Confirmed: The End of Easy Wins
One of the most concrete signs of a shift is the law of diminishing returns. For a long time, the primary method for creating more powerful AI was simply to make the models bigger—feeding them more data and using more computing power. This strategy, known
as scaling, led to incredible leaps in capability. However, top AI labs are now finding that simply increasing a model's size no longer guarantees a proportional jump in performance. While new models still inch past benchmarks, the revolutionary breakthroughs that defined the early 2020s are becoming less frequent. The consensus is growing that just adding more data and compute is not a viable path to the next level of artificial intelligence.
Confirmed: The Data Bottleneck
A major reason for this slowdown is that AI has become a victim of its own success. Foundational models from the biggest players have now been trained on vast swathes of the publicly available internet. Researchers are confronting a new reality: we may be running out of high-quality, human-generated data. Using lower-quality or AI-generated synthetic data to train new models introduces the risk of 'model collapse', where performance can actually degrade over time. This data scarcity is a fundamental barrier, forcing a move away from brute-force data consumption toward more sophisticated training techniques.
Confirmed: Hardware and Energy Walls
The physical constraints of AI development are also becoming impossible to ignore. Training state-of-the-art models consumes enormous amounts of electricity, driving up both financial and environmental costs. Furthermore, the hardware itself is hitting a 'memory wall'. Processors can often compute faster than the system can feed them data, meaning incredibly expensive GPUs sit idle while waiting for information. This bottleneck between memory and processing is a critical constraint that slows down development and inflates budgets, making the current path of ever-larger models economically and physically unsustainable for many.
Unclear: Is True Innovation Stalling?
While the returns on scaling are diminishing, it's far from clear that overall innovation has stalled. Many experts argue this isn't an end, but a maturation. The industry is shifting focus from making models bigger to making them smarter, more efficient, and more capable of reasoning. Research is now pouring into new architectures beyond the dominant transformer model, such as state space models and neuro-symbolic AI. It is uncertain whether we are in a temporary plateau before the next big architectural breakthrough or facing a more fundamental, long-term slowdown. The answer will determine the trajectory of AI for the next decade.
Unclear: The Investment Picture
Despite talk of a slowdown, venture capital continues to flood into the AI sector at an astonishing rate. In the first half of 2026, AI companies captured a staggering 86% of all US venture capital investment. However, the nature of this investment is changing. While mega-rounds for a few foundational model companies like OpenAI and Anthropic dominate headlines, many investors are shifting focus to the application layer. They are funding startups that use existing AI to solve specific business problems rather than trying to build the next giant model from scratch. It remains to be seen if this capital will continue to concentrate at the top or spread out to foster a new wave of innovation in specialised AI.
