00Opening read
The AI buildout is becoming more than a supply chain. It is becoming a financing network.
Broadcom has agreed to lend Anthropic as much as $42 billion to help finance compute tied to Broadcom-designed chips. Micron's customers have committed $32 billion under long-term supply agreements, mostly through cash deposits. Nvidia wants banks to treat AI accelerators as durable collateral, although lenders are asking for stronger guarantees.
Power is moving the same way. Japan's JERA is pairing a future 400 MW data center with land beside an existing power station and a long-duration electricity commitment. Amazon has signed a 20-year agreement supporting new investment at a Maryland nuclear plant.
The physical AI economy is starting to pull future demand into today's balance sheets.
That can make scarce capacity easier to build. It also means the stack is becoming more financially connected to itself.
01What moved
Broadcom finances the customer buying its compute
Broadcom's relationship with Anthropic now spans chips, compute, equipment leases and financing.
According to Anthropic's IPO filing, Broadcom has agreed to lend the AI company up to $42 billion. The facility could finance roughly a third of Anthropic's five-year TPU lease commitment.
This is a useful picture of how the AI buildout is changing.
Suppliers are no longer waiting for customers to arrive with financing already arranged. They can help create the financing that makes the demand possible.
That accelerates deployment, but it also concentrates risk.
Micron's customers put cash behind the shortage
Micron says customer commitments under long-term supply agreements have risen to $32 billion. Most are backed by cash deposits.
Its future contracted revenue under those agreements has climbed to roughly $150 billion.
The message is stronger than another bullish memory forecast.
Customers are reserving production before it exists.
That gives Micron more confidence to build capacity, while customers get a better chance of securing HBM and other memory in a tight market.
Memory is starting to behave like industrial offtake.
Wall Street asks what an old GPU is worth
Nvidia is trying to expand financing around AI infrastructure using chips as collateral.
Lenders are not fully convinced.
Reuters reports that some banks and asset managers want stronger guarantees because they are uncertain how long today's accelerators will produce enough revenue to support long-term debt.
That question sits underneath a large part of the AI capital cycle.
A GPU can be extraordinarily productive today and still depreciate quickly if the next architecture changes the economics.
Compute has become an asset class. Its residual value is still being negotiated.
Japan builds the data center around the power station
JERA, Dell and RHAELM plan to begin with a $15 billion AI campus in Chiba.
The project is designed around an existing energy node. JERA will provide land next to its thermal power station and supply 400 MW for as long as 25 years.
That reverses the normal development sequence.
Instead of choosing a data-center site and then searching for electricity, the consortium starts with power and builds compute around it.
In a world of grid queues, that may become a repeatable model.
AI demand reaches the nuclear balance sheet
Amazon and Constellation signed a 20-year power agreement tied to the Calvert Cliffs nuclear plant in Maryland.
The arrangement is expected to support more than $3 billion in infrastructure investment and about 190 MW of new generating capacity.
This is another form of infrastructure finance.
A hyperscaler's long-duration electricity demand can give a power producer the revenue confidence to extend and expand an asset.
The AI load is becoming the counterparty.
02The throughline
The AI infrastructure boom is becoming circular by design.
Customers reserve memory with deposits. Suppliers finance compute buyers. Power companies attach long contracts to new capacity. Hardware vendors ask lenders to treat accelerators as infrastructure.
Each structure solves a real problem: too much capital has to be committed years before the capacity produces revenue.
But the contracts also connect risks that once sat in different places.
If utilization disappoints, the effect can move from the model company to the chip supplier, lender, data center and power asset.
The next RCW question is not simply whether a project is financed.
It is what the financing depends on.
03What to watch
- October 6 as Firmus begins the planned bookbuild for a roughly $5 billion Australian AI-infrastructure IPO.
- Mid-2027 for the first wafers from Micron's new capacity and whether customer commitments continue outrunning supply.
- 2028 for the first phase of JERA's power-station-adjacent Chiba AI campus.
- 2030 to 2032 for the new nuclear capacity supported by Amazon's Calvert Cliffs agreement.
04Source trail
Reuters, Broadcom and Anthropic financing
https://www.reuters.com/business/broadcom-lend-anthropic-up-42-billion-lease-its-chips-filing-says-2026-10-01/
Reuters, Nvidia chip-backed financing
https://www.reuters.com/legal/transactional/nvidias-bet-that-its-chips-can-finance-ai-boom-gets-wall-street-reality-check-2026-10-01/
Reuters, Micron memory commitments
https://www.reuters.com/business/micron-forecasts-quarterly-revenue-above-estimates-2026-09-30/
Reuters, South Korean exports
https://www.reuters.com/world/asia-pacific/south-korea-september-exports-rise-835-yy-monthly-record-2026-10-01/
Reuters, JERA / Dell / RHAELM
https://www.reuters.com/business/energy/jera-teams-up-with-dell-rhaelm-ai-infrastructure-development-japan-2026-10-01/
Reuters, Amazon / Constellation
https://www.reuters.com/business/energy/constellation-amazon-sign-20-year-power-deal-support-maryland-nuclear-plant-2026-09-30/
Reuters, Tencent / Oracle, citing Financial Times reporting
https://www.reuters.com/world/china/chinas-tencent-leases-100000-chips-oracle-accelerate-ai-push-ft-reports-2026-10-01/
Reuters, Synopsys / OpenAI
https://www.reuters.com/business/synopsys-openai-strike-deal-develop-ai-model-chip-design-work-2026-09-30/