A few thoughts on DeepSeek

and why you can't trust what you read

Good Morning Team.

For those of you not in the know, there’s a new player in the AI space - termed ‘DeepSeek’ - which has rocketed to number one on the App Store.

I’ve tried this new app, and the bottom line is that for several functions it works better than both OpenAI’s ChatGPT and Alphabet’s Gemini. The question is one of how this app was developed, and which narrative you choose to believe.

But did this really happen organically? The idea that this came to No1 in the App store overnight without massive marketing support is surely a bad joke?

Let’s dive in.

The plucky start-up?

Three things could have happened here:

1. DeepSeek reverse-engineered an API from Llama or another open-source LLM to train their model, effectively leveraging that LLM's training logic without their own original development.

2. They actually used advanced Nvidia H100 GPUs, which could potentially have been funded by the CCP, meaning these resources might have come at no cost to them.

3. They discovered an innovative method to scale an LLM using older, low-cost chips, potentially upending forecasts of trillions of dollars in AI capital expenditures.

The first two explanations are in my view the most likely, but if the third is true, the AI bubble is going to implode.

The key context to understand when making your own mind up is that China can easily access Nvidia chips through Singapore (in the same way as Europe buys Russian commodities through the Middle East).

Nvidia has tried to cover itself in this regard - arguing that ‘shipments to Singapore are insignificant.’ But more than a fifth of billings in its latest quarter were to Singapore, which promptly ships the chips to the Chinese (likely at a healthy premium).

So DeepSeek could easily be a psy-op designed to scare US-based AI investors - many of whom are looking for reasons for the bubble to pop given the premium valuations at stake.

DeepSeek is making the case that it was built in less than two months with around $6 million in capex, outdated chips, and 200 employees.

Interestingly, these people are predominantly younger, many of them PHD students at top Chinese universities - and the vast majority have no academic or work background from overseas. The argument here is that China is now creating graduates with a university education rivalling the older power, which has implications for decades to come.

Or not, because again, this could all be bullshit.

DeepSeek CEO questions

Here’s the thing to understand. There is no such thing as private finance in China - everything is controlled by the CCP - and if you don’t believe that, Google any Chinese CEO who has ever put a toe out of the party line.

And here’s what they want you to believe:

DeepSeek CEO Liang Wenfeng studied machine vision at Zhejiang University, and then launched his High-Flyer quant hedge fund in 2015 (at 30 years old).

This fund now boasts $8 billion in assets under management, and this CEO then decided to build an AI model as a side gig. He buys 10,000 now outdated H800 chips in 2021 and brings over his top hedge fund employees, for their experience in optimising Nvidia chip usage.

DeepSeek launches in 2023, then gets hiring all the top PHDs, creates a LLM on a shoestring that is at least as good as vastly better-funded operators, and open sources its method.

Why open source? To prove it can be done -or to undermine the US tech stocks.

As ever with Chinese apps, censorship regarding Chinese politics is present - ask it about Tiananmen Square and it develops a sudden case of Covid-19 - but with the US pouring $500 billion into AI, NASDAQ investors get a bit scared and we have today’s sell-off.

The problem with this ‘plucky start-up’ narrative is that China has also recently released its New AI Industry Development Action Plan, which will provide one trillion Yuan ($137 billion) to support its native AI industry over the next five years.

Clearly the CCP is not convinced that you can beat the US with a few million quid and a yǒng gǎn attitude, and yet this is the narrative being pushed all over social media.

China has plenty of reasons to lie. And it has done so many times in the past.

Perhaps most importantly, this may just be another salvo in the race to AGI (artificial general intelligence). If China can convince the markets it is winning the race, it may create an investment feedback loop that makes this true, diverting capital from what is arguably an overvalued NASDAQ to undervalued Chinese stocks.

LLMs are impressive, but the real focus is on AGI products that could well change the world. For perspective, I never use any of the current chatbots in my own work - they just aren’t capable of producing high quality work yet - but AGI could change this overnight.

And secondary to this, China may be trying to convince US policymakers that GPU export controls are ineffective in impeding Chinese AI progress.

If you think this is a bit farfetched, then consider the lies of the past:

• The Straddling Bus scam

• Jiaolong Submersible

• The Hanxin Microchip Scandal

• Faster-than-light communications

• Hongxin Semiconductor (HSMC)

• Brain-reading AI

• The 1000km/h Maglev train

• Quantum satellite Micius

• The three-second battery

• Tokamak Reactor

Google ‘em. The country is utterly shameless - yes some of these are real achievements, but every single one has been hugely exaggerated :)

There has been no independent or international audits conducted on the $6 million development cost - which means that believing the Chinese claim of old chips and a shoestring budget is a choice.

For context, ChatGPT has not yet proven to be profitable, and more widely, LLM revenue remains insignificant for the magnificent seven (even if it has an outsized impact on sentiment). There is a massive gap between revenue and AI investment.

It’s worth noting that Meta already released Llama as open-source earlier this year and that more widely, Chinese tech alternatives has historically always been cheaper - but shitter.

Yes, Chinese deflation may hit AI, and the larger operators might be scared of intense price competition - and justifying capex. But there is more to this story than the CCP is letting out.

But what if it’s genuine?

The $500 billion Stargate AI project recently unveiled by Trump may be a massive waste of money.

If what you’re really talking about to develop top-level AI is around 2,000 H800 GPUs (maybe $50 million) and then a few million in training costs, then the capex spend across the Mag7 cannot be justified - and they should crash (because again, AI sentiment is driving record valuations, even though it only represents a tiny portion of revenue).

For Nvidia in particular, this could echo Cisco’s crisis during the dot-com bubble crash.

During that last tech bubble, Cisco dominated networking equipment but suffered after the burst, with the market flooded by second-hand equipment.

Nvidia now dominates the AI hardware market, with a gold rush for GPUs from startups and tech giants. If DeepSeek disrupts hardware demand or triggers start-up failures, a similar flood of second-hand Nvidia GPUs could occur.

For context, AI start-ups with large NVIDIA GPU inventories could collapse, forcing liquidation and driving down new GPU prices. Start-ups renting NVIDIA GPUs could easily fail to generate enough returns, leading to more unused GPUs on the market. And if the Mag7 slow or even stop GPU purchases, there will immediately be thousands of GPUs sitting in excess in warehouses.

Both Cisco and Nvidia scaled up hardware production during speculative booms, risking oversaturation when demand collapses. Cisco depended on dot-com companies and Nvidia depends on high-risk AI startups. I’d venture a guess that these speculative companies may even be being run by the same people…

Cisco’s dominance was followed by years of slow growth. Nvidia could face the same fate if demand weakens. If the Chinese are telling the truth, start-ups and the larger tech companies may well pivot, reducing Nvidia’s relevance - and the titan’s market cap could face a sharp correction - which will likely have a contagion effect on the rest.

This could be the black swan.

I did say the bubble would pop this year.

But it might just be a CCP Trojan Horse filled with Nvidia’s most advanced hardware, masquerading as a breakthrough.