UBS Survey: Enterprise AI Spending Holds Strong, but Build-Your-Own Trend Pressures SaaS Vendors
Taylor Wilson
A UBS enterprise AI survey published August 10 finds companies shifting from "maximize tokens" to "optimize tokens" — yet spending has not pulled back. The real pressure falls on SaaS vendors, as large enterprises prefer to build AI in-house.
Companies are watching ROI now — so why is spending still rising?
UBS analyst Karl Keirstead's core finding: enterprise AI logic is moving from "token maximization" to "token optimization," with efficiency and ROI now the top priority.
This means → "calculating ROI" is not the same as "cutting AI spend." Companies fear missing the window and prefer to deploy first, add guardrails later.
Cloud vendors posted strong Q2 results; demand for frontier models from OpenAI and Anthropic has not visibly contracted. In plain terms = the spending pace hasn't slowed — but the spending *pattern* has gotten more deliberate.
How are companies controlling costs? They've already started
The central concern has shifted from "can AI work?" to "who's using it, how much, and is it worth it?" Several firms have deployed usage-tracking systems that auto-switch requests to cheaper models once an employee hits a quota.
On the model-call side, optimization is underway: feeding models leaner context instead of filling the full context window, reducing repeated calls to RAG systems — retrieval-augmented generation, a technique that has the model look up data before answering — and using scripts and prompt compression to cut per-task costs.
One telling case: a company offered customers access to a frontier model at a flat price. Model-usage costs ate into margins. This means → the pricing structure will have to change — "unlimited access to the strongest model" is not sustainable.
Why do large enterprises prefer to build rather than buy SaaS?
Salesforce founder Marc Benioff has repeatedly predicted that in-house AI efforts will fail and companies will turn to mature SaaS products. UBS's survey suggests reality has not yet sided with SaaS vendors.
Fortune 500 companies cite three reasons for building in-house: ① internal workflows are highly customized and off-the-shelf SaaS AI may not fit; ② building preserves full control over data and agent systems; ③ some vendor products are overpriced or under-deliver.
In plain terms = one company said outright: "SaaS vendors have no AI feature worth buying — we can build plug-ins ourselves." This reflects a deeper signal: the AI monetization inflection point for application-software companies has not arrived.
What does one build-your-own case reveal?
One enterprise plans to use Turtle files to build its own ontology layer — a structure that organizes corporate data by semantic relationships — piping SAP and PTC Windchill data in real time into a Databricks data lake, cleaning it through multiple stages, then exposing it to AI agents.
The goal is to replace Palantir, citing lower cost, easier hiring, and the option to eventually open the technology to customers.
This means → companies are not just building applications in-house — they are building foundational data architecture. For platform companies like Palantir, that is a deeper threat.
Data layer safe, application layer anxious — who holds the strongest position?
Databricks was cited repeatedly, spanning real-time data lakes, data governance, and AI tool infrastructure. Snowflake is being used to build agents; respondents said explicitly that LLMs cannot replace it — LLMs handle language well but struggle with math operations across millions of data points.
Palantir's feedback is mixed: it still holds a place in enterprise AI and ontology layers, but some clients are experimenting with AI-built ontology layers as replacements. A large defense contractor noted that internal teams have already begun scaling back Palantir usage. UBS flags these cases as early signals, not yet a trend.
The database layer faces pressure too: some customers chose Redis and Postgres for AI workloads, intensifying scrutiny of whether MongoDB can fully participate in AI-driven demand.
Small vendors winning deals, cloud landscape unchanged — what else stands out?
Listed application-software companies appeared infrequently in the survey. Atlassian, CrowdStrike, and Salesforce were all mentioned but not prominently. Instead, a wave of small, private, AI-native companies showed up more actively: CodeRabbit (code review), Opik (LLM observability), Onyx (agent security) — each targeting a very specific pain point.
This means → enterprise AI procurement is bypassing large SaaS platforms in favor of vertical, purpose-built tools.
On cloud deployment, AI infrastructure remains heavily concentrated on AWS, Azure, and Google Cloud. No substantial customer-side evidence supports an "AI repatriation" trend. One company plans to migrate entirely from Azure to Google Cloud on pricing grounds, but the migration will take at least a year.
Content is for reference only, not financial advice.