Barclays: AI capital expenditure forecasts for 2027 and 2028 may be underestimated by $225 billion

Investing.com - According to Barclays’ analysis, the spending by mega-scale tech companies on AI infrastructure may be severely underestimated, with current market forecasts far below the actual investment needed in the coming years.

Get AI breaking news with InvestingPro - Enjoy up to 50% discount now

Barclays analyzed financial disclosures related to OpenAI and Anthropic to estimate the implied chip expenditures of mega-scale cloud service providers, rather than relying on models based on token usage or query demand.

The research shows that the investment cycle is much larger than market consensus expectations.

Tom O’Malley and his team of analysts wrote: “Our analysis indicates that the capital expenditure upcycle will last at least until 2028 and may be several orders of magnitude higher than market consensus (exceeding $225 billion in 2027 and 2028).”

This means that capital spending by mega-scale cloud providers in 2027 and 2028 could be more than $225 billion higher than current Wall Street forecasts. The analysts emphasized that this gap would be a “substantial positive” for AI semiconductor companies. They also noted that Nvidia (NASDAQ: NVDA) currently appears to be priced based on the expectation that capital expenditure by mega-scale cloud providers will peak in 2027, but their framework indicates spending will continue to grow through 2028.

The bank’s framework also suggests that the peak of AI infrastructure spending may occur later than many investors currently expect.

The analysts wrote: “Under this framework, the peak of mega-scale cloud provider capital expenditure will occur in 2028, based on the expected timing of recursive self-improvement in AI labs, which will increase spending efficiency and reduce training costs.”

Barclays assumes that training operational costs will peak in 2029, meaning capital expenditure will peak the year before, as infrastructure must be deployed before workloads increase.

The bank also noted that its base model assumptions may still be conservative. The analysis assumes that starting in 2027, existing training chips will be sufficient for most inference workloads, meaning additional spending on inference-specific chips has not yet been reflected in the forecasts.

In addition to OpenAI and Anthropic, Barclays expects other AI developers and platforms to gradually increase their share of global computing demand, potentially further expanding future infrastructure spending.

The analysts stated that OpenAI and Anthropic may currently account for about two-thirds of the computing demand, but other AI labs like Gemini, Grok, and sovereign AI projects could increase their share over time.

This article was translated with the assistance of artificial intelligence. For more information, please see our Terms of Use.