On June 5, the startup Macrocosmos announced that it has successfully trained a 100‑billion‑parameter language model—named Orion‑100B—by linking Nvidia A100 GPUs located in disparate data centers around the world. The achievement challenges the prevailing assumption that such massive models must be built inside a single, multi‑billion‑dollar facility.
The core of the effort rests on the Bittensor network, which Macrocosmos leveraged to create a distributed training framework called IOTA. Rather than requiring each participant to host the entire model, IOTA partitions the network‑wide model into 16 pipeline‑parallel stages. Each stage runs on a separate machine, allowing the overall workload to be spread across many geographically separated GPUs.
Running a model of this size on a scattered infrastructure presented several technical obstacles. The team had to contend with heavy inter‑GPU traffic, the instability of nodes that could drop out, and the heterogeneity of hardware capabilities across participants.
To mitigate the bandwidth burden, Macrocosmos applied a compression scheme that reduced the data exchanged between stages from roughly 150 MB per step to just 2.2 MB, dramatically easing network strain. Despite the complexity, the researchers reported that the distributed system achieved more than 30 percent FLOP utilization for the model and reached about 65 percent of the efficiency seen in a comparable, centralized data‑center configuration. While the performance does not yet match the raw throughput of hyperscaler clusters, the results demonstrate that large‑scale AI training can be carried out without a monolithic super‑computer.
The implications of this work extend beyond a single technical milestone. By turning idle GPUs into valuable compute assets, a market could emerge that rewards owners who contribute unused processing power.
This could democratize access to the resources needed for training cutting‑edge AI, enabling a broader range of organizations and researchers to participate in developing massive models. Macrocosmos cautions that the current implementation is not meant to replace the existing hyperscale infrastructure for production‑grade workloads. Nonetheless, the proof‑of‑concept underscores a viable path toward more distributed AI training ecosystems.
As the approach matures, it may spur further innovations in model parallelism, compression, and fault‑tolerant networking. Looking ahead, the success of Orion‑100B suggests that the barrier to entry for building large language models could be lowered, potentially accelerating innovation across the field.
More participants with access to idle GPU resources might contribute to future breakthroughs, expanding the pool of talent and ideas working on AI development. Macrocosmos plans to refine its IOTA framework and explore larger models and more diverse hardware configurations in the coming months. Observers will be watching to see how the distributed training paradigm evolves and whether it can eventually rival the performance and cost‑effectiveness of traditional, centralized data‑center solutions.
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “Global AI Model Trained Across Geographically Dispersed GPUs”, “description”: “On June 5, the startup Macrocosmos announced that it has successfully trained a 100‑billion‑parameter language model—named Orion‑100B—by linking Nvidia A100 GPU”, “datePublished”: “2026-06-26T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Muhammad Asim”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}, “image”: “https://americansuntimes.com/wp-content/uploads/sites/19/2026/06/data-center-2-scaled-1.jpg”}
