News: 1738086311

  ARM Give a man a fire and he's warm for a day, but set fire to him and he's warm for the rest of his life (Terry Pratchett, Jingo)

US AI shares battered, bruised, and holding after yesterday's DeepSeek beating

(2025/01/28)


US tech shares, rattled yesterday by the release of a supposedly more efficient AI model by Chinese outfit DeepSeek, appear to have staunched the bleeding, but not recovered.

After [1]losing enough market cap on January 29 to set a new record, Nvidia has pick up by a few percent, though not enough to erase the losses. The same goes for firms like Oracle, Marvell and Supermicro, all of which lost considerable cash and all of which recovered - slightly - overnight. Shares in the businesses dipped again when trading began in the US this morning, but were hovering near their low-points when markets closed yesterday.

Similar chaos hit tech stocks in Japan for a [2]second day as concern over whether all those expensive chips are needed to develop cutting-edge AI: After all, if China can do it [3]on the cheap in the face of [4]sanctions keeping it from getting its hands on the latest and greatest AI chips, why is all that high-end Nvidia hardware necessary?

[5]

Nvidia has explained the advancement as an "excellent AI advancement" in the realm of scaling resources to allow a model to think harder, known as test-time scaling that [6]leveraged "widely available models and compute that is fully export control compliant," the company told The Register in a statement.

[7]

[8]

But don't go thinking Nvidia is admitting defeat: It still thinks you're going to need a lot of its chips.

"Inference requires significant numbers of Nvidia GPUs and high-performance networking," the company said yesterday. "We now have three scaling laws: pre-training and post-training, which continue, and new test-time scaling."

[9]

OpenAI chief Sam Altman also [10]celebrated DeepSeek's new R1 model as "impressive … particularly around what they're able to deliver for the price." Altman said it is "legit invigorating to have a new competitor," but assured backers his company "will obviously deliver much better models" as it works to advance its AGI mission.

[11]DeepSeek isn't done yet with OpenAI – image-maker Janus Pro is gunning for DALL-E 3

[12]DeepSeek limits new accounts amid cyberattack

[13]How Did DeepSeek Train Its AI Model On A Lot Less – And Crippled – Hardware?

[14]Tech stocks tank as US AI dominance no longer a sure bet

"[We] believe more compute is more important now than ever before to succeed at our mission," Altman said. In other words, American AI firms will still be buying Jensen's chips as fast as he can assemble them. That confidence isn't lost on US retail investors, who bought a [15]record amount of Nvidia shares following the company's losses yesterday.

US President Trump also weighed in on the chip stock battering, [16]saying yesterday that DeepSeek "should be a wake-up call" for US tech companies.

David Sacks, Trump's AI and crypto czar, took the opportunity to say DeepSeek proved his boss right in [17]scrapping Biden's AI safety executive order, which he said in a post on X "hamstrung American AI companies without asking whether China would do the same."

"DeepSeek R1 shows that the AI race will be very competitive and that President Trump was right to rescind the Biden EO," Sacks said. "I'm confident in the US but we can't be complacent." ®

Get our [18]Tech Resources



[1] https://www.theregister.com/2025/01/27/tech_stocks_tank_as_us/

[2] https://www.bloomberg.com/news/articles/2025-01-28/japan-s-tech-stocks-extend-slide-as-deepseek-pressures-ai-sector

[3] https://www.theregister.com/2025/01/26/deepseek_r1_ai_cot/

[4] https://www.theregister.com/2023/10/18/america_china_sanctions/

[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Z5lhj1PLBgOPLAjC-o6ZVgAAAFA&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0

[6] https://www.theregister.com/2025/01/27/deepseek_r1_identity/

[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z5lhj1PLBgOPLAjC-o6ZVgAAAFA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0

[8] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z5lhj1PLBgOPLAjC-o6ZVgAAAFA&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0

[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z5lhj1PLBgOPLAjC-o6ZVgAAAFA&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0

[10] https://x.com/sama/status/1884066337103962416?ref_src=twsrc%5Egoogle%7Ctwcamp%5Eserp%7Ctwgr%5Etweet

[11] https://www.theregister.com/2025/01/27/deepseek_image_openai/

[12] https://www.theregister.com/2025/01/27/deepseek_suspends_new_registrations_amid/

[13] https://www.nextplatform.com/2025/01/27/how-did-deepseek-train-its-ai-model-on-a-lot-less-and-crippled-hardware/

[14] https://www.theregister.com/2025/01/27/tech_stocks_tank_as_us/

[15] https://www.reuters.com/technology/retail-investors-bought-record-amount-nvidia-stock-deepseek-rout-2025-01-28/

[16] https://www.nbcnews.com/tech/innovation/trump-china-deepseek-ai-wake-call-rcna189526

[17] https://www.theregister.com/2025/01/21/trump_eliminates_biden_ai_order/

[18] https://whitepapers.theregister.com/



Doctor Syntax

"After all, if China can do it on the cheap in the face of sanctions"

Not in the face of sanctions but because of them. If you force someone to go their own way you shouldn't be surprised if they do. It's amazing how legislators never work out how the populus is going to respond in reality instead of in theory.

BartyFartsLast

Exactly, necessity is the mother of invention and if you cut off access then people develop their own alternatives

David 132

Whoever imposed the sanctions clearly hadn't read Asimov's Foundation . Wherein, if I recall - it's been years - the exiled scientists, forced to work with very limited resources, made ultra-miniaturized power cells and other devices, to the wonderment of another civilisation. Who, rich in resources, had power cells that filled an entire room rather than fit in a hand.

Anonymous Coward

Yep, see Ukrainian drone manufacture...

beast666

All sanctions on Russia should be revoked.

TReko

Sanctions have a long history of failure, from South Africa developing a first world armaments industry in the 1980's to Iran and North Korea now supplying war supplies to Russia in the Ukraine war.

The inference (not the training) for Deepseek's models is apparently done on Huawei GPU chips.

Sanctions are worse than just failures, they backfire.

Paul Crawford

Perhaps it is really showing the world there is less to AI than huge costs would suggest, and some market reality is due to come home to roost?

Anonymous Coward

It shows that brute forcing is not efficient. Surprise!

A ten year old can fully understand speech and text. Just think of how much speech, text,band energy said ten year old has consumed.

American AI needs more speech and text than the combined population of a mid sized country consume in their live times to reach that level of language command. Using more energy than such country uses.

Something tells me that could be done more efficiently.

Anonymous Coward

The way AI works reflects the way its creators think...

How long before...

Mentat74

Hair furor imposes tariffs on A.I. ?

Awwwww

sarusa

So sorry to hear ChatGPT's lost its job to AI.

Oh joy

Oh Homer

An even cheaper way to make humanity redundant.

Re: Oh joy

Strahd Ivarius

It needs to be very cheap to make all Chinese redundant, there are so many of them.

Now waiting to see the Indian version...

Too early...

Dostoevsky

...to say, "Ding, dong, the witch is dead?" I'll be glad when the world outside El Reg realizes stochastic cow manure generators aren't profitable, or even useful.

Re: Too early...

Anonymous Coward

Sorry, but applications like translation of text and speech and writing aids that produce good text are already changing the world.

Students all over the world already recognized that writing "good" prose and correct grammar has become as easy as using correct orthography. That won't go away.

The same is happening in visual arts.

Remember that Gutenberg went bankrupt from inventing the printing press. Commercial failure often follows technical breakthroughs.

Those companies at the cradle of AI will probably all crash and burn. Others still will get rich.

Just like the current internet billionaires did not contribute to its initial development.

Re: Too early...

Anonymous Coward

One of the most useful feature I got with AI is the ability to take notes and summarize when I got an only meeting.

And it makes me laugh each time I get a notification asking me what language is the speaker is using when it is one of my European colleague with a strong accent...

Re: Too early...

Chet Mannly

"Students all over the world already recognized that writing "good" prose and correct grammar has become as easy as using correct orthography."

Those students haven't learned enything except gto get a computer to do their work for them.

And I'd hardly call what GPT produces as good prose. It's grammar is hit and miss as well...

"will obviously deliver much better models"

Pascal Monett

Yes, of course.

But you don't need $500 billion any more, now do you ?

Think harder.

Re: "will obviously deliver much better models"

Anonymous Coward

"Think harder."

PHB: No, develop an AI to think harder.

Re: "will obviously deliver much better models"

Strahd Ivarius

Of course not, the need is now for $1000 billion because you need to be able to automatically translate location names between the MAGA version and the international one by detecting the user's location, nationality and political leaning

Moving the wall...

Jason Hindle

I agree that what Nvidia makes will still be in demand. Last year, the talk was of LLM development hitting a plateau of diminishing returns, or a wall. The compute cost of o1 seemed to bear that out... DeepSeek potentially moves that wall into the distance. And it's FOSS. Anyone can take it, study its guts and improve it. I think it also points to OpenAI in particular being a massive fake it 'til you make it scam.

"Another world, another day, another dawn. "