Updates: Snowflake, Databricks & Palantir- Controlling The MDS And GenAI PaaS Future (Pt.3)

Summary

• In Part 3 we discuss AI PaaS (Platform-as-a-Service), comparing Snowflake's potential to rivals attempting to offer end-to-end services for LLM implementations.
• We also discuss the numerous factors and levels of the AI stack to consider, from foundation models, fine-tuning, RAG, governance, and more.
• We have added an additional Part 4, which will be published next week, where we share further specific thoughts on Snowflake and round up this research series.

AI PaaS

This space is witnessing extensive innovation, changes, and ongoing competition. Various AI PaaS players are emerging, including established companies, new vendors from adjacent fields, and startups eyeing this novel opportunity.

Key Competition

Palantir (PLTR)

Our top pick remains PLTR, a company we've been bullish on since late 2020. Its AIP stands out as the leading contender in the AI PaaS field, boasting strong data integration, privacy and security measures, data science capabilities, and a no/low code development interface. PLTR's robust field engineering service (Forward Deployed Engineers) serves as a critical value-add, addressing unique enterprise demands while maintaining 80%+ gross margins. PLTR has no intention of joining the foundation model war, instead focusing solely on the PaaS layer and areas where it can provide value without significant capex involvement.

Microsoft (MSFT)

MSFT has taken the most capital-intensive route, investing $13bn in OpenAI, spending $100bn+ in capex to build GPU clusters, and integrating AI across every stack. As OpenAI's exclusive partner, MSFT had early access to GPT-4, giving it a significant head start over hyperscaler peers in developing a mature and comprehensive AI PaaS. This includes a model garden, evaluation and testing tools, fine-tuning capabilities, copilot building features, and data labeling and cleaning services.

Unlike AWS and GCP, which face an AI talent exodus, MSFT has retained most of its talent and even attracted valuable AI specialists from companies like Inflection AI. MSFT remains the best-positioned vendor in AI PaaS, with exclusive access to GPT-4 and the most advanced SLM Phi-3. Through developing these models, MSFT has refined its model training and fine-tuning stack, now available as Azure AI Studio, its AI PaaS platform.

Alphabet (GOOGL)

GCP was initially viewed as a promising investment by the market at large due to GOOGL's extensive AI expertise. However, GOOGL's AI PaaS platform, Vertex AI, has not emerged as a strong competitor to MSFT's Azure AI Studio as anticipated. This is partly due to GOOGL's foundation models, including Gemini, being released later and perceived as less powerful than competitors'.

GOOGL's LLMops (Large Language Model operations) stack is also considered less developed compared to MSFT's, which had been in preparation for several quarters prior to the launch of GPT-3.5-Turbo in November 2022.

Generally, GOOGL has faced challenges in its AI rollout, exacerbated by a pervasive wokism permeating its culture, exemplified by controversies surrounding its image generation capabilities. Despite these setbacks, GOOGL remains a significant player in the AI PaaS space. However, given its enterprise GTM track record and recent execution with Gemini, GOOGL may not emerge as a top-tier winner but rather as a close follower behind the leading players in this field.

Amazon (AMZN)

AWS was initially perceived as lagging in the GenAI era, lacking MSFT's early vision and GOOGL's extensive AI assets. However, AWS has made significant strides with the introduction of Bedrock and a $6bn investment in Anthropic, bringing the GPT-comparable Claude model to the public. This strategic move led to a positive valuation adjustment for AMZN in recent quarters.

Despite this progress, AWS still trails Azure and GCP in copilot building capabilities and foundation models (FM). Industry rumors suggest that AMZN has experienced a higher rate of AI talent attrition compared to its peers, potentially impacting its ability to develop its own FM, Titan, and delaying the general availability of its copilot products. While AWS remains a significant player in the AI PaaS landscape, it may not assume a leadership position if current trends persist. The appointment of the new AWS CEO, Matt Garman, could potentially alter this trajectory, but the impact of this change remains to be seen.

Databricks

Databricks, despite not initially envisioning the LLM wave, has a strong foothold in ML, closely adjacent to DL. Its academic roots at UC Berkeley enabled it to quickly attract key LLM talent and adapt to the latest trends. Consequently, it acquired MosaicML, a promising LLMops startup, for $1.3bn. MosaicML was a leading player in providing an LLMops stack, from model training to fine-tuning and evaluation. It also offered MPT open-source FMs for customers to fine-tune and build applications upon. However, MPT's leadership was short-lived with Mistral's entry and META's release of Llama2 and Llama 3.

Databricks' latest DBRX demonstrated the capability to train a good FM using its AI PaaS platform, but it lags behind Llama 3 in both performance and compute efficiency, and also trails Snowflake's Arctic. Despite FM weaknesses, Databricks has made significant progress in PaaS, particularly in data labeling, MLflow integration, evaluation, DSPM, and training dataset management. It has successfully integrated DL into its ML products and workflows, helping existing ML engineers transition to DL.