From Nvidia to Wall Street: who pays for the AI buildout?

Key takeaways

• AI’s next bottleneck may be capital as infrastructure spending outruns what technology companies want to fund alone.

• CoreWeave tests whether enormous contracted demand can translate into better cash economics as new capacity comes online.

• Follow the AI dollar: chipmakers, cloud operators and financiers can grow quickly, but their returns may differ sharply.

Artificial intelligence (AI) keeps moving its bottlenecks. First, advanced chips were scarce. Then data-centre space, electricity and networking joined the queue. Now money is becoming harder to ignore.

On 10 August 2026, Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create financing platforms designed to mobilise more than 500 billion USD of third-party capital for AI infrastructure. This is not Nvidia writing a 500 billion USD cheque. It is an attempt to make computing infrastructure easier for outside investors to finance.

CoreWeave reports second-quarter results on 11 August after the US market closes, making it a useful live test. The specialist AI cloud company rents powerful computing infrastructure. Demand is enormous. The harder question is what it costs to serve it.

Wall Street joins the server room

Nvidia’s move shows how the AI cycle is changing. Graphics processing units (GPUs) need buildings, power, cooling and networking before customers can use them.

The money therefore travels through a long chain: capital into data centres, then Nvidia chips and networking, then cloud capacity sold to AI developers and enterprises. Every layer can grow, but the economics differ.

Cheaper financing can help Nvidia customers build more infrastructure using its technology. Financiers can earn fees and interest. Cloud operators face a different burden: spend heavily today, then depend on utilisation and pricing to make those assets profitable tomorrow.

CoreWeave turns demand into a financing question

CoreWeave’s first quarter shows why revenue growth is an incomplete scorecard. Revenue more than doubled to 2.08 billion USD and contracted backlog reached 99.4 billion USD. Yet it spent 7.7 billion USD on property and equipment, carried roughly 24.9 billion USD of debt and recorded 536 million USD of net interest expense.

For 2026, CoreWeave has indicated capital spending of roughly 31 billion to 35 billion USD, more than double 2025 levels. That spending creates value only if customers use the capacity, contracts turn into revenue on schedule and financing does not absorb too much profit.

This is where return on invested capital (ROIC) helps. It asks how much operating profit a business produces from the money tied up in it. Booming revenue can still disappoint shareholders if every extra unit of growth requires even more capital.

Five words matter in tonight’s results: demand, capacity, capital, cash and concentration. Is backlog becoming revenue? Is new capacity filling quickly? Is financing becoming cheaper? Is cash generation improving? How much depends on a few large customers?

The same bill arrives in different envelopes

CoreWeave is not alone. Nebius, another AI cloud provider, raised 775 million USD in secured debt on 17 July, backed by GPU infrastructure and contracted cash flows. Super Micro Computer, which builds AI servers, announced up to 7 billion USD of equity and equity-linked financing in June to fund components for its order book.

The structures differ, but the lesson is similar: rapid demand can consume cash before it generates it.

Risks: when the maths stops cooperating

Three risks matter most. First, utilisation or pricing could weaken if capacity grows faster than demand. Watch backlog conversion and margins. Second, financing costs could stay high. Rising interest expense or repeated capital raises are warning signs. Third, hardware improves quickly. If older GPUs lose economic value faster than expected, replacement spending can erode returns.

Investor playbook

• Compare capital spending with revenue and margin progress, not revenue growth alone.
• Watch backlog conversion. Contracts matter most when they become utilised capacity and cash.
• Separate technology suppliers, infrastructure owners and financiers. Their economics can diverge sharply.
• Stress-test customer concentration and hardware cycles before extrapolating today’s demand.

The bill becomes part of the story

The AI race started as a scramble for chips. It became a race for data centres, power and networks. Nvidia’s new Wall Street partnerships show that capital is now becoming part of the infrastructure itself. CoreWeave, Nebius and Supermicro show why: extraordinary demand can require extraordinary funding long before the cash economics mature.

For investors, the useful question is therefore no longer simply who is building the most AI capacity. It is who can turn that capacity into durable cash returns without continually asking lenders or shareholders to refill the tank. In the next phase of AI, scale will matter. So will the return on every dollar used to build it.