Don’t get too attached to DeepSeek – it’ll never survive in the US

Has everyone lost their collective minds? How did DeepSeek go from a relatively unknown Chinese startup AI chatbot to the hottest commodity on smartphones and tangentially the US stock market?

To be clear, I can’t get on DeepSeek and challenge the R1 model. I’ve been trying for two days and the backlog of signups (and some cyber attacks) has completely overwhelmed DeepSeek’s vaunted servers.

I read John-Anthony Disotto excellent DeepSeek hands on which made it clear that while DeepSeek R1 is good, it’s not a world-changer. Yes, he called it an “excellent free reasoning model” that might make you question why you’re paying for OpenAI’s more expensive and also quite good o1 model. That’s a fair question but it doesn’t account for the US stock market panic or borderline irrational response to the emergence of this model.

This smells fishy

At issue here is that not only is DeepSeek-R1 free, but the company claims it trained its models with cheap chips, and for the cost of just $5.6 million US dollars. In my world, that’s a lot of money. In the realm of AI, it’s nothing.

Meta will reportedly spend $65 billion on AI training this year alone. The need for ever-greater AI model processing power turned Nvidia into the most valuable company in the world (it lost that distinction during the sell-off). Now along comes DeepSeek, saying it did the same thing for, essentially, pennies.

Suddenly every investor wondered if they’d made the wrong bet. So, with almost no information or real proof that DeepSeek and its investors are being transparent and truthful, investors have started pulling their AI dollars from the US stock market.

A matter of trust

As I mentioned, DeepSeek is a Chinese app. Not Chinese in the way TikTok is, with a distant parent – ByteDance – maintaining a thready control over a now mostly US-based counterpart. TikTok, which is run in the US by US employees and stores all its data in the US has until April to find a US buyer or be banned, again, inside the US.

DeepSeek makes no bones about being Chinese. In the App Store, its listing is full of Chinese-language characters. Millions of US users are rushing to use the R1 model and send it their text queries – prompts about their lives, work, and plans.

The Chinese government has full access to all that data and now we’re just willingly handing it to them because someone said this R1 model was the “most amazing and promising AI breakthroughs he’d seen.” Really? The app doesn’t even include multimodal support. It’s just text. How does that add up to amazing when I’m talking to all my AIs and showing them pictures they can instantly interpret and react to?

Listening to one activist investor tout a new product for undefined reasons is never a good idea. Do we know if he has a stake in this company? Do we know if he has an axe to grind with US-based AI companies?

What’s clear though, is that he and DeepSeek do not necessarily have US interests or its consumers’ interests at heart.

This won’t last

Let’s say for the sake of argument, DeepSeek’s R1 does turn out to be a breakthrough. There are hints that the company has far more models up its sleeve (it recently released a generative image model to developers). That still doesn’t wipe out its Chinese parentage and continuing ownership.

It doesn’t matter how good it is; this app will not survive in the current US climate.

It’s reasonable to assume, though that if DeepSeek’s R1 and its other models do turn out to be as groundbreaking as promised, and trained at that incredibly low cost, it will reset the AI industry.

Meta, OpenAI (which is promised a $500 billion investment with the US government to build Project Stargate), Microsoft, Nvidia, and others will have to rethink their AI strategies and figure out how to dramatically cut their own costs.

I’m not convinced though that’s what’s coming. There is no reason to trust DeepSeek and its backers’ claims. I think the truth about the time and financial investment it took them to get here will come out. And then DeepSeek and R1 will go through its own kind of reset.

You might also like