72% of Enterprises Increase LLM Budgets as Security Overtakes Model Quality

Paid Enterprise LLMs Win as Buyers Abandon Free Tiers

Nearly three-quarters of enterprises plan to increase spending on large language models in 2025 despite flat or declining overall IT budgets, according to Kong's newly released enterprise GenAI report. The data marks a clear shift: 63% of surveyed companies now prefer paid enterprise versions of LLMs over consumer or free tiers, driven by security requirements that have overtaken model quality as the primary selection criterion.

The findings matter because they establish a quantitative baseline for how enterprises are actually structuring LLM deployments versus how vendors pitch them. Forty percent of respondents expect to invest more than $250,000 in LLM-related projects this year — a figure that transforms generative AI from experimental line item to defensible platform spend.

Security Becomes the Primary Vendor Filter

Thirty-one percent of buyers now cite security as their top factor when choosing an LLM provider, ahead of model performance, cost, or integration ease. This reorders the competitive landscape. Vendors without SOC2 attestations, VPC isolation, or enterprise key management are effectively disqualified before technical evaluations begin.

The preference for paid enterprise SKUs directly benefits Microsoft Azure AI, OpenAI's enterprise offerings, and Google Vertex AI — the three platforms Kong identifies as capturing the majority of current enterprise spend. But the survey also documents meaningful adoption of open-source and hybrid models, particularly for internal workloads where data residency or cost sensitivity outweighs cutting-edge performance.

This split creates a two-tier architecture: closed commercial models for customer-facing applications where quality and support matter, open models for batch processing, internal automation, and regulated environments where control trumps capability. Enterprises are not standardizing on a single provider. They are building routing layers to switch between OpenAI, Azure, Vertex AI, and self-hosted models depending on workload sensitivity and cost.

Multi-Model Strategies Force Abstraction Layer Investments

The documented trend toward provider diversity has immediate architectural consequences. Buyers can no longer hard-code applications to a single LLM API without creating vendor lock-in and technical debt. This drives investment in abstraction layers — tools like Portkey, LiteLLM, or internal routing logic that let teams map different use cases to different models without refactoring code.

Development teams are using LLMs primarily for code generation, automation, and data analysis, all workloads where security and integration with existing dev tooling matter more than raw model benchmarks. A code assistant that cannot integrate with existing CI/CD pipelines, version control, or secret management is unusable regardless of model quality. This raises the bar for vendors positioning LLMs as developer productivity tools: integration depth now determines adoption more than model leaderboard rankings.

Buy Over Build Becomes the Default Posture

Separate research from Andreessen Horowitz surveying 100 enterprise CIOs across 15 industries shows enterprises favoring embedded LLM features inside existing SaaS platforms over building proprietary LLM infrastructure from scratch. This trend strengthens vendors like Salesforce, ServiceNow, and Workday that can bundle LLM capabilities into products buyers already deploy, versus startups requiring greenfield platform adoption.

The implication for budget planning: six-figure annual LLM spend is now defensible with peer benchmarks. When 40% of companies are allocating over $250,000 per year and 72% are increasing LLM budgets while cutting elsewhere, generative AI has crossed from experiment to strategic platform investment. CFOs and boards will approve line items they would have rejected 18 months ago.

What to Watch

Security-first vendor selection will accelerate consolidation among LLM providers. Startups without mature enterprise controls or hyperscaler distribution face shrinking addressable markets as buyers formalize procurement requirements. The companies best positioned are those offering enterprise SKUs with logging, DLP, private networking, and clear data residency commitments — whether through proprietary models or managed open-source offerings.

The rise of multi-model architectures will also create demand for governance tooling that does not yet exist at enterprise scale: centralized model versioning, audit trails across providers, cost allocation by workload, and policy enforcement for which models can be used for which data. Buyers building these capabilities in-house today are creating procurement opportunities for vendors who productize them tomorrow.

Enterprises are no longer asking whether to deploy LLMs. They are allocating budgets, setting security baselines, and designing architectures that assume LLMs are permanent infrastructure. The question is no longer adoption but standardization — and the window for vendors to establish themselves as enterprise-grade providers is closing faster than model capabilities are advancing.