NVIDIA announced a record-breaking 104-qubit simulation of a fault-tolerant quantum computing system on JUPITER, Europe’s first exascale supercomputer. This marks the largest comprehensive simulation of its kind to date, modeling realistic quantum processor performance under noise conditions.

The research validates magic state cultivation, a leading approach to fault tolerance that requires 90% fewer physical qubits than previous methods. By simulating this process with high fidelity, NVIDIA provides actionable insights into error impacts, potentially shortening the development timeline for viable quantum hardware by years.

Scaling from development to exascale

The simulation was executed using a system of NVIDIA GH200 GPUs at the Jülich Supercomputing Centre. The project utilized the NVIDIA Accelerated Quantum Research Center (NVAQC) as a development platform, scaling tests from four GPUs to hundreds before deploying on JUPITER’s 8,192 GPUs. This staged approach reduced what would have been a weekslong simulation effort to days.

Earlier research from Jülich had simulated 50 qubits under ideal, noiseless conditions. The new work advances this by incorporating realistic hardware noise, allowing researchers to predict the performance of future designs before building physical hardware.

Technical infrastructure and software

The simulations leveraged NVIDIA cuQuantum software for optimized performance and scale, alongside the NVIDIA CUDA-Q platform’s fault-tolerant logical layer to express the magic-state circuit. This combination enabled the detailed modeling necessary for such large-scale fault-tolerance analysis.

What the numbers show

The divergence between the previous 50-qubit noiseless benchmark and the current 104-qubit noisy simulation highlights a critical shift in research priorities. While earlier models focused on theoretical limits, the current work prioritizes practical error correction. The validation of an approach requiring 90% fewer qubits suggests that resource efficiency, rather than raw qubit count alone, is becoming the primary bottleneck for scalable fault-tolerant systems. This data allows researchers to refine designs computationally, reducing reliance on costly hardware trial and error.

Nvidia (NASDAQ: NVDA) has transformed AI chips into collateral for Wall Street credit, enabling lenders to finance computing infrastructure using equipment, customer contracts, and expected revenue.

On August 10, 2026, Nvidia announced partnerships with six institutions to establish financing platforms designed to mobilize more than $500 billion in third-party capital for AI infrastructure over time. These preliminary agreements aim to support customers across Nvidia's ecosystem without relying solely on general corporate credit.

CoreWeave sets the benchmark

CoreWeave (CRWV) closed an $8.5 billion GPU-backed loan facility on March 31, 2026, marking the first investment-grade rated GPU-backed financing. The facility, known as DDTL 4.0, was structured as a delayed-draw term loan to finance infrastructure supporting contracted cloud services.

Moody's assigned the facility an A3 rating, while Morningstar DBRS assigned it an A (low) rating. The floating-rate tranche was priced at the secured overnight financing rate (SOFR) plus 2.25%, and the fixed-rate tranche carried costs of approximately 5.90%. The loan matures in March 2032.

Subsequent transactions highlight varying risk profiles:

Lambda expands institutional lending

Lambda secured a $926 million senior secured term loan B facility on August 12, 2026. This loan supported GPU infrastructure dedicated to an investment-grade customer and carried a Moody's Baa2 rating. Pricing tightened to SOFR plus 3.00% during syndication.

On October 1, Lambda announced a separate $1.008 billion delayed-draw term loan with investment-grade ratings. This facility covers three customer deployments involving two investment-grade counterparties across multiple data centers.

How GPU-backed credit works

In typical structures, borrowers raise debt to purchase or deploy AI servers. A special-purpose vehicle may own the equipment and pledge it as collateral, while customer agreements provide the revenue for repayment. Take-or-pay terms can make revenue more predictable but do not eliminate default risk.

Lenders evaluate customer creditworthiness, contract duration, utilization, operating costs, debt-service coverage, and equipment recovery value. Nvidia offers a residual value guarantee of up to 25%, which reduces losses if GPUs need remarketing but does not remove the need for lenders to underwrite customer and contract risk.

What the numbers show

The divergence in pricing between CoreWeave's initial facility and its subsequent loans reveals that collateral quality is not uniform. The March 2026 facility priced at SOFR + 2.25% with an A3 rating, whereas the May and August facilities priced at SOFR + 4.50% and SOFR + 5.50% respectively, with Ba2/BB+ ratings. This spread expansion of 225 bps to 325 bps indicates that lenders differentiate significantly based on specific contract structures and customer profiles, rather than applying a blanket premium to all GPU-backed debt.