By Tomasz Godziek, portfolio manager of the Tech Disruptors fund at J. Safra Sarasin Sustainable Asset Management
DeepSeek’s breakthrough
DeepSeek, a Chinese AI startup, has just released an AI model that performs competitively against top-tier models such as OpenAI o1 and Anthropic’s Claude 3.5 Sonnet on various benchmarks. Over the past weekend, DeepSeek became the most downloaded app on the App Store. Meanwhile, renowned venture capitalist Marc Andreessen described this development as ’AI’s Sputnik moment’ highlighting its significance. According to the DeepSeek team, training their model cost just $5.6 million, significantly less than the amounts spent by U.S. tech giants. Notably, DeepSeek did not use Nvidia’s most advanced chips, as U.S. export restrictions prevent Chinese firms from accessing them. While some skepticism remains regarding the actual training costs, the expenditure was just a fraction of what U.S. firms have been investing.
On Monday, January 27, DeepSeek’s breakthrough sent shockwaves through equity markets, raising concerns about the billions of dollars spent on training large AI models using extensive GPU resources and massive computing clusters. More importantly, DeepSeek employed several innovative techniques – such as multi-head latent attention and a mixture-of-experts approach – to optimize memory usage and reduce inference costs (cost of using the model). DeepSeek’s breakthrough could mark a pivotal moment for AI investment. While it raises concerns for AI infrastructure providers, it could drive significant new opportunities in software, consumer internet, and inference technologies such as autonomous driving. We explore these trends in further detail below.
What does this mean for investors?
This development has sparked key questions among investors: can AI models be trained more efficiently with fewer resources? Is it necessary to spend billions on GPUs and large-scale computing clusters?
Investor sentiment is now divided into two perspectives: One, AI infrastructure – companies like Nvidia and other infrastructure providers have seen significant gains over the past two years. However, this breakthrough raises new concerns about the sustainability of high infrastructure spending. Two, AI beneficiaries (software & consumer internet) – a potential reduction in AI costs could accelerate adoption, benefiting software and consumer internet companies that integrate AI into their products.
If DeepSeek’s innovations lead to cost reductions across the AI industry, it could exemplify the Jevons Paradox, where lower costs drive higher adoption. Ultimately, this could fuel a new wave of AI investment, creating fresh opportunities, particularly in software and inference technologies.
Impact on different segments of the AI value chain
Nvidia (NVDA): Sell-side analysts continue to defend Nvidia’s dominant position, citing Jevons Paradox. DeepSeek itself relied on Nvidia’s technology, including CUDA, chips, and networking, to train its model. However, bearish arguments against Nvidia are gaining momentum. After the ROI debate last summer and scaling laws end of 2024, DeepSeek now raises questions on price and volume: volume concerns – do we really need as many chips as previously thought? Pricing concerns – do we still need Nvidia’s latest and most expensive chips? While Nvidia’s 2025 revenue appears secure, questions about AI infrastructure spending in 2026 and beyond are growing.
Custom ASICs (application-specific integrated circuits): at this stage we believe that this development is neutral for custom ASIC companies such as Broadcom, Marvell. ASIC chips are generally cheaper than Nvidia’s GPUs -e.g., $10K vs. $30K per chip. These custom chips are widely used for AI inference tasks, such as recommendation engines at Meta and Google. However, two caveats exist. Training vs. inference – Broadcom’s CEO has emphasized that for them ’the big money is in training.’ Since DeepSeek’s breakthrough challenges training economics, Broadcom’s stock declined alongside Nvidia’s. Flexibility vs. efficiency – ASICs excel in well-defined workloads, but AI algorithms are evolving rapidly. Nvidia’s more flexible GPUs still have an advantage for now.
Other AI infrastructure components:
· Memory: DeepSeek demonstrated more efficient memory utilization, putting pressure on DRAM memory -’fast memory’- related stocks -Micron, SK Hynix-, while storage memory -’slow memory’- stocks were more resilient as easier model training means more data to be stored.
· Interconnect/networking – Do we still need large computing clusters? If AI training requires less infrastructure, demand for networking solutions inside datacenters could decline. At the same time, there will be more traffic between the datacenters as more models will be trained.
· Semiconductor equipment (SemiCaps) – While most are not direct AI plays, stocks related to advanced memory packaging saw the steepest declines.
Hyperscalers -cloud providers: the hardest-hit stocks on Monday were companies heavily invested in AI training, such as Oracle or Nvidia. This aligns with Microsoft’s recent strategic shift – CEO Satya Nadella has deprioritized AI training workloads, offloading them to Oracle and SoftBank. Smart move by Microsoft! However, Microsoft still owns OpenAI, which remains a financial burden in the near term. Despite this, hyperscalers -Microsoft, Amazon, Alphabet and Chinese cloud giants- stand to benefit from increased AI inference and broader AI adoption. If research advances allow AI models to be developed with fewer resources, hyperscalers could also see reduced capital expenditures, alleviating investor concerns.
Software: the biggest winner?: software companies are the early beneficiaries of these developments:
· Infrastructure software – databases, search, cybersecurity, etc. – more AI activity should drive greater usage of these platforms.
· Application software – process automation, data analytics, marketing software, etc. – lower AI costs could accelerate adoption, enabling software firms to integrate AI more affordably without significant margin pressure.
The Trillion-Dollar question: did DeepSeek just trigger AI’s second ’ChatGPT moment’?
When ChatGPT launched, investors focused on AI Infrastructure – the ’picks and shovels’ of the AI gold rush. This was the right move. Nvidia’s stock has surged 8x since early 2023, even after yesterday’s drawdown, while other AI semiconductor and infrastructure names have also delivered strong returns. The initial investment sequence in AI follows a clear progression:
1. AI infrastructure → 2. AI software enablers (databases, search, cybersecurity) → 3. Software applications (process automation, data analytics, marketing software, etc.).
In other words, capital has first flowed into AI infrastructure, however since Monday, January 27, inflows may now be shifting towards the second phase.
What’s next? How is JSS Sustainable Equity – Tech Disruptors positioned?
After generating significant returns from our AI exposure since early 2023 – particularly in semiconductors and semiconductor capital equipment – we reduced our semiconductor holdings in H2 2024. Semis now represent 22% of AuM. Conversely, we increased our allocation to software following its underperformance versus semis in H1 2024. Software now makes up 34% of AuM. In response to the volatility spike triggered by DeepSeek on Monday, we executed further shifts:
· Reduced exposure to AI infrastructure, trimming positions in Nvidia and Micron (’fast memory’).
· Increased exposure to AI software enablers, adding to Elastic, CrowdStrike, Amazon and Alphabet.
· Initiated a new position in Ciena – optical networking between data centers -, capitalizing on the recent sharp selloff.
Importantly, we do not place all our bets on a single trend—nor have we ever. The fund is much more than just an AI-focused strategy. We maintain a diversified exposure to multiple structural growth themes, including cybersecurity, process automation, new semiconductor architectures, autonomous driving, quantum computing (added exposure in Q4 2024), etc. Looking ahead, we will continue to closely monitor market developments and adjust our positioning accordingly. As such, we eagerly await upcoming earnings calls from tech companies, particularly management commentary on capital expenditures and DeepSeek’s potential impact.