UBS survey: enterprise AI spend holds as cloud capex tops $600B

Enterprises are tightening AI cost controls without cutting budgets, UBS's survey of large companies found, as in-house builds displace traditional SaaS vendors while cloud infrastructure and data management layers hold their ground.

"Enterprises have moved from a 'token maximization' phase to one focused on 'token optimization,' where efficiency and return on investment take priority," UBS analyst Karl Keirstead said in the Aug. 10 report.

Companies are deploying usage-tracking systems that route requests to lower-tier models once employees hit thresholds, trimming context windows and cutting repeated calls to retrieval systems. The shift has not yet translated into budget pullbacks, Keirstead said, with cloud vendors' strong second-quarter results corroborating sustained demand.

The findings support differentiated positioning across the software sector: cloud providers and data management firms such as Databricks, Snowflake and Palantir Technologies remain entrenched, while traditional application vendors including Salesforce and ServiceNow face a delayed AI monetization thesis as Fortune 500 companies build core systems in-house.

In-House Builds Outpace SaaS AI Features

Salesforce founder Marc Benioff has argued that enterprise-built AI will ultimately fail, pushing customers toward mature SaaS applications, and ServiceNow holds a similar view. UBS's survey suggests reality has not yet sided with the vendors. Fortune 500 companies still favor building AI products in-house for three reasons: highly customized internal processes that off-the-shelf SaaS AI features may not fit, full control over data and workflow orchestration, and pricing or delivery shortfalls in external products.

One surveyed company plans to build its own ontology layer using Turtle files, importing SAP or PTC Windchill data in real time into a Databricks data lake for agents to call, aiming to replace Palantir on grounds of lower cost and easier hiring. Another said no SaaS vendor currently offers AI features worth buying, and it could develop its own plugins. Procurement has not disappeared entirely — enterprises still buy mature products for code review, security, observability, SIEM and workflows — but the signal for software investors is whether customers will keep paying for these features rather than building them.

Data Management Layer Holds, MongoDB Faces Pressure

Data management companies show a stronger presence in the enterprise AI stack than application software. Databricks was cited repeatedly for real-time data lakes, governance and AI tooling, while Snowflake is used to build agents. Respondents noted that large language models cannot replace Snowflake — they handle language well but struggle with mathematical processing of millions of data points, making it more sensible to layer models such as Anthropic's Claude on top for anomaly detection.

Palantir drew mixed feedback. It retains a place in enterprise AI and ontology layers, but some customers are attempting to build their own ontology layers to replace it. One large defense contractor said internal teams have begun scaling back Palantir use and expect the company to face different challenges in coming years, though UBS cautioned these cases remain early signals not yet generalizable. In the database layer, customers are choosing Redis and Postgres for AI workloads, intensifying scrutiny over whether MongoDB can participate fully.

Small AI-native vendors are also appearing on procurement lists. Enterprises named CodeRabbit for code review, Opik for LLM observability, Onyx for agent security, ArmorPoint for AI SIEM, Workfabric for digital workflow twins, and LiteLLM with Langfuse for model routing and prompt compression. These target specific pain points rather than broad platform narratives — AI-generated code now outpaces human review capacity, so companies buy CodeRabbit to keep vulnerabilities out of production, and AI is sharpening attack capabilities, pushing security budgets higher on more rigid logic than feature experimentation.

Cloud dominance persists. AI infrastructure remains concentrated on AWS, Azure and Google Cloud, with few surveyed companies planning large-scale on-premises AI hardware. One enterprise plans to migrate all data from Azure to Google Cloud on pricing grounds, a move that would take at least a year and raise compatibility questions for Power BI users and Excel scripts, suggesting Microsoft spending may be trimmed but not eliminated. The survey found no customer-side evidence of a broad shift of AI workloads back on-premises.

The survey supports continued strength for cloud infrastructure and data management companies while pressuring traditional SaaS application vendors. Palantir shares have climbed about 37 percent since its second-quarter release, when it beat revenue guidance by $136 million, and cloud giants are pouring nearly $600 billion into capital expenditure as AI demand surges. MongoDB, by contrast, faces competitive pressure from Redis and Postgres on AI workloads. The rise of AI-native startups in specific niches could pressure incumbent enterprise software valuations, making the build-vs-buy preference the key variable for software investors to watch.

This article is for informational purposes only and does not constitute investment advice.