AI & INDUSTRIAL STRATEGY
What virtual-cell research and NVIDIA’s platform strategy reveal about national AI policy, venture capital, and industrial advantage
Technology / Industrial Policy / Venture Capital / Intellectual Property
I recently watched a video on the Korean YouTube channel SOD about advances in biological AI, particularly virtual-cell research, and NVIDIA’s expanding role in the field. The video’s dramatic headline suggested a breakthrough that had astonished even Jensen Huang. The science was fascinating. What stayed with me, though, was a business question: When these technologies mature, who will capture the value?
One part of the story involved a relatively small research team that achieved an outstanding result by combining deep learning with statistical methods rather than simply throwing more computing power at the problem. Another involved NVIDIA supporting scientific research while extending the reach of its own computing platform. Together, they got me thinking about where South Korea should direct its AI investment—and where Korean businesses have a credible chance to build enduring advantages.
Should we continue chasing the frontier in general-purpose large language models (LLMs), where major technology companies already command enormous resources? Or should we put more effort into the unsolved problems of biology, robotics, manufacturing, and materials science? To explore those questions, I went beyond the video and reviewed the competition organizers’ reports, the relevant research papers, and NVIDIA’s announcements. The discussion that follows is my assessment of what they mean for industrial policy, venture investing, and intellectual property.
1. The general-purpose LLM race and a lesson from operating systems
When governments talk about AI competitiveness, homegrown foundation models are often at the center of the conversation. Model size, access to GPUs, and positions on benchmark leaderboards become shorthand for technological strength. South Korea has similarly supported the development of domestic AI foundation models, in part to reduce dependence on foreign providers and maintain greater control over a technology with strategic importance.
That is a legitimate concern. AI is increasingly embedded in business operations, public services, and national security. Relying entirely on outside suppliers for critical capabilities introduces real risks. Yet the capabilities a nation needs for strategic resilience are not necessarily the same investments that will produce attractive commercial returns for startups.
I am reminded of the history of operating systems. Many companies—and some governments—invested in alternatives to the dominant platforms. Some were technically sophisticated. Still, Windows, Android, and iOS came to dominate much of the personal computer and smartphone markets. The decisive factor was not a simple ranking of technical features. It was the ecosystem: developers wrote applications where users were, and users gravitated toward platforms with the applications they needed.
General-purpose AI is not identical to the operating-system business, but the analogy is worth considering. Leading AI companies are building much more than models. They have development tools, cloud services, distribution channels, enterprise relationships, and users. A newcomer might produce a model with comparable benchmark performance and still face a formidable challenge in reaching customers and developers.
There are important differences. Open-weight models are improving quickly, multiple models can often be used side by side, and switching models may be easier than changing operating systems. Advances in training efficiency can also help challengers. None of this makes domestic foundation-model research pointless. It does, however, make the allocation of scarce resources a question worth asking.
Public investment in basic AI capabilities can serve national security, technological sovereignty, and supply-chain resilience. Venture investment, by contrast, must also examine customers' willingness to pay, the economics of delivering a service, and whether a competitive advantage can last. Those are distinct objectives, and they deserve distinct evaluation criteria.
Developing another world-class general-purpose model is one ambition. Building companies that can use AI to lead an industry—and earn sustainable returns—is another. South Korea needs to consider both without confusing one for the other.
2. The unexpected lesson of the Virtual Cell Challenge
The 2025 Virtual Cell Challenge, organized by the nonprofit Arc Institute and sponsored by NVIDIA, 10x Genomics, and Ultima Genomics, offers a useful case study. More than 5,000 people from 114 countries registered, and over 300 teams made final submissions.
The task was to predict how a cell’s gene-expression profile would change when the activity of a particular gene was suppressed. To make the comparison possible, the organizers assembled single-cell RNA-sequencing profiles from roughly 300,000 human embryonic stem cells, covering 300 CRISPR-interference perturbations.
The problem sounds like a standard machine-learning prediction exercise. In practice, biology adds complications that are easy to underestimate. Genes, proteins, signaling pathways, and metabolic processes operate in interconnected systems. Suppress one gene, and the effects can propagate through multiple pathways; change the cell type or its condition, and the response may differ. Experimental observations also contain biological variation and measurement noise. A model that recognizes correlations in one dataset may not accurately predict what happens after an intervention in a different setting.
When the results were announced in December 2025, first place went to Team BM_xTVC at BioMap Research. Its model, xTrimoSCPerturb, did not win by relying exclusively on a larger neural network. The team improved the scFoundation architecture, incorporated protein-model embeddings and curated public perturbation data, and explicitly included statistical features such as average gene expression and the frequency of differentially expressed genes.
That last decision was revealing. The researchers found that neural-network predictions did not consistently outperform simple statistical baselines. Rather than dismiss those baselines, they used the information the baselines captured and combined it with their learning architecture. The third-place model, TransPert, also made extensive use of statistical methods. The competition illustrated the value of hybrid modeling: deep learning and established statistical tools can complement each other.
It would be a mistake, however, to pronounce deep learning—or scaling laws—defeated. A paper published in Nature Methods in August 2025 reported that several deep-learning models for genetic perturbation prediction failed to outperform simple baselines in the tests examined. A later study published in Nature Biotechnology in October 2026 reached a more qualified conclusion. Looking across 14 datasets and 18 evaluation metrics, its authors found that some commonly used metrics were poorly calibrated for detecting meaningful performance. Under better-calibrated measures, deep-learning models could outperform uninformative baselines.
The competition itself exposed similar issues. Maximizing a score on a particular metric was not always the same as faithfully predicting biology. For the 2026 challenge, Arc Institute broadened its evaluation and introduced a zero-shot task: models must predict perturbation responses in cellular contexts where those responses have not been observed during training.
I do not take these developments as proof that one family of models has won. They tell us something more useful: In some scientific and industrial problems, understanding the data, designing the experiment, and choosing a meaningful test matter as much as the size of the model. More computing power remains valuable. It simply is not the whole answer.
For companies without the resources of a major hyperscaler, that distinction matters. Specialized knowledge, careful statistics, high-quality experimental data, and sound validation can be sources of competitive strength.
3. The real opportunity lies in connecting prediction with experiment
A predictive model becomes commercially valuable when it helps someone make a better decision. Consider a startup developing AI tools to identify cancer-drug candidates. It might screen thousands of compounds using gene-expression data. But a high prediction score is not evidence that a compound will work in an actual experiment, much less in a clinical trial.
Researchers still need to test selected compounds in cells, measure the responses, and compare the results with the predictions. When reality diverges from the model, the discrepancy becomes new information. Feeding those findings into the next round of modeling creates what is often called a closed-loop or lab-in-the-loop system.
Active learning can make this process more useful. Instead of selecting only compounds expected to perform best, a model can propose experiments that are especially informative—ones likely to reduce uncertainty about how the system works. Each cycle can improve both the model and the quality of the organization’s proprietary knowledge.
Over time, the defensible asset may be much broader than the original algorithm. It can include rights to use experimental data, methods for designing studies, validation protocols, software integrated into customer workflows, and accumulated expertise about a particular disease or class of molecules. A competitor may be able to reproduce similar code without being able to reproduce years of experimental learning or customer integration.
Owning a laboratory is not essential. A startup may build an effective feedback loop through partnerships with universities, hospitals, contract research organizations, and testing facilities. What matters is reliable access to the relevant data and sufficient control over the process that turns experiments into better predictions.
The same logic applies beyond biotechnology. Robotics systems need feedback from the physical world, not just simulations. Manufacturing AI must be tested against actual quality outcomes, yields, and equipment performance. Materials models need to be checked against laboratory measurements. South Korean firms have spent decades acquiring process knowledge in semiconductors, automobiles, batteries, precision machinery, and biotechnology. Those assets cannot simply be scraped from the public internet.
Of course, possessing data is not enough. Its quality, legal usability, and relevance to a real customer problem determine whether it creates value. But an established industrial knowledge base gives Korean companies a plausible way to compete without duplicating every general-purpose model developed abroad.
4. NVIDIA’s strategy extends well beyond selling GPUs
NVIDIA’s place in this story is just as instructive. Its advantage cannot be explained by GPU performance alone. Over many years, the company has built an ecosystem around CUDA, including development tools, libraries, and optimized software. As developers build and maintain systems in that environment, moving to an alternative can require rewriting code, retesting performance, and retraining teams. That switching cost strengthens the platform.
NVIDIA is applying a related approach in biological AI. Its BioNeMo platform supports biological data processing, model training, optimization, inference, and deployment. In January 2026, NVIDIA announced an AI drug-discovery co-innovation lab with Eli Lilly and a collaboration with Thermo Fisher Scientific on intelligent and increasingly autonomous laboratory infrastructure.
The ambition is broader than supplying chips to researchers. NVIDIA wants its technology involved as laboratory data are generated, models are trained, and predictions are validated against new experiments. That puts it closer to the operating infrastructure of scientific discovery.
There was a smaller illustration in the 2025 Virtual Cell Challenge. The prize packages included cash and NVIDIA DGX Cloud credits. For researchers, that meant access to powerful computing resources. For NVIDIA, it offered an opportunity to introduce a promising group of users to its cloud environment.
A cloud credit does not, by itself, prove lock-in. Whether researchers stay depends on software portability, compatibility, the ability to move data, and the full cost of running their work elsewhere. Still, the commercial logic is plain. Once a team has developed code, models, and data pipelines in a particular environment, continuing there may be easier than rebuilding them on another platform. Growth in the research program can translate into recurring demand for computing and software.
These economics should not be confused with the token billing common in language-model services. Biological AI workloads can generate revenue through GPU compute time, cloud usage, software licenses, and inference services. The pricing units differ, but the commercial objective is similar: become part of a workflow customers use repeatedly.
What interests me most is that NVIDIA does not need to bet on one virtual-cell model becoming the winner. If many competing research groups rely on its infrastructure, the company may benefit from growth across the sector. It is trying to occupy a position from which the success of an entire category supports its own business.
That position is not guaranteed. Open-source software, alternative hardware, rival cloud providers, and interoperable standards create competitive pressure. But the strategic lesson is worth absorbing. Making excellent technology matters. Giving customers durable reasons to keep using it may be just as valuable.
5. What should Korea’s public venture funds look for?
The next question is how to translate these lessons into investment policy. The Korea Fund of Funds, managed through Korea Venture Investment Corporation, generally invests in venture funds rather than choosing every startup directly. Government capital sets broad policy priorities while specialized managers make company-level investment decisions.
That structure makes sense. Policymakers should not tell every AI startup to build the same kind of model. Investments intended to maintain domestic foundation-model capabilities should be assessed differently from investments intended to help specialized companies grow.
If I were evaluating an industrial AI startup, I would spend less time on parameter counts and leaderboard positions and more time on three questions.
First, does the company have access to information competitors cannot easily obtain—and the legal right to use it? A service built entirely from public datasets and open models may be copied relatively quickly. Reliable access to proprietary experimental or manufacturing data can be much harder to replicate. But “proprietary data” is not a magic phrase. Investors must ask whether those data actually improve performance, whether the supply will continue, and whether the company has enforceable rights to use them.
Second, can the company demonstrate that its technology solves a real problem better than the alternatives? The complexity of a neural network is not a business result. Accuracy against a credible baseline, performance in unfamiliar conditions, reproducibility, and the cost of obtaining a useful answer may matter more. If a simpler model delivers the same practical results at a fraction of the computing expense, it may be the better commercial product.
Third, why would customers keep using the product? A company that completes one-off analysis projects has different economics from one whose software becomes part of a customer’s daily research or production process. A virtual-cell tool might begin with predictions, then expand into experiment tracking, compound comparison, and study design. As customers accumulate validated results and tailored workflows, the product can become increasingly useful—and more costly to replace.
That is often called platform lock-in. I would be cautious about treating lock-in as an end in itself. Deliberately trapping data or making migration unreasonably difficult can undermine trust and discourage adoption. A better business is one customers continue to choose because it consistently creates more value than the alternatives.
Not every successful AI company needs to operate a giant platform. A specialized component supplier or software vendor may build an excellent business. The point is to identify the source of economic value and determine whether competitors can easily substitute for it.
6. Public support should extend to testing and validation
Another implication concerns the gap between a promising model and a product customers can trust. Predicting a cellular response does not establish clinical efficacy. A robot’s success in simulation does not prove it will behave safely on a factory floor. The difficult, expensive work of validation often begins after the model has performed well in the lab.
Young companies may not have the resources to carry that risk alone. Governments and public research institutions can help by developing shared testing facilities, trustworthy benchmarks, and methods for independently evaluating performance. Biotech firms need experimental replication; robotics firms need real-world testing under varied conditions; manufacturers and materials developers need common measurement and verification infrastructure.
That support must respect trade secrets, personal data, and customers’ confidential information. A productive innovation ecosystem does not require every dataset to be made public. It requires credible common infrastructure that firms can use without surrendering the assets on which their businesses depend.
A useful division of labor would assign long-horizon research risks to public R&D programs, shared evaluation and testing infrastructure to public institutions, and commercialization and growth decisions to professional investors working alongside private capital. Funding milestones could move from independent technical performance to reproducible results, measurable customer value, paying customers, and recurring revenue.
The goal is not merely to announce that a model was built. It is to find out whether public money helped create companies whose products work in the market.
7. Intellectual property strategy has to look beyond the model
From an intellectual property perspective, these business models raise questions that are easy to miss if the conversation stops at algorithms. Novel model architectures, training techniques, data-processing methods, and AI integrated with particular devices or industrial processes may be patentable, subject to the applicable legal requirements. But patents alone rarely capture the whole value of an industrial AI platform.
Data-use rights, accumulated experimental know-how, validation procedures, customer contracts, and integration with existing systems can be equally important to its commercial position.
Imagine a biotech AI startup working with a pharmaceutical company. Allocating patent ownership for inventions arising from the collaboration is only the beginning. Who may use the experimental data? May the startup train models for other customers on the results? Who controls improvements to the model? What happens to the data and derived systems when the agreement ends?
For a platform that continuously processes customers’ research or production data, privacy, trade-secret protection, cybersecurity, and data licensing all become part of business strategy. The phrase “data ownership” may hide more than it explains: legal protection depends on the nature of the data, how they were obtained, the applicable jurisdiction, and contractual terms.
Investors should ask not just how many patents a company owns, but whether those patents cover commercially important features, how easily competitors can design around them, and whether the company has freedom to operate without infringing third-party rights. Copyright, trade secrets, patents, and contract rights need to work together.
A team can be exceptionally good at building a model and still lack the rights needed to operate the business at scale. Good IP strategy is about closing that gap.
8. Where should South Korea build its AI advantage?
There are sound reasons for South Korea to maintain domestic foundation-model capabilities and secure reliable computing infrastructure. Technological sovereignty and supply-chain resilience matter. But if we measure AI competitiveness only by model benchmarks or the number of GPUs installed, we will miss much of the industrial opportunity.
South Korea already has deep expertise in semiconductors, automobiles, batteries, robotics, biotechnology, precision machinery, and materials. These industries hold process data, experimental experience, and engineering know-how that are not readily available on the open internet. Companies can use global AI models as building blocks while developing proprietary capabilities around these hard-won assets.
Participation in NVIDIA’s ecosystem need not be treated as a threat in itself. Using world-class infrastructure can be sensible, provided a company retains control of the data, validation capabilities, and customer relationships that make its business distinctive. Portability, interoperability, supplier dependence, and long-term cost should nevertheless be managed deliberately.
Nor should specialized, or vertical, AI be romanticized as an easy blue ocean. High-quality data are expensive and difficult to obtain. Experiments take time. Enterprise customers may be slow to adopt new tools. Established industrial firms may decide to develop their own AI systems, shrinking the market available to independent startups. A specialized label is not evidence of a viable business.
What matters is whether the company can solve a consequential problem, prove the results, and convert that performance into durable customer relationships and earnings.
The Virtual Cell Challenge showed that difficult scientific problems remain unsolved and that progress may come from thoughtful combinations of learning, statistics, and domain knowledge. NVIDIA’s strategy showed why the infrastructure serving many competing innovators can be commercially valuable. Looking at those two developments together, I believe Korean policymakers should more clearly distinguish scientific achievement from industrial success.
Building frontier AI models matters. It does not automatically create enduring businesses. Companies also need to bring those models into laboratories, factories, and engineering workflows, learn from the resulting data, and develop products customers have reason to keep using. Public investment should be evaluated in part by whether it helps such companies emerge and grow.
The question is no longer simply what AI model we should build. It is what industries we can help lead by applying AI well.
Will South Korea be content to develop AI technology, or will it use that technology to shape the industries and ecosystems of the future?
That is the question I believe our industrial and investment policies need to answer.
Sources and further reading
- SOD, Korean-language video on virtual cells and NVIDIA, YouTube — the inspiration for this essay.
- Arc Institute, Virtual Cell Challenge 2025 Wrap-Up: Winners and Reflections (December 6, 2025) — competition results and the BM_xTVC approach.
- Arc Institute, The 2026 Virtual Cell Challenge (August 20, 2026) — zero-shot evaluation.
- Ahlmann-Eltze, Huber & Anders, Nature Methods (2025) — perturbation models versus simple baselines.
- Miller et al., Nature Biotechnology (October 1, 2026) — model performance under calibrated metrics.
- NVIDIA, BioNeMo Platform Adopted by Life Sciences Leaders (January 12, 2026) — NVIDIA’s biotechnology platform and partnerships.
- Korean Ministry of Science and ICT, Independent AI Foundation Model Project: First-Stage Evaluation (January 15, 2026; in Korean) — South Korea’s foundation-model policy.
- Korea Venture Investment Corporation, Korea Fund of Funds Overview (in Korean) — public venture-investment structure.
