Thematic 4 of 2025 – JUST BLEW UP! The third digital age ramps

Since writing our Thematic paper for 2025 two of the four major themes have already shook markets in 2025. One has been the inauguration of Donald Trump as president of the United States of America and he nationalistic policies the second is the AI revolution taking a massive left hand turn. 

With the release of DeepSeek, the AI story may actually become the biggest theme of the year (big call considering it’s still January). We also need to rethink demand and the flow of funds in the wake of DeepSeek R1 disruption. Because if open-source DeepSeek R1 model does deliver performance comparable to OpenAI’s o1 reasoning model at a fraction of the cost (VentureBeat, Jan 20), it raises critical questions for not just AI, but periphery players in the AI chains as well. Here’s why:

DeepSeek R1 is reshaping investor perceptions of the AI compute investment cycle. Reports suggest that the model achieved competitive performance using significantly fewer computing resources and lower inference costs. These efficiency gains have cast doubts on the future scale of AI semiconductor and equipment investments, leading to selloffs across semiconductor stocks. Look at NVIDIA, Meta and closer to home NextDC and the like.

We also know that in China, DeepSeek v3 has already driven AI compute cost deflation. The R1 model leverages advanced techniques like multi-head latent attention (MLA) and mixture of experts (MOE), enabling more efficient training by breaking workloads into smaller parts. This is something o1 has only just begun doing – but at a higher cost load. While these innovations may not apply universally to all models, they signal a shift that could impact global AI training strategies. 

Now the caveat the tech boffins point out here is that the greater training efficiency is unlikely to reduce overall compute spending in the near term, as improvements typically fuel more inference demand. But and it is a but, economies of scale always come home to roost in the end. The consumer will benefit – the investor needs to get picky – who is going to lead and who is going to fall by the wayside?

Lets look at the Tech Gorillas like Amazon, Google, and Meta, DeepSeek R1’s efficiencies create a mixed outlook. While these firms develop proprietary models it’s important to remember that they also monetise AI services through platforms like Amazon Bedrock and Google Vertex. Lower training costs could reduce operating and capital expenditures, boosting profitability. Amazon benefits from its partnerships with external developers. Google, meanwhile, leverages its Gemini model for growth.

Meta appears best positioned, as its Llama model generates minimal direct revenue, while its announcement of a 54%-66% year on year increase in capex to $60-$65 billion in 2025 demonstrates its commitment to scaling AI capabilities. This follows news of the Stargate project, which is estimated to add $100 billion in incremental cloud-related capex. These moves underscore that demand for high-performance data centres remains robust, even as cost-efficient AI models emerge.

So what about the hardware developers like NVIDIA (NVDA), Broadcom (AVGO), and Marvell (MRVL) that have been savaged by DeepSeek’s cost efficiencies? While concerns about reduced training investment persist, long-term demand for AI compute remains robust and one player will not be enough to change that. Training requirements, particularly for inference workloads, are expected to drive growth over the next 2-3 years and no model as yet can do that without increased hardware. The question will be margins – will hardware margins get squeezed from cost efficiencies?  

Then there is the memory sector, DRAM demand remains steady despite short-term pressures. TSMC, a critical player in AI accelerators, has also faced declines but remains integral to the semiconductor supply chain. So, no real change here from DeepSeek.

Where does this all leave us?

The emergence of cost-efficient AI models like DeepSeek R1 introduces both opportunities and risks. Yes, declining compute costs have mixed implications for the tech sector. On one hand, they alleviate margin pressures for software companies grappling with expensive AI features, potentially accelerating AI integration across product lines. But the risk to this is, lower barriers to entry could intensify competition from agile AI startups, challenging incumbents. In short these innovations could reshape AI economics, the sustained demand for robust infrastructure, driven by broader AI adoption and multi-cloud trends are likely to overawe the negatives for a positive long-term outlook.

Thus investors should focus on companies well-positioned to benefit from these shifts while remaining cautious of near-term volatility.