Behind every AI model that answers your questions or generates a video clip lies an enormous amount of hardware – processors, servers, data centres, fibre-optic cables. In 2026, demand for all of this has far exceeded the available supply. At Altanet Craiova we believe understanding this infrastructure is essential for anyone who wants to make good technology decisions in the coming years.
How big is the appetite for AI hardware?
The numbers are hard to imagine. At the end of 2025, organisations around the world were spending 37 billion dollars a year on AI infrastructure – three times more than in 2024. This amount represents roughly 6% of total global spending on software as a service (SaaS – that is, applications and platforms accessed over the internet on a monthly or annual subscription).
Demand for computing power exceeded supply in 2025 and shows no sign of levelling off any time soon. Companies reserve processing capacity months or even years in advance.
Who makes the hardware for AI?
Not all processors are the same. Those used for AI are far more specialised than an ordinary laptop processor:
- NVIDIA remains the undisputed market leader, with the H200, B200 and GB200 series – graphics processors (GPU – graphics processing units) optimised specifically for training and running AI models. A single H200 processor costs between 30,000 and 40,000 dollars.
- AMD is advancing with competitive alternatives at lower costs, gaining ground especially in data centres.
- Intel, Google, Amazon, Microsoft have developed their own specialised AI processors (called ASICs – application-specific integrated circuits), optimised for their internal needs.
- Huawei and Cambricon (China) produce competitive processors that circumvent American export restrictions, fuelling AI development in China.
What AI infrastructure spending looks like
The chart below shows the evolution of global AI infrastructure spending and the distribution by hardware type:
$12B
$25B
17%
Actual / estimated data
Projection
Cloud, hybrid or local – where does AI run?
Not all companies can afford or want to buy their own processors. In 2026 there are three main models of access to AI infrastructure:
- Cloud: you pay for what you use, with no upfront investment. Amazon AWS, Microsoft Azure and Google Cloud offer access to top AI processors by subscription. Advantage: flexibility. Disadvantage: costs can climb quickly at high volume.
- Hybrid: part of the processing is done in the cloud, part on your own servers. Estimates show that 75% of the AI used in large companies will run on hybrid infrastructures by 2028.
- Local (Edge AI): small models that run directly on a phone, laptop or industrial equipment, with no internet connection. In 2026, this option has gone from experiment to commercial reality.
What comes next?
Kaoutar El Maghraoui of IBM anticipates the emergence of a new class of processors specialised for AI agents – different from today's GPUs, optimised for the specific tasks of autonomous agents. Competition is no longer about models, but about integrated systems. In 2027, 70% of organisations will prioritise aligning their AI infrastructure investments with measurable business results – they no longer invest just anywhere, but strategically.
If your company uses or plans to use AI and you want to understand which infrastructure suits you best, the Altanet Craiova team can help with a concrete assessment. Visit our contact page and let's talk.
This article is part of the Altanet series on AI trends in 2026. Next article: AI Security: The New Types of Attacks That No Classic Antivirus Protects You From. See also the complete guide to the series.