The Federal Reserve raised rates on September 16 for the first time in thirty-seven months, and the 10-year Treasury printed 5.01% that day — a level last seen in 2007 — before settling at 4.94% on the 17th. The relationship between those two numbers is the one worth keeping: measured on the 17th, the long end is up 88 basis points over twelve months against a policy rate down 50, a 138 basis point widening between the fed funds upper bound and the 10-year. Long-duration capital repriced for reasons that have little to do with monetary policy.
Where that repricing is visible, it is visible in public markets. Goldman Sachs named the mechanism; SoftBank went to the bond market for $10 billion; Nscale filed an S-1 and had to show a $1.02 billion loss against $140.6 million of revenue. It is invisible in what private equity capital paid — Crusoe took an initial close at $30.9 billion, and Hang Ten Systems added $53 million five weeks after its seed. Near a seed round it is simply unobserved: none of the quarterly instruments that price early-stage rounds published this week, and the only early-stage measurement anyone put out was about whether rounds finish, not what they cost. The useful question is not whether capital got more expensive. It is which market you are actually standing in.
The Federal Reserve raised its policy rate by 25 basis points to 3.75–4.00% on September 16, on a unanimous vote. The prior increase was announced in July 2023 — a gap of thirty-seven months. The same day, the 10-year Treasury yield printed 5.01%, its highest daily level since July 2007; the October 2023 peak was 4.98% and never crossed five. It settled at 4.94% on September 17. The more revealing number is the relationship between the two. Over twelve months the policy rate has fallen 50 basis points while the 10-year has risen 88 — so the spread between them swung from −44 bp to +94 bp, a 138 basis point steepening. Long-duration capital repriced for reasons other than monetary tightening. Chair Kevin Warsh gave three: economic strength first, then competition for capital — the hyperscalers “are out in the market raising funding, and so the competition for capital is real and I think it partly explains the increase in yields” — and geopolitics third, adding that “these things tend to be overdetermined.” He also said credit flows “have been robust, particularly for businesses,” and that he would be “hard-pressed to describe broad financial conditions as restrictive.”
AngelList published an analysis of its 2025 seed vehicles on September 18, with returns marked as of August 1, 2026. Rounds started by founders moved from started to wired 26% of the time; rounds with an outside lead wired 65% of the time. The two failures are not the same failure. Of the roughly three-quarters of founder-started rounds that never wire, AngelList writes that it “believe[s] more than 90% of the time that indicates the company failed to fundraise at all” — while the third of lead-backed rounds that fail to wire “could happen for many reasons that have nothing to do with the viability of the underlying company,” such as a lead losing its allocation. The prices tell a different story from the completion rates: founder-started rounds began at a $10 million median pre-money and led rounds at $20 million, but both converged on the same $25 million median once a round actually came together. AngelList is explicit about why: “the meaningful distinction is between rounds that attract sufficient investor support and those that do not” — the lead is a marker of that support, not a demonstrated cause of it. Two limits matter. The bias runs one way: founders who already have a committed lead rarely need a syndication platform, so the sample over-weights unled rounds and the weaker end of them. And these are 2025 vehicles — the cohort predates this week's move entirely, and no measurement of the current one exists yet.
Nscale filed a Form S-1 with the SEC on September 18 to list on the NYSE under “NSCL”. For the six months to June 30, 2026 it reported a net loss of $1.02 billion on revenue of $140.6 million. The prior-year period showed a $368.9 million loss on $10.4 million of revenue. Revenue grew 1,252%, and the unit moved the other way from the headline: loss per dollar of revenue fell from about $35 to about $7. The filing does not break out how much of the $1.02 billion is cash operating loss versus non-cash remeasurement. It reports $56.4 billion of remaining performance obligations — an accounting disclosure with a definition — alongside a company-defined contracted order book of roughly $103 billion, both as of August 31 against financials to June 30. It contains no share count, no price range, no deal size and no timetable — the raise and valuation figures circulating this week are press reports, not filed numbers.
Goldman Sachs published Competition for Capital on September 17, by chief global equity strategist Peter Oppenheimer. The argument is that AI infrastructure spending and government borrowing are drawing on the same pool of savings, pushing the global cost of capital up. The measured figures behind it: capital expenditure at AA-rated issuers rose 65% year-over-year in the second quarter, the tenth consecutive quarter above 35%; US convertible issuance reached $135 billion year-to-date with AI borrowers at 44% of it. Goldman also raised its full-year US investment-grade gross issuance forecast by $200 billion to $2.3 trillion, with AI issuers around a quarter of supply — that last one a forecast, not a measurement. Oppenheimer does not separate how much of the move belongs to AI capex and how much to government borrowing. He presents them as jointly competing and isolates neither.
Crusoe announced on September 17 that it had raised $3.9 billion in a Series F at a $30.9 billion post-money valuation, co-led by Atreides Management, Mubadala Capital and Valor Equity Partners, with Founders Fund, GIC, NVIDIA, the Qatar Investment Authority, Radical Ventures and TPG participating. The company describes it as an initial closing, not a final one. Bloomberg had reported roughly $3 billion at about $30 billion on September 3, two weeks earlier — the company's own figure arrived later and larger. Crusoe states alongside it more than $140 billion of total contracted value, more than 6 GW of gross contracted capacity with 1 GW operational, and more than $100 million of managed-inference ARR; those are self-reported and unaudited.
Hang Ten Systems, founded by former Infosys chief executive Vishal Sikka, announced a $53 million seed extension on September 16, roughly five weeks after closing a $32 million seed — bringing the seed to $85 million in total. The extension was led by Xora, Temasek's early-stage platform, with Mayfield, which led the first seed, and Aramco Ventures participating alongside angels including Intel chief executive Lip-Bu Tan, Micron chief executive Sanjay Mehrotra and Jerry Yang. Sikka declined to state a valuation, describing only “a bump up”. The company sells AI advisory and software-modernization work to enterprises above $10 billion in revenue.
Bloomberg reported on September 15 that investors had approached OpenAI about a round valuing it above $1.2 trillion. CNBC reported no formal discussions were under way and OpenAI declined to comment — this is an approach in early talks, and no money has moved at that mark. The standing mark is $852 billion, set by the roughly $122 billion primary round in March 2026 and carried into an approximately $7 billion employee tender that cleared in August at that mark — a tender price is generally set off the last primary rather than independently discovered, which makes it a held mark, not a new one. Fortune reports that people at OpenAI put its worth internally at at least $1.5 trillion. And on September 18 the Financial Times reported a leaked internal July presentation projecting $278 billion of negative free cash flow across 2026–2030, roughly $856 billion of compute spend through 2030, and the March money exhausted in 2028.
SoftBank launched $10 billion of US dollar senior unsecured notes and €1 billion of euro notes on September 20–21, to fund the $10 billion third tranche of its OpenAI follow-on investment. The bonds were scheduled to price on September 24 and settle on September 29, with the tranche itself expected to close around October 1. Reporting describes it as the largest non-financial corporate bond deal on record in Japan and the wider Asia-Pacific region. It is distinct from the $11.87 billion two-year bank loan from roughly twenty lenders reported on September 14, which sits outside this week's window — and which is bank debt rather than public issuance, so it prices privately and tells you nothing about where the market clears.
Torsten Slok, chief economist at Apollo, published Hyperscaler CDS Widening Is Not a Dealer Inventory Story in the firm's Daily Spark on September 16. His argument, verbatim: “What the market is repricing is hyperscaler credit fundamentals, namely a debt-financed AI capex cycle with rising leverage, negative free cash flow and uncertain payback on depreciating assets.” Apollo's note says bank spreads “have sat flat near 40 bps” and presents the comparison as a chart; the roughly 60 basis point gap that circulated this week is press reporting of that chart rather than a figure Apollo states. Worth knowing what is being compared: Apollo's hyperscaler basket is Amazon, Google, Microsoft and Oracle — the most levered and most AI-debt-exposed of the group — against JPMorgan, Citi, Bank of America, Wells Fargo and US Bancorp.
The Financial Times reported on September 17 that Emulate was in advanced talks to raise up to $700 million at a $3.7 billion valuation, with Index Ventures and Lightspeed expected to co-lead and Creandum participating. The round is not closed. City AM reported explicitly that terms were not finalised and could change, and there is no company announcement. Bloomberg reported the company seeking approximately $3 billion pre-money, which is not a competing figure: $3 billion pre plus a $700 million raise is $3.7 billion post. The two outlets agree to the dollar, standing on opposite sides of the round. Emulate was founded in August 2026 by three former DeepMind researchers behind Genie, and builds world models predicting real-world physics. It has no product, no website and no disclosed headcount.
Bain Capital Ventures announced on September 16 that it had closed Fund XI at $1.6 billion, exceeding an undisclosed target. Fund X closed at $1.4 billion in 2023, making this a 14% step up. The stated focus spans AI infrastructure, applications, physical AI, science, security and services. It is worth separating this from the week's other closes, which are frequently counted alongside it: Radical Ventures announced a first close above $1 billion on a fund explicitly aimed at late-stage AI scale-ups; Morgan Stanley closed a growth equity fund; and Portage closed Fund IV at $600 million, below its predecessor's $655 million. Of the four, only Bain XI is early-stage capital.
Morgan Stanley Investment Management announced the final close of the North Haven Growth & Innovation Fund at $1.3 billion on September 16, describing it as oversubscribed. It is a genuine debut vehicle, already invested in Databricks, Anduril and Ramp. Morgan Stanley, its affiliates and its employees supplied 24% of the capital — more than $300 million. The roughly $280 billion figure that appears in coverage is the firm's alternatives platform, not this fund.
PitchBook reported on September 15, with data as of August 18, that just six climate-specialist venture funds had closed year-to-date, and projected the full year to finish below $1 billion — the first time since 2015. Five years ago the figure was above $10 billion. Two vehicles, Kompas II in Denmark and 2150's second urban-tech sustainability fund in the UK, account for roughly 79% of what has been raised. This is a projection of a year not yet finished, not a closed-year figure.
The obvious read of this week is that money got more expensive. The Fed raised rates for the first time since 2023, and the ten-year printed 5.01% on the sixteenth — a level last reached before the financial crisis. The more useful question is whose money.
Goldman’s Peter Oppenheimer put the mechanism plainly: AI capital expenditure and government borrowing are drawing on the same pool. What he does not do, and what nobody should do on his behalf, is say how much of the move belongs to AI. In issue #026 we found that venture was no longer the marginal dollar in the buildout. This week the markets that are the marginal dollar went out to borrow at the new price — SoftBank for $10 billion of notes, with the coupon due to print on the twenty-fourth.
Whether any of it reached a seed round is a question nobody can currently answer. The firms that see the term sheets report quarterly, and the two most recent — Cooley and Wilson Sonsini — disagree about which direction seed moved last quarter. Meanwhile Hang Ten Systems added $53 million five weeks after closing a $32 million seed. And the Fed Chair called credit flows “robust, particularly for businesses” — a direct challenge to this issue’s frame, from the person with the best data. If he is right, the repricing is narrower than a 5.01% print suggests, confined to the borrowers funding AI capex, and not a cost-of-capital event at all. What a founder can actually observe is not the price.
AngelList published the only measurement anyone made of founders this week, and it is already historical. Of seed rounds started by founders on its platform in 2025, twenty-six percent ever wired; rounds a third-party deal lead brought to the platform wired sixty-five percent. Both converged on the same $25 million median pre-money, so price is not what separated them. AngelList says the meaningful distinction is between rounds that attract enough investor support and those that do not. The lead marks that support. It does not manufacture it.
The wager: that cohort is from 2025 and predates everything above, and no measurement of the current one exists — which is the argument, not a caveat. When AngelList or Carta next reports a started-to-wired rate covering rounds begun after September 2026, if the wire rate falls by more than a fifth while the median seed pre-money holds within ten percent, then the repricing reached seed through completion rather than price, and I read the direction right and the timing wrong. If both fall together, the separation this issue draws between the two markets was wrong.
So do not plan against a price you cannot see. If you cannot attract a lead, stop running a process that only pays out when everyone says yes: take each commitment as final on its own, with no dollar waiting on another. And the cheapest capital in a market this loud is still a customer who pays in advance.