European Supercomputing Proves Viable for Trillion-Parameter AI Training

2026-11-13 digital

Brussels, Friday 13 November 2026
European researchers have trained a landmark one-trillion-parameter model on shared public supercomputers, achieving a fourteenfold throughput boost that proves public infrastructure can rival private hyperscale tech giants.

Architectural Efficiency on JUPITER

The technical demonstration was spearheaded by the EUROPA consortium, uniting Domyn, Fraunhofer IAIS, and Fraunhofer IIS, within the framework of the EuroHPC Frontier AI Grand Challenge as reported on 11 November 2026 [1][4][6]. Conducted on the JUPITER exascale supercomputer hosted at Forschungszentrum Jülich, the initiative scaled pretraining runs for a 951.9-billion-parameter Mixture-of-Experts architecture featuring 32.8 billion active parameters per token across 74 layers [1][8]. The model incorporates a hidden dimension of 6,144, 384 routed experts with top-8 routing, an expert hidden dimension of 1,792, a vocabulary size of 256,000, and an extensive sequence length of 65,536 tokens [1]. Parallelism was orchestrated using a 384-way Fully Sharded Data Parallel (FSDP) configuration paired with 4-way expert parallelism per node, completely eschewing pipeline parallelism while handling large batches of approximately 25 million tokens per training step [1].

Engineering Breakthroughs and Memory Optimisation

Overcoming the communication bottlenecks and high-bandwidth memory (HBM) constraints typical of shared public high-performance computing clusters required substantial software-level innovation across PyTorch and TorchTitan [1][4]. The physical infrastructure utilised 384 NVIDIA GH200 Grace Hopper Superchips per data-parallel replica across 96 nodes, with scaling demonstrated up to 1,536 GPUs across four replicas [1][8]. Each node integrated four GH200 superchips, combining 96 GB of HBM3 with a 72-core Grace CPU and 120 GB of LPDDR5X memory [1]. Throughput was initially constrained at a modest 90 tokens per second per GPU, but deliberate memory and pipeline interventions boosted execution speeds to roughly 1,300 tokens per second per GPU [1][4][6], representing an increase of 1344.444% [1].

Overcoming Systemic Latency Constraints

A pivotal enhancement came from addressing CUDA caching allocator fragmentation via the native FSDP2 unshard_async_op flag [1][8]. In standard operations, FSDP utilised one compute stream alongside four side streams, causing the caching allocator to reserve isolated pools where side streams hoarded 27.3 GiB while only 6.4 GiB was actively in use [8]. Routing the all-gather setup onto the compute stream pool eliminated allocator retries across all 384 ranks, dropping peak reserved memory from 88.1 GiB down to 74.6 GiB, freeing approximately 14 GiB of HBM, and lifting throughput from 996 to 1,210 tokens per second per GPU [1][8]. Furthermore, offloading sequence boundary activations to the Grace CPU over the 900 GB/s NVLink-C2C bus improved sequence handling up to 65,536 tokens and increased per-GPU output from 370 to 508 tokens per second, a gain of 37.297% [1][8]. Simultaneously, running the Newton-Schulz orthogonalisation iterations for the Muon optimiser directly on GPU hardware curtailed step times from 41.7 seconds to 2.2 seconds, elevating throughput by 35% from 441 to 597 tokens per second [1][8].

Sovereign AI and Policy Headwinds

The success of open public architectures lands amidst broader strategic and regulatory reorganisations aimed at establishing digital sovereignty across the European Union [2][3]. Sovereign AI—defined by organisations such as McKinsey and Gartner as the capacity to independently develop, deploy, and govern frontier systems within domestic jurisdictional borders—is becoming a central economic imperative [3]. On 30 July 2026, the EuroHPC Joint Undertaking opened a call to fund seven AI gigafactories, consisting of four medium and three large installations, supported by the €20 billion gigafactory envelope established under the €200 billion InvestAI initiative launched in February 2025 [2][3]. However, as confirmed by the Cloud Infrastructure Services Providers in Europe (CISPE) during the second week of November 2026, all initial submissions across both lots must be refiled following a procedural cancellation caused by technical issues, revising deadlines to 6 December 2026 for the medium tier and early February 2027 for the large tier [2].

Long-Term Economic and Industrial Implications

These regulatory adjustments and open-source breakthroughs arrive at a critical transition point for legacy enterprise digitisation, financial services, and sovereign data residency [3][7]. By showing that open-weight frameworks running across publicly funded supercomputers like JUPITER and Leonardo can achieve commercial throughput targets without reliance on hyperscaler proprietary software stacks, European researchers have laid a replicable foundation for regional public services and industry [1][3][7]. Gartner anticipates that by 2030, upwards of 75% of enterprises across Europe and the Middle East will repatriate their workloads from global public cloud providers back to regional and sovereign configurations [3]. As legislative proposals such as the Cloud and AI Development Act advance, the capability to scale near-trillion-parameter foundation models within local jurisdictions directly supports software supply-chain autonomy, lowering operational lock-in and democratising competitive frontier capabilities for non-hyperscale European cloud ecosystems [2][3].

Sources & Ecosystem Partners

Artificial intelligence High-performance computing