Teradata Q2 2025: Small Language Models Drive AI Innovation & Competitive Edge

Analysis of "Small language models" in Teradata Corporation's Q2 2025 Earnings Transcript

Teradata’s discussion of small language models (SLMs) in the Q2 2025 earnings transcript highlights a strategic focus on leveraging these models within their data analytics and cloud/on-premise platforms. The topic arises primarily in the Q&A section, where the company addresses customer adoption trends and the technical advantages of their infrastructure for running SLMs.

---

Key Themes and Context

1. Customer Adoption and Use Cases
   Teradata observes growing customer interest in deploying small language models, particularly for specialized, domain-specific tasks such as analyzing customer feedback or call center interactions. This indicates that SLMs are seen as practical tools for targeted natural language processing workloads rather than broad, general-purpose AI.

   "In fact, we are seeing customers deploy small language models like Hugging Face to get specialized results. So it could be in terms of looking at customer feedback or call center interactions."

2. Technical Differentiation and Performance
   A significant point made by Stephen McMillan (likely CEO or senior executive) is the performance advantage Teradata’s massively parallel CPU architecture offers for running SLMs. They cite a specific test with a U.S. bank where running a Hugging Face model on Teradata’s architecture outperformed running it on an NVIDIA DPU (Data Processing Unit). This suggests Teradata’s platform is well-optimized for these workloads, which could be a competitive differentiator.

   "Our massively parallel architecture that we deploy on-prem and in the cloud is really well suited to run in the small language models... we actually ran the Hugging Face model on our parallel CPU architecture and it actually ran better than running on NVIDIA DPU."

3. Strategic Positioning: Openness and Flexibility
   Teradata emphasizes an open and connected ecosystem approach, deliberately avoiding lock-in to any single language model provider. This flexibility allows customers to choose the size and brand of language models that best fit their needs, whether deployed behind firewalls or in the cloud.

   "We don't want to be locked into any one particular language model as some of our competitors have done. We want to be able to enable our customers to have a full deployment choice and to use the language models in the right place."

---

Business Implications

• Market Opportunity: The company views SLMs as a growing opportunity aligned with enterprise needs for specialized AI applications. This aligns with broader industry trends where smaller, fine-tuned models are preferred for privacy, cost, and latency reasons in enterprise settings.

• Competitive Advantage: Teradata’s ability to run SLMs efficiently on their existing parallel CPU infrastructure, both on-premises and in the cloud, positions them well against competitors who may rely more heavily on GPU or DPU hardware. This could translate into cost savings and performance benefits for customers.

• Customer Empowerment: By supporting multiple language model options and deployment environments, Teradata aims to attract customers seeking flexibility and control over their AI tools, which may enhance customer retention and platform stickiness.

---

Summary

Teradata’s commentary on small language models in the Q2 2025 earnings call reveals a clear strategic focus on integrating SLMs into their platform offerings. They highlight:

• Active customer adoption of SLMs for specialized tasks.
• Superior performance of their parallel CPU architecture for running these models.
• Commitment to an open ecosystem that avoids vendor lock-in and supports deployment flexibility.

This positions Teradata as a forward-looking player in the enterprise AI space, leveraging their core infrastructure strengths to capitalize on the growing demand for small, efficient language models tailored to specific business needs.

---

Relevant excerpt:

"We are seeing customers deploy small language models like Hugging Face to get specialized results... our massively parallel architecture... actually ran better than running on NVIDIA DPU... We don't want to be locked into any one particular language model... enable our customers to have a full deployment choice." – Stephen McMillan

---

This discussion suggests that Teradata views small language models not just as a technical trend but as a meaningful business opportunity that complements their existing data platform capabilities.