Barclays expects AI capital expenditures to peak at about $1 trillion in 2028

Overview

This is a Barclays report on an AI demand and supply modeling framework for the U.S. internet and semiconductor sectors. The report’s core question is: against the backdrop of accelerating AI capabilities, the spread of Agentic AI, and potential recursive self-improvement, how much training and inference compute will AI labs need in the future, and how will these needs translate into hyperscaler capital expenditures? The report concludes that AI Capex may reach a peak of about $1 trillion in 2028, after which growth may flatten or decline moderately, while overall inference usage continues to grow.

Core views

The report argues that the AI industry is moving from simple Q&A and assisted workflows toward Agentic AI and automated ecosystems, and that the pace of model releases and capability improvements may accelerate further after 2027. Barclays believes the market still underestimates this round of capital expenditure upside, especially since hyperscaler Capex in 2028 could exceed consensus by more than $300 billion. The logic is that the training and research compute costs of frontier AI labs will peak in 2029, while cloud providers need to build the corresponding infrastructure about one year in advance, placing the Capex peak in 2028.

Analysis framework

Rather than starting from supply-chain capacity or traditional annual Capex forecasts, the report uses forecasts for revenue, training compute costs, inference compute costs, and users/query volumes of AI labs such as OpenAI and Anthropic as demand proxies, and then converts them into B200-equivalent GPUs, GW-scale compute capacity, and hyperscaler Capex. The report also assumes that OpenAI and Anthropic currently account for about two-thirds of industry compute demand, while other labs such as Google, Meta, and xAI account for the remaining one-third and will increase their share over time.

Methodology notes

• demand infers supplyAI lab compute demand-driven Capex framework

  Use forecasts of AI lab training and inference compute costs to infer new cloud provider compute buildout.

  The report argues that AI labs are the largest consumers of tokens and compute, and that their internal forecasts better reflect real demand over the next few years than simply observing the supply chain or annual Capex consensus.

• timing mismatchT-minus-1-year Capex logic

  The peak in training compute consumption usually requires capital expenditure buildout in the prior year.

  If training and research compute demand peaks in 2029, cloud providers need to complete data center and next-generation compute buildout around 2028, so the Capex peak will appear earlier.

• compute standardizationNvidia Blackwell-equivalent estimation

  Use B200-equivalent compute to unify the compute benchmark across different GPU, TPU, and ASIC generations.

  The report acknowledges that FLOPS/GPU, GPU/rack, and rack/GW will change in the future, but argues that an equivalent-compute benchmark can more stably estimate the capacity implied by AI lab demand.

• compute waterfalltraining-to-inference compute conversion assumption

  Older-generation compute previously used for frontier training can later be shifted to inference.

  The report assumes that new-generation compute is mainly used for frontier training, and that as newer generations come online, older training compute can be redirected to inference, so incremental inference-dedicated Capex demand may be lower after 2027.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• AMZN AWS

  Potential compute supplier to hyperscalers and AI labs such as OpenAI

  Strengths

  The report estimates AWS AI capacity rising from 8GW in 2025 to 22GW in 2028, and mentions an increased agreement size between Amazon and OpenAI.

  Weaknesses

  High Capex may compress free cash flow, and AI compute returns depend on long-term leases and utilization rates.

  Comparison

  In Barclays' estimates, AWS is one of the major capacity contributors.

  Risks

  Contract execution, data center delivery, power constraints, and lower-than-expected Capex returns.

• MSFT Azure

  One of OpenAI's key cloud compute partners

  Strengths

  Closely tied to the OpenAI ecosystem and may benefit from growth in training and inference compute demand.

  Weaknesses

  The report uses both consensus expectations and Barclays forecasts, showing that 2028 industry Capex demand may exceed market expectations.

  Comparison

  Estimated 2028 AI capacity is 14GW, below AWS and GOOGL but still a core supplier.

  Risks

  Capital expenditure intensity, changes in contract structure, and diversion to competing cloud providers.

• GOOGL

  Both a hyperscaler and an AI lab participant

  Strengths

  The report estimates GOOGL's AI capacity will reach 25GW in 2028, the highest among the listed companies.

  Weaknesses

  It needs to continue investing in proprietary models, TPUs, and cloud infrastructure.

  Comparison

  Its capacity is higher than AWS, MSFT, META, and ORCL in the capacity table.

  Risks

  Pace of AI commercialization, Capex efficiency, and competition with the OpenAI/Anthropic ecosystem.

• META

  Another major AI lab and compute demand participant

  Strengths

  Over the long term, it may increase the industry compute share of labs outside OpenAI and Anthropic.

  Weaknesses

  Its near-term commercialization path is relatively less direct than APIs or enterprise subscriptions.

  Comparison

  Estimated 2028 AI capacity is 15GW.

  Risks

  Returns on open-source models, fit between ad business and AI investment, and cash flow pressure.

• ORCL

  Cloud infrastructure and AI compute leasing supplier

  Strengths

  May benefit from long-term compute contracts with AI labs and demand for new cloud infrastructure.

  Weaknesses

  Its capacity base is relatively small, estimated at 6GW in 2028.

  Comparison

  It has the lowest capacity among the listed hyperscalers, but its incremental elasticity remains in focus.

  Risks

  Delivery capability, financing structure, and customer concentration.

• Semiconductors and semiconductor capital equipment

  Upstream beneficiary chain of rising AI Capex

  Strengths

  Demand for B200-equivalent GPUs, accelerators, servers, and equipment increases with GW capacity expansion.

  Weaknesses

  Barclays rates this group NEUTRAL, implying that valuation or cycle risks may offset part of the demand tailwind.

  Comparison

  Compared with the POSITIVE rating on U.S. Internet, the rating on semiconductors and equipment is more cautious.

  Risks

  Supply bottlenecks, technology generation shifts, customer Capex cuts, and cyclical pullbacks.

Key data

• Peak year for AI Capex2028The report estimates that hyperscaler AI-related capital expenditures will peak in 2028.
• 2028 hyperscaler Capex约$1.0TBarclays estimates about $1.036T, or about 89% of hyperscalers' operating cash flow.
• Gap versus consensus约$300B+The report argues that market consensus underestimates the Capex slope in 2028.
• 2029 industry incremental AI capacity demand约23GWAbout 15.5GW from OpenAI and Anthropic, and about 7.8GW from other AI labs.
• 2029 OpenAI and Anthropic training compute cost约$155BThis figure is the core constraint used to derive the prior-year Capex peak.
• 2030 OpenAI and Anthropic total compute cost约$290BIncluding both training and inference compute costs, sharply up from about $7B in 2024.
• 2030 AI industry WAU assumption约4.3BThe report says this scale is close to the current global internet user base outside China.
• Agentic query compute intensity约为单次查询的7倍The report argues that inference-heavy and Agentic workflows will significantly increase compute consumption per query.

Impact & implications

If the report’s framework holds, the intensity and duration of the AI infrastructure investment cycle will exceed current market expectations. Beneficiaries include cloud infrastructure, AI servers, GPUs/accelerators, data centers, power, and related semiconductor equipment; however, for internet companies and cloud providers, capital expenditures as a share of operating cash flow could rise significantly, and investors need to watch cash flow pressure, payback periods, and the quality of compute leasing contracts.

Risks

• Forecasts for AI lab revenue, ARR, training compute, and inference compute are mainly based on media reports and company disclosures, and could be materially revised in the future.
• There is inconsistency in the compute forecasts for OpenAI and Anthropic: the report says OpenAI's 2029 training and research compute costs exceed Anthropic's by more than three times, which may need to converge over time.
• If recursive self-improvement or Agentic AI diffusion falls short of expectations, the peaks in training compute and Capex could be delayed or revised down.
• If inference demand grows faster than training compute can be repurposed, incremental inference-dedicated Capex could exceed the report's assumptions.
• Power, data centers, chip supply, financing structures, and long-term leasing contracts could all become constraints.
• The report discloses that Barclays and its affiliates may have business relationships with companies under research coverage, and investors should be aware of potential conflicts of interest.

What to watch

• Whether OpenAI and Anthropic ARR, revenue forecasts, cash burn, and compute costs continue to be revised upward.
• Whether the model release cycle compresses materially around 2027 and whether Agentic AI and automation products truly diffuse beyond software development into other industries.
• Changes in hyperscaler 2026-2028 Capex guidance, share of operating cash flow, and AI order backlog.
• Whether training compute costs peak around 2029, and whether inference compute can be adequately absorbed by prior-generation training compute.
• Compute efficiency, rack density, data center power, and construction timelines of B200 and subsequent GPU/TPU/ASIC generations.
• Although sovereign AI and China AI cloud are not in the report's core scope, they may increase global demand for accelerators and infrastructure.