Even Artificial Intelligence Can Be Disrupted As China’s DeepSeek Is Showing

Peter Diamandis calls the latest news about China’s DeepSeek artificial intelligence (AI), Nvidia’s worst nightmare. He’s right in one sense. DeepSeek didn’t use Nvidia state-of-the-art chips to develop its industry-disrupting AI application. Instead of AIs needing to raise billions of dollars for purpose-built data centres populated with computer systems with $30,000 Nvidia chips installed, the Chinese competitor used ones developed for video games to train AI specialists.

DeepSeek provided other surprises for the AI crowd including Elon Musk’s xAI, Mark Zuckerberg’s Meta AI, Google’s Gemini, Amazon’s Anthropic and Microsoft-backed OpenAI. The initial reaction to DeepSeek’s revelation was felt in stock prices for these companies with sharp declines.

What other attributes of the DeepSeek reveal are of interest?

• DeepSeek uses less memory, rather than 32-decimal places in computational requirements, it showed that 8-decimal places were more than enough.
• Instead of training one word at a time, the multi-token system DeepSeek used trains on whole phrases saving time while sacrificing little in the way of recognition accuracy.
• Instead of generalized AI training, DeepSeek divided its AI into specialists for building the application.

The comparison between DeepSeek, OpenAI and the other AI players is stark. For example, OpenAI has 4,500 employees. DeepSeek has 200. OpenAI has raised over $6.6 billion and spends more than $100 million to train new iterations of ChatGPT. DeepSeek spent $5 million to build and launch its AI.

Diamandis describes the results as “staggering” offering this summary:

• DeepSeek costs for AI training slashed from hundreds of millions to less than $6 million when compared to OpenAI making it at least 20 times cheaper if not more.
• Computing requirements for DeepSeek compared to OpenAI dropping from 100,000 to 2,000 GPUs with the latter using graphics processing circuit boards used for video gaming, engineering and mathematical applications.
• A 95% reduction in cost using APIs (application programming interfaces) when comparing DeepSeek to OpenAI.
• Pricing per million tokens, the fundamental data units processed by AI in natural language and machine learning, dropping from OpenAI’s $15 per input and $60 per output to DeepSeek’s $0.55 per input and $2.19 per output.
• A more than 90% cost reduction in GPU costs from $30,000 to $2,650 per unit.
• Instead of the 4,500 employees OpenAI has needed to develop ChatGPT, DeepSeek has 200.
• Instead of the AI application being proprietary, DeepSeek is open source.

By making DeepSeeks’ AI open source, it is opening the application to allow researchers and developers to build on it. This is democratizing AI and should lead to faster development. Open source means DeepSeek’s AI is transparent and can rely on the global community to help make improvements and corrections to it, let alone build niche applications for it. Meanwhile, compared to its Silicon Valley competitors, DeepSeek will remain cheap with application development on it attractive to small and medium-sized businesses that now can find reasons to jump on the AI bandwagon.

Diamandis welcomes the democratizing of AI stating:

• The “only-big-tech-can-play” era is OVER.
• Innovation barriers have been shattered.
• A few good GPUs might be all you need.
• The playing field has been levelled.

For NVIDIA it means retooling rather than relying on AI chips at $30,000 a pop to prop up its earnings. For OpenAI, Meta, Google, Microsoft, Amazon, and xAI it means going back to the drawing board if they want to compete with their more agile, smaller Chinese competitor. Those investors in Silicon Valley AI companies will wonder if they will soon be the victims of an AI bubble just like the dot-com one that burst in the late 1990s.

Diamandis sees the arrival of DeepSeek as a significant course change for AI describing it as “one of those moments we’ll look back on as an inflection point, like when PCs disrupted mainframes or cloud computing changed everything.” He continues, “The efficiency genie is out of the bottle, and there’s no going back.”