Three years into the generative AI boom, a growing number of analysts argue that the gap between what the technology was sold as and what it reliably does is about to matter commercially. Their case rests on two predictions for 2026: that enterprise AI agents will fall short of the automation they promised, and that Nvidia’s hold on AI computing will begin to loosen. Both deserve scrutiny, because both are forecasts rather than facts, but each draws on developments that are already visible.
Problems that were supposed to be temporary
When large language models reached mass audiences in 2023, their tendency to produce confident but false statements, known as hallucination, was widely treated as a bug that scale would fix. It has not disappeared. OpenAI’s own evaluations in 2025 found that some of its newer reasoning models hallucinated more often than earlier ones on certain benchmarks, and researchers continue to debate whether the problem can be eliminated or only managed.
The standard safeguard, a human reviewing the output, has proved weaker in practice than in theory. Studies of workplace AI use repeatedly find that people check model output less carefully as they grow accustomed to it, which means errors pass through unexamined.
Agents meet the enterprise
The 2025 pitch to businesses centered on AI agents: systems that would carry out multistep tasks across software with little supervision. Adoption on major enterprise platforms has been slower than vendors forecast, and surveys through the year reported that a large share of corporate AI pilots never reached production, with trust, accuracy and integration costs cited as the main reasons. Several companies have said publicly that they scaled back deployments after results disappointed.
Klarna is the most cited example. The payments company announced in 2024 that an AI assistant was doing the work of about 700 customer service agents, then acknowledged in 2025 that service quality had suffered and began hiring human agents again, saying customers should always be able to reach a person. Market research firm Forrester has also forecast that many organizations will defer part of their planned AI spending as returns prove harder to demonstrate.
Skeptics draw a historical parallel with the 1966 ALPAC report, a U.S. government assessment that judged machine translation not worth further investment and contributed to a decade-long funding freeze. The comparison is imperfect, since today’s models are in commercial use at enormous scale, but the pattern of inflated expectations followed by retrenchment is familiar.
Challenges to Nvidia
The second prediction concerns hardware. Nvidia’s position rests not only on its chips but on CUDA, the software platform that most AI code is written against. In December 2025, Google and Meta introduced TorchTPU, an open source project that gives PyTorch, the dominant AI programming framework, a native backend for Google’s TPU chips. The stated goal is to let developers train and run models on TPUs with the tools they already use, which removes one of the main reasons to stay on Nvidia hardware. Google has also moved to sell TPUs for installation in customers’ own data centers rather than only through its cloud.
Amazon Web Services and Anthropic, meanwhile, built Project Rainier, a cluster of hundreds of thousands of Amazon’s Trainium chips used to train Anthropic’s Claude models, one of the largest AI training workloads to run on non-Nvidia silicon. In China, export controls have pushed companies toward domestic chips. Custom accelerators are generally cheaper and more power-efficient for specific workloads than general-purpose GPUs, which strengthens the economic case for alternatives as AI spending comes under pressure.
None of this means Nvidia’s business is in immediate danger; its revenue continued to grow sharply through 2025 and its chips remain the default for most training and inference. The argument is narrower: if the CUDA software moat erodes, Nvidia’s pricing power erodes with it.
What a reckoning would look like
If the skeptics are right, 2026 will be marked less by new capabilities than by accountability: tighter budgets, demands for measured returns, and a shift from broad promises to narrow, verifiable deployments. If they are wrong, improvements in model reliability and agent tooling will show up in production numbers. Either way, the companies that survive the adjustment are likely to be those making fewer claims and documenting more results.
