AI capex: A global growth story still in the making

Some American-headquartered hyperscaler companies have tended to dominate the conversation around AI capex. Recently, however, the narrative has gone global, with increased activity from both global and local companies in Europe, the Middle East and Africa (EMEA), Asia-Pacific (APAC) and Latin America (LATAM). Completed and soon-to-be-completed projects across these geographies illustrate this expansion:

• South Africa: Teraco, a data center provider based in South Africa, completed an expansion in Johannesburg that added 30 MW of capacity specifically designed to support high-density AI and cloud workloads.
• Qatar: Meeza, a Qatari data center provider, completed a 4 MW data center expansion project on behalf of an anonymous global hyperscaler, and inked another deal to build 8 MW as of September 2026.
• Malaysia: Vantage Data Centers, a U.S.-based company, operates three primary campuses outside Kuala Lumpur.
• Thailand: Alibaba Cloud opened its second data center in Thailand in 2025, marking a significant expansion of Chinese hyperscaler infrastructure in Southeast Asia.
• Brazil: ByteDance, the owner of TikTok, is spending about $39 billion on one data center on Brazil’s northeast coast, according to a report from Inc. If completed, it will be the largest data center operated by ByteDance outside China.
• Mexico: “Mexico is emerging as a hardware and manufacturing hub within the AI ecosystem,” said Adrian Huerta, head of Mexico Equity Strategy at J.P. Morgan. Huerta noted that tech-related shipments from Mexico have risen from 36% of total country exports in 2023 to 48% today, generating an estimated $10.9 billion trade surplus as of September 2026.

In additional to buildout, global policy is starting to align behind the AI capex wave. In June 2026, the European Commission unveiled the European Technological Sovereignty Package (ETSP), a measure designed to reduce reliance on non-EU suppliers for digital technologies and accelerate AI infrastructure investment. The Cloud and AI Development Act, which was part of the ETSP, specifically aims to triple EU data center capacity within 5–7 years.

The EU has also launched tenders to build up to seven AI gigafactories — massive, purpose-built computing buildings that can house over 100,000 specialized AI chips, draw more than a gigawatt of power and are designed for training and deploying frontier artificial intelligence models. Seven sites have been targeted and over €10 billion in public funding mobilized, with construction expected to start in 2027.

Other factors beyond financing limit the pace of the AI capex buildout

Beyond questions around financing, several factors have emerged as meaningful roadblocks for AI capex:

Power availability and infrastructure:

• Securing sufficient, reliable power remains the primary bottleneck for AI capex, due largely to the amount and density of energy required to power data centers, as well as their need for land and proximity to water. Upgrades to grid infrastructure and access to renewable energy are often slow and costly, impacting project timelines and scalability.
• “The surge in AI workloads is driving a step-change in power demand, with data centers projected to account for 9%–17% of total U.S. electricity demand by 2030, up from 3%–4% today,” said Tarek Hamid, head of the North American Corporate Credit Research team at J.P. Morgan. “Globally, data center electricity consumption is expected to more than double by 2030, with the U.S. and China leading incremental demand.”

Equipment supply and accelerated depreciation

• Long lead times for critical equipment — including graphics processing units, or GPUs, along with memory modules, cooling systems and power solutions — represent another major constraint. Currently, a memory shortage is driving up the cost of DRAM, or dynamic random access memory, which is used in AI inferencing and model training. Amazon’s $20 billion increase in its capex guidance for 2026 was attributed largely to increased memory costs.
• “AI infrastructure relies on rapidly advancing chips and systems,” said Feroli. “As newer generations deliver better performance and lower cost per unit of compute, older hardware loses economic value faster than conventional equipment.” In other words, AI capex can advance so rapidly that chips, hardware and even models become obsolete faster than traditional hardware and machinery, making the true value of these goods more difficult to ascertain.

Construction and labor

• Data center construction cycles typically run 18–30 months, and can be further delayed by shortages of skilled labor, permitting challenges and rising construction costs, slowing the delivery of new capacity and triggering inventory adjustments across the supply chain.

Regulatory developments

• Permitting, environmental regulations and land availability often delay projects, especially in high-demand regions. Increasingly, public backlash and municipal or state moratoriums are also forcing developers to seek alternative sites as public and political debate continues. Much of this opposition is rooted in consumer concerns: as data centers draw on local grids, residents and policymakers are increasingly focused on the effect on household electricity bills and grid reliability. In several markets, that has shifted the debate toward who bears the cost of new capacity — with proposals to place more of the incremental infrastructure expense on developers rather than ratepayers.
• “Data center expansion is no longer just an engineering-scarcity story; it’s a political economy question of who gets power, who funds upgrades, and who accepts the local impacts,” said Hannah Lee, an equity analyst covering APAC Thematic and Sustainable Investing Research at J.P. Morgan. “As those constraints stack, approval timelines and delivered costs become harder to predict and more dispersed.”