Data platforms Snowflake and Databricks acquiring model developers: Reka AI and MosaicML

Profiling the expansion of data platforms Snowflake and Databricks into model development via their acquisitions of Reka AI ($1B) and MosaicML ($1.3B)

Bloomberg yesterday announced that Snowflake is acquiring Reka AI for $1 billion. Founded in 2022 by former researchers from Alphabet and Deepmind, Reka AI specializes in LLM development and has developed advanced AI technologies like Reka Core, a multimodal LLM. The acquisition seemingly will enhance Snowflake’s model development team, extending its work with models like Arctic to leverage its immense customer datasets with AI capabilities.

Having model development inside the data platform allows Snowflake or Databricks to enable customers to securely use their enterprise data to build, fine-tune, and augment machine learning and generative AI models. This approach ensures that sensitive data and intellectual property remain within the organization's control, enhancing privacy and security — and importantly for Snowflake and Databricks, their moat and lock-in.

In many ways, Snowflake's acquisition of Reka AI parallels Databricks' $1.3 billion acquisition of MosaicML. Both deals involved significant investments in AI-first infrastructure companies that focus on leveraging private enterprise data. Both teams were Silicon Valley-based model developers. MosaicML had 62 employees at acquisition, representing $21M per employee; while Reka AI has just 24 employees representing over $41M per employee. Elite model developer talent is clearly scarce and of high value.

Both Snowflake and Databricks aim to use these proprietary models to improve the value of customer data with data-centric AI.

Snowflake’s $1B acquisition of Reka AI

Reka AI is a team of 24 founded just two years ago in Silicon Valley. Reka specializes in multimodal language models that can process text, images, videos, and audio — perhaps suggesting these other data formats of unstructured data are of high strategic value for Snowflake. Reka’s flagship model, Reka Core, offers advanced capabilities in reasoning, coding, and flexible deployment options. It particularly excels in multi-modality, as evidenced by its human evaluation scoring in multimodal prompts against peers, placing 2nd only to OpenAI:

Reka Core can handle prompts with up to 128,000 tokens and has demonstrated competitive performance compared to GPT-4 and Claude 3 Opus. It excels in multimodal data processing, outperforming Google's Gemini Ultra in certain video analysis tasks. Additionally, Reka AI has developed proprietary model distillation technology that transfers knowledge from advanced LLMs to smaller, more efficient models. This reduces hardware requirements and costs, making advanced AI capabilities more accessible and cost-effective.

Reka’s engineers compared the LLM against GPT-4 and Anthropic’s top-end Claude 3 Opus model in a series of internal tests. One of the evaluations used MMLU, a benchmark datasets that comprises several thousand questions spanning dozens of topics. Reka Core answered 83.2% of the questions correctly, which put it about 3% behind GPT-4 and Claude 3 Opus:

The Reka AI showcase page demonstrates the capabilities of Reka Core, its multimodal language model. It highlights various applications such as detailed image captioning, landmark recognition, and chart understanding, showcasing the model's ability to process and interpret text, images, videos, and audio. These examples illustrate how Reka Core can handle complex data tasks (perhaps even data labelling) and provide valuable insights, making it a powerful tool for enhancing data platforms like Snowflake's. Below is the Reka Core showcase, showing its capabilities for tasks:

Reka offers Reka Core alongside two less capable, but more hardware-efficient models. The first is called Reka Edge and features a lightweight architecture that allows it to run on devices with limited onboard computing capacity. A third model, Reka Flash, is situated between the company’s two other LLMs on the price-performance scale.

Snowflake’s venture arm had previously invested in Reka AI, valuing it at $300 million. Other investors were Radical Ventures and DST Global. In the past year, Snowflake has made at least 8 acquisitions — many of which are bolstering its AI capabilities.

Before acquiring Reka AI, Snowflake announced a partnership with NVIDIA to allow customers to build AI applications inside Snowflake Data Cloud. This collaboration allows enterprises to develop and fine-tune AI models, such as chatbots and search tools, without sending their data to external services.

Strategic Implications:

• Enterprise Focus: Snowflake aims to integrate Reka’s technology to enhance its platform's capabilities in handling complex enterprise data tasks, such as text-to-SQL queries.

• Open Source Initiative: Snowflake recently open-sourced its LLM, Arctic, which features a mix of dense and MoE (Mixture of Experts) architecture, optimizing inference speed and efficiency.

• Leadership and Vision: The acquisition aligns with Snowflake's strategic shift under its new CEO, emphasizing AI and machine learning as key growth areas.

• RAG applications: By integrating Reka AI's advanced models into its data platform, Snowflake can enable Retrieval-Augmented Generation (RAG) applications, allowing users to efficiently query and generate insights from vast amounts of enterprise data without compromising data privacy. This is further enabled by Streamlit, which can deploy the AI application to production.

• Competition: This move positions Snowflake to offer its own unique models, leveraging Reka’s technology with its world-class enterprise data infrastructure and unique customer data access.

Databricks’ $1.3B acquisition of MosaicML

Hanlin Tang, Mosaic ML co-founder and current chief technology officer; Ali Ghodsi, co-founder and chief executive of Databricks; and Naveen Rao, MosaicML’s CEO and co-founder. [Fortune]

Databricks acquired MosaicML in July 2023 for $1.3 billion, aiming to bolster its AI model development and training capabilities. MosaicML provides tools for efficient AI training and model development, aligning with Databricks’ mission to unify data and AI. Founded in 2020, MosaicML focuses on simplifying the use of AI by providing efficient and scalable solutions for training and deploying large-scale machine learning models.

Databricks acquired MosaicML to enhance its data intelligence platform by integrating MosaicML's advanced AI model training and optimization capabilities, enabling faster and more efficient machine learning deployments.

Strategic Implications:

• AI Training Toolkit: Databricks has launched the Mosaic AI Training toolkit, which enhances its platform’s ability to train and deploy advanced AI models inside customer’s Databricks environment.

• DBRX Model: The DBRX open-source LLM developed post-acquisition focuses on efficient training and inference, making it suitable for various enterprise applications. Databricks’ focus on developing AI models that excel in tasks like text-to-SQL queries differentiates its offerings from competitors.

• Comprehensive AI Solutions: Databricks provides a full-stack solution for AI model development, training, and deployment, streamlining workflows and improving productivity for its users. Having this all in one platform reduces complexity, improves efficiency and performance for customers.

  • Unstructured data: Databricks acquired Lilac to enhance its ability to manage unstructured data for generative AI applications. Lilac's scalable tools simplify the process of searching, clustering, and analyzing text datasets, helping Databricks customers develop high-quality AI models more efficiently.

• Data Privacy: By integrating MosaicML, Databricks allows customers to use their enterprise data securely, avoiding the need to send sensitive information to third-party services.

• RAG applications: By integrating MosaicML's model development tools with its data platform, Databricks can enable Retrieval-Augmented Generation (RAG) applications, allowing enterprises to efficiently retrieve and generate insights from their data.

Conclusion: Integrated data platforms and model development

These acquisitions by Snowflake and Databricks highlight their strategic moves to integrate model development directly into their platforms, leveraging their biggest asset—customer data. This approach reduces reliance on third-party model developers like OpenAI, Anthropic, Google, and Meta, allowing for greater control and customization. Both teams are expected to build proprietary models tailored to their platforms' specific tasks, such as text-to-SQL, enhancing their data processing and AI capabilities.

We may see other data platforms like Oracle, MongoDB, Salesforce or SAP acquire model developers to build models specific to their unique enterprise customer data. Model developer talent is scarce and highly valued clearly…