LONDON—After years of soaring expectations, the AI industry is approaching a moment of reckoning. Despite record investment, investors, executives, and corporate boards are growing increasingly skeptical that AI will deliver significant productivity gains anytime soon.
AI skepticism is driven by several factors. For starters, the cost of using large language models continues to climb, often with few measurable benefits. According to the Silicon Data Token Expenditure Index, LLM spending has doubled since 2025, even though the cost per token has fallen by more than 90% over the past three years.
In other words, AI usage has grown so rapidly that lower token prices have failed to reduce overall spending, leaving companies paying more than ever. Economy-wide, the Federal Reserve Bank of Atlanta estimates that firms will increase AI spending by 50% in 2026, to $280 billion.
This spending surge has exposed a fundamental asymmetry. As companies come under growing pressure to justify rising AI costs, LLM providers have become increasingly dependent on token consumption to sustain revenue growth and support lofty market valuations.
The obvious question is when—and if—these investments will begin generating meaningful productivity increases or cost savings. A 2025 MIT report found that 95% of generative AI pilot programs, including many intended to accelerate revenue growth, have failed to achieve their stated objectives.
Another source of concern is the growing risk that companies deploying AI tools may inadvertently hand over their proprietary intellectual property, data, and business know-how, accumulated over decades, to major model providers like OpenAI and Anthropic. While licensing agreements and contractual safeguards limit exposure, LLM providers may nevertheless gain access to information that reveals how companies create, price, and capture value.
The dispute between Anthropic and Figma illustrates this tension. Figma partnered with Anthropic to develop AI-powered design assistants, but the relationship came under strain when reports emerged that Anthropic was developing a competing tool, which it subsequently launched as Claude Design.
This is not a new phenomenon. History is replete with examples of technology companies leveraging their scale, capabilities, and market position to enter adjacent markets, most notably Microsoft’s displacement of Lotus 1-2-3 and WordPerfect with Excel and Word in the 1990s. Companies would be wise to keep that in mind as they deepen their reliance on AI.
Beyond such commercial risks, AI is also likely to impose broader economic and societal costs over the longer term. These include rising demand for electricity, water, and data-center infrastructure, as well as growing job displacement and inequality as AI automates an ever-larger share of routine work.
Addressing these challenges will require greater public investment and, ultimately, higher taxes on businesses, potentially including new forms of social support such as a universal basic income. Many business leaders fear that the private sector will increasingly be held responsible for AI’s unintended consequences, creating pressure to raise corporate taxes.
Taken together, these concerns help explain why many business leaders are adopting a more cautious approach to AI adoption. Palantir CEO Alex Karp recently captured this sentiment in an interview with CNBC, accusing OpenAI and Anthropic of “stealing the weights and alpha of my business” through token-based pricing.
Against this backdrop, business leaders should consider four steps as they develop their AI strategies. First, companies should review and restructure AI contracts so that model providers have more skin in the game. If LLM providers are confident they can deliver billions of dollars in productivity gains and cost savings, they should be willing to tie at least part of their compensation to the value they create rather than rely solely on usage-based pricing.
Business leaders should also continually compare the cost and time required for AI to complete a task with the cost and time required for humans to perform it, using clear operational and financial metrics to assess AI’s economic value.
Second, companies should avoid becoming overly dependent on any single model provider. Many firms are already experimenting with multiple models, assigning different providers to different functions or deploying AI tools within operational sandboxes before rolling them out company-wide.
More broadly, companies should consider open-source alternatives. Many of today’s leading open-source models have been developed by Chinese firms and arguably outperform their Western counterparts. They also come at a fraction of the cost. Token prices for MiniMax M3 and DeepSeek V4 Pro, for example, are less than one-tenth of those for Claude Opus 4.8. That is why companies like DoorDash and Airbnb are turning to Chinese models. For many Western companies, however, geopolitical and cybersecurity concerns are likely to outweigh those advantages.
That brings us to the third step business leaders should take. Greater reliance on AI will inevitably require stronger cyber defenses.
Beyond the cybersecurity risks associated with today’s LLMs, advances in encryption, decryption, and quantum computing are likely to make cyber threats more frequent and severe, accelerating the fragmentation of the global technology ecosystem and making it increasingly difficult to operate across borders.
These threats are already attracting the attention of policymakers. In its latest Financial Stability Report, the Bank of England warned that rapid advances in frontier AI capabilities have heightened cyber and operational risks, increasing the likelihood of faster, more coordinated disruptions.
Finally, to justify the enormous sums being invested, business leaders must set timelines for seeing clear evidence that AI is delivering on its promise. With global AI spending projected to reach $2.5 trillion in 2026 and corporate token bills up 320% since 2023, the pressure to prove AI’s economic value is bound to increase.
If that evidence fails to materialize, growing investor impatience could trigger a dramatic repricing of AI-related assets. The longer expectations outpace results, the greater the risk of a painful market correction.
Dambisa Moyo, an international economist, is the author of Edge of Chaos: Why Democracy Is Failing to Deliver Economic Growth—and How to Fix It (Basic Books, 2018).
Copyright: Project Syndicate, 2026.
www.project-syndicate.org
