OP 2020: Chips go cold in AI winter

Since the depths of the Great Recession, semiconductors stocks have been star performers, with the Philadelphia Semiconductor Index (SOX Index) returning 24% annualised compared to 16.6% for the S&P 500, including dividend reinvestments. 

Outside of widening deployment in consumer products, semiconductor growth has been driven by an explosion of investment in everything from cloud-based infrastructure and cryptocurrency mining to artificial intelligence (AI), big-data processing and “deep learning”. But in 2020, diminishing returns will mean that semiconductors are set to hit the wall.

In fact, a new “AI winter” is likely coming soon, following the two previous AI winters in the 1960s and 1980s, where actual results for semiconductor applications failed to live up to the hype. This time around, the hype bar has once again been set very high, with AI researchers such as Andrew Ng claiming that AI is the new electricity. 

On the physical side, transistor density and clock speeds in semiconductor chips were approaching theoretical limits years ago. Those constraints have meant diminishing returns on R&D efforts. Intel has spent a cumulative $86bn in capital expenditures since 2010, with 2019 at three times 2010 levels. But neither shareholder value nor operating profits have scaled with this huge increase in spending.  

On the application side, “deep learning” algorithms have created the last leg of hype and the number of researchers in deep learning is around double the level seen in 2014. But according to Francois Chollet, one the of leading figures in the deep learning community, the rate of progress in deep learning applications is the slowest in five years. The AI industry also has a scaling problem: it costs more and more in energy to train their ever-larger models, with less and less marginal improvement. Only the largest companies such as Facebook, Microsoft and Google can keep up. A higher bar of entry will eventually dry up venture capital and innovation for new AI ideas. 

Then there is valuation: while earnings in the MSCI World Semiconductor Index are down 17% since 2018, the index hit a new all-time high in October 2019. As reality sets in on the limitations of AI, the SOX Index collapses 50% with deteriorating earnings growth as investments freeze in a new AI winter.