Barclays Calls $225 Billion Capex Blind Spot, Says Nvidia Undervalued Through 2028

Barclays equity research published this week projects hyperscaler capital expenditure will exceed Street forecasts by at least $225 billion in 2027 and 2028 combined, a gap that positions Nvidia and semiconductor peers as materially undervalued despite recent rallies. The analysis, based on financial data from OpenAI and Anthropic, argues the AI investment cycle will peak later and run longer than current valuations assume, with direct implications for energy infrastructure planning and chip supply chain geopolitics.

Analyst Tom O’Malley wrote in a note to clients that consensus hyperscale capex is at least $225 billion “too low” in 2027 and 2028, warning the market significantly underestimates spending by Microsoft, Alphabet, and other tech giants. Barclays assumes training operating expenses peak in 2029, which implies Capital Expenditure peaks a year earlier as infrastructure must be deployed before training workloads ramp. The thesis directly contradicts assumptions embedded in current semiconductor valuations and suggests Nvidia trades at roughly 17.5x calendar year 2027 earnings estimates today, but at just 14.5x 2028 numbers even assuming conservative earnings growth of 44% and 11% respectively—a penalty the bank considers “simply too great”.

AI Lab Financials Point to Sustained Demand

The Barclays model derives its forecasts from revenue and operating expense trajectories at frontier AI labs rather than top-down GDP analogies. Based on analysis of financials from AI leaders OpenAI and Anthropic published in The Information, Barclays analysts believe the market significantly underestimates how much tech Hyperscalers will need to spend in the next few years. OpenAI ended 2025 with approximately $20 billion in annual recurring revenue, a threefold increase from the prior year, while Anthropic reached a $9 billion run rate with 9x year-over-year growth—though both figures represent roughly 3% of projected 2026 hyperscaler capex.

Hyperscaler capex for Amazon, Alphabet, Microsoft, Meta, and Oracle is now widely forecast to exceed $600 billion in 2026, a 36% increase over 2025, with roughly 75% or $450 billion directly tied to AI Infrastructure including servers, GPUs, and data centers. Yet consensus capex estimates have proven too low for two years running—at the start of both 2024 and 2025, Street projections implied roughly 20% growth, but actual spending exceeded 50% in both years, per Goldman Sachs research.

Recursive Self-Improvement Sets 2028 Peak

In the Barclays framework, hyperscale capex peaks in 2028 based on when AI labs are expected to hit recursive self-improvement, which would increase the efficiency of spend and reduce training cost, with training operating expenses peaking in 2029. The analysis assumes existing training chips will handle most inference workloads beginning in 2027, meaning additional spending on inference-specific silicon is not yet reflected in projections, and other AI developers beyond OpenAI and Anthropic will gradually expand their share of global compute demand, which could further increase infrastructure spending.

Capital intensity has surged to previously unthinkable levels—Oracle at 57% of revenue, Microsoft at 45%—and is expected to increase further in 2026, with approximately 75% of aggregate hyperscaler capex dedicated to AI infrastructure. Hyperscalers are increasingly leaning on debt markets to bridge the gap between rapidly rising AI capex budgets and internal free cash flow, with aggregate capex for the big five now above projected cash flows after buybacks and dividends, necessitating external funding, according to CreditSights.

Semiconductor Valuation Repricing Ahead

The research arrives as Nvidia shares have been down less than 1% year to date after the stock’s blockbuster performance in recent years, with underperformance driven by broader rotation away from megacap technology stocks amid concerns about high valuations and the longer-term viability of AI infrastructure spending. Investors largely ignored Nvidia’s blowout earnings report and strong guidance in late February, yet O’Malley pointed out that hyperscalers’ total AI spending could climb even further as tech giants move to next-generation hardware.

Barclays’ current estimates assume only Nvidia’s Blackwell-architecture GPUs, but the firm said Nvidia’s newer solutions—including the Vera Rubin, Vera Rubin Ultra, and Feynman chip families—could lead to total capex expansion. Nvidia anticipates the AI infrastructure build-out will continue for several more years, with data center capital expenditures projected to reach $650 billion in 2026, according to company guidance.

Energy Demand Forecasting Recalibrated

The upward revision to semiconductor capex carries downstream consequences for power grid planning. Electricity consumption in accelerated servers, mainly driven by AI adoption, is projected to grow 30% annually, with accelerated servers accounting for almost half the net increase in global data center electricity consumption, per the International Energy Agency. Global power demand from data centers is forecast to increase 50% by 2027 and by as much as 165% by 2030 compared with 2023, according to Goldman Sachs.

U.S. data centers consumed 183 terawatt-hours of electricity in 2024, more than 4% of total U.S. consumption and roughly equivalent to Pakistan’s annual demand, with projections for 133% growth to 426 TWh by 2030, per Pew Research. BloombergNEF forecasts U.S. data center power demand will more than double by 2035, rising from almost 35 gigawatts in 2024 to 78 gigawatts, while actual energy consumption growth will be steeper with average hourly electricity demand nearly tripling from 16 gigawatt-hours to 49 gigawatt-hours.

Supply Chain Geopolitics Amplified

The extended capex cycle sharpens strategic questions around semiconductor manufacturing concentration. Taiwan Semiconductor Manufacturing captures nearly all highest-end chip production work as the largest manufacturer by revenue and leader in delivering the most advanced chips, with Broadcom, AMD, and Nvidia working with TSMC for cutting-edge chips in an industry nearly impossible to disrupt. This concentration means GPU suppliers like Nvidia, memory vendors including SK Hynix, Samsung, and Micron, and data center infrastructure providers face unprecedented demand, with the $180 billion GPU/accelerator spend representing roughly 6 million GPUs at approximately $30,000 average price.

Hyperscalers are on track to spend between $500 billion and $600 billion annually on AI infrastructure, yet the market is becoming impatient with return on investment. The current dip in Nvidia’s valuation highlights a growing concern known as the “Monetization Gap,” with industry data suggesting that while billions are poured into GPUs and data centers, direct revenue from AI services is estimated between only $25 billion and $100 billion.

What to Watch

Hyperscaler Q1 and Q2 2026 capex guidance will test whether spending tracks Barclays’ elevated forecasts or converges toward consensus. Power purchase agreement announcements from Microsoft, Amazon, and Google signal whether energy procurement matches projected compute expansion. TSMC capacity allocation and lead times for 3nm and 2nm nodes will reveal supply chain stress points. Nvidia’s Vera Rubin and Feynman rollout timelines in late 2026 provide a real-time test of whether next-generation architecture shifts extend the replacement cycle. Watch whether energy regulators in Virginia, Ohio, and Texas approve grid expansion plans sized to elevated demand rather than consensus forecasts. The monetization gap between infrastructure spend and AI service revenue remains the central risk; if direct AI revenues fail to scale proportionally through 2027, the thesis unwinds regardless of labs’ training budgets.