Goldman Sachs expects hyperscalers to spend $1.14 trillion on AI data centers in 2027, and roughly a third of that gets borrowed rather than earned. Some of that borrowing has already moved onto crypto rails (save this). Those same firms spent $405 billion in 2025, and they're tracking about $750 billion this year, so the bill nearly triples in two years. Growth is expected to cool off, from roughly 85% this year to about 52% next year, but that's still the fastest spending ramp on record. And the count only covers the hyperscalers. Goldman tallied another $412 billion of AI-linked borrowing this year from chipmakers and data center builders. The market is laser focused on the assets/sectors where that $1T+ is headed, which is Nvidia, memory and power. Very few people, by comparison, are watching the stack of lenders who front that cash. The first in the stack is the investment-grade bond market. Hyperscalers sold $108 billion of those bonds in 2025, about 26% of their capex. They sold $194 billion in the first half of this year alone. Goldman is expecting $400 billion against 2027 capex. Alphabet posted its first negative free cash flow quarter since 2004 in July. Meta handed BlackRock an 80% stake in its El Paso site, with $12.5 billion of debt behind it. Second in the stack is private credit, and it exists because those bonds only serve borrowers rated AAA or AA. Goldman says project finance and asset-backed lending are the channels soaking up the rest. The same five firms have already disclosed about $1.2 trillion of lease commitments. Third in the stack is where you come in, because smaller GPU operators can't reach either of the first two. USD(dot)AI lends against Nvidia hardware tokenized as collateral. By its own count it originated its first $100 million of GPU-backed loans in Q1, and the yield comes from borrower repayments rather than token emissions. Figure runs that same plumbing at real size. It reported $4.3 billion of consumer loan volume in Q2, up 132% on the year, with 65% of it now crossing its tokenized marketplace. $1.14 trillion has to get borrowed from somewhere, and it splits across bonds, private credit and onchain rails. We're tracking all three, and the onchain one is barely off the ground. Don't miss our next entry. Follow us for more @MilkRoadDaily, and track each one of our investments in real time with Milk Road PRO (prices go from $25/m to $39/m next Weds): link.milkroad.com/zwu5a9

Goldman Sachs expects hyperscalers to spend $1.14 trillion on AI data centers in 2027, and roughly a third of that gets borrowed rather than earned. Some of that borrowing has already moved onto crypto rails (save this). Those same firms spent $405 billion in 2025, and they're tracking about $750 billion this year, so the bill nearly triples in two years. Growth is expected to cool off, from roughly 85% this year to about 52% next year, but that's still the fastest spending ramp on record. And the count only covers the hyperscalers. Goldman tallied another $412 billion of AI-linked borrowing this year from chipmakers and data center builders. The market is laser focused on the assets/sectors where that $1T+ is headed, which is Nvidia, memory and power. Very few people, by comparison, are watching the stack of lenders who front that cash. The first in the stack is the investment-grade bond market. Hyperscalers sold $108 billion of those bonds in 2025, about 26% of their capex. They sold $194 billion in the first half of this year alone. Goldman is expecting $400 billion against 2027 capex. Alphabet posted its first negative free cash flow quarter since 2004 in July. Meta handed BlackRock an 80% stake in its El Paso site, with $12.5 billion of debt behind it. Second in the stack is private credit, and it exists because those bonds only serve borrowers rated AAA or AA. Goldman says project finance and asset-backed lending are the channels soaking up the rest. The same five firms have already disclosed about $1.2 trillion of lease commitments. Third in the stack is where you come in, because smaller GPU operators can't reach either of the first two. USD(dot)AI lends against Nvidia hardware tokenized as collateral. By its own count it originated its first $100 million of GPU-backed loans in Q1, and the yield comes from borrower repayments rather than token emissions. Figure runs that same plumbing at real size. It reported $4.3 billion of consumer loan volume in Q2, up 132% on the year, with 65% of it now crossing its tokenized marketplace. $1.14 trillion has to get borrowed from somewhere, and it splits across bonds, private credit and onchain rails. We're tracking all three, and the onchain one is barely off the ground. Don't miss our next entry. Follow us for more @MilkRoadDaily, and track each one of our investments in real time with Milk Road PRO (prices go from $25/m to $39/m next Weds): link.milkroad.com/zwu5a9

Source:https://x.com/USDai_Official/status/2092002800905437375