Carta published its second-quarter fund data on September 23. Nine years in, the typical venture fund raised in 2017 has returned 0.37x in cash, against a value of 1.89x at Carta’s last public reading — about a fifth of what it says it is worth. The typical 2018 fund has returned 0.15x. The best 2017 funds are marked at 4.14x, up from 3.31x two years ago, even as their internal rates of return slipped from 28.7% to 25.7%: gains that are real on paper and arriving later than planned. In Carta’s words, they are “still on paper, not realized into distributions.”
The same week showed both sides of that gap in motion. Prices kept rising where someone was buying: Instinct raised $1 billion at four times its August price, and Modal and Fal are reportedly in talks at $15 billion each. The door that turns a price into cash stayed narrow: of the $161.4 billion US listings have raised this year, $86.2 billion was SpaceX, and there have been three debuts since Labor Day. What the best funds are short of is not value but time, and this week time got more expensive — the 30-year Treasury reached its highest since 2004, and SoftBank priced notes at up to 9.75% to fund its last OpenAI tranche. The paper is real. The cash is late.
Carta published its fund-performance data for the second quarter on September 23, covering venture funds that use it as their administrator. Nine years in, the median 2017 fund has returned 0.37x in cash; the median 2018 fund 0.15x and the median 2019 fund 0.04x. At Carta’s last public reading of value, for the end of 2025, that same median 2017 fund was marked at 1.89x — so the typical fund has paid out about a fifth of what it says it is worth. At the top, the gains are larger and slower: the 90th-percentile 2017 fund is marked at 4.14x, up from 3.31x two years earlier, while its IRR fell from 28.7% to 25.7%. Because IRR is tied to timing and TVPI is not, Carta reads that divergence as gains arriving on a longer timeline than expected — “in many cases, those gains are still on paper, not realized into distributions.” Two limits travel with the numbers. Carta’s platform leans toward smaller and emerging managers, so its medians may sit below the industry’s. And the 1.89x value comes from its previous public report, two quarters before the 0.37x.
Bloomberg counted only three debuts in the stretch after Labor Day, usually one of the busiest of the year, and found that of more than two dozen companies that filed publicly since July with a top Wall Street bank, nine had gone public. CVC-backed Bamboo Insurance postponed its offering on September 22, within days of Holtec Nuclear; both cited market conditions. The year is not small: US listings have raised $161.4 billion excluding blank-check and other financial vehicles. But $86.2 billion of that is a single listing, SpaceX’s, and five of the year’s ten largest listings trade below their offer price. This is the door that turns a venture mark into cash — and even a listing pays early holders only once its lockup ends. This month it has admitted very few.
Instinct, the personal AI assistant from Spear Street Technology that users reach by phone or text, announced on September 28 a $1 billion Series C from Sequoia, Benchmark and Coatue at a $10 billion valuation. In late August it was valued at $2.5 billion. Each step is a new price set by a new buyer of a minority stake. None of it has returned a dollar to anyone who bought in earlier, and each step raises the value that an eventual sale or listing must reach to make the newest investor whole.
Oracle sent a force majeure notice on Project Jupiter, the 2.45-gigawatt Stargate campus in Santa Teresa, New Mexico, to Blue Owl, owner of the developer STACK Infrastructure — first reported by Bloomberg on September 24. The notice would let Oracle delay payments if the site misses its 2028 target, after a gas pipeline it depends on slipped to February 2027. Oracle says the project “remains on our planned schedule.” The market read it on its own terms: the stock fell 3.48% that day, the yield on Oracle’s 6.7% bonds due 2056 moved above 8% for the first time, and its credit default swaps reached a record. The “$165 billion” often attached to the project is the county’s ceiling for industrial revenue bonds, not its cost. Part of the bond move is the rate move — the 30-year Treasury reached its highest since 2004 the same day — but not the record in the swaps. A commitment on paper, and a market deciding what it is worth in cash, is the same gap this issue is about, one floor up.
SoftBank Group set terms on September 24 for the notes it launched the week before: $1 billion at 8.625% due 2030, $4.5 billion at 9.25% due 2032, $4.5 billion at 9.75% due 2034, and €500 million each at 7.125% and 8.0%, all at par and settling on September 29. The deal was not upsized. The proceeds fund the $10 billion third and final tranche of its $30 billion follow-on investment in OpenAI, expected to close on October 1. Reuters described it as the largest high-yield bond sale by an Asia-Pacific issuer on record. At the very top of the market, carrying a private stake on borrowed money now costs between 8.6% and 9.75% a year in dollars, on a stake that pays nothing until it is sold.
Nscale announced on September 25 a $3.36 billion convertible financing led by Third Point: $2.36 billion at closing and a $1 billion Nvidia commitment that funds in mid-November, with the notes converting automatically when its IPO completes. No price range has been filed since its September 18 S-1. Earlier in September it was reported to be seeking about $3.5 billion; the release confirms how much closed and how much is still a commitment. Separately, Fortune reported, from the filing and the Financial Times, that Spring, a ByteDance subsidiary, rented 2,304 Nvidia B200 chips at Nscale’s Norway site and accounted for 73% of its 2025 revenue.
Bessemer Venture Partners announced on September 23 that it had raised $5.75 billion in a single close — $1.75 billion for seed and early stage and $4 billion for growth — up from $4.6 billion across its 2022 funds. A day later TechCrunch reported, from an investor letter, that Lightspeed India Partners V is targeting $250 million, about 80% committed, half the size of its $500 million 2022 predecessor. Carta’s data the same week shows the pattern across its platform: 64% of 2025 fund capital went to funds above $100 million, against 38% in 2017, and the median fund of $100–250 million now has 45% fewer investors than four years ago.
Heidi, the Melbourne company that makes AI documentation tools for clinicians, announced a $340 million round on September 22 at a $900 million valuation, nearly double its October 2025 Series B. The split matters more than the total. $100 million is a Series C equity round led by Blackbird; $240 million is growth financing from General Catalyst’s Customer Value Fund, repaid out of revenue rather than exchanged for shares. About 29% of the headline set a price on the company. The rest is a claim on future revenue, which only works for a company that has revenue to claim — TNW reports Heidi went from $1 million to $50 million of annual recurring revenue in two years.
Ando came out of stealth on September 24 with a $20 million seed co-led by Accel and Index Ventures, for team messaging in which AI agents take part with their own identity and permissions. Ande announced $52 million across a combined seed and Series A from Lightspeed, Redpoint, Duration and Sierra, for managing corporate entertainment spend, and Sela $21 million across a combined seed and Series A led by Costanoa, for voice agents that, in the company’s words, “help originate” more than $1 billion of mortgages a month. Feather Robotics raised a $7.6 million pre-seed on SAFEs led by Gradient Ventures; one investor, the LSE-listed Seed Innovations, disclosed a $60 million valuation cap in its own filing.
Island, which makes an enterprise browser, raised a $400 million Series F led by Evolution Equity Partners at a $6.4 billion valuation on September 24, which it says more than doubles its 2024 price. Two days earlier Cyera, which sells data-security software, added $400 million led by Goldman Sachs Alternatives as an extension of its June Series G, at a valuation TNW reports as flat at more than $12 billion. Evolution Equity led both Island’s round and Cyera’s original Series G. One company took a step-up; the other took the cash at the price it already had.
Bloomberg reported on September 23 that Modal Labs, which runs AI workloads for other companies, is in talks at about $15 billion, roughly three times its May valuation, with Accel in talks to lead. The Information reported that Fal is negotiating at a $15 billion target on about $800 million of annual revenue. And on September 22 Bloomberg reported that Mirendil, a startup founded by former Anthropic researchers, is in talks at $5 billion, three months after a seed at $1 billion. None of the three has closed, and each is sourced to people familiar with the talks.
The Information reported on September 23 that DeepSeek founder Liang Wenfeng told investors its annualized revenue had passed $1 billion, up from under $500 million a few months earlier, and that customers held after its August price increase. Reuters had reported that increase in August: prices for its V4-Pro and V4-Flash models rose by 50% to 1,100% depending on model, token type and time of day. DeepSeek has not commented publicly, and an earlier report put its revenue for the first seven months of 2026 near $71 million, so the figure describes a recent month annualized, not a year of sales. The same week, OpenAI launched GPT-6 Sol and Luna below the prices of the models they follow — while no existing list price was cut.
A week after the Federal Reserve’s September 16 increase, the 10-year Treasury climbed to roughly 5.2%, its highest since 2007, and the 30-year reached about 5.47% on September 24, its highest since 2004. New York Fed President John Williams said “another rate hike may be appropriate by the end of the year,” and Governor Michael Barr and Philadelphia Fed President Anna Paulson also signaled further increases; CME FedWatch put the odds of an October hike at 66.4% on September 25. A Treasury buyback of about $4 billion of longer-dated securities on September 24 did not stop the rise. A venture mark is a claim on a payout years away, and a higher long rate lowers what that payout is worth today.
Databricks announced on September 24 that it had acquired Row Zero, a Seattle company whose cloud spreadsheet handles datasets of up to about a billion rows, to give its Genie AI assistant a governed spreadsheet interface. Row Zero had raised about $13 million, and TechCrunch reports Databricks is looking for more companies to buy. Terms were not disclosed, and Databricks is private, so it is not known whether Row Zero’s investors received cash or Databricks stock — which would be one mark exchanged for another. By count, a sale to a strategic buyer is how most venture-backed exits happen, and it does not wait for an IPO window.
The obvious read of this week is that venture is flush. Bessemer closed $5.75 billion, Instinct’s price quadrupled in a month, and Modal and Fal are reportedly in talks at $15 billion apiece. The more useful read is that only one of those numbers is a price anyone has paid — and it too is a mark, set by the latest buyer of a small stake. In the same week Carta published what older marks have turned into.
The typical fund raised in 2017, nine years in, has sent back thirty-seven cents on the dollar in cash against $1.89 of value for every dollar paid in — a fifth of what it says it is worth. The door that converts a mark into money is open mostly at the top: of the $161.4 billion US listings have raised this year, $86.2 billion was SpaceX, and three companies have debuted since Labor Day. In issue #029 I noted that SoftBank’s coupon was due to print on the twenty-fourth. It printed at between 8.625% and 9.75% — the price, at the top of the market, of carrying a stake that pays nothing until it sells.
None of this makes the marks fictional. The best 2017 funds may yet return four times their money, most venture exits by count are sales rather than listings, and Bessemer’s investors just committed $5.75 billion more, which suggests some of them have been paid. My claim is narrower: for the typical fund the cash is late, and it stays late through next year. The people who wait longest for it are the ones holding common — your team, whose shares turn into money only when someone buys the company.
So before the next raise, ask about the fund before the one writing your check — when it was raised and what it has paid back — and if the answer is vague, treat that as the answer. Then decide which part of your round the future has to justify. Heidi priced only $100 million of its $340 million round at its new valuation and borrowed the rest against customers who pay every month, where SoftBank borrowed against a stake that pays nothing until it sells. That is not timidity. It is choosing what your price has to carry.
The wager: at Carta’s last readings the typical 2017 fund had paid out about a fifth of its value, 0.37x against 1.89x. If its reports through mid-2027, or the latest published by October 31, 2027, put that share at 30% or more — through cash arriving or marks falling — the gap between price and payment is closing faster than I argued, and the clock I asked you to plan against is shorter than I said.
Marks are how belief gets priced. Cash is how it gets proven. In a market where funds are waiting on cash, the company that produces its own holds the leverage — so build the one whose customers pay it before any window opens.