Large scientific models should not belong only to organizations that can afford a single, tightly coupled supercluster. This concern is not only philosophical: modern AI research has become increasingly compute-intensive, and unequal access to specialized compute has been described as a “compute divide” in AI research [1]. In biology and chemistry, the most valuable data is often distributed across labs, instruments, clinical partners, industrial facilities, and autonomous experimental platforms. Moving all of that data into one central training site is expensive, slow, and often impossible for privacy, sovereignty, or operational reasons, especially in biomedical settings where clinical data remains siloed across institutions [2]. PCP, the Planetary Compute Protocol, is our answer: a distributed compute fabric for training, reinforcement learning, inference, and lab-connected scientific automation across heterogeneous infrastructure.

PCP began from the motivation behind SPRIND’s Composite Learning Challenge. SPRIND framed composite learning as a combination of distributed, decentralized, and federated learning: AI training across diverse systems without depending on one centralized data center [3]. The challenge asks teams to build robust frameworks that train efficiently across different locations, data regimes, and hardware, from GPUs to CPUs and devices from different manufacturers, while tolerating outages and changing resource availability [3]. That maps closely to PCP’s core design: useful learning across machines that do not look identical, fail together, or live in the same administrative domain. This direction also builds on the broader federated-learning literature, where shared models are trained from decentralized data through locally computed updates and orchestrated aggregation rather than raw-data centralization [4], [5].

We extended that scope in two directions. First, PCP is not only a training scheduler. It is a scientific execution layer for training, reinforcement learning, and inference. Bio/chem workflows need closed loops: generate a molecule, predict a property, decide the next experiment, observe the result, and improve the policy. This closed-loop view is aligned with the emerging self-driving laboratory literature, where automation, machine learning, active learning, and robotic or remote experiment execution are combined to accelerate discovery [6], [7]. It also matches reinforcement-learning formulations of molecular optimization, where molecule generation can be treated as a sequential decision problem over chemical space [8]. PCP supports RL rollouts and inference services alongside model training, using the same worker-fabric abstraction and capability-aware scheduling. This is consistent with distributed AI systems such as Ray and RLlib, which showed how task-parallel, actor-based, and reinforcement-learning workloads can be coordinated through scalable distributed execution primitives [9], [10].

Second, PCP is the infrastructure side of DeltaWave OS. In DeltaWave OS, Eve orchestrates agents, projects, artifacts, provenance, and lab-facing workflows [11]. PCP provides the compute fabric underneath: workers register their backend, target architecture, capacity, and capabilities; the gateway routes training, inference, and RL jobs; Kubernetes owns pod placement, secrets, storage, and device allocation [11], [12]. This separation is deliberate. DeltaWave OS is the operating layer for scientists and labs; PCP is the disaggregated compute layer that lets the system use whatever hardware is available without hard-coding the science workflow to one cluster.

Today PCP supports a modular set of execution backends and learning modes. It can run CPU, NVIDIA CUDA, AMD ROCm, Apple Metal, and Vulkan/IREE paths, with Tenstorrent/RISC-V planned as a next backend lane [13]–[16]. It supports local and gateway-driven training, inference endpoints, RL job routing, and JEPA-style scientific training flows, drawing on the broader joint-embedding predictive architecture literature in self-supervised representation learning and world-modeling [17], [18]. For distributed optimization, PCP can select among classic DiLoCo, Streaming DiLoCo, and our newest integration in the series: Decoupled DiLoCo. DiLoCo introduced low-communication training across poorly connected islands of devices [19]. Streaming DiLoCo reduced peak bandwidth by synchronizing parameter subsets sequentially and allowing workers to continue training during synchronization [20]. Google DeepMind introduced Decoupled DiLoCo as a way to train across distant data centers with lower bandwidth and higher resilience by separating training into asynchronous learner islands [21]. PCP’s modular design lets us integrate that progress quickly: aggregation strategy, backend, target architecture, and worker placement are configuration and protocol concerns, not a platform rewrite.

This is also where PCP connects to the broader movement toward disaggregated AI systems. In inference, disaggregation often means separating prefill and decode across different devices or resource pools, because prompt processing and token generation have different compute, memory, latency, and throughput profiles [22]–[24]. More generally, it means treating heterogeneous compute, memory, network, and scheduling as one coordinated system, as seen in large-scale accelerator orchestration systems such as Pathways [25]. PCP applies the same idea to scientific learning: a lab GPU, a cloud H100 pool, an Apple workstation, an AMD node, and eventually Tenstorrent hardware can all become parts of one learning fabric where the scheduler understands their capabilities.

The future direction is clear. We will expand backend support, harden large-scale RL model rollouts for bio/chem use cases, and deepen lab integration through DeltaWave OS and Eve. That includes laboratories on Earth and, through SpacePharma, literal space labs: autonomous microgravity lab-on-a-chip systems that can be monitored and controlled remotely [26], [27]. Our goal is a platform where models do not just learn from static datasets, but participate in the scientific loop: proposing, executing, observing, and improving experiments across distributed labs and distributed compute.

References

  1. Ahmed, N., & Wahed, M. (2020). The de-democratization of AI: Deep learning and the compute divide in artificial intelligence research. arXiv. https://arxiv.org/abs/2010.15581
  2. Rieke, N., Hancox, J., Li, W., Milletari, F., Roth, H. R., Albarqouni, S., Bakas, S., Galtier, M. N., Landman, B. A., Maier-Hein, K., Ourselin, S., Sheller, M., Summers, R. M., Trask, A., Xu, D., Baust, M., & Cardoso, M. J. (2020). The future of digital health with federated learning. npj Digital Medicine, 3, Article 119. https://doi.org/10.1038/s41746-020-00323-1
  3. SPRIND. (n.d.). Composite Learning Challenge. https://www.sprind.org/en/actions/challenges/composite-learning
  4. McMahan, H. B., Moore, E., Ramage, D., Hampson, S., & Agüera y Arcas, B. (2017). Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (pp. 1273–1282). PMLR. https://proceedings.mlr.press/v54/mcmahan17a.html
  5. Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., Bennis, M., Bhagoji, A. N., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R., D’Oliveira, R. G. L., Eichner, H., El Rouayheb, S., Evans, D., Gardner, J., Garrett, Z., Gascón, A., Ghazi, B., Gibbons, P. B., ... Zhao, S. (2021). Advances and open problems in federated learning. Foundations and Trends in Machine Learning, 14(1–2), 1–210. https://doi.org/10.1561/2200000083
  6. Tom, G., Schmid, S. P., Baird, S. G., Cao, Y., Darvish, K., Hao, H., Lo, S., Pablo-García, S., Rajaonson, E. M., Skreta, M., Yoshikawa, N., & Aspuru-Guzik, A. (2024). Self-driving laboratories for chemistry and materials science. Chemical Reviews. https://doi.org/10.1021/acs.chemrev.4c00055
  7. Kusne, A. G., Yu, H., Wu, C., Zhang, H., Hattrick-Simpers, J., DeCost, B., Sarker, S., Oses, C., Toher, C., Curtarolo, S., Davydov, A. V., Agarwal, R., Bendersky, L. A., Li, M., Mehta, A., & Takeuchi, I. (2020). On-the-fly closed-loop materials discovery via Bayesian active learning. Nature Communications, 11, Article 5966. https://doi.org/10.1038/s41467-020-19597-w
  8. Zhou, Z., Kearnes, S., Li, L., Zare, R. N., & Riley, P. (2019). Optimization of molecules via deep reinforcement learning. Scientific Reports, 9, Article 10752. https://doi.org/10.1038/s41598-019-47148-x
  9. Moritz, P., Nishihara, R., Wang, S., Tumanov, A., Liaw, R., Liang, E., Elibol, M., Yang, Z., Paul, W., Jordan, M. I., & Stoica, I. (2018). Ray: A distributed framework for emerging AI applications. In Proceedings of the 13th USENIX Symposium on Operating Systems Design and Implementation (pp. 561–577). USENIX Association. https://www.usenix.org/conference/osdi18/presentation/moritz
  10. Liang, E., Liaw, R., Moritz, P., Nishihara, R., Fox, R., Goldberg, K., Gonzalez, J. E., Jordan, M. I., & Stoica, I. (2018). RLlib: Abstractions for distributed reinforcement learning. In Proceedings of the 35th International Conference on Machine Learning (pp. 3053–3062). PMLR. https://proceedings.mlr.press/v80/liang18b.html
  11. DeltaWave. (2026). PCP technical specification and DeltaWave OS integration notes [Internal technical documentation].
  12. Kubernetes. (2024). Schedule GPUs. Kubernetes Documentation. https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus/
  13. IREE. (n.d.). Deployment configurations. https://iree.dev/guides/deployment-configurations/
  14. Apple. (n.d.). Accelerated PyTorch training on Mac. Apple Developer Documentation. https://developer.apple.com/metal/pytorch/
  15. AMD. (n.d.). AMD GPU device plugin for Kubernetes. AMD Instinct Documentation. https://instinct.docs.amd.com/projects/k8s-device-plugin/en/latest/
  16. Tenstorrent. (n.d.). TT-Metalium software stack overview. Tenstorrent Documentation. https://docs.tenstorrent.com/getting-started/tt-software-stack.html
  17. Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., & Ballas, N. (2023). Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 15619–15629). https://arxiv.org/abs/2301.08243
  18. Bardes, A., Garrido, Q., Ponce, J., Chen, X., Rabbat, M., LeCun, Y., Assran, M., & Ballas, N. (2024). Revisiting feature prediction for learning visual representations from video. arXiv. https://arxiv.org/abs/2404.08471
  19. Douillard, A., Feng, Q., Rusu, A. A., Chhaparia, R., Donchev, Y., Kuncoro, A., Ranzato, M., Szlam, A., & Shen, J. (2023). DiLoCo: Distributed low-communication training of language models. arXiv. https://arxiv.org/abs/2311.08105
  20. Douillard, A., Donchev, Y., Rush, K., Kale, S., Charles, Z., Garrett, Z., Teston, G., Lacey, D., McIlroy, R., Shen, J., Ramé, A., Szlam, A., Ranzato, M., & Barham, P. (2025). Streaming DiLoCo with overlapping communication: Towards a distributed free lunch. arXiv. https://arxiv.org/abs/2501.18512
  21. Douillard, A., Rush, K., Donchev, Y., Charles, Z., Fallen, N., Dubey, A., Gog, I., Dean, J., Woodworth, B., Garrett, Z., Keating, N., Bishop, J., Prior, H., Yvinec, E., Szlam, A., Ranzato, M., & Dean, J. (2026). Decoupled DiLoCo for resilient distributed pre-training. arXiv. https://arxiv.org/abs/2604.21428
  22. Zhong, Y., Liu, S., Chen, J., Hu, J., Zhu, Y., Liu, X., Jin, X., & Zhang, H. (2024). DistServe: Disaggregating prefill and decoding for goodput-optimized large language model serving. In Proceedings of the 18th USENIX Symposium on Operating Systems Design and Implementation (pp. 193–210). USENIX Association. https://www.usenix.org/conference/osdi24/presentation/zhong-yinmin
  23. Patel, P., Choukse, E., Zhang, C., Shah, A., Goiri, Í., Maleki, S., & Bianchini, R. (2024). Splitwise: Efficient generative LLM inference using phase splitting. In Proceedings of the 51st Annual International Symposium on Computer Architecture. https://arxiv.org/abs/2311.18677
  24. Qin, R., Li, Z., He, W., Zhang, M., Wu, Y., Zheng, W., & Xu, X. (2024). Mooncake: A KVCache-centric disaggregated architecture for LLM serving. arXiv. https://arxiv.org/abs/2407.00079
  25. Barham, P., Chowdhery, A., Dean, J., Ghemawat, S., Hand, S., Hurt, D., Isard, M., Lim, H., Pang, R., Roy, S., Saeta, B., Schuh, P., Sepassi, R., Shafey, L. E., Thekkath, C. A., & Wu, Y. (2022). Pathways: Asynchronous distributed dataflow for ML. Proceedings of Machine Learning and Systems, 4, 430–449. https://arxiv.org/abs/2203.12533
  26. SpacePharma. (n.d.). Lab-on-a-chip technology for microgravity research. https://www.spacepharma.health/
  27. Amselem, S. (2019). Remote controlled autonomous microgravity lab platforms for drug research in space. Pharmaceutical Research, 36, Article 183. https://doi.org/10.1007/s11095-019-2716-5