Venture, inverted

PETER THIEL has, over twenty-three years, built two firms whose business is concentration. Palantir, the data-analytics company he co-founded in 2003, sells a small number of very large contracts to a small number of governments. Founders Fund, the venture firm he started in 2005, has spent the past decade moving away from the diversified portfolios most of its peers maintain and toward a structure that resembles an industrial holding company more than a venture fund. On May 1st the firm closed its fourth growth-stage vehicle with $6 billion in commitments — its largest haul ever and the fastest succession in its two-decade history. The fund will back roughly a dozen companies over two to three years. $4.5 billion came from outside limited partners, sovereign wealth funds prominent among them; $1.5 billion came from the partners themselves. That last figure, equal to a quarter of the fund's capital, is the disclosure worth lingering on.

Founders Fund's prior vehicle — a $4.6 billion fund closed less than a year ago — backed seven companies with average checks of roughly $600m. Of that, $1.25 billion went into Anthropic, the AI lab behind Claude, at a $350 billion valuation, the firm's first direct check into the Claude maker. $1 billion went into Anduril Industries, a defense-technology firm co-founded by Trae Stephens, a Founders Fund partner. The two checks together accounted for nearly half the deployed fund. The remainder funded follow-ons into Stripe, Ramp, OpenAI and Cognition AI, a coding upstart. By any conventional reading of venture portfolio construction, this is a strange-shaped fund. By the new reading, it is the only shape that works.

The category is moving in concert. In February, Thrive Capital, Josh Kushner's investment firm, closed a $10 billion vehicle, its largest ever. In April, Sequoia Capital, a Sand Hill Road stalwart, closed about $7 billion for what its new leaders described as biggest-bets investing. Andreessen Horowitz, a Menlo Park venture firm, closed a $6.75 billion growth fund in January. The four most recent late-stage growth vehicles raised by these four firms total $29.75 billion in dry powder. Assuming each backs roughly a dozen names, the implied capital per investment exceeds $600m — more than twice the average late-stage round in 2026. None of the four is structured to deploy capital the way late-stage venture has been deployed for the past decade. They are designed to do something else.

Bidding farewell

Start with the mechanic. Founders Fund deployed its previous growth fund in under twelve months — a third the typical pace for the firm, and faster than any vehicle of comparable size in the industry. According to people familiar with the deployments, the firm rushed through the capital by approaching companies to take on more capital before they had even kicked off formal fundraising. The phrasing is worth lingering on. The traditional architecture of late-stage venture is the auction: a company decides it wants to raise, hires a banker, runs a process, and accepts the cleanest term sheet at the highest valuation. Pre-emption inverts that sequence. Capital arrives before the round exists, on insider terms, with an offer the company has neither solicited nor benchmarked. Whether that offer is competitive in the auction sense is unanswerable, because there is no auction to compete against.

What that looks like in practice is straightforward. A late-stage company with strong growth metrics — Stripe, Anduril, OpenAI all qualify — does not need to raise capital this quarter. The firm approaches anyway, with a term sheet, a valuation, and a check ready to wire. The company is not auctioning anything; it is being offered something. The decision is bilateral, the price is set in conversation, and the cap table absorbs $500m or more of new equity from a single allocator before any other firm has been told a round is happening. That is what "before they had even kicked off formal fundraising" means.

The mechanic favors capital in three ways at once. The investing firm avoids the runup in valuation that competitive bidding produces. The company avoids the disclosure burden of a wide process. The existing investors avoid having to share information with strangers. All three constituencies prefer pre-emption to the auction once they have access to it. The auction was tolerated for thirty years because no firm had the balance sheet to consistently pre-empt. By the start of 2026 four did.

Take the math. A $4.6 billion fund deployed across seven companies works out to roughly $657m per name. A $6 billion fund across twelve names averages $500m. The combined late-stage stack — Founders Fund, Sequoia, Thrive, a16z — is sized to deploy hundreds of millions per check into a universe of perhaps thirty to fifty companies. That universe is not coincidentally small. It is the entire population of late-stage privates in which a $500m check can be absorbed without breaking the cap table — a population mostly defined by the arrival of frontier artificial intelligence and its capital appetites.

The orthodoxy this displaces is forty years old. The venture model that produced Sequoia's early bet on Apple, Kleiner's bet on Genentech and Benchmark's bet on eBay was built around the assumption that no one knew, ex ante, which name would become the outlier return. The portfolio was the answer: thirty to fifty checks per fund, most of which would write to zero, one or two of which would return the entire vehicle. Power-law diversification was a confession of ignorance disguised as a strategy. Twelve names is a different confession. It is the confession that the universe is now small enough to be picked.

Consider the general partner commitment. Across all venture funds tracked by Carta, a fund-administration platform, the median GP entity commitment is 1.7% of fund size; for mega-funds above $250m, the median is 1.5%, with the 75th percentile at 2.1%. Founders Fund's fourth growth vehicle is committed at 25% — more than ten times the upper-quartile benchmark. There is no public precedent for a venture fund of this scale with this much partner capital. The structural implication is that the $1.5 billion is not signalling alignment in the conventional sense, where a GP puts a token amount of personal money in to assure LPs that the fund is being underwritten as if it were the GP's. At 25%, the partners are no longer signalling alignment; they are constructing a personal investment vehicle and inviting outside capital to ride alongside. The carry calculus changes. The governance calculus changes. The question of whose interests the fund actually represents changes. Fund-of-one structures exist at this commit ratio in private equity. They are not, by tradition, called venture.

The Anduril check from Fund III illustrates the texture. Trae Stephens co-founded Anduril and remains a general partner at Founders Fund. The $1 billion that flowed from Fund III into Anduril's Series G in June 2025 — at the time, the largest single investment in the firm's history — was capital provided by LPs and partners into a company partly built by one of those partners. At a 25% GP commit, roughly $250m of any analogous Fund IV check would be partner capital flowing toward partner-adjacent equity. Conventional venture conflict frameworks were not designed for ratios at this scale.

The crux is the compute. The five largest American hyperscalers — Alphabet, Amazon, Meta, Microsoft and Oracle — are projected to spend between $660 billion and $690 billion on capital expenditure in 2026, according to Futurum, a research firm, with the vast majority of that deployment directed at AI compute, data centers and networking. Anthropic, which posted run-rate revenue of $19 billion in March 2026, up from $1 billion fifteen months earlier, has committed $50 billion to American AI infrastructure on top of partnerships with Google and Broadcom. OpenAI is projected to lose $14 billion in 2026, with cumulative losses through 2028 reaching $44 billion. Dario Amodei, Anthropic's chief executive, has said that "there is no hedge on earth" against the risk of overbuying compute. The remark applies with equal force to the firms financing the buying.

The companies that can absorb late-stage checks of $500m and up are, with rare exceptions, companies whose capital intensity is set by the cost of training and serving frontier models, the cost of building defense-grade autonomy stacks, or the cost of putting hardware into orbit. SpaceX, Founders Fund's flagship holding and an expected initial public offering candidate later this year, is the senior member of that cohort. Anthropic, OpenAI and the small number of competing labs are the second tier. Anduril and a handful of defense peers are the third. Crusoe, an AI-focused cloud provider that Founders Fund backed early, is the fourth. The list of names able to absorb a $500m primary check without distortion is, in 2026, perhaps twenty to thirty long. The mega-funds know it.

That is what makes pre-emption rational. In a universe of twenty-five names, an auction is fatal. If a firm runs a process and compares term sheets, it concedes the round to whoever bids highest, and the winner pays a premium to the cohort of also-rans. The mega-funds have collectively decided to skip the bid by paying earlier — at insider terms negotiated before the round becomes a process. The mechanic only works if every major late-stage allocator tacitly agrees to play the same game; otherwise the first to break ranks wins by reverting to the auction. So far, none has broken ranks. Thrive's $10 billion, Sequoia's $7 billion and Founders Fund's $6 billion have moved through their target lists without an auction worth the name.

Consider the LPs. Sovereign wealth funds account for the bulk of Founders Fund's $4.5 billion outside capital, the firm's investors said. The pattern is broadly true of the cohort: Gulf, Asian and European sovereigns are now the dominant allocator into American late-stage venture. The American AI buildout — Anthropic, OpenAI, Anduril, SpaceX — is structurally financed by foreign capital intermediated through a small number of US-based growth managers. The Committee on Foreign Investment in the United States, which screens direct foreign investment for national-security risk, does not at present screen LP capital flowing into private growth funds.

The exit math has changed in step. SpaceX's expected IPO is, in this telling, an exception. Anthropic's path to public markets is at the company's discretion; with positive cash flow projected for 2027 and an enterprise customer base producing $19 billion in run-rate revenue, the question of when to list is no longer a financing question. Anduril, Stripe and Ramp can each raise multi-billion-dollar private rounds for as long as the mega-funds want to write them. The IPO has been demoted from a financing necessity to a discretionary liquidity event. That demotion is what shrinks the universe of names: most of the companies that would, in another era, have gone public to raise growth capital simply have not, because they have not needed to.

The arithmetic is brutal in concept and unforgiving in execution. Twelve names, $500m apiece, no auction, sovereign LP base, partner capital at 25%. If the names are right, the fund returns at multiples that make the 1.7% benchmark look quaint. If the names are wrong — if Anthropic and OpenAI converge into a duopoly one of them loses, if Anduril hits a procurement wall, if SpaceX's IPO underperforms — the concentration that delivered the upside delivers the drawdown. Power-law diversification, the orthodoxy of venture portfolio construction for forty years, was an answer to the problem of not knowing which names would win. The mega-funds have decided they know.

Improving model capability takes months, whereas building a new generation of compute infrastructure takes years. The capital is committed; the silicon, the power, the talent and the policy are not. Twelve names is a small number of bets to place. It is a smaller number still if one of them is the bet on whether twelve was the right number.