This Weekend Marks the Launch of DeepSeek AI from China

https://iframe.iono.fm/e/1522210″ width=”100%” height=”126″ frameborder=”0

You can also listen to this podcast on iono.fm here.

SIMON BROWN: I’m speaking with Viv Govender from Rand Swiss. Viv, thanks for joining me. We often discuss tech and AI; you’re my go-to expert on AI. Recently, I explored DeepSeek, which launched from China over the weekend. It’s open-source, and I found it quite impressive, albeit a bit slower. Reportedly, it was developed with just around $6 million.

First off, they claim they couldn’t acquire H100s from Nvidia due to sanctions. Do we trust their assertion that they didn’t utilize H100s for training?

VIV GOVENDER: It’s likely they’ve used more hardware than they let on, mainly because the Chinese have managed to circumvent some of the sanctions imposed by the US. Even with the tightening of sanctions, there is still likely to be some leakage. However, it’s probable that initial training was completed under the previous sanctions.

It’s essential to recognize their advancements, but remember they haven’t developed this entirely from the ground up. They may have leveraged existing frameworks, and intriguingly, this model sometimes identifies itself as ChatGPT, suggesting a foundational model was incorporated during development.

Nonetheless, there’s a significant enhancement in efficiency. It’s worth noting that increased efficiency leads to reduced inference costs. With smaller models resulting in lower inference expenses. Jensen Huang from Nvidia has repeatedly stated that the inference market could vastly outstrip the training market, potentially by a millionfold. This indicates that if they develop something that is over 90% more efficient, the inference market could be approximately one-tenth the size, or in terms of costs, one-tenth what it could be. This signifies a substantial technological leap.

SIMON BROWN: They’ve adopted a unique strategy compared to many others; instead of aggregating more hardware, they focused on optimizing their software approach. I don’t want to delve too deeply into specifics, but as we both know from various articles, they’ve innovatively navigated the software landscape.

VIV GOVENDER: Absolutely – and with a relatively small team of around 200 individuals. This advancement should not be downplayed; it’s a remarkable feat, especially considering its impact on major players like Nvidia and ASML over recent days. It has certainly reshaped market perceptions regarding future chip demand.

SIMON BROWN: That leads to my next inquiry. They likely possess some H100s, but the future quantity remains uncertain. Let’s set that aside. The product is strong. How detrimental is this for companies like Nvidia and ASML? As you mentioned, they’ve faced considerable pressure, being leaders in their domains.

VIV GOVENDER: TSMC’s strategy remains unaffected by these events. However, Nvidia has shed over half a trillion dollars in market cap due to this situation. In the short term, this drop seems justifiable, but I believe that if the technology continues to progress towards greater efficiency, it will ultimately drive increased utilization.

Concerns have emerged regarding the O1 and O3 models from OpenAI, primarily due to the costs associated with every query processed. Even with the $200 subscription, operational costs for these models have been significant.

On the other hand, this new technological advancement may create higher profitability for these companies, leading to increased demand for chips in the long run.

SIMON BROWN: I recall our previous discussions concerning Nvidia’s CEO at CES. His vision encompasses much more than just AI.

What about companies like Meta and Microsoft OpenAI, which, although not publicly traded, have invested heavily in this domain, acquiring chips and establishing large data centers? A new entrant like DeepSeek could disrupt their operations.

VIV GOVENDER: Indeed, but keep in mind that Meta is focusing on creating open-source solutions. A more cost-effective approach would greatly benefit them since their strategy isn’t about selling these products for profit but rather utilizing them for internal purposes.

For Meta, finding cheaper and more efficient AI solutions is a win. In contrast, OpenAI’s business model involves charges for their services. I think Microsoft might not experience severe adverse effects, considering that their future AI ambitions will heavily rely on their infrastructure, especially with their planned $80 billion annual investment in Azure.

While they might face short-term challenges, I believe this could lead to better profit margins for their operations in the future.

SIMON BROWN: You’re right. Their focus is less on whether I’m using ChatGPT or DeepSeek; they just want effective solutions hosted on their servers, which is where they generate revenue.

Thank you, Viv Govender from Rand Swiss, for your valuable insights.