Custom Generative AI: Snowflake and NVIDIA Collaboration

Generative AI

June 26, 2023, marked an important day in the cloud technology sector as Snowflake and NVIDIA announced their partnership aimed at providing businesses with an accelerated path to create customized Generative AI applications using their proprietary data, all within the confines of the Snowflake Data Cloud​​.

This collaborative venture embodies a significant strategic advantage for both companies, as they leverage their positioning in the data analytics and AI ecosystems to offer an alternative to hyperscalers’ bespoke architectural designs and customizations.

A New Frontier in Generative AI

Snowflake and NVIDIA‘s partnership aims to bring high-performance machine learning and artificial intelligence capabilities to Snowflake’s vast volumes of proprietary and structured enterprise data. This initiative sets a new frontier in leveraging data to derive insights, predictions, and prescriptions for the global business world​.

The collaboration enables enterprises to use their proprietary data — which can range from hundreds of terabytes to petabytes of raw and curated business information — to create and fine-tune custom Large Language Models (LLMs) that power business-specific applications and services​​. The integration of NVIDIA’s NeMo platform with Snowflake will allow businesses to make custom LLMs for custom Generative AI services, including chatbots, search, and summarization​​.

This partnership has significant implications for industries such as healthcare, retail, and financial services, among others. With over 8,000 customers worldwide, the Snowflake Data Cloud offers enterprises the ability to unify, integrate, analyze, and share data across their organizations. The collaboration between Snowflake and NVIDIA will further enable customers to transform these industries by bringing customized generative AI applications to different verticals with the Data Cloud​​.

Why Custom Generative AI Matters

In today’s digital business landscape, the application of generative AI, particularly Language Learning Models (LLMs), is garnering significant interest. However, the implementation of these models within an enterprise isn’t as simple as plugging in an LLM and expecting instantaneous results. The current usage of generative AI has its roots in a B2C context with popularity and awareness propelled by ChatGPT.

Current LLMs are trained on publicly available web data, but lack the nuances of enterprise data. Using generative AI effectively within enterprises isn’t about predictions, which these models aren’t designed for, but more about gleaning insights from the rich tapestry of proprietary enterprise knowledge.

To make effective use of generative AI, businesses need to create data pipelines that pull proprietary datasets into data platforms, and then fit machine learning models on this data to obtain valuable insights specifically tailored to their business. The role of LLMs becomes crucial here, as they can process queries such as ‘what are my sales going to be next month’, and leverage an underlying predictive model to provide an estimate based on factors like pipeline generation and conversion rates.

The rising interest in generative AI has created a situation where demand is outpacing the availability of practical solutions. Enterprises, and even application vendors, are feeling the pressure to incorporate generative AI into their operations. The initial step many vendors take is to use an open-source or commercially available model from one of the hyperscalers, providing a one-size-fits-all solution that may not be optimal.

To truly harness the power of generative AI in business, the need for fine-tuning Language Learning Models (LLMs) and the use of Custom Generative AI models becomes crucial. Each enterprise has its unique challenges and data landscapes. Therefore, a one-size-fits-all model is unlikely to deliver great results. Fine-tuning LLMs and customizing generative AI models ensures that these tools align with the specific enterprise’s context, thus delivering more accurate and relevant insights, leading to improved decision-making and business performance.

Verticalized Generative AI Models: An Emerging Market Trend

The vertical approach to Generative AI and Large Language Models has been gaining substantial attention recently.  Verticalization is simply another form of customization.  Unless architectural changes to a model are needed, it’s often more beneficial to either directly use an existing pre-trained LLM and fine-tune it, or use the weights of an existing model as a starting point and continue pre-training.  Here is an approach to build verticalized custom Generative AI models:

  1. Start with Open Source or Commercially Available Models: Pre-trained models available in the public domain as well as through hyperscalers such as Microsoft and Google are becoming the starting point for further fine-tuning, reducing both the developmental timeline and cost. It’s a trend that could significantly enhance the speed of AI deployment across industries.
  2. Pre-train Using Proprietary Data: A distinctive feature of customizing open source models involves focusing on discrete data collection with direct relevance to problems within specific verticals. This sector-specific data allows for the development of highly targeted models, optimally tuned to address unique industry challenges.
  3. Integrate with Existing Workflows and Apps: Furthermore, this approach emphasizes the seamless integration of AI models into workflows and applications. This enables an efficient, user-oriented AI solution that can meaningfully enhance productivity within a specific industry or vertical.

The above approach also presents an opportunity for third-party AI vendors to derive revenue from the refinement and customization of base models for their business clients.  Training services come into play particularly when a company needs to modify an existing model’s architecture or training dataset. Changes could involve vocabulary size, the number of hidden dimensions, attention heads, or layers — each catering to unique business needs.

Customize an Existing Generative AI Model vs. Build Your Own

When considering whether to utilize an off-the-shelf Large Language Model or build your own, it’s crucial to assess your specific needs. If you don’t intend to significantly alter the underlying model architecture, it’s generally more advantageous to leverage an existing pre-trained LLM and fine-tune it to suit your requirements.

Alternatively, using the weights of an existing pre-trained LLM as a foundation and continuing its pre-training is another efficient approach. This strategy not only saves time and resources but also provides a solid starting point for most users, facilitating easier entry into the field of AI.

In other cases, where LLMs are a core part of a company’s technology moat, there is a significant appetite to train and maintain expensive models on an ongoing basis. This is expensive but could be valuable.

Swift and Strategic Execution

The Snowflake-Nvidia partnership offers businesses a framework to customize LLMs for their unique data and context.  Importantly, it does so in a way that is agnostic to which public cloud vendor an enterprise happens to use (e.g., Azure, Google Cloud, or Amazon Web Services).   What also makes the partnership between noteworthy is the speed at which both Snowflake and NVIDIA managed to operationalize their collaboration. In a rapidly evolving sector like Generative AI, the ability to execute quickly  is a significant competitive advantage.

Implications for Snowflake

For Snowflake, this is a smart move that furthers its value proposition as a cross-platform data warehouse that offers interoperability across public cloud vendors. It demonstrates Snowflake’s understanding of the modern business landscape, where flexibility and adaptability are key. This partnership with NVIDIA adds another layer of depth to Snowflake’s offerings, furthering their objective of providing their customers with the tools and resources needed to leverage their data effectively.

In doing so, Snowflake advances its position in data warehousing but also pioneers a new path in data-driven decision making,  analytics applications, and business intelligence use cases. .

Implications for NVIDIA

For NVIDIA, this partnership signifies a strategically astute move. With the looming possibility of hyperscalers developing AI workload-specific chips for inference and training, akin to Tensor Processing Units (TPUs), NVIDIA’s involvement in this partnership potentially mitigates risks associated with future specialized AI hardware developments by hyperscalers.

NVIDIA’s NeMo is a cloud-native enterprise platform for building, customizing, and deploying generative AI models with billions of parameters. Through the partnership, Snowflake plans to host and run NeMo in the Data Cloud, enabling customers to build, customize, and deploy custom LLMs used for generative AI applications, such as chatbots and intelligent search​​.

Market Implications and Potential Challenges

While this collaboration presents considerable opportunities for Snowflake and NVIDIA, it’s also important to consider its potential ramifications for other players in the market, such as Salesforce, ServiceNow, and Workday. The Snowflake-NVIDIA partnership could pose challenges for these entities, considering it offers an alternative to their existing Generative AI and Custom LLM roadmaps.

However, as with any strategic decision, it’s crucial to consider that the ultimate impact of this collaboration will depend on its execution and the evolving dynamics of the  market.

Final Thoughts

In conclusion, the partnership between Snowflake and NVIDIA symbolizes a significant step forward in enabling companies to build and deploy customized LLMs.  This partnership offers a viable alternative to traditional hyperscalers and pushing the boundaries of what’s possible in AI and machine learning. It underscores the potential of strategic collaborations in navigating the complexities of the AI sector and setting the stage for future innovations.