UBS Corporate Survey: AI Spending Remains Robust, "In-House" Software Becomes a Trend, Data Management Layer Is Crucial

The logic behind corporate AI investment is undergoing a structural shift, but this transformation has not led to budget contractions.

According to Zhui Feng Trading Desk, Karl Keirstead, an analyst at UBS Group AG, pointed out in a corporate AI survey report released on August 10 that many companies have moved from the initial phase of "maximizing tokens" to a new phase characterized by token optimization, with efficiency and return on investment (ROI) taking higher priority. The key point is that this change "has not yet translated into a pullback in AI spending." This sends an important signal to the market: Companies are starting to crunch the numbers, but they haven't hit the brakes yet.

For frontier model providers like OpenAI and Anthropic, as well as the cloud vendors behind them, demand has not significantly shrunk due to cost anxiety. The strong performance of cloud vendors in the second quarter corroborates this. The ones truly under pressure are traditional application software vendors attempting to bundle AI features for sale to enterprises—companies still generally prefer to build core AI systems in-house, and the turning point awaited by traditional SaaS vendors has not yet arrived.

From the perspective of supply chain division of labor, divergence is already evident: The cloud infrastructure layer and data management layer hold relatively stable positions, with Databricks, Snowflake, and Palantir continuing to occupy key spots in the enterprise AI stack; the application software layer faces the most pressure, with AI features from large SaaS vendors having limited presence on corporate procurement lists. Instead, a batch of small AI-native companies focusing on specific pain points have entered actual deployment scenarios.

Cost Control Has Started, But Budgets Have Not Receded

The core concern for enterprises has shifted from "Can we use AI?" to "Who is using it, how much are they using, and is it worth it?"

Many companies have deployed usage tracking systems. Once employees reach a certain quota, requests are automatically switched to lower-tier models with slower response speeds. Some companies have begun restricting employees' permissions to create agents. Sensitivity to consumption-based billing models is also rising, as it directly links AI usage to bills.

More granular optimizations are occurring at the model invocation stage: Providing models with more refined context rather than filling the entire context window; reducing repeated calls to RAG systems; and lowering the cost per task through scripting, routing, and prompt compression. One company offered frontier models to customers at a fixed price, only to find that model usage costs eroded profit margins, potentially forcing a future adjustment to its pricing structure.

However, there has been no significant contraction in overall budgets. Companies generally fear missing the window of opportunity and prefer to push forward with deployment first, adding guardrails gradually. Karl Keirstead emphasized in the report that "calculating ROI" is not synonymous with "stopping AI investment," at least among this sample of surveyed companies.

Rise of Model Routing, Chinese Open-Source Models Not Yet Mainstream in Large Enterprises

A natural path to reducing AI invocation costs is to route some tasks from frontier closed-source models to open-source models. However, in this survey sample dominated by large non-tech enterprises, OpenAI's GPT and Anthropic's Claude remain the mainstream choices.

Some companies explicitly stated they are still using frontier models. While tech-native companies show higher acceptance of Chinese open-source models, large traditional enterprises place greater emphasis on security, compliance, and controllability, especially those in government, defense, and related supply chains.

Currently, cost reduction for models is proceeding along three parallel paths: Distributing different tasks to different models, using small models to handle low-value requests, and optimizing invocation chains to reduce token waste. Chinese open-source models may be an option, but they have not yet become the priority choice for most large enterprises.

"In-House Build" Logic Dominates, Turning Point for Traditional SaaS Vendors Has Not Arrived

Salesforce founder Marc Benioff has repeatedly expressed the same judgment: enterprises building AI in-house will ultimately fail and turn to purchasing AI applications from mature SaaS vendors. ServiceNow holds similar logic. However, the UBS survey shows that reality has not yet sided with SaaS vendors.

Many Fortune 500 companies still prefer to build AI products in-house, primarily for three reasons: Internal corporate processes are highly customized, and off-the-shelf SaaS AI features may not be suitable; building in-house retains full control over data, workflow orchestration layers (harness), and agent systems; and some external products are priced too high or deliver results that fall short of expectations.

A typical case from the survey: A company plans to use Turtle files to build its own ontology layer, importing SAP or PTC Windchill data in real time into a Databricks data lake. After multi-layer cleaning, the data will be available for agent calls to build an action engine, aiming to replace the Palantir solution. Reasons include lower costs, easier talent recruitment, and the potential to offer the technology to customers in the future. Another company stated directly that current SaaS vendors do not have AI features worth buying, and they can develop plugins themselves.

This does not mean procurement has disappeared entirely. In specific scenarios such as code review, security, observability, SIEM, and workflows, companies still purchase mature products. But for software investors, the real signal lies in whether companies are willing to pay continuously for these features and abandon the in-house build option. If this preference does not reverse, AI products from application software companies will struggle to achieve high ROI.

Data Management Layer Position Is Solid, Databricks Frequently Mentioned

In the enterprise AI technology stack, the presence of data management companies is significantly stronger than that of the application software layer. Databricks was mentioned multiple times in the survey, commonly in scenarios such as real-time data lakes, data governance, and AI tool foundations; Snowflake is used to build agents, and surveyed companies explicitly pointed out that LLMs cannot directly replace Snowflake—LLMs excel at processing language but are not good at mathematical processing and transformation of millions of data points. A more reasonable approach is to place models like Claude on top of Snowflake for anomaly observation.

Feedback on Palantir is relatively complex. On one hand, it still holds a place in enterprise AI and the ontology layer; on the other hand, some customers are trying to replace Palantir by building their own ontology layers with AI. A large defense contractor mentioned that internal teams have begun reducing their use of Palantir and predict the company will face different challenges in the coming years. However, the UBS report notes that such cases are currently early signals and insufficient to extrapolate as a general trend.

Competitive pressure has also emerged at the database layer. Some customers choose Redis and Postgres for AI workloads, intensifying external concerns about whether MongoDB can fully participate in AI workloads.

AI-Native Small Firms Have Entered Corporate Lists, Large SaaS Presence Is Limited

A noteworthy detail in the survey is that the number of listed application software companies mentioned by enterprises is limited. Atlassian, CrowdStrike, and Salesforce all appeared, but not frequently. Instead, a batch of small private AI-native companies are more active on procurement lists: CodeRabbit for code review, Opik for LLM observability, Onyx for agent security, ArmorPoint for AI SIEM, Workfabric for digital workflow twins, and LiteLLM connecting Langfuse for model routing and prompt compression.

These companies are not targeting the macro narrative of "enterprise AI platforms," but rather addressing very specific business pain points. Taking code review as an example, the speed of AI-generated code far exceeds human review capabilities, so companies choose CodeRabbit to prevent security vulnerabilities from entering production environments. In the security field, AI can prioritize vulnerabilities by combining information such as firewall configurations and user permissions, helping security teams identify truly critical threats.

The survey also pointed out that AI is rapidly enhancing attack capabilities, and corporate security-related spending is expected to increase accordingly. Unlike other AI application scenarios, the budget logic for security spending is more rigid, driven by the need to prevent the expansion of the attack surface rather than the desire to experience new features.

Cloud-Dominated Landscape Unchanged, On-Premise Deployment Narrative Lacks Customer-Side Support

Enterprise AI infrastructure remains highly concentrated on AWS, Azure, and Google Cloud. Surveyed companies rarely mentioned large-scale expansion of on-premise AI hardware and software stacks. Automotive data, predictive maintenance, agent applications, data lakes, and AI workflows on Snowflake and Databricks are mostly built around the three major clouds.

One company in the survey plans to migrate all its data from Azure to Google Cloud, citing the latter's more competitive pricing. However, migration itself is not simple, involving a data relocation cycle of at least one year, as well as issues such as Power BI users, Excel scripts, and Google Sheets compatibility. Microsoft-related spending is expected to be reduced but not eliminated. This indicates that while there is price competition among cloud vendors, the substantial trend of AI moving back from the cloud to on-premise environments is not yet sufficiently supported by customer-side evidence.

The UBS report points out that although some partners are still evaluating the feasibility of "AI factories" and moving models back to on-premise environments, based on the current customer sample, the main battlefield for AI deployment remains in the cloud.