AI Reflections: DeepSeek’s Lesson – A Playbook for Assessing the AI Cycle

Summary

• The ‘DeepSeek moment’ selloff earlier this year was primarily driven by the unwind of crowded momentum stocks and de-grossing events, rather than the popular narratives at the time concerning DeepSeek’s model development or fears about GPU requirements.
• AI demand continues to soar, evidenced by an exponential increase in token processing (e.g., Google processing nearly 1,000 trillion tokens within a month) and record numbers of firms planning AI adoption
• Our framework for evaluating AI cycle risks identifies four key underpinnings of Hyperscaler capex spending: AI stock valuations, fears of limited ROI, slowing model improvement, and infrastructure oversupply.
• For now, investor concentration in AI stocks remains the largest risk. However, the broader AI upcycle remains robust due to strong medium-term metrics.

Recent Selloff Echoes Early Stages of the ‘DeepSeek Moment’

Tech stocks sold off last week, with Tuesday proving particularly brutal. The FT reported a muted reception to GPT-5’s launch may have catalysed the decline, while others pointed to an MIT study suggesting 95% of AI pilots deliver little-to-no value.

The rapid change in narrative partly echoes the ‘DeepSeek moment’ from earlier this year when tech stocks sold off rapidly, but not for the reasons many might think.

An ex-post review of the event reveals an unwind of crowded momentum stocks primarily drove the selloff. That is because the AI theme was extremely crowded in January, as was positioning within large-cap US equities that held up despite a rising US 10-year yield.

By February, declining yields triggered a rebound in crowded shorts, forcing a de-grossing event in the most crowded names and driving AI stock sales.

Several other events occurred at the time feeding the narrative including:

• Concerns over slowing AI model development as pre-training scaling laws reach their peak
• Satya Nadella highlighting fears of potential AI infrastructure oversupply but was ‘good for his $80bn’.
• DeepSeek’s ‘v3’ model launched in December, followed by its ‘r1’ reasoning model before Davos, which Nadella praised for making solid progress.

Despite awareness of DeepSeek during Davos, the selloff only accelerated once the model gained broader adoption a week later – with a fierce reaction.

A basket of the 12 most prominent AI stocks fell from a 40% premium to the S&P 500 in January to trading at par by March. They have since recovered back to a 25% premium.

Semiconductor stocks also saw broad-based declines, with nearly 70% trading below their 200-day moving average, a metric we highlighted as typically consistent with semi-cycle downturns and predictive of broader slowdowns. By ‘Liberation Day’, almost all semiconductor stocks traded below their 200dma. Today, 70% trade above their 200dma.

In both cases, the market sent false signals to participants about the nature of the ‘DeepSeek moment’, made clearer when examining the popular narratives at the time.

Among the most egregious claims surrounded Nvidia. Fears erupted that their expensive GPUs were no longer required to develop SOTA models, and the investment case is less valid as compute resources would shift away from pre-training to inference. Both proved wrong as it became clear DeepSeek had access to many H100s and H20s. Lastly, Nvidia’s new Blackwell chip (or a 1.2-ton rack more specifically), which was in development at the time, was designed with inference in mind!

Granted, Nvidia is not AI, but the key takeaway is that the prevailing narrative was mostly wrong—and in ways that were identifiable ex-ante! This raises a key question: if AI positioning was less crowded, would the ‘DeepSeek moment’ have occurred at all? I’m not so sure!

AI Demand Continues Soaring

Where do we stand today?

First, AI demand continues rising sharply, in some cases exponentially. The emergence of agentic AI, advanced reasoning models (outlined here), and multi-modal capabilities explains much of this surge. These technologies require orders of magnitude more token processing than prior generations and therefore need far greater compute capacity than earlier projections anticipated. We see this in two metrics:

• Token processing explosion. Google processed 480 trillion tokens in May 2025, up from just 10 trillion twelve months earlier. This figure doubled within a month, approaching 1,000 trillion tokens.
• Rising business adoption. The Census BTOS survey shows record numbers of firms planning AI adoption over the next six months. Over the past year, firms planning to use AI for goods and services production increased by more than three percentage points to over 12% of surveyed companies. Notably, this survey only covers production use but not broader operations such as back office or finance.

Second, DeepSeek’s market share among consumers has shrunk since jumping to second place after its launch six months ago. Today, DeepSeek lags in fourth place behind OpenAI’s ChatGPT, Google’s Gemini, and xAI’s Grok. Meanwhile, Google searches for ‘DeepSeek’ are now at a fraction of their January highs, while ChatGPT continues reaching new highs. This highlights ChatGPT’s dominance over the consumer market, and the temporary nature of DeepSeek’s popularity.

Lastly, despite their rising capital intensity, ROICs at the largest Hyperscalers continues to rise suggesting these businesses are generating more profit for each dollar they invest. This would not occur if AI investments produced no revenue or lost money. A rising ROIC also helps to explain why the market maintains elevated multiples for these stocks.

A Framework for Evaluating Where We Are in the AI Cycle

Perhaps the key lesson from DeepSeek is that investors need a framework for pinpointing true AI risks from narrative-driven head fakes. Afterall, the stakes are high and rising. It’s not just because nearly 40% of the S&P 500 is now partly or entirely AI-driven but also because of the ~$400bn in capex Hyperscalers plan for next year, with further increases expected in 2027.

Therefore, even a slight reduction in spending versus expectations can trigger major stock price reverberations given the elevated expectations baked into these names. We identify four underpinnings to determine when Hyperscaler capex may be at risk:

• Positioning/AI stock valuations. As this episode has shown, lofty valuations and crowded positioning create the tinder required to fuel large selloffs. Therefore, monitoring positioning can help avoid drawdowns. Also, lower stock valuations can be seen as the market rebels against ever more capex, which we have not seen yet!
• Limited ROI – the perennial fear. What if all this spending leads to nothing, or what if AI is just a fad? However, Hyperscaler ROIC continues rising while there is also more anecdata suggesting AI use cases are rising. Our favourites include: WMT’s four ‘super agents’, Morgan Stanely modernisation of legacy code, KR delivering a personalised experience for customers, and DHL’s voicebot.
• Slowing model improvement. Differentiating today’s AI investment cycle from railroads or fibre is the technology’s continues improvement. However, AI investment has a far shorter expected lifespan. The prospect of superior models with increasing capabilities fuels further investment in chips, data centres and infrastructure. Despite GPT-5’s frosty reception (which we believe was largely misguided), the broader trend remains positive, though data availability concerns persist.
• Infrastructure oversupply. Matching datacentre and AI capacity buildout to demand will prove incredibly difficult. Demand is rising rapidly, making accurate forecasting nearly impossible, while chips, power, water and other factors constrain supply. Planning for this demand is exceptionally challenging, even small forecast shortfalls can dramatically shift supply-demand dynamics, pressuring AI service pricing. Today, the four largest Hyperscalers report AI demand exceeding supply, supporting the investment case for next year (see below).

‘In the rapidly evolving world of generative AI, AWS continues to build a large fast growing triple digit year-over-year percentage, multi-billion dollar business with more demand than we have supply for at the moment.’ – Andy Jassy, Amazon CEO

‘Revenue from Azure AI services was generally in line with expectations. And while we brought additional data centre capacity online this quarter, demand remains higher than supply.’ – Amy Hood, Microsoft CFO

‘In Cloud, as I mentioned, the demand for our products is high as evidenced by the continued revenue growth and the Cloud backlog of $106 billion. While we have been working hard to increase capacity and have improved the pace of server deployment, we expect to remain in a tight demand-supply environment going into 2026.’ – Anat Ashkenazi, Alphabet CFO

‘We actually, currently are still waving off customers from – or scheduling them out into the future so that we have enough supply to meet demand. This is a situation that we have not seen in our history, and the numbers themselves are so enormous, and the reason is because our technology is different.’ – Safra A. Catz, Oracle CFO

Table 2: Framework for Assessing Risks to the AI Cycle

For now, investor concentration in AI stocks poses the largest risk. Whether GPT-5 performance fears materialise remains unclear, but conditions for a larger selloff appear present. However, given strong medium-term metrics, including rising AI use cases, adoption and the favourable supply-demand mix, we should have more confidence in advance that it’s a blip within a stronger uptrend that remains intact.

(The commentary contained in the above article does not constitute an offer or a solicitation, or a recommendation to implement or liquidate an investment or to carry out any other transaction. It should not be used as a basis for any investment decision or other decision. Any investment decision should be based on appropriate professional advice specific to your needs. You are not permitted to publish, transmit, or otherwise reproduce this information, in whole or in part, in any format to any third party without the express written consent of Macro Hive. This includes providing or reproducing this information, in whole or in part, as a prompt.)