AI Computing as a Driver of Scientific Innovation: A Practical Path for New R&D Institutions

1. Background and Positioning: From Research Platform to Strategic Technology Capability

In September 2017, a new R&D institution was established in the core area of Hangzhou’s Western Science and Technology Innovation Corridor. The institution was led by a provincial government, supported by universities, and joined by industry. In 2021, it was incorporated into the national laboratory system.

This timeline reflects a broader trend: research infrastructure for intelligent computing is evolving from single-discipline experimental facilities into cross-disciplinary and cross-institutional public technology platforms.

Compared with traditional laboratories, this type of R&D institution has three defining characteristics:

1. Demand-driven research: It directly addresses three categories of needs—national strategic priorities, emerging changes in scientific and technological innovation, and innovation in strategic industries.
2. System-level innovation: Rather than pursuing isolated breakthroughs, it builds an integrated technology stack around computing power, models, and sensing.
3. Open collaboration: It delivers capabilities through public computing platforms, standards initiatives, and participation in international organizations.

From an engineering perspective, the central challenge can be summarized as follows:

How can computing capabilities be converted into scientific discoveries and industrial value in a stable, efficient, and reproducible way?

This is not only an infrastructure challenge. It is also a software engineering and organizational collaboration challenge.

2. The Technology Foundation: Coordinating Hardware and Software Around Three Frontiers

The institution has identified three frontier areas: the limits of computing power, intelligence, and sensing. It uses breakthroughs in critical hardware and software technologies to establish a foundation for intelligent computing. These areas are not independent; they form a progressive chain.

• The limits of computing power concern the scale and efficiency boundaries of computing systems. Relevant engineering problems include high-performance computing clusters, heterogeneous resource scheduling, and energy-efficiency optimization.
• The limits of intelligence concern the capabilities of models, including training stability, inference efficiency, domain adaptation, and deployability of large models.
• The limits of sensing concern data acquisition and the connection between computing systems and the physical world. Typical applications include remote sensing, astronomy, geoscience, and other Earth-observation and space-science workloads.

The corresponding system goals are high performance, high efficiency, and high energy efficiency. In engineering terms, these refer respectively to throughput, resource utilization, and output per unit of energy. Together, they are essential indicators of whether a computing system can operate sustainably.

Achieving these goals requires hardware-software co-design. At the hardware level, systems must be optimized for representative workloads. At the software level, compilers, schedulers, runtimes, and model frameworks must work together closely.

A representative layered architecture is shown below:

Application layer   Scientific foundation models / Domain models / Data products
Model layer         Training frameworks / Fine-tuning toolchains / Inference engines /
                    Model deployment and validation
Platform layer      Compute scheduling / Data governance / Experiment management /
                    Monitoring and alerting
Resource layer      Compute clusters / Storage services / High-speed networks /
                    In-orbit computing nodes

This layered structure clarifies responsibilities:

• The resource layer ensures availability.
• The platform layer provides scheduling and observability.
• The model layer supports reproducible development and evaluation.
• The application layer ensures that capabilities can be delivered to users.

3. Platform Capabilities: Public Computing and Standards First

Platformization is critical to the practical adoption of intelligent computing. In July 2023, the institution became one of the first national open innovation platforms for next-generation artificial intelligence public computing in China. In April 2024, it completed the only national technology standards innovation base in the computing field focused on intelligent computing.

Together, these capabilities address a common engineering principle:

Computing resources must be usable, and standards must come first.

The value of a public computing platform does not come simply from accumulating hardware. It depends on three capabilities:

1. Unified access: Standardized interfaces can hide differences among heterogeneous resources, allowing researchers to submit jobs through a consistent workflow.
2. Workload orchestration: Training, inference, and data-processing workloads can be scheduled together to improve overall resource utilization.
3. Observability: End-to-end monitoring of job queues, resource consumption, and fault recovery reduces the risk of interruptions during long-running training jobs.

The standards innovation base serves a different but complementary purpose. It turns interface specifications, evaluation methods, and data formats developed through engineering practice into reusable standards. This helps organizations avoid repeatedly solving the same integration problems.

For a scientific computing platform that serves multiple disciplines, standards are the final safeguard for reducing collaboration costs.

4. Model Engineering: From General Capabilities to Scientific Applications

The institution’s work on models illustrates a common pattern in scientific AI. On one hand, it has advanced a geoscience large language model and completed national algorithm filing. On the other, in October 2025, it jointly released what it described as the world’s first deployable general-purpose genomic foundation model at the tens-of-billions-of-parameters scale.

Together, these examples highlight an important engineering principle:

For scientific models, value depends more on deployability than on parameter count.

A genomic foundation model presents several major engineering challenges:

• Heterogeneous data: Genomic sequences are long and structurally complex, requiring specialized tokenization and positional encoding strategies.
• Computing constraints: Training and running a model with tens of billions of parameters on limited resources requires parallelization strategies and mixed-precision optimization.
• Deployment requirements: The model must run with acceptable latency in real research environments, which involves quantization, distillation, and service-oriented packaging.

A typical engineering pipeline can be summarized as follows:

# Example: engineering workflow for a domain model (pseudocode)
dataset = load_domain_corpus()          # Data cleaning and deduplication
tokenizer = build_domain_tokenizer()    # Domain-specific tokenizer
model = init_base_model(params="10B")   # Base model initialization
model = pretrain(model, dataset)        # Distributed pretraining
model = finetune(model, task_data)      # Task-specific fine-tuning
model = quantize(model, bits=8)         # Quantization and compression
serve(model, endpoint="inference")      # Service deployment
evaluate(model, metrics=["acc", "latency"])

Deployability also requires complete evaluation and version-management mechanisms. Verifying the correctness of a scientific model often requires domain experts. Experiment management platforms therefore need to associate data versions, model versions, and evaluation results so that the entire process can be traced and reproduced.

5. Space Computing and Constellation Networking: Extending Computing to the Edge

In May 2025, the institution initiated the “Three-Body Computing Constellation” and completed the formation of China’s first fully orbit-connected space computing constellation. Subsequent breakthroughs in inter-satellite networking increased the number of satellites in orbit, while dozens of in-orbit model deployment validations were completed.

The technical significance of this work is that it moves computing from ground-based data centers to edge nodes and tests the engineering feasibility of in-orbit inference.

Space computing imposes several special constraints on system design:

• Limited resources: Onboard computing capacity, power, and thermal conditions are far more constrained than those of ground-based data centers. Models must therefore be lightweight.
• Intermittent connectivity: Space-to-ground links are available only during specific communication windows. Job orchestration must support checkpointed execution and result caching.
• High reliability requirements: In-orbit nodes are difficult to maintain manually, so automated fault detection and recovery are essential.

The goal of in-orbit model deployment validation is not simply to run the largest possible model. It is to validate the combination of small models, efficient scheduling, and reliable communications.

This experience is also relevant to terrestrial edge computing. As computing moves closer to the network edge, lightweight software stacks and autonomous operation can be more important than peak performance.

6. Cross-Disciplinary Collaboration: Scientific Foundation Models and Open Cooperation

The institution has presented its geoscience large language model and astronomical science foundation model at multiple international conferences and exhibitions. Some of its work has also been selected for a major science program under UNESCO’s “Science Decade.”

Behind these activities is a reusable collaboration model:

1. Joint domain development: Domain scientists define the research questions and evaluation criteria, while engineering teams build the data pipelines and training infrastructure.
2. Platform reuse: The same computing scheduler and model toolchain serve multiple disciplines, reducing marginal development costs.
3. Open delivery: Capabilities are shared through public platforms, standards organizations, and international cooperation channels, creating a feedback loop for further improvement.

The key is to turn scientific cooperation into clearly defined engineering interfaces. Data delivery, model evaluation, and result reproduction all require explicit agreements. Without them, cross-institutional collaboration can quickly degrade into manual, case-by-case operations.

7. Talent and Organization: The Engineering Value of a Young Team

As of November 2024, the institution had more than 1,700 full-time employees, with people under the age of 35 accounting for more than 70 percent of the workforce.

A young team can offer clear advantages in a rapidly changing field such as intelligent computing. Its members may adapt quickly to new frameworks and hardware and may be more willing to adopt hands-on engineering practices. At the same time, this structure places higher demands on the organization.

Mature code standards, review processes, and knowledge-management systems are needed to prevent expertise from being lost as personnel change.

In practice, this often involves three measures:

• Standardized project templates and scaffolding
• Mandatory code reviews and automated testing
• Structured onboarding documentation for new team members

For research-oriented teams, treating reproducibility as part of the engineering culture provides more long-term value than simply maximizing development speed.

8. Conclusion: Practical Principles for Computing-Driven Innovation

This development path can be summarized through four transferable principles:

• Define the technology stack according to layered needs: Computing, models, and sensing should each have clear engineering objectives and evaluation metrics.
• Reduce collaboration costs through platformization: Public computing platforms and standards are prerequisites for large-scale collaboration.
• Measure model value by deployability: The success of a scientific model depends on whether it can run reliably in real environments.
• Use edge scenarios to test system resilience: Extreme environments such as space computing can drive the software stack toward greater lightweight efficiency and autonomy.

Computing-driven innovation is not merely a slogan. It is an integrated engineering system that spans resource scheduling, model serving, standards development, and cross-disciplinary collaboration.

For teams building similar research infrastructure, the layered architecture and practical principles described above offer a directly applicable reference.