The artificial intelligence industry has long been dominated by Nvidia, whose graphics processing units power everything from large language models to self-driving cars. But a team at the University of Cambridge has just demonstrated that the monopoly is not absolute. They have built a planetary-scale foundation model for the Earth, and every chip that trained and ran it came from AMD, not Nvidia.
The model is called TESSERA, an acronym that stands for Transformer for Earth Surface and Environmental Representation and Analysis. Its design borrows from large language models. Where an LLM learns from text, TESSERA learns from space. It ingests years of imagery from the European Space Agency’s Sentinel satellites, both radar and optical, the team announced. It then compresses each 10-metre square of the planet’s land into a compact 128-number “fingerprint,” or embedding.
That may sound abstract, but the payoff is concrete. Normally, mapping crops, forests, or floods from satellite data means building a bespoke model and hand-labelling thousands of examples for each task. With TESSERA’s fingerprints already computed, researchers can build those tools with far less data, often on an ordinary CPU. The team says it needs about 30 times less labelled data than starting from raw imagery. For instance, a farmer monitoring crop health might only need a handful of annotated satellite images to calibrate a yield prediction model, rather than thousands. This reduction in data requirements is a game-changer for regions where expertise and computing resources are scarce.
Training and inference on AMD hardware
The more unusual part of the project is what it ran on. Cambridge’s Energy and Environment group trained TESSERA on 16 AMD Instinct MI300X GPUs, using roughly 6,200 GPU-hours and AMD’s open ROCm software stack. Training a model is the easy bit. Running it across every 10-metre pixel on Earth—about 1.5 trillion a year—is the hard part. For that, the team turned to Vultr, an independent cloud firm that sponsored the compute, using six of its bare-metal servers. Each server packs eight AMD Instinct MI325X GPUs. Together they churn out roughly three terabytes of compressed data a day. The researchers liken that to covering the landmass of Italy, daily. A single year of global coverage takes months of continuous processing.
The use of AMD hardware is significant because it challenges the conventional wisdom that serious AI workloads require Nvidia’s CUDA ecosystem. AMD’s ROCm is an open-source alternative that has historically lagged in software maturity and community support. But projects like TESSERA show that with enough engineering effort, AMD GPUs can perform comparably for large-scale geospatial AI tasks. The team noted that they had to optimize the training pipeline for ROCm, but the performance was sufficient to train a frontier-scale model on a modest budget.
Applications across agriculture, conservation, and energy
The point of TESSERA is openness. Cambridge is releasing the embeddings free under a Creative Commons licence, alongside the full training recipe. Any government, researcher, or startup can build on them. Professor Anil Madhavapeddy, who leads the work, says the goal is to “democratise access to planetary-scale environmental monitoring.” The uses are concrete.
Farmers can track crop health and forecast yields at 10-metre resolution, a boon for smallholders in developing regions. For example, in sub-Saharan Africa, where ground-based surveys are expensive and infrequent, TESSERA’s embeddings could allow local agricultural offices to detect pest outbreaks or water stress weeks sooner than traditional methods. Conservationists can watch habitats shift, from tropical forests to the UK hedgerows that shelter hedgehogs. Energy planners can map solar and wind sites to guide the transition. All of it runs off the same shared fingerprints.
Moreover, TESSERA’s pre-computed embeddings enable rapid experimentation. Researchers can run a random forest or simple neural network on the embeddings to classify land cover types, rather than training a deep learning model from scratch on raw pixels. That lowers the barrier to entry for scientists in disciplines like ecology or urban planning who may not have deep AI expertise.
Broader context: the AI hardware landscape
TESSERA is a small story with a large subtext. A public university built a frontier-scale AI model, made it free, and did it on the hardware everyone says you cannot use. The field is convinced that serious AI means Nvidia GPUs and private labs. A planet’s worth of embeddings, trained on AMD and given away, is a quiet rebuttal. The development also highlights the growing importance of open-source AI stacks. As concerns mount over vendor lock-in and the environmental cost of training massive models, initiatives like TESSERA demonstrate that alternative paths are viable.
Cambridge’s choice of AMD was not merely ideological. The team considered Nvidia options but found that AMD offered better price-to-performance for their specific workload—large-scale matrix multiplications on high-resolution satellite data. Additionally, because ROCm is open source, the team could inspect and modify the driver-level code to optimize memory usage for the 10-metre pixel processing pipeline. This level of customization is harder to achieve with Nvidia’s proprietary CUDA ecosystem.
The implications extend beyond Earth observation. If AMD’s Instinct line can handle planetary-scale geospatial AI, it may also be suitable for other data-intensive fields like climate modeling, genomics, and physics simulation. The success of TESSERA could encourage other research groups to explore AMD hardware, potentially fostering a more competitive and diverse AI hardware market. That, in turn, could lower costs and accelerate innovation across the board.
Looking ahead, the Cambridge team plans to extend TESSERA to include time-series forecasting—predicting crop yields or deforestation months in advance based on the embeddings. They are also exploring partnerships with humanitarian organizations to deploy TESSERA-based tools for disaster response. For instance, after a flood, the embeddings could be used to quickly map inundated areas without needing manual annotation of new satellite imagery. The team emphasizes that all future work will remain open-source, ensuring that the benefits are widely shared.
In a field often dominated by billion-dollar corporate labs, TESSERA stands as a reminder that public research institutions can still punch above their weight. By choosing AMD and an open licence, Cambridge has not only advanced geospatial AI but also sent a message that the future of AI need not be locked into a single vendor.