Cambridge creates planetary AI model with AMD chips, challenging Nvidia

What happened?

Researchers at the University of Cambridge have unveiled TESSERA (Terra Earth System Science and Earth Resources Analysis), a foundation AI model designed to analyze Earth observation data on a planetary scale. What makes TESSERA unique is that it was trained and run exclusively on AMD processors, completely bypassing Nvidia chips, which dominate the AI market. The model was trained on the Vultr cloud, using clusters of AMD Instinct MI250 and MI300 GPUs. According to The Next Web, the Cambridge team needed only 48 hours to train the model, highlighting the efficiency of AMD hardware for this specific task. TESSERA is designed to process petabytes of satellite data, such as from the ESA's Copernicus program, and can be applied to monitoring deforestation, climate change, agriculture, and natural disasters.

Why is it important?

Nvidia controls roughly 80% of the AI accelerator market, and its CUDA platform is the de facto standard for deep learning model development. The TESSERA project demonstrates that it is possible to build cutting-edge models with alternative hardware, specifically AMD GPUs and the ROCm (Radeon Open Compute) software. This has significant implications:

• Hardware competition: It breaks Nvidia's monopoly and validates AMD as a viable alternative for large-scale AI workloads. The Instinct MI250 and MI300 GPUs, based on CDNA 2 and 3 architectures, offer competitive performance in FP16/FP32, though historically they have lagged behind Nvidia in software and ecosystem.
• Technological sovereignty: Governments and companies seeking independence from a single supplier can consider AMD options. For example, the European Union has shown interest in diversifying its chip supply chain, and projects like TESSERA could accelerate that transition.
• Software innovation: ROCm, AMD's software ecosystem for high-performance computing, demonstrates sufficient maturity for complex projects. Although ROCm has had compatibility and performance issues in the past, recent versions have improved significantly, and TESSERA is proof that it can scale to a planetary level.

What consequences will it have?

In the short term, TESSERA is a proof of concept that could accelerate AMD adoption in AI data centers. In the long term, it could pressure Nvidia to reduce prices or innovate more aggressively. It could also boost other manufacturers like Intel (with its Ponte Vecchio GPUs) or startups like Cerebras and Graphcore. For users, more competition means more options and potentially lower compute costs. However, we should not underestimate Nvidia's software advantage: CUDA has a library of optimizations and tools (cuDNN, TensorRT) that ROCm does not yet match. Additionally, AMD's AI performance is not uniform; in benchmarks like MLPerf, Nvidia GPUs often lead in training large models. But for specific tasks like Earth observation, AMD has proven sufficient. In terms of market share, AMD's share of data center GPUs was around 10% in 2024, but it could grow if projects like TESSERA inspire confidence. Companies like Microsoft are already testing Instinct for certain workloads, and Vultr, the cloud provider used, benefits from offering an alternative to Nvidia clusters.

What should readers know?

TESSERA is not only a technical achievement but also a symbol that the AI ecosystem is not tied to a single vendor. However, there are nuances: AMD's AI performance still does not match Nvidia's in all benchmarks, and the maturity of ROCm software is lower than CUDA. Additionally, TESSERA focuses on Earth observation, not generative tasks like GPT, so its applicability to other domains remains to be seen. Readers should closely follow independent benchmarks and enterprise adoptions of AMD in AI. For example, the Cambridge team used a technique called 'contrastive pretraining' to learn representations of Earth without labels, similar to how LLMs learn from unstructured text. This could inspire other researchers to explore alternative hardware. Compared to previous events, such as Google's use of TPUs for language models, TESSERA shows that hardware diversification is possible and beneficial. However, Nvidia is not standing still: its upcoming Blackwell GPUs promise a performance leap, and its software remains the most robust. For enterprise users, the decision between Nvidia and AMD will depend on total cost of ownership, availability of optimized software, and scalability. In summary, TESSERA is a milestone, but not an immediate revolution.

“The AI industry runs on Nvidia. A team at the University of Cambridge has just shown it does not have to.” — The Next Web