Preferred Networks Launches On-Premise Enterprise LLM Platform

Preferred Networks (PFN), one of Japan's most prominent AI companies, has launched a new on-premise enterprise Large Language Model (LLM) platform designed to let businesses deploy powerful AI capabilities entirely within their own infrastructure. The move positions PFN as a direct competitor to cloud-based LLM providers like OpenAI, Google, and Microsoft in the rapidly growing enterprise AI market.

The platform targets organizations in regulated industries — finance, healthcare, manufacturing, and government — where data sovereignty, security, and compliance concerns make cloud-based AI solutions impractical or outright prohibited. Unlike API-based services such as OpenAI's ChatGPT Enterprise or Google's Vertex AI, PFN's solution operates entirely on a company's own servers, ensuring sensitive data never leaves the premises.

Key Takeaways at a Glance

• Fully on-premise deployment: No data leaves the enterprise's own infrastructure
• Built on PFN's PLaMo models: Leverages the company's proprietary large language models optimized for Japanese and English
• Enterprise-grade security: Designed for regulated industries including finance, healthcare, and defense
• Custom fine-tuning capabilities: Organizations can train models on their proprietary data
• Hardware flexibility: Runs on NVIDIA GPU clusters with support for various configurations
• Estimated pricing: Custom enterprise licensing, reportedly starting at several hundred thousand dollars annually

PFN Bets Big on Data Sovereignty Demand

Preferred Networks, founded in 2014 and headquartered in Tokyo, has long been recognized as one of Asia's leading AI research firms. The company has built deep partnerships with industrial giants like Toyota, FANUC, and NTT, and once operated one of the world's most energy-efficient supercomputers, the MN-3, which topped the Green500 list.

The new enterprise LLM platform builds on PFN's PLaMo (Preferred Language Model) series, which the company has been developing as a bilingual Japanese-English foundation model. PLaMo models have demonstrated competitive performance against international benchmarks, particularly in Japanese language tasks where they outperform many Western-developed alternatives.

PFN's timing is strategic. Enterprises worldwide are increasingly concerned about sending proprietary data to third-party cloud providers. A 2024 survey by Cisco found that 92% of organizations consider AI-related data privacy a more significant issue than general data privacy, while Gartner projects that by 2027, over 50% of enterprise AI workloads will run on-premise or in private clouds — up from less than 15% in 2023.

What the Platform Actually Delivers

The on-premise LLM solution is not simply a downloadable model — it is a full-stack platform that includes model serving infrastructure, fine-tuning pipelines, monitoring tools, and enterprise administration capabilities. This distinguishes it from open-source alternatives like Meta's Llama 3 or Mistral models, which require significant engineering effort to deploy at enterprise scale.

Key technical features include:

• Retrieval-Augmented Generation (RAG) integration for connecting LLMs to internal knowledge bases and document repositories
• Role-based access controls allowing different departments to manage their own AI workflows
• Model fine-tuning tools that let non-ML engineers customize model behavior using proprietary datasets
• Inference optimization for running large models efficiently on limited GPU hardware
• Audit logging and compliance reporting built for regulatory requirements in financial services and healthcare

The platform supports deployment on NVIDIA A100 and H100 GPU clusters, with PFN providing optimization layers that reportedly reduce inference costs by 30-40% compared to running standard open-source models on equivalent hardware. This cost efficiency is critical for enterprises evaluating the total cost of ownership against cloud-based alternatives.

How PFN Compares to Western Competitors

The enterprise on-premise LLM market is heating up globally. In the United States, IBM offers its watsonx platform with on-premise options, while Databricks acquired MosaicML to provide enterprise model training capabilities. Palantir has integrated LLM functionality into its government and enterprise platforms, and even Dell Technologies has partnered with multiple AI companies to sell turnkey on-premise AI infrastructure.

PFN's differentiators center on 3 core areas. First, its deep expertise in Japanese language processing gives it an edge in Japan's $4.7 trillion economy — the world's 4th largest. Many global enterprises operating in Japan struggle with LLMs that underperform in Japanese compared to English.

Second, PFN's industrial AI heritage means the platform is designed with manufacturing and robotics use cases in mind, not just document processing and chatbots. The company's existing relationships with factory automation leaders like FANUC provide credibility that pure-play software companies lack.

Third, PFN's approach contrasts sharply with the 'AI-as-a-service' model championed by OpenAI and Google. While those companies benefit from centralized infrastructure and economies of scale, they require customers to trust external providers with their most sensitive data. For defense contractors, central banks, and pharmaceutical companies, that trade-off is often unacceptable.

The Growing Enterprise Appetite for Private AI

PFN's launch reflects a broader industry trend that is accelerating throughout 2024 and into 2025. The initial excitement around cloud-based generative AI tools like ChatGPT is giving way to a more nuanced enterprise reality where data governance, regulatory compliance, and intellectual property protection are non-negotiable requirements.

Samsung famously banned employees from using ChatGPT after sensitive semiconductor data was accidentally uploaded to the platform. JPMorgan Chase, Goldman Sachs, and several major banks restricted employee use of external AI tools. The European Union's AI Act, which began phased implementation in 2024, imposes strict requirements on how AI systems handle personal data — requirements that are far easier to meet with on-premise deployments.

Japan's own regulatory environment further supports PFN's approach. The Japanese government has been actively promoting AI sovereignty — the idea that the nation should not be entirely dependent on American AI infrastructure. Japan's Ministry of Economy, Trade and Industry (METI) has allocated billions of yen in subsidies for domestic AI development, and PFN has been a primary beneficiary of these initiatives.

This trend extends beyond Japan. Governments in the EU, South Korea, and the Middle East are all investing in domestic AI capabilities, creating a growing market for on-premise solutions that can operate independently of U.S. cloud providers.

What This Means for Enterprise AI Buyers

For CIOs and CTOs evaluating enterprise AI strategies, PFN's platform launch adds another serious option to the consideration set. The decision between cloud-based and on-premise AI deployment increasingly depends on industry-specific factors rather than pure technical capability.

Organizations should consider several factors when evaluating on-premise LLM platforms:

• Total cost of ownership: On-premise solutions require upfront hardware investment but eliminate per-token API fees that can escalate rapidly at scale
• Data sensitivity: Industries handling classified, regulated, or proprietary data benefit most from on-premise deployment
• Customization needs: Companies requiring deep model customization on proprietary datasets gain more control with on-premise solutions
• IT infrastructure maturity: Organizations need GPU-capable infrastructure and ML engineering talent to operate on-premise platforms effectively
• Vendor lock-in risk: On-premise solutions based on open architectures may offer more flexibility than proprietary cloud APIs

The economics are shifting in favor of on-premise deployment for large-scale users. a]6z (Andreessen Horowitz) published analysis in 2024 showing that enterprises spending more than $500,000 annually on cloud AI inference often achieve lower costs by bringing workloads in-house — a threshold that many large enterprises already exceed.

Looking Ahead: PFN's Global Ambitions

Preferred Networks has signaled that while Japan remains its primary market, the on-premise LLM platform is designed for global deployment. The company has been expanding its English-language model capabilities and exploring partnerships with hardware vendors and system integrators in North America and Europe.

The competitive landscape will intensify throughout 2025. NVIDIA itself is pushing into enterprise AI software with its NIM (NVIDIA Inference Microservices) platform, while Oracle, SAP, and Snowflake are all embedding LLM capabilities into their enterprise software stacks. Open-source alternatives continue to improve rapidly, with Meta's Llama models and the DeepSeek series from China offering increasingly capable free options.

PFN's success will likely depend on execution in 3 key areas: demonstrating model quality that matches or exceeds cloud-based alternatives, building a partner ecosystem for deployment and support, and pricing competitively against the growing field of on-premise options. If the company can deliver on these fronts, it stands to capture a meaningful share of what McKinsey estimates could be a $100 billion enterprise AI market by 2030.

For now, PFN's launch sends a clear signal: the future of enterprise AI is not exclusively in the cloud. As data sovereignty concerns grow and regulatory requirements tighten, on-premise LLM solutions are moving from niche to mainstream — and Preferred Networks intends to lead that transition.