Mumbai: The AI boom has reached a curious stage. The question is no longer simply who can build the most powerful model. It is becoming who can afford the infrastructure required to keep that model running.
That distinction matters. Amazon is reportedly exploring a structure to move about $8 billion of Nvidia AI chips into a special-purpose vehicle and lease them back. Broadcom has agreed to provide up to $42 billion in financing tied to Anthropic’s infrastructure plans, while Anthropic has disclosed hundreds of billions of dollars in long-term computing and infrastructure commitments. OpenAI, meanwhile, closed a $122 billion funding round in March and says durable access to compute is a strategic advantage.
AI, it seems, has discovered the ancient corporate truth: ambition is wonderful until someone sends the invoice.
The Capital Machine Behind AI
The scale is extraordinary because frontier AI requires an unusual combination of assets. Chips have to be bought or leased, data centres constructed, electricity secured and networking infrastructure expanded before the resulting AI services can generate corresponding revenue.
OpenAI’s March funding round valued the company at $852 billion post-money, with $122 billion in committed capital. The company says its infrastructure strategy now spans cloud providers, Nvidia and AMD chips, custom silicon, data centres and energy-related partnerships.
OpenAI has also said its Stargate programme had already surpassed its original 10 GW U.S. infrastructure target for 2029, with more than 3 GW added in a 90-day period.
This is no longer merely venture funding.
It is becoming an industrial financing ecosystem.
Amazon Is Testing A Different Way To Pay
The reported Amazon transaction offers an intriguing glimpse into that evolution.
Amazon is discussing transferring approximately $8 billion of Nvidia Grace Blackwell chips into an SPV funded by outside investors. Amazon would then lease the chips back rather than retaining them directly on its balance sheet. The processors are reportedly already deployed across data centres in the U.S.
The attraction is obvious: enormous AI hardware expenditure can be converted into a financing arrangement rather than being carried entirely as owned equipment.
But leasing does not make the economic cost disappear. It changes its form.
The structure also raises questions around depreciation, financing costs, asset ownership and how much AI infrastructure expenditure ultimately sits outside traditional capital-expenditure headlines.
Anthropic Is Turning Compute Into A Long-Term Commitment
Anthropic’s numbers are even more striking.
According to its latest filing, the company has disclosed around $518 billion in future infrastructure commitments, with roughly 80% described as binding regardless of actual usage. Major commitments include approximately $111.1 billion with Google, $110 billion with Amazon and $161.2 billion in non-cancelable Broadcom lease obligations, alongside other arrangements.
Broadcom is also expected to play a particularly unusual role: supplier, infrastructure partner and financier.
Broadcom and Google previously announced that Anthropic would access approximately 3.5 GW of next-generation TPU capacity through Broadcom from 2027, as part of a much larger compute arrangement.
The advantage is capacity certainty. The drawback is equally clear: large fixed commitments become much less comfortable if AI demand, pricing or model economics change.
Why Investors Aren’t Running Away
There is another side to this story.
Temasek has explicitly identified AI, core-plus infrastructure and private credit as areas for increased investment. Its AI-related exposure stood at about 6% of portfolio value in March 2026, with a target of up to 15% by 2031. It plans to invest across energy and data centres, semiconductors, cloud providers, foundation models and AI applications.
Temasek’s reasoning is not simply that AI companies will keep spending. Its investment framework places considerable emphasis on AI adoption across the wider economy, rather than only betting on companies developing frontier models.
That is significant.
If AI becomes genuinely useful across finance, healthcare, manufacturing, software, logistics and other sectors, infrastructure spending can eventually be supported by a much broader economic base.
The Bill Has Benefits — And Risks
The positive case is fairly tangible:
- Massive infrastructure investment expands computing capacity and supports AI adoption.
- Chipmakers, data-centre operators, energy companies and networking suppliers gain new demand.
- Financing structures can allow companies to deploy infrastructure without paying the entire cost upfront.
- Competition among OpenAI, Anthropic and other AI developers encourages investment in alternative chips and computing architectures.
But the negative side deserves equal attention.
The largest risk is mismatch: infrastructure commitments can run for years, while AI technology changes in months. A facility built around one generation of processors can face depreciation or technological obsolescence before the original economics have fully played out.
There is also concentration risk. Anthropic itself has acknowledged dependence on major technology companies that can simultaneously be investors, infrastructure providers and competitors.
And when suppliers start financing customers who then use those suppliers’ hardware, the ecosystem becomes wonderfully efficient—and rather complicated.
The Real AI Bubble Question Is About Cash Flow
Calling every large AI investment a bubble would be premature. There is genuine demand: OpenAI says enterprise already represents more than 40% of its revenue, while Anthropic has been securing additional capacity across AWS, Google, Microsoft, Nvidia, SpaceX and other infrastructure partners.
But capital intensity changes the question.
The next phase of AI will be judged not merely by how much money can be raised, but by whether revenue, productivity and utilisation can eventually justify the extraordinary infrastructure being financed today.
That is the uncomfortable arithmetic hiding underneath the AI excitement.
The models may become smarter.
The financial models now have to become smarter too.
