Morgan Stanley: $33.5B Can't Buy the Power AI Needs
Thirty-three and a half billion dollars. That is how much venture money has poured into American startups whose entire job is making sure AI data centres do not run out of electricity — or run out of clever ways to use it. The figure covers funding raised by early October 2026, and it is nearly double what the same corner of the market managed in all of 2025.
And the bank that counted it all thinks the pile of money may already be the wrong scoreboard. In a research note relayed by ANI on October 13, Morgan Stanley warns that the next phase of this trade will not be judged by capital raised. It will be judged by capacity deployed. Cash, it turns out, cannot buy megawatts — and 2026 is the year Silicon Valley is learning that the hard way.
The note lands at a moment when the AI buildout's money story is splitting into two streams. We have been tracking both on this page: the credit markets, where the giants are borrowing at a scale Wall Street has never seen, and the equity layer, where startups are now raising at a pace that would have been unthinkable two years ago. Morgan Stanley's new numbers suggest the second stream is accelerating even as the first one starts to strain. The question the report quietly asks is whether either stream is actually buying the one thing the buildout needs most — electrons, on schedule.
Where the $33.5 billion actually went
The headline figure is startling, but the shape of it matters more. "Venture funding for this universe has accelerated sharply in 2026, reaching $33.5 billion by early October, nearly double the 2025 total," the report says, describing a universe of US startups built specifically around AI data-centre power and computing constraints.
Those startups have now raised more than $50 billion since the beginning of 2024, spread across four areas: expanding power supply, improving computing performance per unit of electricity, integrating power and computing systems, and developing new data-centre designs and locations. More than 80 per cent of the 2026 money went to the first two — companies trying to expand power availability and squeeze more computing out of every watt.
The bank's three-word diagnosis of the whole sector is blunt: "Power is the key constraint." Not chips, not talent, not demand. Power. And the next phase of AI infrastructure, the note argues, is shifting beyond simple access to computing capacity toward integrated systems that coordinate power, computing, networking and cooling as a single problem. That is a genuine strategic turn. For three years the industry treated electricity as a procurement detail — something the utility handled. It is now the product.
The strangest stat in the note: power and compute investors are merging
Buried in the report is a figure that tells you how fast the mental model is changing. In 2026, 73 per cent of funding rounds in the power and computing segments included investors who had already backed companies in the other segment. In 2024 that figure was 49 per cent. Before 2021, it was 16 per cent — power and compute investors have effectively merged into one trade, and the money is following.
The note also flags the uncomfortable spread of timelines inside the $33.5 billion. Investors are funding everything from cloud providers deploying computing capacity today to advanced nuclear and fusion projects that may take years to generate any commercial output. Some of this money is buying servers. Some of it is buying a bet that physics will cooperate on a decade-long schedule. The report does not pretend those are the same risk. Neither should you.
The $1 billion fringe: data centres in orbit
Then there is the part of the note that reads like science fiction until you check the cap tables. Companies developing orbital and modular data centres — including Starcloud, Cowboy Space and Armada — together raised nearly $1 billion in 2026, according to the report. One billion dollars, for the proposition that the grid problem is so hard it might be easier to leave the planet.
Starcloud is the furthest along, and its story is worth a closer look because it shows what "capacity deployed" actually means in practice. In August, the Redmond, Washington startup raised $250 million at a $2.3 billion valuation in a round led by Manhattan West, with Nvidia and Cisco Investments among the new investors, according to Reuters. The raise brought its total funding to $450 million since its 2024 founding. The company's pitch is disarmingly simple: terrestrial data centres face rising power costs, grid constraints and hostile permitting; space offers uninterrupted sunlight and a vacuum that sheds heat through radiation.
"This fresh capital empowers us to build the infrastructure to launch many more of Nvidia's most advanced GPUs into space," Starcloud co-founder and CEO Philip Johnston said in a news release announcing the round, via GeekWire. And the company has already put hardware where its pitch is: last November it flew an Nvidia H100 GPU to orbit for the first time and used it to train a small language model, NanoGPT, in space. In February it filed with the FCC for a constellation of up to 88,000 satellites targeting roughly 20 gigawatts of orbital computing capacity.
The fringe gets funded first. It gets validated last — which is exactly Morgan Stanley's point.
"From capital raised to capacity deployed"
The report's central warning deserves to be quoted in full, because it is the sentence the whole $33.5 billion will be judged against: "The next phase of validation should increasingly shift from capital raised to capacity deployed."
What that means, in the note's telling, is that the scorecard changes. Commercial agreements, regulatory approvals, initial deployments and evidence of competitive operating costs become more important than the size of the last round. It is a polite way of saying the pitch-deck phase is over. The firms that can show signed customers, permitted sites, energised racks and unit economics that survive a utility bill will keep getting funded. The rest will discover that $33.5 billion was a high-water mark, not a floor.
The report is unsparing about what stands between the money and the megawatts: licensing requirements, construction delays, grid connections and long procurement cycles. It adds that local power availability, transmission constraints and regulations are likely to make the economics of data-centre projects increasingly dependent on where they are built. Location, in other words, is becoming a moat — or a grave. A startup with a brilliant cooling design and no interconnection agreement is worth less than a mediocre one sitting next to a substation with spare capacity.
We pulled the ANI wire copy of this report and cross-checked its figures against three independent relays — the $33.5 billion, the $50 billion, the 73 per cent, the $1 billion orbital figure all held to the decimal. We also pulled Starcloud's August raise terms from the Reuters relay to ground the orbital example, and checked GE Vernova's turbine backlog numbers against Utility Dive's earnings coverage. The figures in this article are the ones that survived that cross-check.
The pipes are the real bottleneck
If you want to understand why Morgan Stanley is right to be nervous, look at the companies that sell the actual equipment. GE Vernova — the power-equipment giant spun out of General Electric — is now sitting on roughly $176 billion in total order backlog. Its gas-turbine backlog alone has climbed to 116 gigawatts, up from 100 gigawatts as recently as the first quarter, and management expects at least 125 gigawatts under contract by year-end, according to Utility Dive's earnings coverage. In the second quarter the company signed 20 gigawatts of turbine orders and slot reservations from about 100 customers across 26 countries — and roughly one in five of those customers is a data-centre developer.
"If we begin with the end markets, I would tell you we continue to see very strong and durable demand," CEO Scott Strazik told investors at the Morgan Stanley Laguna conference in September. He described the company's once-ambitious $200 billion backlog target as, in his words, "a very humble milestone." Turbine delivery slots are now reportedly tightening through 2030 as data centres compete with utilities for manufacturing capacity that simply does not expand quickly.
The demand side of that equation is just as stark. The Wall Street Journal's relay of a Morgan Stanley analysis, carried by FN News this month, puts global data-centre electricity demand at about 56 gigawatts next year against just 51 gigawatts of newly available power capacity — and by 2028, demand of 86 gigawatts against 68 gigawatts of supply, an 18-gigawatt shortfall. The cumulative electricity needed over the next two years, the analysis says, is equivalent to the consumption of roughly 118 million households. Morgan Stanley's own sustainability research now frames it plainly: energy supply has become the primary bottleneck for large-scale data-centre deployment, overtaking semiconductor availability.
The industry spent two years fighting over GPUs. The fight that decides the next two years is over gas turbines, transformers and grid connections — equipment with lead times measured in years, not quarters.
The debt story's next chapter
This is the startup-and-equity layer of a story this publication has been tracking from the credit side for weeks. Earlier this month we reported, via the Financial Times, that AI-related borrowing collapsed to $23 billion in September — less than half of August's pace and far below June's $113 billion peak — while banks including JPMorgan, Morgan Stanley, SMBC and MUFG quietly began offloading data-centre debt through risk transfers. Some $38 billion of Oracle-leased construction debt in Texas and Wisconsin has been stuck on bank balance sheets for more than six months. The Bank of England's Financial Policy Committee put the year's AI-linked borrowing at $450 billion and flagged "circular trading" it cannot fully see through.
The equity layer Morgan Stanley has now measured is the mirror image. Where the credit markets funded the incumbents' buildout, venture money is funding the workarounds — the startups promising to make each megawatt go further or find new ones entirely. And both layers are now colliding with the same wall. Broadcom CEO Hock Tan said it plainly on a Singapore panel this week: "capital, not memory chips, is the biggest constraint on AI growth," as Reuters reported. Bain's Bridge Data Centres is out hunting $6.6 billion in debt for its Asian expansion. Anthropic has signed data-centre leases worth an estimated $530 billion across more than 10 clouds. Everybody is buying. The question Morgan Stanley is now asking is what, exactly, the buying has bought.
Related: AI's $113B Debt Machine Just Hit a Wall in September · Hock Tan: AI's Biggest Problem Isn't Chips — It's Cash · Bain's Bridge Seeks $6.6B Debt as Thailand Freezes 49 Builds
Our take: the scoreboard has to change
Here is the uncomfortable truth inside Morgan Stanley's polite analyst language: venture capital is very good at funding things and very bad at permitting them. The $33.5 billion is real money, and the 73-per-cent convergence stat shows the smartest investors understand the assignment — power and compute are one trade now. But a meaningful share of that capital is pointed at fusion timelines and orbital constellations that will not produce a single commercial megawatt this decade, while the unglamorous work — interconnection queues, turbine slots, substation upgrades — goes begging for attention it cannot buy its way out of.
The AI trade has been priced as a software trade: fast, scalable, asset-light. Morgan Stanley's note is a reminder that it is actually a utility trade with software margins — and utilities get built with permits, transformers and ten-year equipment contracts, not pitch decks. The startups that survive the "capacity deployed" era will be the ones that own electrons, not just architectures. The rest will be very expensive lessons in what happens when capital meets physics. Bet on the electrons.
Sources: ANI wire report on the Morgan Stanley research note (via The Economic Times, ETCIO and Lokmat Times, October 13, 2026); Reuters via SRN News on Starcloud's $250 million August raise; GeekWire on Starcloud's orbital programme and the Philip Johnston statement; Utility Dive on GE Vernova's second-quarter 2026 earnings and turbine backlog; BigGo Finance relay of CEO Scott Strazik's remarks at the Morgan Stanley Laguna conference; FN News relay of the Wall Street Journal's coverage of Morgan Stanley's power-shortfall analysis; Morgan Stanley Investment Management's "Navigating 2026" research themes; TechPulse USA's own prior reporting on AI data-centre debt, Hock Tan's capital warning and Bridge Data Centres.
Daniel Rao is a TechPulse USA reporter covering AI infrastructure and markets — the money, the megawatts, and the companies trying to turn one into the other.
Comments
Post a Comment