72% of Enterprises Raise LLM Budgets as Multi-Provider Strategy Becomes Default

Multi-Provider Becomes the Enterprise Standard

Enterprises are abandoning single-vendor LLM strategies. Kong's 2025 report on generative AI in the enterprise shows usage distributed across Microsoft Azure AI, OpenAI, Google Vertex AI, and open-source models — a sharp departure from the vendor consolidation pattern that defined earlier IT infrastructure cycles. This matters because it changes what you buy: LLM routers, governance layers, and observability tools now justify budget where monolithic platform fees once dominated.

Seventy-two percent of surveyed companies plan to increase LLM spending this year despite broader IT cost pressure. Nearly 40% expect to invest more than $250,000 in LLM initiatives in 2025. That number establishes a peer benchmark for CFOs evaluating whether their own LLM programs are under- or over-indexed. For vendors, it confirms that enterprise LLM spend has moved from pilot-scale to material budget line.

Security Drives Provider Selection, Not Model Performance

Thirty-one percent of enterprises cite security as their top factor when selecting LLM providers — more than cost, more than model capability. Sixty-three percent prefer paid enterprise versions over free or consumer offerings, creating explicit demand for SLAs, support, and governance tooling as part of licensing agreements.

This security-first selection dynamic penalizes vendors without data isolation guarantees, audit trails, and compliance certifications. It also strengthens the case for multi-provider strategies: enterprises hedge not just against cost and performance, but against security incidents and outages at any single provider. OpenAI's June 10 outage, referenced in recent enterprise platform discussions, reinforced that dependency on a single LLM endpoint creates unacceptable availability risk for production systems.

Provider Strategies Diverge on Cost, Safety, and Control

The four major enterprise LLM providers now occupy distinct positions that affect deployment strategy:

OpenAI offers breadth — chat, agents, video, and adaptive compute models like GPT-5's Instant and Thinking variants that allocate compute dynamically. The ChatGPT Agent feature adds multi-step task execution with activity logs and approval workflows, positioning OpenAI as the platform for governed agentic workflows. The risk: rapid model changes require strong internal version control and change management.

Google's Gemini 2.5 delivers multimodal reasoning but at high computational intensity, raising cost and sustainability questions for large-scale deployment. Enterprises running high-volume inference workloads should model Gemini 2.5's compute cost against alternatives before committing.

Anthropic's Claude Sonnet 4.5 prioritizes safety and predictable behavior, with explicit design for long-running agent tasks and regulated environments. It ships with Claude Code and a native VS Code extension, targeting enterprise developer workflows. The trade-off: narrower feature scope and slower adoption of aggressive capabilities compared to OpenAI.

Mistral positions open-weight efficiency as its enterprise value proposition. Medium 3 offers competitive reasoning at lower cost, while Large and Codestral are accessible through Google Vertex AI and AWS Bedrock — enabling open-weight deployment without building infrastructure from scratch. Enterprises choosing Mistral must own more of the stack, but gain transparency and multi-cloud portability.

What This Means for Your LLM Platform

The Kong data and provider differentiation create three immediate implications:

First, budget for platform teams that manage multi-provider routing and governance, not just API integrations. The shift to multiple LLM providers increases the value of abstraction layers that make switching easier and reduce lock-in. If you are still treating LLM selection as a binary OpenAI-versus-Google decision, you are solving the wrong problem.

Second, security and auditability now justify premium pricing. Vendors without enterprise-grade security controls will lose deals to competitors who offer data isolation, compliance certifications, and audit trails — even if their models benchmark lower. Use the 63% preference for paid enterprise versions as leverage when negotiating SLAs and support terms.

Third, cost and compute efficiency matter more as usage scales. Google's high-intensity Gemini 2.5 and Anthropic's safety-focused Claude Sonnet 4.5 represent opposite ends of the cost-capability curve. If you are deploying high-volume inference workloads, model compute cost per token against your expected query load before signing annual commitments. Mistral's open-weight models offer a hedge: competitive reasoning at lower cost, with deployment flexibility through Vertex AI or Bedrock.

What to Watch

The $250,000+ spending tier is now common enough to set internal benchmarks, but it also creates pressure to show ROI. Expect tighter scrutiny on LLM program outcomes in 2025 as finance teams distinguish between pilot-stage experimentation and production-scale value delivery.

Provider consolidation remains possible despite current multi-vendor patterns. If one provider solves the governance, cost, and security trade-offs decisively, enterprises may re-consolidate. Until then, treat multi-provider deployment as the default and build your platform architecture accordingly.