AI’s Next Bottleneck Is the Price of Money
Broadcom and Intel say capital is now the constraint, banks are backing buildings but not GPUs, Nebius anchors a Rotterdam campus, and an emergency order lets PJM call on data center backup power this week.
Today’s Signals
- Broadcom CEO Hock Tan said capital “might be the biggest constraint” on AI growth, and chip stocks fell in the US, Korea and Europe.
- JPMorgan and Goldman Sachs have reportedly underwritten only $1.3 billion of neocloud Volta’s $5 billion loan sale, leaving a $3.7 billion GPU loan priced near 11% to investors.
- Nebius has signed as the first customer of Dutch developer VOLT’s planned 7 billion euro Rotterdam AI campus, leasing up to 200 megawatts, according to Dutch press reports.
- NTT Data is reportedly in final talks to build Anthropic’s first data center in Japan, an Osaka site of up to 500 megawatts that could cost up to 1 trillion yen.
- A US emergency order lets PJM direct data centers to run backup generation as a last resort through October 16 as grid reserves tighten.
The Big Story
For two years the AI build-out has been limited by chips and power. This week the people selling the chips are pointing at a different limit: money. Broadcom CEO Hock Tan said Monday that capital “might be the biggest constraint” on the industry, even with demand for compute still ahead of supply. Intel’s Lip-Bu Tan agreed and added that investors eventually want to see a return. Markets heard it: the Philadelphia Semiconductor Index fell 2.7 percent on Monday, and SK hynix dropped more than 10 percent in Seoul on Tuesday. Lenders are already drawing lines. JPMorgan and Goldman Sachs have underwritten only $1.3 billion of neocloud Volta Infrastructure’s $5 billion loan sale, Bloomberg reports. That piece covers lease payments for a Norway data center; a $3.7 billion loan for about 36,000 Nvidia GPUs is left to investor appetite at a yield near 11 percent. The banks will stand behind the building, not the chips. Even the record $38 billion loan for Oracle-leased campuses is still placing its last slice after six months, with some lenders selling at a discount. The money isn’t drying up. It is getting priced: projects with a creditworthy tenant still get financed, and chip risk without one costs more.
AI and Data Centers
If capital is getting choosier, the chip industry’s answer is to make each dollar of capacity faster to deploy. At the Open Compute Project summit, Nvidia opened its Vera Rubin rack designs to the industry, supporting racks of up to 250 kilowatts. Shared designs let equipment makers build without waiting on each customer’s custom plans. Demand still looks deep. NTT Data is in final talks to build Anthropic’s first dedicated data center in Japan, Nikkei reports: an Osaka site of up to 500 megawatts costing up to 1 trillion yen, about $6.3 billion, with partial operations as early as 2028. Scaleway, meanwhile, plans a Copenhagen GPU zone with several thousand Nvidia B300 chips for customers that want compute kept under EU jurisdiction. This week’s earnings test whether that demand reaches the supply chain. ASML plans to lift EUV machine output from about 65 this year to roughly 85 in 2027, so any change to that target is a direct read on customer commitments.
Money
The deals that work have a name on the lease. Dutch developer VOLT, a separate company from Volta, has signed Nebius as the first customer for its planned 7 billion euro Rotterdam AI campus, De Telegraaf and Financieele Dagblad report. Nebius would lease capacity for 100,000 GPUs and up to 200 megawatts, the anchor contract VOLT needed before approaching lenders. About 4.5 billion euros of that budget is chips, and the lease is what makes that equipment bill bankable. Volta shows the flip side: banks backed the lease payments and stepped back from the GPUs, which can lose value before the debt is repaid. Expect more deals that split the two, with the chip slice priced well above the real estate.
Energy & Power
Money is one constraint; the grid on a bad day is another. A US Department of Energy emergency order lets PJM, the grid operator for 13 states and Washington, D.C., run specified plants beyond normal limits and, as a last resort before rolling outages, direct data centers to switch to backup generation. The order runs October 13 to 16, as PJM projects 61,591 megawatts of generation outages during autumn maintenance. Nothing yet confirms any data center has been asked to switch. The industry wants to look like a better neighbor. Google, Meta, Microsoft and Nvidia published an OCP specification saying AI sites should ride through grid disturbances and recover to near-full load in under two seconds. It is a design target, not a utility rule, but it gives operators something to write into connection agreements. Greece is setting its own terms: a proposed bill requires at least 60 percent renewable power and guarantees of 200,000 euros per megawatt, after applications jumped from 5 to 9 gigawatts in two months.
Government Power
Washington’s next move is about who pays. The House is expected to vote Wednesday on the bipartisan Ratepayer Protection Act, Reuters reports, which would make new data centers of at least 20 megawatts cover the grid costs they create and let states require them to bring their own supply. Amazon says it could unlock investment; one industry official called parts of it discriminatory. It lands days after a FERC-approved PJM registry of large loads took effect. Australia wants more data centers, on its terms. A bipartisan parliamentary AI committee urged investment under national standards on resources and new generation, compute reserved for national security and research, and independent experts inside frontier labs. Both capitals are saying the same thing: build here, but pay your way.
Startup and Small Business Signal
Tighter capital opens specific doors. Lenders splitting building risk from chip risk need better ways to value used GPUs, so residual-value analytics, GPU-backed lending platforms and secondary markets for older accelerators have buyers now. The two-second ride-through target and PJM’s order also reward controls and battery systems that prove how a site behaves in a disturbance. For smaller neoclouds, Rotterdam is the playbook: sign the tenant before you go to the bank.
AMP Bottom Line
The AMP2 read: the AI build-out is turning from a supply problem into a balance-sheet problem. The projects that get built next will show lenders a signed tenant, show grid operators a site that behaves under stress, and show investors a return. Expect the cost of money, not the supply of chips, to decide who keeps building, with this week’s ASML and TSMC guidance as the first test.