Where Enterprise LLM Spend Actually Sits Right Now, And What It Tells You About The Next Three Years
Enterprise LLM spend hit 37 billion dollars in 2025. But 95 percent of AI pilots fail to reach production. That gap is where the next three years of strategy lives.
The numbers floating around AI spend are big enough to feel meaningless, so let me walk you through the ones that actually matter and what they imply for how you plan.
Menlo Ventures pegged enterprise LLM spend at 37 billion dollars in 2025. That is just the model layer. Hyperscaler capex on AI infrastructure crossed 650 billion dollars this year. Goldman expects it to clear a trillion in 2027. McKinsey models 6.7 trillion in cumulative AI infrastructure spend by 2030. Pick whichever number feels least insane to you. The direction is the same.
Now the number that should make you uncomfortable. MIT research published last year found that 95 percent of enterprise AI pilots fail to reach production. Ninety five. Inside companies spending tens of millions on AI, fewer than one in twenty pilots ship.
Sit with that for a second. The capex curve goes vertical. The deployment curve barely moves. That gap is where the next three years of strategy lives.
Here is what is actually happening underneath. The 37 billion dollars in model spend is going mostly to a handful of foundation labs and a small set of hyperscaler offerings. But 27 percent of enterprise LLM usage now comes through what Menlo calls product led growth, meaning individual employees expensing Claude, ChatGPT, Cursor, Perplexity directly. Shadow AI. The CFO did not approve it. The CISO did not review it. The work is happening anyway and the value is real, which is why finance keeps signing the expense reports.
So what does this mean for how you plan.
One. The spend is not the moat. Anyone with a credit card has access to the same models you do. The moat is the workflow you wrap around the model and the data you feed it. Stop benchmarking your AI strategy by budget. Start benchmarking it by deployed workflows.
Two. The 95 percent failure rate is a process problem, not a technology problem. The pilots that ship have three things in common. A specific operator who owns the outcome. A measurable before and after. Permission to kill it in 90 days if it does not work. Pilots that fail are usually owned by a committee, measured by feelings, and immortal.
Three. The shadow AI line tells you where the demand actually is. Your employees are voting with expense reports. If 27 percent of your real usage is unsanctioned, that is not a security problem to suppress. It is a product roadmap your own people just handed you for free. Pave the cow paths.
The next three years will not be decided by who spent the most. They will be decided by who closed the gap between capex and deployment fastest. Right now almost nobody is.
Sources: Menlo Ventures State of Generative AI in the Enterprise 2025, MIT research on AI pilot deployment rates, Goldman Sachs and McKinsey AI infrastructure forecasts.