So far this week, TPC25 has brought together researchers who share one goal: to turn frontier-scale AI into a practical scientific tool. Tuesday’s afternoon talks captured the range of work now underway. Satoshi Matsuoka of RIKEN showed how progress in exascale systems can pave the way for models tuned to scientific problems. Steve Clark of Quantinuum showed how quantum circuits and transformers can learn from each other. Hyperion Research CEO Earl Joseph shared new survey data on how fast AI use is spreading across high performance computing sites. RIKEN’s Jens Domke then pulled the focus back to risk, warning that data leaks and insecure software could stall progress unless research groups build safer on-prem solutions. The afternoon plenary session offered a clear view of both the promise and the practical hurdles that lie ahead.
Why Exascale Alone Won’t Deliver Science-Ready AI
Satoshi Matsuoka, director of RIKEN’s Center for Computational Science, used his TPC25 talk to explain why today’s commercial foundation models are only a starting point for scientific work. Speaking on behalf of RIKEN’s AI for Science Team, he outlined several gaps that must be closed in data handling, model design, and workflow orchestration before large language and learning models can serve as reliable scientific instruments.
Matsuoka began with an update on RIKEN’s infrastructure. Fugaku remains one of the world’s fastest supercomputers with 60,000 CPU nodes, yet RIKEN is adding a GPU complex that will host roughly 1,500 Nvidia Blackwell accelerators and several hundred additional GPUs and TPUs. The center also runs three quantum systems: a 156-qubit IBM Q System 2, a 20-qubit Quantinuum machine scheduled to grow to 156 qubits, and a 64-qubit superconducting machine built in-house. He also described a future system he expects to one day reach zettascale, or 1021 op/s.
“We’re designing our next machine for 2029 with various vendors and other partners, and also working with entities like the DOE, such that in 2029 we will achieve what one might call zettascale computing, both in AI and conventional,” he said.
The Fugaku supercomputer. (Source: RIKEN)
With the hardware roadmap set, Matsuoka shifted his focus from peak performance to practical use. Raw speed, he warned, is only useful if the models running on that speed can understand scientific data and workflows. For example, commercial generative models target general users, so they often lack the features vital for physics, chemistry, and biology. Scientific data is a mix of text, equations, images, and sensor streams, often at terabyte scale. Current tokenizers handle web text but fail on domain symbols, units, and very long sequences. Scientific images like high-resolution microscopy scans often overwhelm standard vision transformers. To address this, Matsuoka says scientific AI systems need features like custom token vocabularies, sparse attention, and physics-aware decoders capable of context windows far beyond the million-token range.
To meet these needs, RIKEN is testing practical methods that let large models use less memory and better grasp scientific data. RIKEN teams are testing quad-tree tiling, space-filling curves, and other adaptive patching schemes that reduce sequence length for ultra-high-resolution images while preserving spatial coherence. Early results suggest exponential savings in compute without losing accuracy, yet they still demand new compiler support and memory layouts. For multimodal fusion, the group is exploring operators that combine neural surrogates with partial differential equation solvers.
These efficiency experiments reflect a bigger change in model strategy, with researchers moving away from huge models to a more varied mix tuned for specific tasks. Matsuoka now sees value in a spectrum of models. He highlighted trends toward mixture-of-experts architectures, fine-tuned domain models, and agents for AI inference. Reasoning during inference, rather than endless parameter growth, can lower compute costs while improving robustness.
Matsuoka invited researchers to contribute to these efforts through the TPC initiative and announced a dedicated TPC track at HPC Asia 2026 in Osaka this January. The goal is to showcase working systems that embed AI agents throughout experimental and simulation pipelines, moving closer to an environment where hypothesis generation, experiment planning, simulation, analysis, and paper drafting can all be accelerated by trustworthy, science-ready AI.
Enabling Scientific Discovery with Generative Quantum AI
What happens when two of the most transformative technologies, quantum computing and AI, begin to work together? For Steve Clark, Head of AI at Quantinuum, the result is a powerful new approach to scientific discovery. In his talk, he explained how his team is developing systems where quantum and AI not only support each other but also open doors to capabilities neither could reach alone.
To make this convergence real, Clark shared how his team at Quantinuum is putting the pieces together through what he called a strategy for generative quantum AI. He organized the talk around three main directions: using AI to optimize quantum computing, exploring how quantum systems might power new kinds of AI, and training machine learning models on data generated by quantum computers. Each area brings its own challenges and tools, but taken together, they reflect how his team is approaching this emerging space.
The first piece of the puzzle, Clark explained, is using AI to make quantum computing more practical. Even the best quantum hardware today still struggles with real-world issues like noise, control precision, and error rates.
That is where machine learning comes in. Clark and his team are using techniques like deep reinforcement learning to help with circuit compilation, which is how quantum algorithms are translated into hardware instructions. In one case, the AI learned how to reduce the number of costly quantum gates, making circuits more efficient to run.
They are also applying AI to problems like optimal control and error correction. These are areas where small improvements can make a big difference. As Clark put it, “We are applying it to hardware and getting really promising results.” These tools are already being tested on actual quantum machines, not just in simulation.
The second direction in Clark’s strategy flips the equation and looks at how quantum computing might unlock new types of AI. His team is exploring what happens when you redesign neural networks to work natively on quantum systems, not just simulate them. Instead of using standard attention mechanisms like in classical transformers, these models tap into quantum properties like superposition to process information differently.
As Clark explained, this is “a completely new type of model that doesn’t really have a classical analog,” and they’ve already tested early versions on real quantum hardware.
The third part of Clark’s approach brings quantum and AI into full collaboration. Instead of training models on classical data alone, his team is feeding AI systems data generated by quantum machines. This gives the models a way to learn patterns that classical systems cannot produce.
One example is the Generative Quantum Eigensolver, used to find a molecule’s ground state. A transformer model suggests quantum circuits, measures their energy, and adjusts to improve the result. In an early run, it found the ground state of hydrogen.
Clark noted this method could be useful beyond chemistry, in areas like materials science, optimization, or even fraud detection. This is where classical and quantum systems might truly learn from each other.
AI Goes Mainstream in HPC, but Data Quality and Budget Pressures Persist
At TPC25, Hyperion Research CEO Earl C. Joseph presented findings from Hyperion’s most recent surveys on artificial intelligence in technical computing, underscoring the rapid pace of change and the challenges ahead for organizations deploying AI in science and engineering.
Joseph noted that AI adoption across HPC has accelerated dramatically in recent years, forcing Hyperion to conduct new surveys every two to three months to keep pace. While roughly one-third of HPC sites reported some use of AI in 2020, that figure rose to over 90% by 2024. Today, AI for scientific workloads is no longer experimental—government labs, academic centers, and industry users are implementing AI for tasks ranging from simulation enhancement to data analysis at scale.
Earl Joseph, CEO of Hyperion Research, speaks at TPC25.
Cloud adoption is tracking closely with this surge. Joseph said organizations are increasingly turning to the cloud to offset challenges associated with procuring and maintaining leading-edge hardware, particularly GPUs. With each new generation of NVIDIA accelerators arriving faster—and at rising costs, especially when factoring in power and cooling—many organizations find it impractical to invest in systems they must keep for four to five years. Cloud adoption continues to grow rapidly as organizations seek access to current-generation hardware and flexibility in exploratory phases.
Despite the growth, Joseph emphasized several barriers that remain pervasive. The most frequently cited challenge was training data quality. Hyperion has documented numerous stalled AI projects due to poor or inconsistent datasets, and Joseph pointed to Mayo Clinic as an example of an organization mitigating this risk by exclusively using its own vetted data to train smaller, high-quality language models.
Other barriers include shortages of in-house AI expertise, insufficient scale of training data, and the overall complexity of integrating AI into HPC environments. Joseph predicted that these complexity issues will fuel the growth of a new market for domain-specific AI software and consulting services.
Hyperion’s studies show that 97% of surveyed organizations plan to expand AI use, scaling up operations, data volumes, and staffing despite the rising costs. He also said that budgets must grow significantly as AI infrastructure becomes more expensive, both for on-premises systems and cloud deployments. The top three challenges Hyperion is tracking for 2025 are complexity, cost, and hardware availability.
Joseph closed by previewing Hyperion’s upcoming study on return on investment (ROI) and “return on research.” Beyond profitability and break-even timelines, the research will examine the scientific impact of AI deployments across government, academia, and industry, measuring not just financial returns but also discoveries and innovations.
RIKEN is Building On-Prem Approach to Mitigating Gen AI Risk
In the rush to leverage AI for scientific use, risk mitigation is too often a distant second thought, cautioned Jens Domke, team leader of the Supercomputing Performance Research Team at the RIKEN Center for Computational Science.
“Everyone is rushing for the gold, but eventually we have to take care of the security aspects as well,” said Domke, who spent time examining five risk factors: 1) humans; 2) AI software; 3) software supply chain; 4) the models themselves; and 5) “lawyers and thieves.”
Jens Domke of RIKEN presents at TPC25.
“People have uploaded all sorts of things into cloud offerings of OpenAI and so forth. So there was an important case of Samsung leaking confidential data about their technology because people were just carelessly using OpenAI. OpenAI has data breaches as well. So they have lost data over the years. Multiple credentials have been leaked from OpenAI. Then we had DeepSeek coming out, and we have seen that there were some warnings about U.S. data being leaked to the Chinese authorities via DeepSeek,” he said.
His five slides detailing risk issues for each factor were a stark reminder that with new AI capabilities come new risks.
About AI software, he said, “Everyone is rushing to bring you the latest and greatest, but no one was really looking into the security aspects. For example, the Model Context Protocol was introduced quite recently, and then some of the security researchers looked at it and basically found that those MCP servers are just listening on all interfaces, on all ports, and just waiting for connectivity. So if you have that running on your laptop and you have some network you’re connected to, be mindful that things can happen. So, MCP is not secure, and other software infrastructure is not secure either.”
He reminded the audience that running models often requires 20 or 30 software packages, complicating and expanding risk. His talk was a good primer on AI risk.
He next dug into RIKEN’s decision to build its own AI management capability, intended to run on-premise, and which Domke argues is a viable solution for most research organizations.
“Basically, there are lots of issues that’s why we decided that we want to have our own on-prem solution, which is basically a secure, privatized, OpenAI alternative, where we can reproduce most of the stuff we are doing with OpenAI and other services, but we do it in house, without having the risk of data leakage, without having the risk of being hacked, without having the risk of, like, our data being exfiltrated to some capacity,” he said.
The two slides below provide an overview.
“We are planning to build our infrastructure essentially on open source, and we have, currently, the concept of multi-tiered security enclaves. One is the semi-open one, where people can broadly use it, similar to OpenAI’s offering, but in-house behind the firewall, which is very secure, at least to some extent,” said Domke. “Then we have even further security layers, where we have high-security instances for highly confidential operations, medical research, research related to all sorts of things within RIKEN. So different tiers that we can build and then set up access patterns depending on what people need, what people are comfortable with. The basic concept is don’t trust anything.
“We use containers for downloading models, which are different from the containers we are serving with. Everything is basically containerized and on private networks. All the containers are connected to a private network and basically just have a reverse proxy access to, like the web UI or to the API, which you’re using to connect with your favorite editor, for example, VS Code or some other. There are a few other nice little features we want to build into that, but that’s the overall concept,” said Domke.
He noted it is very similar to OpenAI. “You have your different chats, you can do translation. You can switch the models out depending on what’s installed by the administrator. So we can easily pull all the models from Hugging Face. We can deploy our own fine-tuned models. We have no restriction to that, similar to OpenAI, where you’re basically bound to OpenAI’s model, but we don’t have that. Everything kind of works quite nicely with open source. This is open web UI, if you’re interested. So, yeah, you can disable models easily. You can have sign-up schemes, group management and permission schemes are all built in.”
This is RIKEN’s long-term vision, said Domke, an on-prem secure environment in which data doesn’t have to be shipped to the cloud. “We are trying to build that right now with a prototype with Spring Eight.”
The Takeaway
Across hardware, algorithms, user adoption, and security, the message was consistent: scale alone is not enough. Scientific AI will depend on domain-tuned models, hybrid classical-quantum workflows, high-quality data, and careful risk controls. The next twelve months will test how quickly researchers can translate these ideas into shared tools and community standards. If the progress outlined at TPC25 continues, the field will move closer to AI systems that speed discovery without compromising trust.
To hear more examples and see where this research is headed next, don’t miss the full talks. The full session videos and transcripts will be available soon at tpc25.org. Stay tuned for our continuing coverage of TPC25!
Contributing to this article were Ali Azhar, Doug Eadline, Jaime Hampton, Drew Jolly, and John Russell