Everything About AI
NextFin News - Private equity is betting the next phase of the artificial-intelligence boom will be built with its own capital, and the size of the wager is testing how much risk limited partners are willing to absorb. AI companies took $355.9 billion of the $412.7 billion invested in U.S. venture capital in the first half of 2026 - 86% of every venture dollar deployed in six months - while private equity firms chased the same theme through data centers, power assets, and the enterprise software portfolios they already own. The question investors now face is not whether AI is real. It is whether the price being paid for it can ever be validated.
The Scale of the Bet
The concentration is historic. Rounds of $100 million or more accounted for 87.5% of everything deployed in the first half of the year. Deals below that threshold - which still make up most of the market by count - drew just $51.4 billion, and their share of total value has slid from 43.8% in 2024 to 33.1% in 2025 to 12.5% this year. Capital is not spreading across the AI ecosystem; it is pooling at the very top.
The flagship deal illustrates the pace. Anthropic closed a $65 billion round that lifted its post-money valuation to $965 billion, moving it ahead of OpenAI. The step-up was steep: three months earlier the company had raised at a $350 billion pre-money valuation, and this round marked it at $900 billion pre, a 157% jump. About $15 billion of the headline figure had been committed in earlier pledges, so less new money actually entered than the $65 billion implied - but the mark was real.
Private equity is not sitting this out. It is entering through the physical stack that AI requires. EQT partnered with data-center builder EdgeConneX to launch an AI investing strategy, drawing on a $100 billion digital infrastructure and energy portfolio. Blackstone acquired a significant minority stake in Rowan Digital Infrastructure, a hyperscale data-center developer. EQT's sixth infrastructure fund joined a consortium including BlackRock's Global Infrastructure Partners, CalPERS, and the Qatar Investment Authority to take power producer AES Corporation private in a deal valuing it at $33.4 billion.
The biggest players are building dedicated platforms rather than making one-off bets. KKR launched Helix Digital Infrastructure with more than $10 billion in committed capital from Nvidia, the Kuwait Investment Authority, and power company Vistra, targeting data centers, power generation, and connectivity for AI workloads. Nvidia has also partnered with six Wall Street firms - Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR - to raise more than $500 billion for AI infrastructure, with the chipmaker holding an option to backstop up to $125 billion, or 25% of potential deals. "Compute has become a critical infrastructure asset," KKR co-chief executives Joe Bae and Scott Nuttall said in a joint statement. The financing platforms are designed to let third-party investors treat AI compute as its own asset class.
The spillover into the trades that feed the buildout is measurable. Datacenter construction spending jumped 46% year over year in the second quarter of 2026, and HVAC contractors alone logged more private equity deals in the first half of the year than in all of 2025 combined. Overall PE deal count in the construction and engineering sector reached an estimated record 529 for the quarter - even as total deal value declined, a sign sponsors are chasing smaller, more numerous opportunities in the specialty trades rather than a handful of marquee platforms.
"Modeling exit multiples has always involved a lot of guesswork, but now it feels like throwing at a dartboard blindfolded," said one private equity veteran.
The quote captures the bind. Private equity is a long-term asset class, typically holding companies for three to four years. AI has made the far end of that holding period nearly impossible to underwrite, "almost regardless of sector," industry participants said at the Milken Global Conference in May. The irony is sharp: the same technology that is driving the deal flow is destroying the visibility that deal flow requires.
Why This Is Different From a Normal Cycle
The first cut is to separate what is structural from what is cyclical - because conflating the two produces the wrong conclusion. The two forces are running in opposite directions at the same time, and most commentary blends them into a single, muddy verdict.
The structural leg is the buildout itself. AI is rewriting the cost base of software and creating a new physical infrastructure layer: data centers, power transmission, cooling, and the grid upgrades that computing demand requires. That demand does not depend on which foundation model wins. A hyperscaler needs the same megawatts whether it trains on one vendor's chips or another's. Capital flowing into contracted infrastructure cash flows is a regime shift, not a sentiment swing. It will not revert on its own. The evidence for a permanent regime change is in the counterparties: hyperscalers signing long-term capacity agreements, utilities planning decade-long grid expansions, and sovereign wealth funds committing permanent capital rather than opportunistic funds.
The cyclical leg is the pricing. Series D+ AI companies now trade at 6.6 times the valuation of their non-AI peers. In the first half of 2026, $0.90 of capital demanded by a venture-growth-stage AI startup was met by $1 in supply - more money chasing deals than the deals can absorb, giving the hottest companies leverage to be highly selective about who gets on their cap tables. That is a liquidity condition, and liquidity conditions reverse. When they do, the 6.6x premium is the first thing to compress. A cyclical claim needs a demonstrated mean-reversion pattern, and the record provides one: of the 10 largest U.S. tech IPOs excluding SpaceX and Cerebras, only three traded higher a year after listing. Private capital that prices off the last round is systematically late to both peaks and troughs.
The distinction matters because it determines which half of the AI binge survives. The structural demand supports the volume of capital deployed; the cyclical premium determines how much of that capital will earn a return. Investors who treat the entire AI binge as a bubble will miss the infrastructure trades that persist through it. Investors who treat the entire binge as a permanent regime will overpay for assets whose cash flows have not yet been stress-tested. Both are right about half the market and wrong about the other half.
There is also a second-order problem specific to private equity: the disruptor is already inside the portfolio. Funds that bought enterprise software companies at pandemic-era multiples now own products that general-purpose AI threatens to displace. The modeling challenge is not abstract - it is a direct hit to the cash flows backing existing funds. This is the mechanism that separates PE from public-market investors: a public investor can sell an AI-disrupted software stock in milliseconds. A PE fund holds it for three to four years, marks it quarterly, and distributes it only when a buyer can be found.
The financial engineering that filled the gap is now a risk factor of its own. Moody's Ratings noted that PE-backed companies in the U.S. raised approximately $94 billion in borrowing last year to fund dividend payouts - a practice that heightens financial risk at the portfolio level precisely when operating visibility is falling. Dividend recapitalizations do not create cash flow; they move claims on it forward in time, ahead of the operating turnaround that may never arrive.
The Validation Gate the Market Cannot Avoid
Here is the second-order question most investors are not asking: who sets the price anchor for this entire complex? In public markets, the answer is the daily auction. In private markets, the answer has been the last round - until the last round was set by a handful of buyers with more capital than alternatives.
That changes when Anthropic and OpenAI list. Both have filed confidentially to go public, and their debuts will be the first genuine public-market test of the private AI thesis. Analysts put the stakes plainly: the VC-backed AI thesis and its elevated valuations have not been validated by the public markets yet, and these offerings will soon put it to the test. The upcoming quarters will reveal whether the rest of the market is willing to pay as much for AI as venture capital has - or whether a significant repricing is needed for the asset class.
The warning signs around that gate are already visible. SpaceX's $1.7 trillion initial public offering in the second quarter raised $75 billion and was the largest IPO of all time, yet its stock fell below its IPO price. Cerebras completed a $34.3 billion IPO after canceling a 2025 attempt; its shares opened at more than double the offer price before sliding back below it. The pattern is consistent: private marks exceed what public buyers will pay, and the gap closes after listing, not before.
Concentration compounds the risk. Analysts cautioned that a market this dependent on a single theme faces a broad correction if AI growth or returns disappoint - and even if AI delivers, the power law means returns will concentrate in a few winners, leaving a wide field of companies that raised at elevated prices exposed. In the first half of 2026, three venture firms - Andreessen Horowitz, Thrive Capital, and Founders Fund - gathered $34.8 billion, or 48% of every fundraising dollar in the period. Corporate venture arms took part in just 21% of deals, their smallest share in 10 years, as parent companies hoarded cash to cover their own climbing AI bills. When the strategic buyers step back, the financial buyers are left holding the theme.
The same concentration is visible in the public markets that will set the anchor. The top 10 stocks now command a larger share of the S&P 500 than at any recent point in history - above the roughly 25% they held during the dot-com era and the roughly 15% of 1980, according to UBS. A private equity industry selling assets into a public market that is itself historically concentrated has fewer exits and less pricing power than it assumes.
The Counter-Case: This Is Not 2000
The strongest argument against the bubble label is that the AI buildout rests on real balance sheets and real revenue, not dot-com-era promises. Gene Goldman, chief investment officer at Cetera Financial Group, has argued that AI stocks are not in a bubble because today's AI giants carry lower debt-to-earnings ratios than WorldCom did, and companies like Nvidia and Meta Platforms are already reporting strong profit growth from AI - something largely absent in the speculative era 25 years ago.
There is force to that view. Data-center infrastructure carries contracted, long-dated cash flows from hyperscale customers. A power purchase agreement is not a pro forma user-growth model. And Bridgewater founder Ray Dalio, who called the AI boom an "early bubble" in January 2026, was describing a phase of a cycle, not a verdict on the technology's durability. The physical stack has customers with investment-grade credit ratings and contracts measured in years, not quarters.
But the counter-case answers the wrong question. The issue is not whether AI creates value. It is whether the value accrues to the assets private equity is buying at 2026 prices. A structural shift can still produce terrible investment returns if the entry price assumes perfection. The dot-com era produced real winners and permanently changed the economy; it also produced a decade of losses for capital that arrived late. The two outcomes are not contradictory - they are the normal distribution of a technological revolution.
There is also a narrower version of the counter-case that deserves attention. Partners Group's Anastasia Amoroso argued in May 2026 that private credit exposure to AI-disrupted software is closer to 7%, not the 20% that some estimates suggest - implying the direct risk is smaller than the panic assumes. If she is right, the contagion channel from AI disruption into private credit is contained. But that argument addresses the size of the existing exposure, not the price being paid for new exposure today. A small position bought at a high price can still lose money.
Who Benefits, Who Is Exposed
The asymmetry is clearest when sorted by asset type rather than by theme. Beneficiaries sit in the physical stack: data-center developers, power producers, cooling and electrical contractors, and the infrastructure funds that can underwrite contracted cash flows with long-dated liabilities to match. Select managers with exposure to AI-enabling technologies have already posted windfall gains - Bain Capital, Silverlake, and KKR among them - and one investment that listed above a $20 billion valuation delivered multi-billion-dollar returns for Advent and co-investors, reflecting demand for energy and infrastructure solutions tied to data-center expansion.
The exposed sit in two places. First, funds holding enterprise software bought at pandemic multiples, where AI is a demand risk to the very products backing the fund. Second, any capital that paid the 6.6x AI premium expecting public-market validation that never arrives. The illiquidity trap makes rotation difficult: an estimated 32,000 private equity-backed companies remain unsold, carrying roughly $8 trillion in valuation, while PE fundraising fell to a seven-year low of $414.2 billion in 2025. Buyout funds made up just 33% of private capital raised through drawdown vehicles last year, down from 41% in 2024. Limited partners who want to reduce exposure cannot simply sell; they must wait for a buyer who faces the same modeling problem.
The exit environment reinforces the trap. PE asset sales slowed sharply in early 2026 as AI-driven volatility and geopolitical tensions strained an already fragile market, and elevated entry valuations from the pandemic deal boom left many firms reluctant to sell at discounted prices. In the UK, nine mega-deals represented about 58% of private equity exit value in the first half of the year - a polarized market where only the highest-quality assets move and the rest wait.
LPs are responding by becoming more selective rather than by exiting. An analysis of co-investment found that until AI exits prove returns at scale, limited partners should apply the same fee discipline to direct AI exposure as they do to the rest of their co-investment programs. That is a quiet form of rationing: the capital is still available, but the terms are tightening for the riskiest slice of the theme.
What to Watch
The forward look splits cleanly by horizon. In the short term - the next six to twelve months - momentum should persist. Fundraising will concentrate in mega-funds, exits will remain frozen, and dividend recapitalizations will continue as managers hunt for distributions. The IPO backlog, not deal flow, is the constraint.
Over the medium term - one to three years - the validation window opens. Anthropic's and OpenAI's listings will set the public anchor. If either prices below its last private round, or trades below its IPO price for a sustained period of three months, the 6.6x premium compresses toward peers and data-center deals entered at 2026 multiples face write-downs. That is the single falsifying signal for the structural-premium thesis: a quantifiable threshold, not a vague warning to "watch the IPOs."
Over the long term - three years and beyond - the structural demand for compute, power, and cooling persists regardless of which model vendors win. Capital that bought contracted infrastructure cash flows survives the repricing. Capital that bought AI software at venture premiums does not. The technology outlives the valuation.
The base case is a split outcome: infrastructure holds, software reprices, and dispersion between managers widens into the widest gap the industry has seen in a decade. The upside case is that the IPOs validate private marks and the binge proves prescient - Anthropic and OpenAI price at or above their last private rounds and trade up, pulling the entire private complex higher with them. The downside case is that public markets refuse the anchor, the premium collapses, and the illiquidity trap turns mark-to-model valuations into mark-to-reality losses as LPs discover they cannot exit positions they no longer want.
Private equity's AI binge is not a bet on whether the technology works. It is a bet on who gets paid for it - and at what price. The technology will almost certainly work. The price is the part that has not yet been tested.
Explore more exclusive insights at nextfin.ai.
