AI Model Churn: Enterprise Spending Not Sticky

Enterprises Flock to New AI Models

The rapid release cycle of frontier AI models has created a distinctly fickle enterprise customer base. According to a recent survey by Menlo Ventures, a staggering 85% of enterprises have switched their primary large language model vendor at least once in the past year. This churn is not driven by dissatisfaction with existing tools, but rather by the relentless pursuit of cutting-edge capability. When OpenAI releases a new GPT iteration, or Anthropic ships a Claude upgrade, corporate procurement teams are quick to re-evaluate their contracts.

This behavior signals a profound lack of loyalty in the market. Unlike traditional software, where switching costs are high and inertia is strong, AI model adoption is treated as a plug-and-play utility. The survey found that the median enterprise now evaluates a new model every few weeks, often running side-by-side benchmarks on internal tasks before making a swap. The primary drivers are threefold:

• Performance gains: New models routinely deliver double-digit improvements in reasoning, coding accuracy, and context handling.
• Pricing pressure: Labs frequently undercut rivals on per-token costs to win market share, making migration financially compelling.
• Feature gaps: Early access to multimodal input, tool use, or longer context windows can be a decisive factor.

The result is a hyper-competitive landscape where no lab can rest on its laurels. For example, when Menlo Ventures tracked enterprise usage, they found that OpenAI’s share of primary models dropped from 50% to 34% in a single year, while Anthropic’s rose from 12% to 24%. Smaller labs like Mistral and Cohere also saw gains. This fluidity is a direct consequence of enterprises treating model access as a commodity—they will happily switch to whichever lab offers the best benchmark score on their specific workloads, even if that means abandoning a vendor they used just months prior.

Volatility in AI Spending

The recent pattern of enterprise AI spending is best described as a series of sharp pivots rather than a steady climb. One quarter, a company publicly commits to a major cloud provider’s AI stack, only to shift workloads to a rival platform the next quarter based on a single benchmark improvement or a pricing change. This back-and-forth behavior suggests that enterprise AI spending is not 'sticky', which should give investors in both companies pause.

Sticky spending—where a customer remains locked into a vendor for years due to switching costs, data gravity, or workflow integration—is the bedrock of predictable revenue. In contrast, the current AI market shows low switching costs. Enterprises are treating AI models as interchangeable commodities, easily swapped out when a newer, cheaper, or more capable option appears. For instance, a financial services firm might deploy a model for fraud detection, then re-evaluate its choice within six months if a competitor releases a fine-tuned variant with marginally better latency. This is not loyalty; it is a spot-market mentality.

What This Means for Investors

For investors in the major AI infrastructure providers, the lack of stickiness undermines the "land and expand" thesis. If a customer can leave as easily as they arrived, then the revenue booked today is not a reliable indicator of future cash flows. The volatility manifests in unpredictable churn rates and sudden drops in usage-based consumption, making quarterly guidance a guessing game.

Moreover, the constant switching forces vendors to compete on price and raw performance rather than on ecosystem lock-in. This erodes margins over time, as seen in the recent trend of aggressive discounting for bulk token purchases. The absence of a durable competitive moat means that today’s market leader could be tomorrow’s laggard, regardless of current market share. For investors, the prudent takeaway is that AI spending is a cyclical, opportunistic allocation of budgets—not a structural commitment—and should be valued accordingly, with a higher risk premium attached to any growth projections built on assumed customer permanence. The only certainty is that the next model release will trigger another round of migration.