AI just had its Sputnik moment. 

Entrepreneur Marc Andreessen made that bold claim on X, the social media platform formerly known as Twitter, this past Sunday. Silicon Valley, along with the stock market and online prognosticators, are all reeling from what seems to be seismic-level activity in the AI space.

DeepSeek AI, a new AI model from China that’s jumped to the top of the Apple App Store, is sending reverberations throughout Silicon Valley. DeepSeek claims its AI competes with, and in some cases outperforms, OpenAI’s o1 reasoning model at a fraction of the cost. Not only that, DeepSeek’s R1 model is completely open source, meaning the code is openly accessible and anyone can use it for free. 

A key differentiator between DeepSeek-R1 and OpenAI’s o1 is that R1 allows you to see its chain of thought. It’s incredible insight into how the AI “thinks.” You can actually see it trying to answer questions about Tiananmen Square before it cancels its response, midway. Nvidia, the company making the chips powering the AI revolution, saw its stock plunge 18% and lose a record $600 billion after DeepSeek’s weekend ascent. It makes sense. If what DeepSeek says is true, it’s achieving near o1-level performance on apparently older Nvidia chips while spending a small percentage of the cost. 

Commenters online are still trying to make sense of DeepSeek’s sudden emergence in the AI marketplace. Is it actually performant with o1 at a lower cost? To what extent can claims by DeepSeek and China be true regarding efficiencies? Do the cost savings come from a major technical unlock, or are other areas in China’s supply chain making it cheaper to use? 

Regardless, R1 is impressive.

“This affordability opens the door for smaller companies and startups to leverage advanced AI technology that was previously inaccessible,” said Mel Morris, CEO of Corpora AI, an AI research engine, in a statement to CNET. Morris added that DeepSeek poses competition to established AI players and its “presence is likely to spur faster advancements in AI technology, leading to more efficient and accessible solutions to meet the growing demand.”

It could be why OpenAI CEO cut prices for its near-top-end o3 mini queries on Saturday.

As Big Tech is continually throws billions of dollars, processing power and energy at AI, DeepSeek’s efficiency unlock could be akin to the kind of leap we saw when cars went from carburetors to fuel injection systems. Unlike OpenAI, DeepSeek’s R1 model is open source, meaning anyone can use the technology. It’s a major disruption to the marketplace, currently dominated by OpenAI’s ChatGPT and Google’s Gemini, both of which are closed and require users to pay to gain full access to their suite of features. 

In the AI race between the US and China, America has stayed ahead thanks to Silicon Valley’s massive investment dump and the government’s blockade on Nvidia selling the latest AI chips to China. However, that blockade might have only incentivized China to make its own chips faster. Money, plus protectionism, was seen as a way to keep China in second place, making the world reliant on American technology. That dynamic may have shifted. Now, consumers and corporations worldwide have access to a highly performant “reasoning” model at a fraction of the cost. Not only that, TikTok parent company ByteDance released an even cheaper model to R1. 

As markets and social media react to new developments out of China, it might be too early to say America has been beaten. But at the very least, China is catching up quickly. 
“China has produced GPT-4 quality models already, but there was a longer lag in time — like it took a year, a year and a half, something like that. But now there is a Chinese model, which perhaps is only six months behind, and I think that is a difference,” said Lucas Hansen, co-founder of CivAI, a nonprofit that uses software to demonstrate what AI is capable of. “So, the US still has a lead, but it’s not as large as it previously was.”

One thing that’ll certainly help AI companies in catching up to OpenAI is R1’s ability for users to read its chain of thought. Even if R1 doesn’t get every answer right, being able to see how it reasons can better help develop it. The “shock and awe” people are feeling with R1 comes from the ability to read its chain of thought, according to Hansen. It’s insight OpenAI hasn’t given access to with its o1 model, as hiding the secret sauce keeps people shelling out a monthly subscription cost for access. 

Still, there’s a level of skepticism that should be taken with R1’s cost-to-performance ratio. The white paper that DeepSeek published had more than 100 co-authors. That’s a lot of brainpower to train an AI for the low cost of $5.5 million. That $5.5 million cost might just be the energy costs to train the model, minus every researchers’ individual salaries, but China hasn’t been fully transparent on how it calculated these energy costs. The cost of setting up a data center in China likely differs from setting up one in the US. And, it’s uncertain if costs were subsidized by a cloud provider or the Chinese government itself, according to Hansen. 

There’s also skepticism on the chips DeepSeek used to train its model. Is the firm actually using older Nvidia A100 and H800 chips or is China accessing the latest H100 chips through other means, as said by Alexandr Wang, CEO of Scale AI.

Even if we take that $5.5 million figure as a highly conservative estimate, it’s still significantly less than the $100 million it cost OpenAI to train GPT-4, the companies previous AI model. OpenAI hasn’t released figures on what it cost to build o1, but given its much higher token cost for customers, it was likely more expensive.

“With data center load in the United States expected to double or triple by 2030, any efficiency savings can have a significant impact,” said Mark James, interim director of the Institute of Energy and the Environment at Vermont Law and Graduate School in a statement. Already, utilities are being stressed by the high energy demands of AI. If DeepSeek’s claims are correct, then it could greatly lighten the potential electricity load, easing stress on both consumers and the environment. “On the flip side,” James said, “more efficient models could unlock even more growth in the sector, which would mitigate efficiency savings and exacerbate the stress on our grid.”

Claims that the US has lost the AI war might be premature. At the very least, the landscape has instantly become more competitive and there’s room for continued innovation. DeepSeek also doesn’t mean that the world is on the precipice of achieving artificial general intelligence, or super advanced AI that’s smarter than humans and can teach itself.

“I don’t think DeepSeek brings us one millimeter closer to Artificial General Intelligence (AGI), but I do think it brings us closer to commercially viable large language model (LLM) applications which is fantastic,” said Ben Goertzel CEO of the Artificial Superintelligence (ASI) Alliance and the Founder of SingularityNET. DeepSeek still has the same cognitive limitations as other AI models. Despite that, DeepSeek’s efficiencies could democratize AI further.

AI just had its Sputnik moment. 

Entrepreneur Marc Andreessen made that bold claim on X, the social media platform formerly known as Twitter, this past Sunday. Silicon Valley, along with the stock market and online prognosticators, are all reeling from what seems to be seismic-level activity in the AI space.

DeepSeek AI, a new AI model from China that’s jumped to the top of the Apple App Store, is sending reverberations throughout Silicon Valley. DeepSeek claims its AI competes with, and in some cases outperforms, OpenAI’s o1 reasoning model at a fraction of the cost. Not only that, DeepSeek’s R1 model is completely open source, meaning the code is openly accessible and anyone can use it for free. 

A key differentiator between DeepSeek-R1 and OpenAI’s o1 is that R1 allows you to see its chain of thought. It’s incredible insight into how the AI “thinks.” You can actually see it trying to answer questions about Tiananmen Square before it cancels its response, midway. Nvidia, the company making the chips powering the AI revolution, saw its stock plunge 18% and lose a record $600 billion after DeepSeek’s weekend ascent. It makes sense. If what DeepSeek says is true, it’s achieving near o1-level performance on apparently older Nvidia chips while spending a small percentage of the cost. 

Commenters online are still trying to make sense of DeepSeek’s sudden emergence in the AI marketplace. Is it actually performant with o1 at a lower cost? To what extent can claims by DeepSeek and China be true regarding efficiencies? Do the cost savings come from a major technical unlock, or are other areas in China’s supply chain making it cheaper to use? 

Regardless, R1 is impressive.

“This affordability opens the door for smaller companies and startups to leverage advanced AI technology that was previously inaccessible,” said Mel Morris, CEO of Corpora AI, an AI research engine, in a statement to CNET. Morris added that DeepSeek poses competition to established AI players and its “presence is likely to spur faster advancements in AI technology, leading to more efficient and accessible solutions to meet the growing demand.”

It could be why OpenAI CEO cut prices for its near-top-end o3 mini queries on Saturday.

As Big Tech is continually throws billions of dollars, processing power and energy at AI, DeepSeek’s efficiency unlock could be akin to the kind of leap we saw when cars went from carburetors to fuel injection systems. Unlike OpenAI, DeepSeek’s R1 model is open source, meaning anyone can use the technology. It’s a major disruption to the marketplace, currently dominated by OpenAI’s ChatGPT and Google’s Gemini, both of which are closed and require users to pay to gain full access to their suite of features. 

In the AI race between the US and China, America has stayed ahead thanks to Silicon Valley’s massive investment dump and the government’s blockade on Nvidia selling the latest AI chips to China. However, that blockade might have only incentivized China to make its own chips faster. Money, plus protectionism, was seen as a way to keep China in second place, making the world reliant on American technology. That dynamic may have shifted. Now, consumers and corporations worldwide have access to a highly performant “reasoning” model at a fraction of the cost. Not only that, TikTok parent company ByteDance released an even cheaper model to R1. 

As markets and social media react to new developments out of China, it might be too early to say America has been beaten. But at the very least, China is catching up quickly. 
“China has produced GPT-4 quality models already, but there was a longer lag in time — like it took a year, a year and a half, something like that. But now there is a Chinese model, which perhaps is only six months behind, and I think that is a difference,” said Lucas Hansen, co-founder of CivAI, a nonprofit that uses software to demonstrate what AI is capable of. “So, the US still has a lead, but it’s not as large as it previously was.”

One thing that’ll certainly help AI companies in catching up to OpenAI is R1’s ability for users to read its chain of thought. Even if R1 doesn’t get every answer right, being able to see how it reasons can better help develop it. The “shock and awe” people are feeling with R1 comes from the ability to read its chain of thought, according to Hansen. It’s insight OpenAI hasn’t given access to with its o1 model, as hiding the secret sauce keeps people shelling out a monthly subscription cost for access. 

Still, there’s a level of skepticism that should be taken with R1’s cost-to-performance ratio. The white paper that DeepSeek published had more than 100 co-authors. That’s a lot of brainpower to train an AI for the low cost of $5.5 million. That $5.5 million cost might just be the energy costs to train the model, minus every researchers’ individual salaries, but China hasn’t been fully transparent on how it calculated these energy costs. The cost of setting up a data center in China likely differs from setting up one in the US. And, it’s uncertain if costs were subsidized by a cloud provider or the Chinese government itself, according to Hansen. 

There’s also skepticism on the chips DeepSeek used to train its model. Is the firm actually using older Nvidia A100 and H800 chips or is China accessing the latest H100 chips through other means, as said by Alexandr Wang, CEO of Scale AI.

Even if we take that $5.5 million figure as a highly conservative estimate, it’s still significantly less than the $100 million it cost OpenAI to train GPT-4, the companies previous AI model. OpenAI hasn’t released figures on what it cost to build o1, but given its much higher token cost for customers, it was likely more expensive.

“With data center load in the United States expected to double or triple by 2030, any efficiency savings can have a significant impact,” said Mark James, interim director of the Institute of Energy and the Environment at Vermont Law and Graduate School in a statement. Already, utilities are being stressed by the high energy demands of AI. If DeepSeek’s claims are correct, then it could greatly lighten the potential electricity load, easing stress on both consumers and the environment. “On the flip side,” James said, “more efficient models could unlock even more growth in the sector, which would mitigate efficiency savings and exacerbate the stress on our grid.”

Claims that the US has lost the AI war might be premature. At the very least, the landscape has instantly become more competitive and there’s room for continued innovation. DeepSeek also doesn’t mean that the world is on the precipice of achieving artificial general intelligence, or super advanced AI that’s smarter than humans and can teach itself.

“I don’t think DeepSeek brings us one millimeter closer to Artificial General Intelligence (AGI), but I do think it brings us closer to commercially viable large language model (LLM) applications which is fantastic,” said Ben Goertzel CEO of the Artificial Superintelligence (ASI) Alliance and the Founder of SingularityNET. DeepSeek still has the same cognitive limitations as other AI models. Despite that, DeepSeek’s efficiencies could democratize AI further.

AI just had its Sputnik moment. 

Entrepreneur Marc Andreessen made that bold claim on X, the social media platform formerly known as Twitter, this past Sunday. Silicon Valley, along with the stock market and online prognosticators, are all reeling from what seems to be seismic-level activity in the AI space.

DeepSeek AI, a new AI model from China that’s jumped to the top of the Apple App Store, is sending reverberations throughout Silicon Valley. DeepSeek claims its AI competes with, and in some cases outperforms, OpenAI’s o1 reasoning model at a fraction of the cost. Not only that, DeepSeek’s R1 model is completely open source, meaning the code is openly accessible and anyone can use it for free. 

A key differentiator between DeepSeek-R1 and OpenAI’s o1 is that R1 allows you to see its chain of thought. It’s incredible insight into how the AI “thinks.” You can actually see it trying to answer questions about Tiananmen Square before it cancels its response, midway. Nvidia, the company making the chips powering the AI revolution, saw its stock plunge 18% and lose a record $600 billion after DeepSeek’s weekend ascent. It makes sense. If what DeepSeek says is true, it’s achieving near o1-level performance on apparently older Nvidia chips while spending a small percentage of the cost. 

Commenters online are still trying to make sense of DeepSeek’s sudden emergence in the AI marketplace. Is it actually performant with o1 at a lower cost? To what extent can claims by DeepSeek and China be true regarding efficiencies? Do the cost savings come from a major technical unlock, or are other areas in China’s supply chain making it cheaper to use? 

Regardless, R1 is impressive.

“This affordability opens the door for smaller companies and startups to leverage advanced AI technology that was previously inaccessible,” said Mel Morris, CEO of Corpora AI, an AI research engine, in a statement to CNET. Morris added that DeepSeek poses competition to established AI players and its “presence is likely to spur faster advancements in AI technology, leading to more efficient and accessible solutions to meet the growing demand.”

It could be why OpenAI CEO cut prices for its near-top-end o3 mini queries on Saturday.

As Big Tech is continually throws billions of dollars, processing power and energy at AI, DeepSeek’s efficiency unlock could be akin to the kind of leap we saw when cars went from carburetors to fuel injection systems. Unlike OpenAI, DeepSeek’s R1 model is open source, meaning anyone can use the technology. It’s a major disruption to the marketplace, currently dominated by OpenAI’s ChatGPT and Google’s Gemini, both of which are closed and require users to pay to gain full access to their suite of features. 

In the AI race between the US and China, America has stayed ahead thanks to Silicon Valley’s massive investment dump and the government’s blockade on Nvidia selling the latest AI chips to China. However, that blockade might have only incentivized China to make its own chips faster. Money, plus protectionism, was seen as a way to keep China in second place, making the world reliant on American technology. That dynamic may have shifted. Now, consumers and corporations worldwide have access to a highly performant “reasoning” model at a fraction of the cost. Not only that, TikTok parent company ByteDance released an even cheaper model to R1. 

As markets and social media react to new developments out of China, it might be too early to say America has been beaten. But at the very least, China is catching up quickly. 
“China has produced GPT-4 quality models already, but there was a longer lag in time — like it took a year, a year and a half, something like that. But now there is a Chinese model, which perhaps is only six months behind, and I think that is a difference,” said Lucas Hansen, co-founder of CivAI, a nonprofit that uses software to demonstrate what AI is capable of. “So, the US still has a lead, but it’s not as large as it previously was.”

One thing that’ll certainly help AI companies in catching up to OpenAI is R1’s ability for users to read its chain of thought. Even if R1 doesn’t get every answer right, being able to see how it reasons can better help develop it. The “shock and awe” people are feeling with R1 comes from the ability to read its chain of thought, according to Hansen. It’s insight OpenAI hasn’t given access to with its o1 model, as hiding the secret sauce keeps people shelling out a monthly subscription cost for access. 

Still, there’s a level of skepticism that should be taken with R1’s cost-to-performance ratio. The white paper that DeepSeek published had more than 100 co-authors. That’s a lot of brainpower to train an AI for the low cost of $5.5 million. That $5.5 million cost might just be the energy costs to train the model, minus every researchers’ individual salaries, but China hasn’t been fully transparent on how it calculated these energy costs. The cost of setting up a data center in China likely differs from setting up one in the US. And, it’s uncertain if costs were subsidized by a cloud provider or the Chinese government itself, according to Hansen. 

There’s also skepticism on the chips DeepSeek used to train its model. Is the firm actually using older Nvidia A100 and H800 chips or is China accessing the latest H100 chips through other means, as said by Alexandr Wang, CEO of Scale AI.

Even if we take that $5.5 million figure as a highly conservative estimate, it’s still significantly less than the $100 million it cost OpenAI to train GPT-4, the companies previous AI model. OpenAI hasn’t released figures on what it cost to build o1, but given its much higher token cost for customers, it was likely more expensive.

“With data center load in the United States expected to double or triple by 2030, any efficiency savings can have a significant impact,” said Mark James, interim director of the Institute of Energy and the Environment at Vermont Law and Graduate School in a statement. Already, utilities are being stressed by the high energy demands of AI. If DeepSeek’s claims are correct, then it could greatly lighten the potential electricity load, easing stress on both consumers and the environment. “On the flip side,” James said, “more efficient models could unlock even more growth in the sector, which would mitigate efficiency savings and exacerbate the stress on our grid.”

Claims that the US has lost the AI war might be premature. At the very least, the landscape has instantly become more competitive and there’s room for continued innovation. DeepSeek also doesn’t mean that the world is on the precipice of achieving artificial general intelligence, or super advanced AI that’s smarter than humans and can teach itself.

“I don’t think DeepSeek brings us one millimeter closer to Artificial General Intelligence (AGI), but I do think it brings us closer to commercially viable large language model (LLM) applications which is fantastic,” said Ben Goertzel CEO of the Artificial Superintelligence (ASI) Alliance and the Founder of SingularityNET. DeepSeek still has the same cognitive limitations as other AI models. Despite that, DeepSeek’s efficiencies could democratize AI further.

AI just had its Sputnik moment. 

Entrepreneur Marc Andreessen made that bold claim on X, the social media platform formerly known as Twitter, this past Sunday. Silicon Valley, along with the stock market and online prognosticators, are all reeling from what seems to be seismic-level activity in the AI space.

DeepSeek AI, a new AI model from China that’s jumped to the top of the Apple App Store, is sending reverberations throughout Silicon Valley. DeepSeek claims its AI competes with, and in some cases outperforms, OpenAI’s o1 reasoning model at a fraction of the cost. Not only that, DeepSeek’s R1 model is completely open source, meaning the code is openly accessible and anyone can use it for free. 

A key differentiator between DeepSeek-R1 and OpenAI’s o1 is that R1 allows you to see its chain of thought. It’s incredible insight into how the AI “thinks.” You can actually see it trying to answer questions about Tiananmen Square before it cancels its response, midway. Nvidia, the company making the chips powering the AI revolution, saw its stock plunge 18% and lose a record $600 billion after DeepSeek’s weekend ascent. It makes sense. If what DeepSeek says is true, it’s achieving near o1-level performance on apparently older Nvidia chips while spending a small percentage of the cost. 

Commenters online are still trying to make sense of DeepSeek’s sudden emergence in the AI marketplace. Is it actually performant with o1 at a lower cost? To what extent can claims by DeepSeek and China be true regarding efficiencies? Do the cost savings come from a major technical unlock, or are other areas in China’s supply chain making it cheaper to use? 

Regardless, R1 is impressive.

“This affordability opens the door for smaller companies and startups to leverage advanced AI technology that was previously inaccessible,” said Mel Morris, CEO of Corpora AI, an AI research engine, in a statement to CNET. Morris added that DeepSeek poses competition to established AI players and its “presence is likely to spur faster advancements in AI technology, leading to more efficient and accessible solutions to meet the growing demand.”

It could be why OpenAI CEO cut prices for its near-top-end o3 mini queries on Saturday.

As Big Tech is continually throws billions of dollars, processing power and energy at AI, DeepSeek’s efficiency unlock could be akin to the kind of leap we saw when cars went from carburetors to fuel injection systems. Unlike OpenAI, DeepSeek’s R1 model is open source, meaning anyone can use the technology. It’s a major disruption to the marketplace, currently dominated by OpenAI’s ChatGPT and Google’s Gemini, both of which are closed and require users to pay to gain full access to their suite of features. 

In the AI race between the US and China, America has stayed ahead thanks to Silicon Valley’s massive investment dump and the government’s blockade on Nvidia selling the latest AI chips to China. However, that blockade might have only incentivized China to make its own chips faster. Money, plus protectionism, was seen as a way to keep China in second place, making the world reliant on American technology. That dynamic may have shifted. Now, consumers and corporations worldwide have access to a highly performant “reasoning” model at a fraction of the cost. Not only that, TikTok parent company ByteDance released an even cheaper model to R1. 

As markets and social media react to new developments out of China, it might be too early to say America has been beaten. But at the very least, China is catching up quickly. 
“China has produced GPT-4 quality models already, but there was a longer lag in time — like it took a year, a year and a half, something like that. But now there is a Chinese model, which perhaps is only six months behind, and I think that is a difference,” said Lucas Hansen, co-founder of CivAI, a nonprofit that uses software to demonstrate what AI is capable of. “So, the US still has a lead, but it’s not as large as it previously was.”

One thing that’ll certainly help AI companies in catching up to OpenAI is R1’s ability for users to read its chain of thought. Even if R1 doesn’t get every answer right, being able to see how it reasons can better help develop it. The “shock and awe” people are feeling with R1 comes from the ability to read its chain of thought, according to Hansen. It’s insight OpenAI hasn’t given access to with its o1 model, as hiding the secret sauce keeps people shelling out a monthly subscription cost for access. 

Still, there’s a level of skepticism that should be taken with R1’s cost-to-performance ratio. The white paper that DeepSeek published had more than 100 co-authors. That’s a lot of brainpower to train an AI for the low cost of $5.5 million. That $5.5 million cost might just be the energy costs to train the model, minus every researchers’ individual salaries, but China hasn’t been fully transparent on how it calculated these energy costs. The cost of setting up a data center in China likely differs from setting up one in the US. And, it’s uncertain if costs were subsidized by a cloud provider or the Chinese government itself, according to Hansen. 

There’s also skepticism on the chips DeepSeek used to train its model. Is the firm actually using older Nvidia A100 and H800 chips or is China accessing the latest H100 chips through other means, as said by Alexandr Wang, CEO of Scale AI.

Even if we take that $5.5 million figure as a highly conservative estimate, it’s still significantly less than the $100 million it cost OpenAI to train GPT-4, the companies previous AI model. OpenAI hasn’t released figures on what it cost to build o1, but given its much higher token cost for customers, it was likely more expensive.

“With data center load in the United States expected to double or triple by 2030, any efficiency savings can have a significant impact,” said Mark James, interim director of the Institute of Energy and the Environment at Vermont Law and Graduate School in a statement. Already, utilities are being stressed by the high energy demands of AI. If DeepSeek’s claims are correct, then it could greatly lighten the potential electricity load, easing stress on both consumers and the environment. “On the flip side,” James said, “more efficient models could unlock even more growth in the sector, which would mitigate efficiency savings and exacerbate the stress on our grid.”

Claims that the US has lost the AI war might be premature. At the very least, the landscape has instantly become more competitive and there’s room for continued innovation. DeepSeek also doesn’t mean that the world is on the precipice of achieving artificial general intelligence, or super advanced AI that’s smarter than humans and can teach itself.

“I don’t think DeepSeek brings us one millimeter closer to Artificial General Intelligence (AGI), but I do think it brings us closer to commercially viable large language model (LLM) applications which is fantastic,” said Ben Goertzel CEO of the Artificial Superintelligence (ASI) Alliance and the Founder of SingularityNET. DeepSeek still has the same cognitive limitations as other AI models. Despite that, DeepSeek’s efficiencies could democratize AI further.