What’s Going On With AI Today? Disruptions, Market Shifts & Growing Concerns

Things move fast in generative AI, but the sector’s breakneck start to 2025 has been something to behold. In the two weeks since China’s DeepSeek smashed its way onto the scene, a series of furious reactions and knock-on effects have called into question some of the biggest assumptions about AI creation, investment, and pricing.

With its (apparently) cheap and cheerful development and (sort-of) open-source model, corporate customers are turning DeepSeek curious while investors in AI stocks are rethinking their portfolios.

In short, we’re watching tech’s hottest market being re-configured in real time – and what a show it is! Here are the highlights so far.

Table of Contents Table of Contents

1. Key Takeaways
2. Low Cost, Open-Source AI?
3. GenAI Doesn't Have to Cost the Earth
4. AI Firms Are Getting Hammered
5. People Have … Concerns
6. The Bottom Line
7. FAQs
8. References

Show Full Guide

Table of Contents

1. Key Takeaways
2. Low Cost, Open-Source AI?
3. GenAI Doesn't Have to Cost the Earth
Show Full Guide
4. AI Firms Are Getting Hammered
5. People Have … Concerns
6. The Bottom Line
7. FAQs
8. References

Low Cost, Open-Source AI?

The buzz around a new and inexpensive Chinese competitor to ChatGPT began to build in December 2024 when DeepSeek placed its V3 model on GitHub and told testers to have a go. It quickly jumped to the top 10 in the University of Berkeley’s Chatbot Arena, a platform where computer science researchers rate different AI chatbots. Then in January the hedge fund-backed startup launched its high performance R1 large language model (LLM) for more sophisticated problem-solving – to rave reviews.

Deepseek R1 is one of the most amazing and impressive breakthroughs I’ve ever seen — and as open source, a profound gift to the world. 🤖🫡

— Marc Andreessen 🇺🇸 (@pmarca) January 24, 2025

Such high praise from the man who invented Netscape and one of Silicon Valley’s leading AI VCs kicked off a market and media hailstorm that continues to this day.

GenAI Doesn’t Have to Cost the Earth

DeepSeek’s claim to fame is being roughly equal to established GenAI chatbots at a much lower price per token. The startup’s development model also suggests new AI tools can be created faster, and for less.

Some notable breakthroughs allowed DeepSeek to wring more computational power out of basic model Nvidia chips, meaning its LLMs can be trained faster, with reduced computing requirements, and without the massive dev costs racked up by Silicon Valley tech giants.

The fact that these innovations came out of China has also put paid to the idea that US export restrictions have dented the country’s ability to innovate and excel.

A team of Berkeley researchers now claim to have replicated R1 for a cost of about $30, ‘scaling’ development costs using the number puzzle algorithm from popular game show Countdown.

We reproduced DeepSeek R1-Zero in the CountDown game, and it just works

Through RL, the 3B base LM develops self-verification and search abilities all on its own

You can experience the Ahah moment yourself for < $30
Code: https://t.co/B2IsN1PrXV

Here's what we learned 🧵 pic.twitter.com/43BVYMmS8X

— Jiayi Pan (@jiayi_pirate) January 24, 2025

AI Firms Are Getting Hammered

Can GenAI development really be so easy? Inquiring tech investors want to know.

On January 27 – immediately following Andreesen’s tweet – they crashed the AI market by close to a trillion dollars.

As DeepSeek rocketed to the top of the app store charts, Nvidia stock sank by 17% or nearly $600 billion dollars, roughly equivalent to losing both Nike and Netflix in a single day. It sank further in the following days and was just above its January 27th low at the time of writing.

Markets are still pondering what comes next. DeepSeek has said it spent $5.6 million training R1 versus the $100 million to $1 billion range suggested by some leading AI developers.

If an export-restricted Chinese company using out-of-date chips can create a high-performing LLM on the cheap, demand for expensive high-end AI processors may be a lot less than forecast.

There are also questions around the startup’s use of ‘distillation’ as a way to shortcut model training, basically asking an existing tool like ChatGPT a zillion questions and analyzing the answers.

Incumbents naturally have objections to this technique, but it still undermines Silicon Valley’s approach to AI. The cat is out of the bag.

Assuming distillation is found to be legally defensible, does it still make sense for tech’s leading AI developers, who cumulatively hemorrhage billions of dollars every year on development costs and show no signs of being profitable anytime soon, to keep to the same path?

Distillation opens the door to a new AI startup boom that could wipe away even more of their market value.

OpenAI stole from the whole internet to make itself richer, DeepSeek stole from them and give it back to the masses for free

I think there is a certain british folktale about this https://t.co/ERdt1odGF1

— Baithoven 🎣 (@SuspendedRobot) January 23, 2025

People Have … Concerns

As good as DeepSeek is, critics have noted some worrying traits. One is the tool’s seeming inability to process prompts on topics relating to Chinese government policy or history:

😆 pic.twitter.com/Rr4iJ1mzGz

— Marc Andreessen 🇺🇸 (@pmarca) February 1, 2025

And there are the usual fears about sourcing key technologies (advanced processors, telecommunication infrastructure, etc.) from Chinese vendors that might hand Beijing access to Western data and intelligence. Some governments have decided the risk is simply too great to ignore.

Italy and Taiwan have imposed outright bans on DeepSeek while a rising chorus of Western politicians in the UK, Europe, South Korea and Australia have warned citizens against using it.

DeepSeek’s chatbot app has already been delisted on the iOS and Google Play app stores in Italy and Ireland, with more expected to follow suit.

In the US, MAGA-adjacent Senator Josh Hawley has tabled legislation that would prohibit US companies from buying, selling, or using Chinese AI technologies, effectively banning DeepSeek in the US on national security grounds.

Security vendors have heard the call and started compiling lists of potential DeepSeek vulnerabilities. A report by Enkrypt AI found that R1 was “11 times more likely” to generate harmful outputs (e.g., untrue or biased answers or buggy code) than OpenAI’s recent o1 release.

Meanwhile testers have got busy uncovering the flaws in DeepSeek’s reasoning. A study published in Nature found that R1 drops the ball on some fairly simple asks; for example, counting the number of US state names that contain the letter W.

It’s not clear yet if these issues are worse than those already found in established tools like ChatGPT. The researchers noted that similar fails bedevil other large language models, too.

The Bottom Line

Silicon Valley is still fashioning its response to DeepSeek’s challenge and getting to grips with its implications. Microsoft and OpenAI announced on January 29 that they were investigating the possible exfiltration of MS data from a group linked to DeepSeek – before suddenly making it available on Azure a day later.

Open AI’s Sam Altman also seems to be of two minds. On January 31, he told a Reddit AMA that the answer to DeepSeek might be to give ChatGPT away for free.

For now, watch this space. The latest great AI transformation is underway and shows no signs of slowing down.

FAQs