The most interesting takeaway from these charts is not that AI spending is rising. It is that AI spending is rising despite AI becoming dramatically cheaper.
Historically, technology follows a familiar pattern. As costs fall, spending often falls as well. AI is doing the opposite. According to Ramp’s data, monthly AI spend per employee has exploded across every cohort. The top 1% of firms have increased AI spending to roughly US$7,500 per employee per month, the top 10% are spending more than US$600, and even the median company has seen AI spending rise several-fold over the past two years.
This is happening during a period when inference costs have been collapsing and model performance has been improving rapidly. That combination is extremely important because it suggests AI is exhibiting classic Jevons Paradox behavior. When a technology becomes cheaper and more efficient, demand often increases so much that total spending rises rather than falls.
Many investors expected cheaper models from OpenAI, Anthropic, Google, DeepSeek, Alibaba, and others to trigger a race to zero pricing and compress industry revenues. What appears to be happening instead is that lower costs are expanding the number of use cases faster than prices are declining.
Companies are not using AI to save a few dollars on software subscriptions. They are increasingly embedding AI into coding, customer service, research, workflow automation, content generation, sales processes, and internal operations. As AI becomes more capable, firms simply consume more of it.
The second chart is equally revealing. Despite all the excitement around AI, the median company still spends only about US$11 per employee per month on AI compared with nearly US$16,000 per month for a software engineer. Even among the most aggressive adopters, AI spending remains well below human compensation costs. This suggests we are still very early in the adoption curve.
The market narrative often assumes AI spending has already reached unsustainable levels. The data suggests the opposite. For most companies, AI remains a rounding error relative to payroll expenses. If AI can reliably replace even a small portion of repetitive knowledge work, there is enormous room for spending to rise further before AI budgets begin approaching labor budgets.
From an investment perspective, this reinforces why AI infrastructure remains attractive. If model costs fall but usage grows faster than prices decline, total compute demand continues rising. More AI agents, more inference, more reasoning loops, and more automation all require more GPUs, more networking equipment, more memory, more power generation, and more data center capacity.
The biggest mistake investors continue to make is assuming cheaper AI is bearish for the AI ecosystem. The evidence increasingly suggests the opposite. Cheaper AI appears to be accelerating adoption, and accelerating adoption is driving higher aggregate spending. That is precisely what a technology supercycle looks like in its early stages.