The Week's AI Money Went to Plumbing, Not Labs
No frontier lab raised. The largest transaction was a bond. What that tells you about where your inference bill is going.
No frontier lab raised. The largest transaction was a bond. What that tells you about where your inference bill is going. What happened Between 17 and 23 August, StartuphHub tallied $11B across 60 AI rounds — the aggregator's own count, methodology not disclosed. The prior week it counted $10B across 40 rounds. More rounds, barely more money: the average cheque shrank by more than a quarter. The money did not go to frontier labs. It went to the layer that runs models rather than the layer that makes them. - Nebius closed $5.75B in convertible senior notes on 24 August — debt, not equity — split into $3.45B of 0.50% notes due 2030 and $2.3B of 4.50% notes due 2034, for data centre construction and GPU procurement. Several roundups still carry $4.5B; that is the pre-option-exercise figure and it is wrong. - Domyn raised roughly $1.1B, about 90% of it debt, for a supercomputer in southern Italy built on nearly 6,000 Grace Blackwell chips (The Next Web, citing Sifted). - Castelion raised a $1B Series C — $800M equity plus a $250M revolving credit facility — at a $13B valuation, co-led by JPMorganChase's Strategic Investment Group, Andreessen Horowitz and Carlyle, against more than $500M in US military contracts over 18 months. - Etched raised $700M at a reported $21B valuation led by Jane Street, roughly a month after $300M at $10.3B — figures from StartuphHub's analysis, not a first-party announcement. The same analysis reports Jane Street validated the chips against its own workloads before investing; treat that as reported, not confirmed. The widely repeated figure that 87.5% of US venture dollars now go to AI comes from Fortune, citing PitchBook . ⚠️ We could not retrieve that primary source, and we do not know the denominator — total VC, total tech VC, or a stage band. We are not treating it as fact, and neither should your board deck. Two other rounds in the week's lists — Fractile at $600M and Temporal at $500M — we could not verify at all. Why it matters Start with the arithmetic nobody in the coverage did. Of the four rounds we could verify, roughly $7B was debt and roughly $1.6B was equity — about 81% debt . Strip Nebius's notes out and the week's equity total falls under $3B, which is ordinary. The headline is a bond offering wearing a funding round's clothes. That matters in two opposite directions at once, and both are actionable. For a CEO or CPO buying inference: debt only gets underwritten against revenue somebody believes is recurring. The lenders behind Nebius and Domyn are betting that GPU capacity has predictable cash flow — and every dollar of it converts into capacity that has to be filled. Capacity that must be filled gets discounted. If your 2027 plan assumes today's $/token, you are planning against the wrong number, and probably in the safe direction. For a founder raising: the cheques in this week's list went to silicon, clouds and missiles. Nothing went to the model layer, and nothing large went to applications. Your next