The horizon of artificial intelligence: LLMs and hardware, its current pillars

© 2026. All rights reserved. | Juan Antonio Lloret Egea | ORCID: 0000-0002-6634-3351 | Instituto Tecnológico Virtual de la Inteligencia Artificial para el Español™ (ITVIA) | IA eñ™ — La Biblia de la IA™ | ISSN 2695-6411 | CC BY-NC-ND 4.0 | Article 006 · 11 November 2026

Spanish original in ITVIA · DOI: 10.21428/39829d0b.2598a051

Competition between accelerators expands AI’s possibilities, but capacity contracts and dependence on their tools also shape which paths can be pursued.

Anthropic, the company behind Claude, offers a revealing picture of the race for artificial intelligence. It has committed to spending more than $100 billion on Amazon technologies over ten years [1]; agreed to purchase $30 billion of Azure capacity, powered by NVIDIA [2]; and announced five gigawatts of capacity with Google and Broadcom, within a TPU expansion expected to come online from 2027 [3, 4]. It thus combines Trainium, GPUs and TPUs. These are commitments with different delivery schedules, not capacity already available in full. Securing somewhere to run models has become a decision capable of shaping which models can be built.

This diversification reflects real differences. NVIDIA and AMD graphics processing units (GPUs) support a broad range of parallel computing, from simulation and scientific analysis to artificial intelligence [5, 6]. Google tensor processing units (TPUs) specialise in matrix operations, which are central to neural networks [7]. Generality makes it easier to experiment with varied tasks; specialisation allows computation to be organised around particular workloads. Both families support training and inference, and some GPUs also contain specialised units [7, 8]. The advantage depends on the workload, memory, communications and cost of the complete system. A peak performance figure alone cannot compare the accuracy, response times and actual utilisation of two installations.

NVIDIA’s position becomes clearer when we examine CUDA, a computing platform and programming model with libraries and tools accumulated over many years [5]. That investment becomes embedded in programs, teams’ experience and their verification procedures. Moving them to another manufacturer can require adaptation, optimisation and fresh testing [6]. This creates a competitive barrier that a TPU’s greater efficiency on a particular task does not automatically remove. Even incorporating Rust—a language I strongly recommend—into CUDA, as NVIDIA announced in September, expands the ways to program its GPUs without removing dependence on its platform [9].

Yet that barrier is being challenged. Google is developing TorchTPU to run PyTorch programs on its accelerators with fewer changes [10]. AMD combines its Instinct GPUs with ROCm and has an agreement with OpenAI to deploy six gigawatts over several years [6, 11]. Commercial boundaries are not absolute either: Google’s TPU 8tx system, announced on 9 November, incorporates AMD central processors [12]. Competing over the accelerator and collaborating on the rest of the system can be parts of the same business. The cost of switching suppliers is therefore itself a target for innovation, rather than an immutable property of the market.

Large language models (LLMs) feed this pressure. Training them requires processing vast quantities of data; using them sustains ongoing demand. Long conversations, elaborate responses and systems that perform sequences of tasks add computation and memory requirements [12, 13]. Each operation can become cheaper while aggregate consumption rises. The International Energy Agency records efficiency improvements alongside constraints on electricity supply, semiconductor manufacturing and high-bandwidth memory [13]. Pressure consequently reaches manufacturers of memory, networking and electrical equipment. The constraint extends beyond the model to the infrastructure needed to use it.

My reading is that the market is selecting part of AI’s future before that future can be scientifically assessed. Amazon, NVIDIA and Microsoft have announced investments in Anthropic [1, 2]. Suppliers thus finance the companies buying their capacity, while buyers diversify to secure supply and negotiate terms. These agreements enable advances, but also commit resources for years. We should ask which research directions this structure favours and which become difficult to finance. The ability to contract more computing power does not, by itself, measure the quality of the reasoning obtained.

As director of several publications, I believe we should also investigate artificial intelligence that does not make large-scale training a condition of its existence. The book Artificial Intelligence of the Vectorial System develops this alternative from Image-Based Language and its relationship with a new programming language: the Vectorial System’s Computing Language [14, 15]. Learning is conceived as incorporating explicit knowledge; reasoning, as verifiable transformations of visual representations, without verbal prediction or opaque weight adjustment. This architecture does not conceptually require a data centre [16]. Its complete implementation and cost on comparable tasks must be demonstrated [15, 16], but it allows us to question whether LLM requirements are universal necessities for AI.

Competition between processors can expand the possibilities of artificial intelligence, provided alternatives remain open. Progress should be judged by the cost of obtaining a valid result, what makes it verifiable and who can reproduce it. If access to capital also determines what deserves to be called intelligence, we will have handed the market a decision that still requires scientific judgement.