Mayur Khandelwal is a Vice President at EXL, where he leads the Data and AI Practice for Life & Annuities and Group Benefits insurance.
getty
Every quarter, I sit across from senior executives at major enterprises who are asking the same question, just with increasing urgency: Which AI model should we bet on?
It's a reasonable question and an exhausting one. The large language model (LLM) market today presents a paradox: The abundance of choice has become its own form of risk. At last count, enterprises could select from dozens of commercially available foundation models, spanning OpenAI, Anthropic, Google, Meta, Mistral, Cohere and a growing list of regional and domain-specific entrants. According to a 2025 survey by Global Market Insights, the top five providers—Anthropic, AWS, Google, Microsoft and OpenAI—collectively held 78% of the enterprise LLM market share in 2024, a figure that signals the beginning of concentration, not the end of fragmentation.
Additionally, a 2025 Andreessen Horowitz analysis found that 37% of enterprises now run five or more models in production simultaneously. And a Menlo Ventures 2025 report noted that enterprise LLM API spending more than doubled from $3.5 billion to $8.4 billion in less than a year. Organizations are spending faster than they're strategizing, and the multimodel complexity is catching up with them.
I've watched this pattern play out in regulated industries particularly sharply. When every business unit is free to procure its own model, you end up with fragmented governance, inconsistent outputs and data residency risks that no compliance team signed off on. The sprawl looks like optionality. In practice, it's technical debt accumulating at the speed of procurement.
History offers a useful lens here. The cloud infrastructure market in 2008 looked much like the LLM market does today: a proliferation of providers, overlapping capabilities and an enterprise buyer base that was genuinely confused about where to place long-term bets. By 2015, three providers—AWS, Azure and Google Cloud—had captured the dominant share. The rest either specialized, merged or exited.
Database markets followed the same arc. So did ERP. The pattern is consistent: Early markets reward experimentation, then economics and ecosystem lock-in force consolidation around a handful of durable platforms. The winners tend to share a few common traits: deep distribution through existing enterprise relationships; a surrounding ecosystem of tools and integrations; and a trust advantage in regulated or sensitive environments.
Looking at the current LLM landscape through that lens, the trajectory is becoming clearer. Enterprises are beginning to weight compliance readiness, audit trails, data sovereignty and the depth of a vendor's enterprise support model—criteria that favor established platforms over agile newcomers.
The practical implication isn't to wait for the market to settle before acting. Consolidation creates winners and losers among the organizations that prepared for it, not just among the vendors.
Three considerations tend to separate companies that navigate this well from those that don't:
1. The model isn't the strategy. Organizations that have invested in model-agnostic orchestration layers, where business logic, memory and workflow live independently of any single provider, will be far better positioned to switch or blend models as the market shifts. Binding business-critical processes tightly to a single vendor's API is a structural risk, not just a procurement one.
2. Contract structures matter now. Multiyear enterprise AI agreements being signed today will span the consolidation period. Executives negotiating these deals should push for exit flexibility, data portability provisions and clarity on how pricing evolves as model capabilities change. What looks like favorable unit economics in 2026 may look very different after the next wave of capability jumps.
3. Regulated industries require a different calculus. In financial services and insurance, the compliance architecture surrounding a model—data handling, explainability and audit capability—often matters more than raw performance. Domain-specific or compliance-ready models may maintain durable market positions even as general-purpose consolidation accelerates.
One principle that tends to get underweighted in these conversations is adaptability. The LLM landscape is evolving faster than most enterprise architecture cycles. Models that lead benchmarks today are routinely displaced within months. An architecture defined around a specific model's strengths is already aging. What endures is the layer above it—the orchestration logic, the data contracts and the evaluation frameworks—built to swap what's underneath without disrupting what's above.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。