Remember a year ago? The release of the Chinese AI model DeepSeek triggered a sharp sell-off in AI-related stocks, including Nvidia. DeepSeek introduced architectural and training optimisations that appeared to improve efficiency significantly. The obvious conclusion was that higher efficiency would translate into lower demand for compute capacity. The media quickly amplified this view and shares of Nvidia and other AI infrastructure companies fell sharply in the weeks that followed.
Not so fast, we said in an investment insight we published this time last year. These improvements are part of the natural progression of technological development. Efficiency does not necessarily reduce demand for compute capacity. On the contrary, it can expand it. Typically, lower costs and better performance increase adoption, not suppress it. That is what happened. Compute demand did not collapse. Instead, these stocks rebounded and ultimately flourished. It happens all the time in equity markets. Negative headlines spread quickly and investors shoot first and ask questions later.
Fast-forward a year and a similar dynamic is playing out again. We have lost count of the reports predicting gloom and doom. Recent advances, particularly from Claude, the generative AI tool created by Anthropic, have triggered a broad decline in share prices across software and many other sectors. The selling has been largely indiscriminate. Volatility like this is uncomfortable but it can also create opportunities for long-term investors.
The progress made by Anthropic and other large language model providers is impressive. They continue to improve at an astonishing pace and we are excited about what may be possible as models are increasingly trained and deployed on Nvidia’s Blackwell architecture.
We have been positive about AI’s long-term prospects for more than three years and continue to believe it represents the next industrial paradigm. Yet there is a curious paradox in the market today.
Two Conflicting Bear Cases
At the end of last year, the dominant concern was that AI infrastructure had become a bubble. Critics said that hyperscalers were overspending on capital expenditure without a clear path to adequate returns. Questions were raised about the accounting treatment of depreciation schedules and circular financing structures. The underlying claim was straightforward: enterprise adoption of AI is limited, so investments are excessive.An oft-cited MIT survey from last summer suggested that 95% of enterprise AI projects had failed. The implication was that leading model providers such as OpenAI and Anthropic would struggle to justify their investment plans, leaving hyperscalers with underutilised data centre capacity. In short, the addressable market was deemed insufficient to support the current high level of spending.
Read Full Article
