The initial panic triggered by DeepShitSeek is starting to subside as more information comes to light. Recent reports indicate that the true cost of their training process was closer to $500 million, not the $6 million initially claimed. Additionally, it appears that DeepSeek’s LLM performance isn’t as groundbreaking as first suggested, with revelations that the company used OpenAI’s foundation model (via distillation, which some are calling a form of piracy) and maybe even its dataset to develop its own solution.
As we pointed out in a previous post, while DeepSeek’s “breakthrough” brought forward some innovative methods for improving LLM training efficiency, these approaches are freely available, given that the model is open source. This has sparked intense interest from top engineers at hyperscale companies, who are currently exploring these techniques in so-called “war rooms.”
The continuous development of new machine learning algorithms is one of the three key pillars of AI, alongside infrastructure and data. On the infrastructure front, significant disruption is unlikely, as trends in computing power, networking speed, and memory storage follow well-established roadmaps. The same holds true for software, as major AI libraries like Pytorch and TensorFlow are open-source and don’t provide a major differentiation edge. In terms of data, hyperscalers have largely exhausted available corpora and are now turning to synthetic data to enhance their models.
As such, the biggest breakthroughs are likely to come from the modeling side of AI. The last major surprise came in 2017 with the introduction of Transformer models, which sparked the GenAI wave we’re experiencing today. Moving forward, we can expect more “DeepSeek moments,” where lesser-known companies or researchers unveil revolutionary model architectures, improved training processes, or better initialization methods just to name a few. These disruptions are key to innovation and will help boost AI’s accuracy, efficiency, and speed, ultimately expanding its reach across all sectors of the economy. This ongoing progress will keep demand strong for GPUs, memory, networking equipment, and data center infrastructure.
This statement was confirmed by the latest quarterly results from Microsoft and Meta, both of which reaffirmed their massive capex guidance and showed that DeepSeek won’t derail data center growth. Meta projected full-year 2025 capex in the range of $60-$65 billion, a 60% increase from last year, and highlighted that significant infrastructure capex would provide a strategic edge over time.
At Microsoft, whose earnings were globally mixed, AI-related business was a source of strength with revenue surging to over $13 billion in annual run rate, up 175% from the previous year, and a projection to invest more than $80 billion in data center infrastructure in FY25.
On the semiconductor manufacturing/equipment side, the ASML’s CEO also reacted to DeepSeek, noting that while LLMs may require less computing power during development, high-performance chips remain essential for running AI applications, highlighting the ongoing need for powerful computers and high-bandwidth memory. This comment confirms TSMC’s – the world’s largest contract chipmaker – view that AI would be the primary driver of its revenue growth in the coming years.
Lastly, reports about SoftBank being in talks to invest up to $25 billion in OpenAI are another reassuring sign for the broader AI industry and confirm our view that the DeepSeek threat to AI capex and US dominance has largely been overblown.
