‘Experts’ fear China’s DeepSeek AI. Econ 101 says they shouldn’t worry

If you care about national security, artificial intelligence (AI) or the index funds in your retirement account, you have likely heard of DeepSeek. Chinese AI model DeepSeek’s release late January caused a $969 billion stock market selloff and prompted responses from AI leaders like President Donald Trump, NVIDIA, venture capitalist Marc Andreessen and OpenAI CEO Sam Altman.   

DeepSeek is a Chinese AI model similar to those which run famous programs like ChatGPT. The difference? This one is reportedly 40-50 times cheaper to train, which some think implies we have invested 40-50 times too much in AI chips and infrastructure.   

If panicked stock sellers and pessimists are to be believed, DeepSeek is some sort of kryptonite whose existence implies China will do to American AI what it did to American toy manufacturing.  

THERE IS A ‘WAKE-UP CALL’ FOR US TO BE THE LEADER IN AI, SAYS WHITE HOUSE AI AND CRYPTO ‘CZAR’

This is nonsense. DeepSeek’s figures are bullish for American AI if true and bullish for American AI if false. Here’s why:  

1. DeepSeek runs on American AI. DeepSeek relied on American chips and models for its development and training, and America has the expertise, infrastructure and lead to continue winning.
2. If DeepSeek’s numbers are true, they do not tell the whole story. DeepSeek cost hundreds of millions more than the numbers suggest.
3. Even if DeepSeek lowered demand for compute (computational power) long-term, the benefits to American AI producers and consumers would far outweigh any damage to companies like NVIDIA, which also produces massive efficiency gains with each new run of GPUs. Just as cheaper gas, wheat and eggs ultimately benefit all Americans, so would cheaper AI.

And since when do lower costs and lower barriers to entry mean lower demand? 

DeepSeek runs on American AI. The headline numbers don’t tell the full story  

DeepSeek runs on the following layers:  

• The physical chips used to run the computations which train the model.
• The other models used to train the program (DeepSeek is a small model built using big models).
• The final model brought to market.

China’s DeepSeek AI relied on chips from NVIDIA, even though it’s banned from export to China. (Justin Sullivan/Getty Images)

The physical chips used were NVIDIA H800s, a downgraded version of the popular H100 chip. Both have been banned by the U.S. for export to China, proving again that President Joe Biden’s export bans are useless and merely force China to become more innovative.  

The other models used to train DeepSeek reflect the sustained need for massive compute and reliance on American models. Per Rand’s Lennart Heim,   

“DeepSeek operated Asia’s first 10,000 Nvidia A100 cluster, reportedly maintains 50,000 ‘Hoppers’ (which could be Nvidia’s H100, H800, or H20), and has additional unlimited access to Chinese and foreign cloud providers (which is not export-controlled). This extensive compute access was likely crucial for developing their efficiency techniques through trial and error and for serving their models to customers.”  

In addition, there is mounting evidence that DeepSeek is a distillation (in essence a copy) of American AI leader OpenAI’s latest models. OpenAI says it is looking into this possibility.  

This means that the remaining component, the final model brought to market, also would depend on American AI. On top of early incidental finds by AI developers independently investigating the matter, Microsoft has joined OpenAI in looking into the depth of DeepSeek’s alleged copying.  

Even if China suddenly decided it likes telling the truth and DeepSeek did cost less than $6 million to train, it required indirect access to nearly a billion dollars of American compute. Analysis by SemiAnalysis suggests DeepSeek’s capital expenditure was closer to $1.6 billion, with $944 million of this cost dedicated to the cost of compute. Hardly the garage band operation the headlines would suggest. 

The $6M number is misleading by many orders of magnitude. And even if it wasn’t, DeepSeek would be no NVIDIA-pocalypse unless we thought all future models would be mere distillations of existing ones. DeepSeek is an affirmation of American AI leadership and a reminder that value lies in AI infrastructure and customer-facing products themselves, not in closed-source models.  

So, what do we do with this information?  

In addition, there is mounting evidence that DeepSeek is a distillation (in essence a copy) of American AI leader OpenAI’s latest models. OpenAI says it is looking into this possibility.  

As Andreessen said, this is AI’s Sputnik moment. America received a reminder that a rival is further ahead in a critical technology than we had previously thought, that innovations lead to lower barriers to entry and lower costs to catch up, and that we should not take our leadership for granted.  

Here’s what Americans should take away:  

1. Biden’s AI restrictions are useless.
2. Where they do lower access (the black market is booming), they only forcefeed China’s innovation and hard-won market share.
3. This is a Manhattan Project moment, not an F-35 moment.

Biden’s AI restrictions were a failure and DeepSeek’s reliance on U.S. tech proves it. (AP Photo/Carolyn Kaster)

This competition runs on open-source (publicly available) science. We do not have a technical moat and will win only through a continued emphasis on speed and quality. Export controls only work for high-complexity products with decades-long development cycles.   

America’s AI industry is safe if it continues its momentum.   

CLICK HERE FOR MORE FOX NEWS OPINION

Others still depend directly and indirectly on American models and chips. It is beneficial to have our rivals depend on us. Resource deprivation breeds innovation and expansionism. Just ask Imperial Japan  

Meanwhile, AI costs will come down for everyone. Computers and machine learning algorithms become cheaper over time, and there is no credible reason to bank on a ceiling or “peak AI.”  

America received a reminder that a rival is further ahead in a critical technology than we had previously thought, that innovations lead to lower barriers to entry and lower costs to catch up, and that we should not take our leadership for granted.  

DeepSeek’s only clear innovations are aspects of its training methods, and we have the scale to make superior use of them if we have the will and humility to do so. If DeepSeek can derive a workable copy from a larger model for less than $6 million, imagine how this capability will compound and accelerate model development for companies like OpenAI and Google ready to deploy hundreds of millions of dollars.  

CLICK HERE TO GET THE FOX NEWS APP

In the words of Commerce Secretary nominee Howard Lutnick, “We need to make sure we set standards that the world meets. And I think that’s the way we’ll keep our lead. Do it the American way, which we know is the winning way.”  

DeepSeek is a fire drill and a reminder to prepare for the real thing. In the meantime, we should heed the reminder that we can’t be complacent with our current leadership. Instead, we must do what we do best: focus on winning. We have a head start, the best scientists and chips in the world, and a president prepared to fight for American AI. We now must execute.  

CLICK HERE TO READ MORE FROM AARON GINN