Enterprise AI adoption faces fundamental tensions. Organizations need sophisticated language models but are hesitant about the infrastructure costs and energy consumption of frontier systems.
NTT Inc.’s recently announced tsusum 2, a lightweight large language model (LLM) that runs on a single GPU, demonstrates how enterprises are solving this constraint, with early deployments showing performance comparable to larger models at a fraction of the operating cost.
The business case is simple. Traditional large-scale language models require tens or hundreds of GPUs, creating power consumption and operating cost barriers that make AI adoption impractical for many organizations.
For companies operating in markets with limited power infrastructure or operating budgets, these requirements preclude AI as a viable option. The company’s press release explains the practical considerations that will drive adoption of the lightweight LLM with the introduction of Tokyo Online University.
The university operates an on-premises platform that stores student and staff data within the campus network. This is a common data sovereignty requirement for educational institutions and regulated industries.
After validating that tsusum 2 could handle complex context understanding and processing long documents at a production-ready level, the university implemented it to enhance course Q&A, support material development, and provide individualized student guidance.
Single GPU operation allows universities to avoid both the capital investment and ongoing power costs of GPU clusters. More importantly, on-premises deployments solve data privacy issues that prevent many educational institutions from using cloud-based AI services to handle sensitive student information.
Performance without scale: technical economics
NTT’s internal evaluation of financial system inquiry processing has shown that tsusum 2 matches or exceeds leading external models, despite having significantly lower infrastructure requirements. This performance-to-resource ratio determines the feasibility of AI implementation for companies where total cost of ownership drives decisions.
The model delivers Japanese language performance that NTT characterizes as “world-class results among models of comparable size,” and is particularly strong in business areas that prioritize knowledge, analysis, following instructions, and safety.
For companies operating primarily in the Japanese market, this language optimization reduces the need to deploy large multilingual models that require significantly more computational resources.
Developed based on customer demand, enhanced knowledge in finance, healthcare, and public sectors enables domain-specific deployments without extensive fine-tuning.
The model’s RAG (Search Augmentation Generation) and fine-tuning capabilities allow you to efficiently develop specialized applications for companies with proprietary knowledge bases or industry-specific terminology where generic models would perform poorly.
Data sovereignty and security as a business driver
Beyond cost considerations, data sovereignty will drive adoption of lightweight LLMs across regulated industries. Organizations that handle sensitive information face exposure when processing data through external AI services that are subject to foreign jurisdictions.
In fact, NTT positions Tsusum 2 as a “purely domestic model” developed from scratch in Japan, and can run on-premises or in a private cloud. This addresses concerns prevalent across the Asia-Pacific market regarding data residency, regulatory compliance, and information security.
Fujifilm Business Innovation’s partnership with NTT Docomo Business shows how companies can combine lightweight models with their existing data infrastructure. Fujifilm’s REiLI technology transforms unstructured corporate data, such as contracts, proposals, and mixed text and images, into structured information.
Integrating tsusum 2’s generation capabilities enables advanced document analysis without sending sensitive corporate information to external AI providers. This architectural approach, which combines a lightweight model with on-premises data processing, represents a practical enterprise AI strategy that balances functional requirements with security, compliance, and cost constraints.
Multimodal capabilities and enterprise workflows
Tsugami 2 includes built-in multimodal support for processing text, images, and audio within enterprise applications. This is important for business workflows where AI needs to handle multiple data types without introducing separate dedicated models.
Manufacturing quality control, customer service operations, and document processing workflows typically involve text, images, and sometimes voice input. A single model that handles all three reduces integration complexity compared to managing multiple specialized systems with different operational requirements.
Market conditions and implementation considerations
NTT’s lightweight approach stands in contrast to its hyperscaler strategy, which emphasizes large models with a wide range of capabilities. For companies with large AI budgets and sophisticated technical teams, Frontier models from OpenAI, Anthropic, and Google offer cutting-edge performance.
However, this approach excludes organizations that lack these resources in large parts of the enterprise market, particularly across the Asia-Pacific region, where the quality of infrastructure varies. Regional considerations are important.
Power reliability, internet connectivity, data center availability, and regulatory frameworks vary widely by market. A lightweight model that allows for on-premises deployments accommodates these changes better than approaches that require consistent cloud infrastructure access.
Organizations evaluating lightweight LLM implementation should consider several factors.
Domain specialization: While tsugami 2’s enhanced knowledge in finance, healthcare, and public sector addresses specific areas, organizations in other industries should evaluate whether the available domain knowledge meets their requirements.
Language considerations: Optimizing Japanese language processing is beneficial for operating in the Japanese market, but may not be suitable for multilingual companies that require consistent performance across languages.
Integration complexity: On-premises deployments require in-house technical capabilities for installation, maintenance, and updates. Organizations lacking these capabilities may find cloud-based alternatives easier to operate, albeit at a higher cost.
Performance tradeoffs: While tsusum 2 fits larger models in certain areas, frontier models may perform better in edge cases and new applications. Organizations must evaluate whether domain-specific performance is sufficient or whether broader functionality is worth the higher infrastructure cost.
What is the realistic path forward?
NTT’s introduction of tsusum 2 shows that advanced AI implementations don’t require hyperscale infrastructure, at least for organizations whose requirements match the capabilities of a lightweight model. Early enterprise deployments are showing real business value, including reduced operational costs, increased data sovereignty, and production-ready performance in specific domains.
As enterprises move forward with AI adoption, the tension between functional requirements and operational constraints increasingly demands efficient, specialized solutions rather than general-purpose systems that require large infrastructures.
For organizations evaluating AI adoption strategies, the question is not whether a lightweight model is “better” than a frontier system, but whether it is sufficient to address specific business requirements while addressing cost, security, and operational constraints that make alternative approaches impractical.
As the development of Tokyo Online University and Fujifilm Business Innovation shows, the answer is increasingly yes.
See also: How Levi Strauss is leveraging AI for its DTC-first business model
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expo in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events including Cyber Security Expo. Click here for more information.
AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.
