America’s worst fears confirmed after China’s $1 trillion AI nuke, Trump warns of ‘wake-up call’
It’s official — China has leapfrogged the US in the AI arms race. And Donald Trump has reacted with a surprising claim.
It’s official — China has leapfrogged the US in the AI arms race.
The launch of DeepSeek, a low-cost Chinese rival to popular artificial intelligence tools like ChatGPT, sparked a $US1 trillion ($1.6 trillion) meltdown in US tech stocks on Monday amid fears the global AI industry has been turned on its head.
US firms have poured hundreds of billions of dollars into expensive, energy-intensive AI data centres in recent years, and news that a rival technology has been developed at a fraction of the cost sent shockwaves through global markets.
It also sparked allegations of cheating, given the Chinese Communist Party’s long history of using its state power to undercut and force out western competitors in everything from steel to electric vehicles.
Have your say in The Great Aussie Debate. Take the survey here:
Elon Musk, who has invested heavily in Nvidia chips for his company xAI, suspects DeepSeek of secretly accessing banned H100 chips — an accusation also made by the chief executive of ScaleAI, a prominent Silicon Valley start-up backed by Amazon and Meta.
But in a statement on Monday, Nvidia itself poured cold water on those theories, saying DeepSeek was “fully export control compliant” — seemingly confirming America’s worst fears.
President Donald Trump on Monday night said he welcomed the prospect of cheaper AI, while warning it was a “wake-up call” to American companies.
Here’s everything you need to know.
What is DeepSeek?
DeepSeek was spun up in 2023 as an offshoot of High-Flyer, an $US8 billion ($13 billion) Chinese hedge fund led by chief executive and co-founder Liang Wenfeng.
Based in Hangzhou, China, DeepSeek develops open-source AI models, meaning the underlying code is publicly available for use and modification by developers.
DeepSeek has released several versions of its AI assistant since late 2023, but it was the launch of its latest model, DeepSeek-R1, that has thrown the industry into chaos.
Over the weekend, DeepSeek’s chatbot surged to become the most downloaded free app on Apple’s US App Store, displacing OpenAI’s ChatGPT.
How does it compare?
DeepSeek claims its R1 release performs “on par” with the latest model from OpenAI.
“It tops the leaderboard among open-source models and rivals the most advanced closed-source models globally,” the company says.
DeepSeek-R1 has been tested against benchmarks for language and coding capability, showing performance that equals and in some cases exceeds established rivals.
ScaleAI chief executive Alexandr Wang recently told CNBC that DeepSeek is “the top performing, or roughly on par with the best American models”.
DeepSeek-R1 is 20 to 50 times cheaper to use than OpenAI’s o1 model — which also costs $US200 ($320) per month for a ‘Pro Plan’ subscription — according to a post on the company’s official WeChat account.
What makes it different?
DeepSeek claims it developed its latest model at a fraction of the cost that major companies are currently investing in AI development, primarily on expensive Nvidia chips and software.
Most notably, DeepSeek says it used Nvidia’s H800 chips, rather than the most powerful H100s, which are subject to export restrictions introduced under President Joe Biden aimed hampering China’s development of AI.
DeepSeek claims its models are cheaper to operate than its rivals as a result of optimised training and inference processes that require fewer computational resources.
The number that is grabbing the most attention actually accompanied the release of DeepSeek’s previous V3 model over Christmas.
DeepSeek said total training costs of its V3 model amounted to a shockingly low $US5.576 million ($8.86 million).
For comparison, Meta’s total training costs for its Llama 3.1 model were estimated to be around $US120 million ($191 million).
So it’s cheaper all around?
In other words, DeepSeek is both cheaper to use, and cost less to develop.
Gavin Baker, chief investment officer at Atreides Management, said the DeepSeek-R1 hype was “real with important nuances”.
“There were real algorithmic breakthroughs that led to it being dramatically more efficient both to train and inference,” he wrote on X.
“Most important is the fact that r1 is so much cheaper and more efficient to inference than o1, not from the $6m training figure. r1 costs 93% less to *use* than o1 per each API, can be run locally on a high end work station and does not seem to have hit any rate limits which is wild.”
But Mr Baker said the sub-$US6 million training figure was “deeply misleading”, as DeepSeek noted in its technical paper that it did not include “costs associated with prior research and ablation experiments on architectures, algorithms and data”.
“This means that it is possible to train an r1 quality model with a $6m run *if* a lab has already spent hundreds of millions of dollars on prior research and has access to much larger clusters,” he said.
DeepSeek also used a lot of “distillation”, or training an AI model on market leaders like ChatGPT by sending requests and recording the outputs — which is typically against the terms of service but is a widespread practice.
“It is unlikely they could have trained this without unhindered access to GPT-4o and o1,” Mr Baker said, adding it “obviously defeats the purpose of the export restrictions”.
“Why buy the cow when you can get the milk for free?”
Does DeepSeek have H100s?
Nvidia dismissed suggestions that DeepSeek secretly trained its models using banned H100 chips, as suggested by Mr Musk and others.
“DeepSeek is an excellent AI advancement and a perfect example of test-time scaling,” Nvidia said. “DeepSeek’s work illustrates how new models can be created using that technique, leveraging widely-available models and compute that is fully export control compliant. Inference requires significant numbers of Nvidia GPUs and high-performance networking. We now have three scaling laws — pre-training and post-training, which continue, and new test-time scaling.”
ScaleAI chief executive Alexandr Wang claimed on CNBC last week that “the Chinese labs, they have more H100s than people think”.
“My understanding is that DeepSeek has about 50,000 H100s which they can’t talk about obviously because it is against the export controls the US has put in place,” he said.
But such accusations “sound like a rich kids team got outplayed by a poor kids team”, wrote Hong Kong-based investor Jen Zhu Scott on X.
“DeepSeek are not faking the cost of the run,” agreed Emad Mostaque, founder of Stable Diffusion developer StabilityAI.
“It’s pretty much in line with what you’d expect given the data, structure, active parameters and other elements and other models trained by other people. You can run it independently at the same cost. It’s a good lab working hard.”
So how was it done?
Ben Thompson of Stratechery, a closely watched tech newsletter, argued DeepSeek’s innovations in training its AI model “only make sense” if the company was constrained by using the less powerful H800s.
DeepSeek developed a number of innovative techniques to train its AI more efficiently, including selective parameter activation — calling on a few “experts” for a specific task, rather than the whole model — and low-precision computing, prioritising speed at the cost of a small amount of accuracy.
Mr Thompson said the “huge number” of innovations were about overcoming the lack of memory bandwidth in using H800s over H100s, noting DeepSeek engineers went as far as specifically reprogramming the Nvidia GPUs at the hardware level.
“This is an insane level of optimisation that only makes sense if you are using H800s,” he wrote.
“If DeepSeek had access to H100s, they probably would have used a larger training cluster with much fewer optimisations specifically focused on overcoming the lack of bandwidth.”
The MIT Technology Review wrote that the export controls were “driving start-ups like DeepSeek to innovate in ways that prioritise efficiency, resource-pooling, and collaboration”.
Why are stocks crashing?
Shares in Nvidia fell nearly 17 per cent on Wall Street, erasing nearly $US600 billion ($954 billion) of its market value — the most ever for one day on Wall Street.
The tech-rich Nasdaq index finished down more than 3 per cent.
Last week’s release of the latest DeepSeek model initially received limited attention, overshadowed by the inauguration of President Trump on the same day.
Markets only began to fully digest the implications of DeepSeek’s achievement over the weekend, leading to Monday’s bloodbath.
The development is particularly significant given that the AI boom, ignited by ChatGPT’s release in late 2022, propelled Nvidia to become one of the world’s most valuable companies.
The news sent shockwaves through the US tech sector, exposing a critical concern — should tech giants continue to pour hundreds of billions of dollars into AI investment when a Chinese company can apparently produce a comparable model so economically?
“The DeepSeek announcement from China was nothing short of a financial Scud missile aimed at a US market that is much more fragile than most will admit,” said Islanda Capital Investments chief executive Anthony Esposito.
“Unfortunately, I see more downside to come in the coming months.”
Art Hogan, chief market strategist at B. Riley Wealth, described the market’s response on Monday as “shoot first, ask questions later”, noting that some were sceptical of the Chinese company’s assertions.
“Everyone is trying to figure out ‘Can it be believed?’ and ‘What does it mean?’” Mr Hogan said.
Microsoft chief executive Satya Nadella took to social media hours before markets opened to dismiss concerns about cheaply-produced AI, saying less expensive AI was good for everyone.
But last week at the World Economic Forum in Davos, Mr Nadella warned, “We should take the developments out of China very, very seriously.”
Microsoft, an eager adopter of generative AI, plans to invest $US80 billion ($127 billion) in AI this year, while Meta announced at least $US60 billion ($95 billion) in investments on Friday.
Much of that investment goes into the coffers of Nvidia.
What does it mean?
DeepSeek was a poke in the eye to Washington and its priority of thwarting China by maintaining American technological dominance.
“If China is catching up quickly to the US in the AI race, then the economics of AI will be turned on its head,” warned Kathleen Brooks, research director at XTB, in a note to clients.
The development also comes against a background of a US government push to ban Chinese-owned TikTok in the US or force its sale.
David Sacks, Mr Trump’s AI adviser and prominent tech investor, said DeepSeek’s success justified the White House’s decision to reverse Biden-era executive orders that had established safety standards for AI development.
These regulations “would have hamstrung American AI companies without any guarantee that China would follow suit” which they obviously wouldn’t, Mr Sacks wrote on X.
Adam Kovacevich, chief executive of the tech industry trade group Chamber of Progress, echoed this sentiment. “Now the top AI concern has to be ensuring [the US] wins,” he said.
Tech investor and Trump ally Marc Andreessen declared “Deepseek R1 is AI’s Sputnik moment”, referencing the 1957 launch of Earth’s first artificial satellite by the Soviet Union that stunned the western world.
“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,” he wrote on X.
Jim Fan, a senior research manager at Nvidia, added, “We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive — truly open, frontier research that empowers all.”
Professor Geoff Webb from Monash University’s Department of Data Science and AI said until now “it has seemed that billion-dollar investments and access to the latest generation of specialised Nvidia processors were prerequisites for developing state-of-the-art systems”, which “effectively limited control to a small number of leading US-based tech corporations”.
“DeepSeek claims to have developed a new Large Language Model, similar to Chat GPT or Llama, that rivals the state-of-the-art for a fraction of the cost using the less advanced Nvidia processors that are currently available to China,” Prof Webb said in a statement.
“If this is true, it means that the US tech sector no longer has exclusive control of the AI technologies, opening them to wider competition and reducing the prices they can charge for access to and use of their systems.”
Does DeepSeek censor?
The chatbot has already faced allegations of censorship of topics sensitive to the Chinese Communist Party.
Users have shared examples on social media of DeepSeek refusing to answer questions about Chinese President Xi Jinping, the 1989 Tiananmen Square massacre, human rights abuses against the Uighurs or whether Taiwan is a country.
DeepSeek admitted in one response that “my programming and knowledge base are designed to follow China’s laws and regulations, as well as socialist core values”, according to a post by the US House of Representatives select committee on China.
Entrepreneur Arnaud Bertrand said discussions of DeepSeek’s censorship “completely miss the point”. “You’re seeing [censorship of certain topics] if you use the DeepSeek chat agent hosted in China where they obviously have to abide by Chinese regulations on content moderation,” he wrote on X.
“But anyone could just as well download DeepSeek in Open Source and build their own chat agent on top of it without any of this stuff. And that’s precisely why DeepSeek is actually a more open model that offers more freedom than say OpenAI. They’re also censored in their own way and there’s absolutely zero way around it.”
What about Trump?
Mr Trump welcomed DeepSeek’s development of a “faster and much less expensive method of AI” in an address to Republican members of Congress on Monday night.
“And that’s good, because you don’t have to spend as much money,” he said.
“I view that as a positive, as an asset. So I really think, if it’s fact and if it’s true, and nobody really knows if it is, but I view that as a positive because you’ll be doing that too so you won’t be spending as much and you’ll get the same result, hopefully.”
Instead of spending “billions and billions, you’ll spend less”, he said.
But Mr Trump warned the “release of DeepSeek AI from a Chinese company should be a wake-up call for our industries that we need to be laser-focused on competing to win”.
“We have the greatest scientists in the world … this is very unusual, when you hear somebody [else] come up with something,” he said. “We always have the ideas, we’re always first.”
The turmoil comes just days after the President announced a $US500 billion ($795 billion) AI infrastructure project dubbed “Stargate”, led by Japanese giant SoftBank and OpenAI, which he said would create “over 100,000 American jobs almost immediately”.
The Stargate project is committed to invest an initial $US100 billion ($159 billion) in the project, and up to $US500 billion ($795 billion) over the next four years.
SoftBank and OpenAI each plan to commit $US19 billion ($30 billion) of capital to Stargate, while cloud giant Oracle and Abu Dhabi’s AI-focused state fund MGX will contribute about $US7 billion ($11 billion) apiece, according to technology news outlet The Information.
The rest of the money will come from limited partners and debt financing.
“What we want to do is we want to keep it in this country,” Mr Trump told reporters at the White House.
“China is a competitor and others are competitors. We want it to be in this country.”
— with AFP
More Coverage
Originally published as America’s worst fears confirmed after China’s $1 trillion AI nuke, Trump warns of ‘wake-up call’
‘Hell of a mess’: Man attacked by kangaroo
A man was walking to his car when he was attacked by a large kangaroo in what’s been described as a “hell of a mess”.
‘Giant’ hail, supercell storm to smash state
A supercell storm which is forecast to bring “giant-sized” hail and heavy rain is set to batter one Aussie state.