year. (Business Insider, via LetsDataScience) A company with $170 billion to spend could not spend it on schedule because somebody else’s racks were not ready. That is what a supply-constrained economy looks like at the very top of the food chain.
Electricity: Where the Clock Runs Slowest
Money moves in days. Chips move in quarters. Power moves in years, and that mismatch is the deepest of the three.
Three numbers from PJM’s annual capacity auction tell the story. Two years ago the clearing price was about $29 per megawatt-day. Last year it was $329 — the legal ceiling. This July it hit the ceiling again. (PJM auction results, via SMNewsNet) Households in 13 states are absorbing bill increases every year, and that is with a cap that the governors of Pennsylvania and Maryland had to fight for. PJM’s own president summarized July’s results in a sentence: demand is growing faster than supply.
Delivery Year
PJM Capacity Auction Clearing Price ($/MW-day)
2024/25
$28.92
2025/26
$269.92
2026/27
$329.17 (cap)
2028/29
~$325 (cap)
Source: PJM Interconnection (Capacity prices have risen roughly 11x in two years, repeatedly hitting the regulatory ceiling.)
The constraint is not a shortage of ideas. Tens of gigawatts of new generation, mostly gas, have cleared PJM’s queue. But gas plants need turbines, and GE Vernova’s order book now stretches into the next decade. (Forbes) In Texas, the grid operator received connection requests in a single quarter equal to its entire peak load, then froze the queue to audit them. In May, PJM received emergency authority from the Department of Energy to cut off data centers first in a grid emergency — the moment the buildout’s newest customers were formally placed last in line.
That is the context for Monday’s filing. Even the industry it is meant to help is skeptical: data-center trade groups warned PJM’s board in October that combining the queues “could have the opposite effect.” Globally, the International Energy Agency expects data centers to consume more electricity this year than Japan, and to double by 2030. (IEA)
And electricity folds back into capital. The gas plants, the nuclear contracts, the $12 billion bond vehicle for Meta’s gigawatt-scale campus in El Paso — all are financed with the same debt the bond market is now repricing. Power is no longer a utility problem the AI industry can outsource. It is a line on the balance sheet that the ratings agencies are reading.
Why Three Bottlenecks Are Really One
The conventional view treats these as independent shortages, each to be solved by a different industry on its own timeline. The past month says otherwise. Each constraint has become the cause of another.
Capital buys chips, but chips are rationed by memory that is pricing itself upward, which compresses Nvidia’s margin, which was the cash that let Nvidia finance customers who could not otherwise buy chips. Chips need power, but power arrives on a five-year clock, so paid-for GPUs sit in contracted-but-undelivered capacity — exactly why Microsoft could not spend its budget. Power needs capital, and the investors widening spreads on Oracle are the same ones asked to fund turbines that will not ship until 2031.
So the binding constraint rotates. In 2024 it was chips. In 2025 it was power. In the final quarter of 2026 it is capital — and capital is binding only because chips and power have made the returns on it slower to arrive than anyone promised. Gartner’s arithmetic captures the gap: data-center owners need about $8.5 trillion in cumulative revenue by 2030 to justify what they are building, against perhaps $3.6 trillion the market can plausibly produce. (Gartner, via Fortune)
For investors, the question has shifted from “who has the best model” to “who can self-fund the next two years.” For operators, the winners will be whoever owns a bottleneck rather than rents one — which is why Nvidia has secured gigawatts of power in Ohio for OpenAI, why hyperscalers are buying gas turbines and nuclear output outright, and why memory makers are suddenly the most courted companies in semiconductors. For Washington, the politics have inverted: a buildout sold as reindustrialization now has at least eight states moving to push grid costs onto data centers, and the largest grid in America holding authority to switch them off first.
Which brings it back to a CFO who became a CEO on a Thursday morning. Kress’s first act as chief executive was not to promise more. It was to promise more care. That is what a bottleneck sounds like when it finally reaches the top.
The Kicker
The winners of the AI age may not be the companies that build the smartest models. They may be the ones that control the scarcest inputs. In 2023 that meant GPUs. In 2026 it means memory, power and capital. Tomorrow it may mean something else. But the lesson is the same one every industrial revolution has taught: a technology is governed not by what it can do, but by what it cannot get enough of. editor@ictpost.com
K. Neelu writes column on the business, policy and geopolitics of artificial intelligence for ICTpost Global AI Intelligence.
