72% of Enterprises Will Increase LLM Spend in 2025 — Multi-Provider Routing Is Now Default

Budget benchmarks and provider diversity define 2025 LLM deployment

Kong's 2025 Enterprise LLM Adoption Report establishes two concrete planning thresholds for CIOs: 72% of companies will increase LLM spending this year, and nearly 40% expect to invest more than $250,000 in LLM initiatives. That $250,000 line separates experimental pilots from strategic programs, and enterprises below that threshold while targeting broad deployment are likely under-investing relative to surveyed peers.

The report also confirms what procurement teams already suspected — no single provider dominates enterprise LLM deployment. Usage is distributed across Microsoft's Azure OpenAI Service, OpenAI's direct APIs, Google's Gemini, and open-source alternatives including Llama and Mistral. This multi-provider reality makes LLM routing architectures the default, not an optimization.

Open-source models gain parity in regulated environments

SnapLogic's recent "Great LLM Race" analysis segments the enterprise LLM market into three deployment categories: proprietary API models (GPT-4o, Claude 3.5 Sonnet), open-source models (Llama, Mistral), and industry-tuned vertical models (Nabla Copilot for healthcare, Harvey AI for legal). The competitive pressure comes from Chinese open-source models — specifically DeepSeek and Qwen 2.5 — which SnapLogic reports achieve near-parity with leading U.S. providers at a fraction of traditional cost. Qwen 2.5 claims to outperform both DeepSeek and GPT-4o in coding, multilingual tasks, and customer interaction.

For regulated industries, this shifts the calculation. Open-source and open-weight models are now described as "increasingly preferred" where data privacy and regulatory constraints dominate, which pushes more buyers toward on-premise or VPC deployments instead of single-cloud API lock-in. If your compliance requirements previously forced you into a single hyperscaler's managed service, the performance parity of self-hosted open models changes the trade-off between control and capability.

Hybrid deployment becomes architectural baseline

The SnapLogic analysis codifies hybrid model strategy as best practice: combine open-source models for sensitive data with cloud API models for scalability, orchestrated by routing tools that switch models based on task requirements. This is not theoretical — the Kong data shows this pattern already in production across multiple providers.

Concrete cost optimization practices now include caching for frequently used prompts, compression for input text, preferring smaller models for simple tasks, and automatic model routing based on latency, quality, and cost thresholds. These are not vendor pitches — they are documented practices from enterprises managing LLM spend above $250,000 annually.

For security and compliance, the recommendation is Amazon Bedrock, Azure OpenAI Service, or similar managed offerings with built-in audit trails, governance procedures, and documented model-selection criteria. The critical shift is that "managed" no longer means "single provider" — enterprises are running managed services from multiple hyperscalers simultaneously.

Procurement strategy: multi-vendor by default

The Kong data justifies multi-cloud and multi-model strategies to boards and audit committees. Recent OpenAI outages that affected enterprises with single-provider architectures provide the risk case. The budget data provides the business case — if peer enterprises are diversifying across Microsoft, OpenAI, Google, and open-source, single-vendor lock-in becomes a board-level risk discussion.

Procurement teams can now negotiate volume discounts, reserved capacity, or committed spend across several providers instead of locking into one long-term LLM contract. The $250,000 threshold also sets a reference band for contract negotiations — if your planned spend is below that level, you are likely in pilot-tier pricing, not enterprise-tier discounts.

Vertical models create workload-specific competition

Nabla Copilot in healthcare and Harvey AI in legal demonstrate workload-specific competition that can displace or reduce reliance on GPT-4o or Claude for those verticals. This matters for budget planning because vertical models often require different contract structures — per-seat licensing instead of per-token consumption, or fixed annual fees instead of variable API costs.

If your deployment includes regulated or domain-specific workloads, the default assumption should be evaluating vertical models alongside general-purpose APIs. The cost structure will differ, but the SnapLogic data suggests performance parity or advantage in those specific domains.

What to watch

The $250,000 spending threshold will likely rise as LLM deployment moves from departmental pilots to enterprise-wide programs. Track whether your peer group increases that number in 2026 budget cycles — if it does, your board will ask why your spend is flat.

Open-source model performance continues to compress the cost-capability gap with proprietary APIs. If Qwen 2.5 or similar models sustain claimed performance advantages in coding and multilingual tasks, expect procurement pressure to shift more workloads to self-hosted deployments, particularly in cost-sensitive or regulated environments.

Multi-provider routing is now table stakes, but tooling maturity varies. Watch for consolidation in LLM gateway and routing platforms — the vendor that makes multi-model orchestration operationally simple will capture budget that currently goes to custom-built routing logic.