Will DeepSeek make AI industry worthless?

• DeepSeek made waves in late December by launching a free, open-source large language model, shaking investors’ confidence in the sustainability of demand for AI chips

• We expect the debate to persist until Nvidia’s earnings release

• The open-source model could disrupt the industry landscape by making the inference process more efficient and eventually encourage wider adoption in AI usage

• With the latest introduction, more training and inference process is expected from now on in order to grab the first-mover market share, and subsequently support GPU demand.

• By adopting similar cost-effective techniques, Big Tech firms will also be able to get more out of their AI capex, translating into higher profit margins and higher ROI for investors.

• With the sharp drawdown overnight, the upside potential of the chip sector has become more appealing. We reaffirm our positive view towards semiconductors and assign a target price of USD 311 for SMH.

• Although PNQI has gone up recently, the upside potential of Digital Economy remains attractive with our target price of USD 61.

US technology firms faced significant declines after Chinese startup DeepSeek raised concerns about competitiveness in AI and the US' leadership in the sector.

DeepSeek made waves in late December by launching a free, open-source large language model (DeepSeek V3), reportedly developed in just two months for under USD 6 million. The latest open-source DeepSeek R1 that utilises multi-stage training and cold-start data before reinforce learning phase has created a major outrage in the global technology landscape.

Simply put, DeepSeek R1 claims to be able to beat other chatbots like OpenAI’s ChatGPT or Meta’s LLAMA, with a much lower cost of training. This development has sparked debates over the massive investments big tech companies have been pouring into AI models and data centres.

Figure 1: Benchmark performance of DeepSeek R1

Source: GitHub, DeepSeek, iFAST Compilation, 

Data as of 27 Jan 2025.

We expect the debate to persist until Nvidia’s earnings release

DeepSeek has demonstrated that through algorithmic and architectural improvements, it’s possible to achieve similar results using older GPUs and at significantly lower costs.

This poses a challenge to the rationale behind the rapid expansion of AI capital expenditures and it might even be a more serious issue than the recent debates around the "scaling law" plateau.

The scaling law plateau debate only suggested that the marginal returns of adding more GPUs for pretraining were diminishing, which is difficult to prove in the short term and easy to refute. For instance, scaling laws still hold for fine-tuning, and companies like Meta and Grok are currently pretraining large models on 100,000 GPUs.

However, this time, DeepSeek has definitively achieved results under constrained resources.

Given that the previous scaling law plateau debate, combined with the discussions about ASICs versus general-purpose GPUslasted over a month (from November 14 to December 20), there’s no reason to believe that the scepticism raised by DeepSeek’s findings will dissipate in just a few days.

Having said that, investors are eyeing Big Tech earnings (with AMD expected to launch by coming Tuesday while Nvidia is targeted to release by end-Feb), and the reply from these companies on overcoming this issue will be much more crucial. While GPU demand might slightly moderate (relatively to the skyrocketing growth in the past 1 year), we posit the incoming growth is still massive and investors should continue investing in it.

Table 1: Incoming Big Tech earnings calendar

Company	Date
Microsoft	1/29/2025
Meta	1/29/2025
ASML	1/29/2025
AMD	2/4/2025
Alphabet	2/4/2025
Qualcomm	2/5/2025
Nvidia	2/26/2025
Broadcom	3/7/2025 (expected)
TSMC	4/18/2025 (expected)

Source: Bloomberg Finance L.P., iFAST Compilation. Data as of 27 Jan 2025.

Read more: Why 2025 and the years that follow will be even better for chipmakers

Yes, more efficient way. No, a high-end GPU is still needed.

We believe the drawdown was mainly triggered by risk-off actions following a recent rally, especially with big tech earnings and the Fed’s latest monetary path ahead. Investors’ focus has centered around whether there will be a near-term mismatch between market expectations on AI capex and computing demand, in the event of significant improvements in cost/model computing efficiencies.

Yes, we think the open-source model could disrupt the industry landscape by making the inference process more efficient and eventually encourage wider adoption in AI usage such as edge computing. As developers gain greater access to robust AI models without the barrier of expensive licensing fees, we expect this could accelerate the integration of AI into a broader range of devices and applications. This, in turn, could create new opportunities for AI-driven solutions in industries such as healthcare, manufacturing, and smart cities, where real-time processing on local devices becomes increasingly important.

No, we do not think the introduction of DeepSeek will entirely transform the industry and make high-end GPUs less needed. The inference cost of training DeepSeek does not include data scraping and training costs, hence it is not an apple-to-apple comparison to OpenAI which spent hundreds of millions of dollars to train their model from scratch. With the latest introduction, it has put the industry players now standing at the same starting line, where more training and inference process is expected starting from now in order to grab the first-comer market share, and subsequently support GPU demand.

Figure 2: % of the total generative AI market forecasted by Bloomberg

Source: IDC, eMarketer, Statista, Bloomberg Finance L.P., iFAST compilations. 
Data as of end-Nov 2024.

Big Tech firms will also be able to get more out of their AI capex

DeepSeek has demonstrated that it is possible to train models with greater efficiency and lower costs, and we believe that Big Tech firms will respond by integrating DeepSeek’s innovative techniques such as Mixture-of-Experts (MoE) architecture to improve their own models. It is reported that Meta has already assembled war rooms of engineers to evaluate how the company’s AI, Llama,can be reconfigured to adopt DeepSeek's cost-effective and efficient training and inference methods. By adopting such cost-effective techniques, Big Tech firms will also be able to get more out of their AI capex, translating into higher profit margins and higher ROI for investors.

It is important to understand computing costs does not mean that Big Tech firms will spend less on AI capex. On the contrary, they are likely to leverage their access to high end chips and combine them with DeepSeek’s technology to widen the gap between US and China AI technology.

Lastly, DeepSeek could force companies such as OpenAI and Alphabet to lower the prices of their LLM subscriptions. However, lower prices could also translate to higher revenue for these companies as lower prices are offset by widespread adoption across both consumers and businesses.

Not the end of era, instead, it is just a start

After Japanese low-fuel-consumption cars captured half of the North American market, the crude oil consumption of North American households increased rather than decreased. Similarly, when Xiaomi reduced the price of smartphones from 5,000 to 1,999, the profits of supply chain companies also grew significantly. DeepSeek can be understood as a form of democracy—more players are joining the table.

During the dot-com bubble era, the interest and usage in ‘internet’ products surged dramatically, fueled by the excitement around the new digital frontier. Interestingly, instead of the hype discouraging people due to overvaluation concerns, it actually inspired more individuals and businesses to enter the internet space. The widespread adoption and exploration of web technologies created an ecosystem where innovation thrived, and the advancements that followed—like search engines, e-commerce, and social networking—shaped the modern internet as we know it today.

It was a self-reinforcing cycle: the boom drew talent, investment, and ambition into the field, accelerating progress even after the bubble burst. Much like a "democratization" of the internet, the barriers to entry of AI have been lowered, and the table expanded, ensuring the growth of industries far beyond the previous peak.

NVIDIA will maintain its dominance

The GPU market is undergoing a structural adjustment, but NVIDIA's dominance remains intact. DeepSeek's technological innovations may reduce some GPU demand for training, but the ongoing AI arms race and pursuit of Aritificial General Intelligence (AGI) still depend heavily on supercomputing clusters, where NVIDIA excels in high-performance computing, though its strong CUDA ecosystem remains a key advantage. Although DeepSeek's open-source strategy could attract developers, NVIDIA can maintain its influence through strategic collaborations, such as supporting mixed-precision training frameworks.

Investors should not be overblown by the latest developments, and we encourage them to see this as an opportunity to position themselves in the disruptive tide.

26% upside potential for Digital Economy

The Invesco NASDAQ Internet ETF (NASDAQ: PNQI)’s well-diversified sector allocation—including key areas like cloud services, e-commerce, and digital platforms, makes it a compelling complement for investors seeking growth exposure in addition to semiconductors. Additionally, with advancements in AI efficiency, the holdings could benefit from margin expansion which further strengthens PNQI's value proposition.

Based on a fair PE multiple of 30X applied to 2026 estimated earnings, we maintain our target price of USD 61 for PNQI, which translates into an upside potential of 26%. Although PNQI has gone up recently, the upside potential of Digital Economy remains attractive.

Read more: Market Outlook: Why investors should continue to load up on Big Tech and chip stocks

32% upside potential for chips stock

In our previous article, we mentioned our optimism about earnings alongside concerns about lofty valuation – Investors concerned about valuations may choose to wait for a pullback to buy, though there is no guarantee that one will materialise. Instead, we recommend adopting a dollar-cost averaging strategy while keeping an eye out for sharp pullbacks to make lump-sum investments. This approach ensures that if the sector does indeed go on a run, investors are still able to benefit.

With the sharp drawdown overnight, the upside potential of the chip sector has become more appealing. Based on a fair PE multiple of 24X applied to 2026 estimated earnings, we maintain our target price of the VanEck Semiconductor ETF (NASDAQ:SMH) at USD 311, which translates into an upside potential of 32% (compared to 27% previously).

Read more: Why 2025 and the years that follow will be even better for chipmakers