Big companies are getting a harsh lesson in the cost of generative AI. A new survey from Capgemini, a global IT consulting firm, finds that 59% of executives say their AI spending in the past year came in over budget—by 24% on average. That sticker shock is pushing many teams to reconsider which AI models they use, with a growing number planning to lean more on cheaper open-weight alternatives over the next year.
What the survey found
Capgemini polled 500 executives at companies with more than $1 billion in annual revenue across 10 countries. The results paint a picture of an industry where the enthusiasm for AI is colliding with the reality of its price tag. More leaders now list cost as a top concern, and that's changing how they think about the technology.
The key distinction is between closed and open-weight models. Closed models, like those from major AI labs, are typically accessed through paid application programming interfaces (APIs), where you pay per token—the chunks of text the model processes. Open-weight models, by contrast, can be downloaded and run on a company's own infrastructure or a chosen cloud provider, giving businesses more control over where and how they deploy AI.
Capgemini estimates that open-weight tokens are 87% cheaper on average than their closed counterparts. That's a massive difference, and it's already showing up in usage patterns. The firm says open-weight models handled 72% of traffic on OpenRouter, a popular AI routing platform, in September—yet they represented just 27% of AI spending. That suggests teams are already quietly routing everyday tasks to lower-cost tools.
Not a wholesale switch
This doesn't mean companies are abandoning closed models entirely. About half of the executives surveyed said they'll keep closed systems for harder jobs where quality matters more than price. That makes sense: for complex reasoning, high-stakes decisions, or tasks requiring the absolute best output, the premium may still be worth it.
But for routine work—drafting emails, summarizing documents, basic coding assistance—the 'good enough' option is increasingly attractive. Procurement teams, under pressure to control costs, tend to standardize on what works at the lowest price. That dynamic is exactly what open-weight models are designed to exploit.
What it means for investors
For investors, the shift has clear implications for the companies selling AI models. If enterprises are already overspending, they're likely to push back on prices. That could pressure revenue per token and weaken pricing power for closed-model providers. Anthropic, for example, warned in its IPO prospectus that customers may shift to cheaper options—a risk that's now looking more concrete. The company has also been launching faster, cheaper models as it prepares for its public debut, a sign that even the big players are feeling the competitive heat.
But the spending doesn't disappear—it just moves. Because open-weight models can run 'anywhere,' more budget is likely to flow toward deployment, security controls, and integration inside existing software. That's work that systems integrators like Capgemini get paid to deliver, which could be a tailwind for consulting and IT services firms.
For everyday investors, the takeaway is that the AI boom is entering a new phase. The early days of 'spend whatever it takes' are giving way to a more disciplined approach. Companies are still investing heavily in AI, but they're becoming more selective about where the money goes. That's a healthy sign for the long term, even if it means some near-term pain for the most expensive model providers.
As the market digests these trends, keep an eye on how AI companies respond. Some may try to cut prices to stay competitive, while others may focus on differentiating their models for high-value tasks. The winners will likely be those that can offer the best performance at a price that makes sense for budget-conscious enterprises.
