Gartner Projects 96% Increase in AI IaaS Spending by 2026, Reaching $42 Billion, with Inference Surpassing Training

Gartner projected on Monday, August 10, that global spending on AI-optimized Infrastructure as a Service (IaaS) will grow by 96% in 2026, reaching $42 billion. By 2027, this category is expected to reach $66 billion, maintaining a growth rate well above the rest of the public cloud market. For the first time since the boom of LLMs, spending on inference will exceed that on training: $23.3 billion versus $19 billion within the same total.

This inflection point is not trivial. For three years, hyperscalers and AI labs have justified billion-dollar capital expenditures with the promise of training larger models. The crossover in 2026 indicates that revenues will be dominated by production workloads running for end customers, not by internal experiments. Gartner attributes this shift to the operationalization of AI in enterprise applications and the continuous demand for large-scale LLMs.

What the 96% Curve Hides

The average figure is misleading. AWS, Microsoft Azure, and Google Cloud account for the largest share of the increase and capture the bulk of the new spending. Oracle Cloud Infrastructure, having signed multi-year contracts with OpenAI, and CoreWeave, focused on Nvidia GPUs, enter as challengers with growth above the market average. Outside the U.S., the picture becomes fragmented.

In the European Union, the demand for data sovereignty pushes clients towards local providers like OVHcloud and Scaleway in France, and Deutsche Telekom in Germany, which operate data centers in Frankfurt, Paris, and Amsterdam with H100 and H200 clusters. None of them come close to the market share of American hyperscalers, but their contracts include data residency clauses that U.S. providers are still negotiating on a case-by-case basis following the Schrems II ruling. The European Commission accelerated the timeline for implementing the EU Data Act in July, and regulatory pressure today appears as a driver of expenditure fragmentation.

Why Consulting and Integration Are Winning

For global consultancies, the $42 billion figure is worth less than the mix behind it. Deloitte, Accenture, Capgemini, and IBM Consulting have reported in recent earnings cycles that agentic AI projects for enterprise clients have shifted from small pilot-style tickets to multi-year contracts worth millions of dollars. The work in integration, MLOps, and governance is growing faster than compute consumption itself, which supports margins despite compression in consulting fees.

Accenture announced in the last three months a layoff of approximately 11,000 people as part of a restructuring program linked to reallocating capital for this expansion, and CEO Julie Sweet publicly stated that the company will end contracts with anyone who cannot be redeployed into new skill sets. Global headcount dropped from 801,000 in February to 779,000 in August. This is the clearest evidence that infrastructure spending is migrating up the value chain without generating corresponding employment at the base.

What Changes for Buyers in São Paulo and Singapore

In Brazil, large banks have migrated critical risk and anti-fraud workloads to GPU pods in Azure and AWS throughout 2025, contracts that will be renegotiated in the next 18 months. The internal reading among architecture teams is twofold. The cost per hour of A100 and H100 has fallen relative to the peak in 2024, but the total volume has exploded, pushing the monthly bill upward rather than reducing it. The ongoing response combines two levers: adopting open models like Muse Glimmer and Qwen to shift some consumption to local GPUs, and tying long-term reserve contracts with hyperscalers before the next wave of increases arrives.

In Singapore, the landscape is different. DBS and OCBC have exported risk models to offices in Jakarta and Manila and are negotiating inference capacity with GPUs in regional data centers. Alibaba Cloud, despite aggressive discounts, faces compliance restrictions for banks regulated by the MAS. Spending is concentrated on Azure Singapore and contracts with the government itself, which subsidized GPU capacity for sovereign use in July.

What Gartner Leaves Open

The point that the report does not address is the return on this spending. The curve of $66 billion in 2027 hinges on the premise that inference generates growing revenues for those who consume it. If one of the large banks or major retailers discloses in the next earnings cycle that the EBIT of AI projects remains below capex costs, the entire market will have to revise its projections downward. This is the second-order risk that infrastructure investors have yet to price in, and which the next quarter may reveal.