SDSC Lands $10M NSF Grant to Build Expanse2, a Successor Tuned for AI Workloads

The National Science Foundation has handed the San Diego Supercomputer Center (SDSC) $10 million to deploy NSF Expanse2, a next-generation supercomputer that will replace the original Expanse system when it goes into production in 2027. If you've been tracking how NSF-funded HPC centers are retooling for the AI era, this is another data point: less emphasis on traditional batch-only supercomputing, more on hybrid workloads that mix simulation, data analytics, and large-scale AI training and inference.

Expanse, the system being retired, has been running since late 2020 and served "tens of thousands of researchers," according to PI Michael Norman, a computational astrophysicist at UC San Diego who also directed SDSC from 2010 to 2021. Expanse2 is built on those lessons but scaled up considerably, and it's arriving at a moment when NSF is trying to widen the on-ramp to advanced computing rather than just add more FLOPS for existing power users.

Here's what's actually changing under the hood:

• Compute: Expanse2 will use the latest generation AMD processors and accelerators. SDSC hasn't published exact chip specs yet, but the framing (AMD across CPU and GPU) tracks with where a lot of NSF-funded systems have landed lately.
• Networking: An "advanced network" designed specifically for data-intensive workloads, AI/ML training, and modeling and simulation, so the interconnect isn't an afterthought bolted onto a CPU cluster.
• Storage: This is the part worth paying attention to. Expanse2 pairs a high-performance all-NVMe file system optimized for AI/ML pipelines with large-capacity storage for durability and system availability. That's a two-tier design: fast flash for active training and inference jobs, bulk storage for everything else. Anyone who's babysat a GPU cluster starved for I/O will recognize why this matters, undersized storage is a classic bottleneck that leaves expensive accelerators idle waiting on data.

None of this is exotic by 2026 standards, but it's a meaningful upgrade path for a system that's spent six years serving what SDSC calls the "long tail of science," meaning researchers and institutions who aren't at the handful of mega-labs with their own dedicated clusters.

Who Actually Gets to Use It

The more interesting story here isn't the silicon, it's the access model. Expanse2 will plug into the NSF-led National AI Research Resource (NAIRR), the pilot program that spent over two years testing how to give academic researchers access to compute and AI tools without requiring them to already work at a place with a supercomputer. NAIRR just moved from pilot to full operations mode earlier this month, so Expanse2 is landing right as that program needs more capacity to point people at.

SDSC Director Frank Würthwein, who's also a physics professor at UC San Diego, put it plainly: the combination of newer GPU resources and the all-NVMe file system is meant to let the NAIRR community "effectively carry out application of AI both for training and inference." That's a specific technical claim worth unpacking. Training and inference have different I/O and latency profiles, training tends to be throughput-heavy with large sequential reads, inference can be latency-sensitive and bursty. Building a storage architecture that handles both well isn't trivial, and it signals SDSC expects a real mix of workload types rather than one dominant use case.

The system is also explicitly targeting researchers who "have not traditionally incorporated HPC into their workflows," per the announcement, meaning AI researchers coming from a pure cloud or single-GPU-workstation background who've never had to think about job schedulers, allocation policies, or MPI. That's a real onboarding problem. HPC centers have historically optimized for computational scientists who already know how to write a batch script; a growing chunk of the AI research community doesn't, and doesn't necessarily want to learn the hard way.

To smooth that over, SDSC says Expanse2 will offer user-friendly allocation and scheduling policies, Jupyter notebook support for rapid iteration, and a stack of AI/ML frameworks and libraries pre-integrated rather than something users have to build themselves. There's also a dedicated user support team with backgrounds spanning computational science, AI/ML, data-intensive computing, and large-scale systems operations, plus education, outreach, and training programs aimed at universities, community colleges, and high schools. SDSC is framing this partly as workforce development: training the next generation of people who can actually operate this stuff, not just consume it.

The project has a formal bench of co-PIs and senior personnel with NAIRR experience, including Amitava Majumdar, Ilkay Altintas, Subhashini Sivagnanam, Mahidhar Tatineni, Christopher Irving, and Mary Thomas. That's a lot of institutional memory from running Expanse getting folded directly into how Expanse2 gets built and operated.

Katie Antypas, Senior Science Advisor within NSF's Directorate for Computer and Information Science and Engineering, framed the award as part of a broader push to keep US research infrastructure competitive globally. That's not surprising rhetoric for an NSF award announcement, but it lines up with the current federal emphasis on expanding access to advanced computing and accelerating AI-enabled science, which is the same policy backdrop driving NAIRR's expansion.

Key Takeaways

Expanse2 is a $10 million bet that the next wave of HPC demand comes less from traditional simulation-heavy fields and more from AI researchers who need serious compute but don't want to become sysadmins to get it. The AMD-based compute, the dual-tier NVMe storage architecture, and the NAIRR integration all point the same direction: lower the barrier to entry while still delivering real throughput for training and inference at scale. The system enters production in 2027, funded under NSF award number 2614012, and its real test won't be benchmark numbers so much as whether it actually pulls in the "long tail" of AI researchers SDSC is explicitly trying to reach.