DeepSeek R1 has caused an uproar in AI land. ‘Bigly’ as the president would say. It was a flaming salvo out of the dark that took the US AI titans Nvidia, Open AI, Microsoft, Google Meta and Amazon by surprise. Or did it?

What is DeepSeek R1?

DeepSeek R1 is a Chinese created LLM (Large Language Model) that was released on January 20, the same day Trump was sworn in. Coincidence? Fast forward a few days and the DeepSeek R1 LLM had a chance to be fully tested by the tech community and the results were stunning. It was at least on par with OpenAI o1 and in some cases it was better.

Consider this chart of the test output of DeepSeek R1 vs. OpenAI o1.

Here is what the benchmarks mean (it’s OK, I had no idea either):

• AIME 2024: Achieving a top-tier score of 79.8%, surpassing even OpenAI-o1-mini and o1-preview.
• Codeforces (Competitive Programming): A groundbreaking percentile of 96.3%, far exceeding other models.
• GPQA Diamond (General Knowledge Reasoning): With a Pass@1 rate of 71.5%, it showcases superior capability in solving complex queries.
• MATH-500: A robust score of 97.3%, proving its dominance in advanced mathematical problem-solving.
• MMLU: Scoring 90.8%, indicating high reliability in handling diverse knowledge domains.
• SWE-bench Verified: An impressive 49.2%, a significant lead in software engineering benchmarks.

To get LLMs to work properly they need to be trained, which takes a long time and is costly in terms of the raw computing power needed. For example, it took Open AI over five months and cost $220M to train ChatGPT-4. DeepSeek claims that it took less time and cost only $6M to train R1. That is 97% less. Let that sink in. And it was done using hardware that was not state of the art as the US has restricted the sale of high-end processors to China. Consider this:

Why Does This Matter?

Cost. DeepSeek did far more with far less. And this came on the heels of the Project Stargate announcement where the US Government, in conjunction with Oracle, OpenAI and SoftBank, would sink $500B into AI development.

But suddenly DeepSeek had everyone asking, “Is it really necessary for companies to spend so much on ultra high-end AI GPUs?”

DeepSeek, like some AI Jesus, walked right across the previously impassable US AI moat and planted the Chinese flag squarely in Nvidia’s front yard. The jig was up. Stocks plummeted. We lost our advantage.

I Knew This Would Happen Seven Years Ago

Well, not this exactly, but I knew back in 2018 there would be an AI war between China and the US. That is when that I read the book AI Superpowers: China, Silicon Valley, and the New World Order by Kai-fu Lee. The book is absolutely riveting and I highly recommend that you read it.

Lee is a Taiwan born computer scientist and considered one of the foremost AI experts alive. He is also a venture capitalist and former executive for Google, Microsoft and Apple.

In his book, Lee predicted that the US would lead in AI, for a time, but that China would catch up to and perhaps even surpass the US. If you are so inclined, Lee has many fascinating videos on YouTube. One of the more recent videos details how search engines, like Google, are doomed.

Or Did It?

I implied earlier that Silicon Valley, specifically Nvidia, was not as surprised by the emergence of DeepSeek R1 as it may have seemed. Why? They knew there would be competition from China and that China would eventually develop their own models.

I suspect Silicon Valley thought they had more time to open a greater gap with China, but such is not the case. And I suspect that no one, save perhaps Kai-Fu Lee, thought the first Chinese models would be so potent when steps were taken to hobble them. In retrospect this was likely a mistake, as it forced the Chinese to innovate in areas and in ways that our indigenous AI efforts would overcome with enormous budgets and ultra high-end compute. But all is not lost.

DeepSeek R1 is Open Source

It is a real surprise to me that DeepSeek R1 is “open source” – meaning its source code is publicly available – especially since it comes from China. And thank goodness. Right now, engineers across all of Silicon Valley have stripped the model down to its procedures and will doubtless be incorporating all its innovations into future iterations of their own models.

Of course, China knew this would happen, so it does make one wonder what’s next.

The Hardware and Valuation Problems With An Unlikely Answer

However, incorporating DeepSeek’s innovations will not resolve the issue of hardware costs for AI training and execution, nor corporate valuation issues.

Nvidia’s market cap has already fallen by a few hundred billion dollars, valuing the company below its $3T peak. Will Nvidia and other companies with sizable investments in AI flourish again? The answer lies with the existence of DeepSeek itself and a hypothesis known as Jevon’s Paradox.

Jevon’s Paradox, named after the English economist William Stanley Jevons, suggests that increases in efficiency can lead to increased consumption, thereby offsetting the gains made by the efficiency improvements. 

As an example, this phenomenon is likely how improved coal-burning technology led to increased coal consumption in the 19th century. Demand increased as more efficient coal-powered machines spurred industrial growth.

In the context of artificial intelligence, this idea implies that as AI systems become more efficient at performing tasks, they may actually increase the demand for those tasks, leading to increased consumption of resources.

This paradox can already be observed in some aspects of AI usage. In the workplace, AI can automate routine tasks, making them cheaper and more efficient. This efficiency might lead to businesses expanding their operations or venturing into new markets, which could create new roles that didn't previously exist. 

Jevon’s Paradox in the age of generative AI reveals that while these technologies streamline workflows and reduce repetitive tasks, they also reshape work habits and expectations. Increased efficiency often leads to increased consumption of time and effort, potentially fostering a culture of perpetual business.

In summary, Jevon’s Paradox suggests that as AI becomes more efficient and accessible, its use will increase significantly, potentially leading to higher energy consumption and the creation of new jobs and industries, rather than just replacing existing ones. Thus, as AI becomes cheaper and ubiquitous, Nvidia, Microsoft, OpenAI and other firms will do more business.

DeepSeek R1 was a warning shot. It was the first shot in our coming AI war with China. It is a war for the future and one the United States must win.

Forecasts & Trends is published by Halbert Wealth Management, Inc., a Registered Investment Adviser under the Investment Advisers Act of 1940. Information contained herein is taken from sources believed to be reliable but cannot be guaranteed as to its accuracy. Opinions and recommendations herein generally reflect the judgement of the named author and may change at any time without written notice. Market opinions contained herein are intended as general observations and are not intended as specific advice. Readers are urged to check with their financial counselors before making any decisions. This does not constitute an offer of sale of any securities. Halbert Wealth Management, Inc., and its affiliated companies, its officers, directors and/or employees may or may not have their own money in markets or programs mentioned herein. Past results are not necessarily indicative of future results. All investments have a risk of loss. Be sure to read all offering materials and disclosures before making a decision to invest. Reprinting for family or friends is allowed with proper credit. However, republishing (written or electronically) in its entirety or through the use of extensive quotes is prohibited without prior written consent.