On January 20, the start-up of Chinese artificial intelligence Deepseek published its first generation reasoning models. In the press release, the company made amazing allegations.
Firstly, Deepseek said that his Deepseek-R1 model reaches performance comparable to Openai-O1, which is largely considered to be the most efficient model available in most areas. Since the Chinese company works with material clearly worse than Optaai and other American companies, it is certainly remarkable.
Even more impressive is that the company claims to have obtained these results at an incredibly low cost. R1 was built on the model of large V3 tongue of Deepseek, released in December. The company estimates that the cost of calculation for V3 training has only reached $ 5.6 million. To put this in perspective, the OPENAI GPT-4 cost $ 100 million to train.
Deepseek has achieved performance similar to a cost fraction. And as it is completely open source, allowing anyone to copy its techniques, it will have lasting implications throughout the industry.
Two companies, in particular, are in good position to benefit from Deepseek innovations.
Image source: Getty Images.
The long-term impact of Deepseek-V3 and R1
Deepseek focused on maximizing the effectiveness of its limited material capacities. Due to the export restrictions of the AI fleas, Nvidia is unable to sell your most powerful H100 GPUs in China. Instead, he sells H800 GPUs, which are specially designed to comply with American regulations. The H800 reduces the chip chip transfer rate, reducing the speed at which it is possible to form large models of AI.
Due to these limits, Deepseek has developed processes that allow him to reduce the amount of data he needs to transfer throughout the system at any time. For example, his “mixture of experts” or Deepseekmoe, introduced last year, made it activate only part of the model to answer questions.
Deepseek is not the only company to use this method, but its new approach has also made its training more effective. Most methods involve more Training of general costs in order to reduce the cost of inference later.
Starting has also developed methods to reduce the amount of memory required for IA inference by compressing important data before storing and transmitting it. He has brought new approaches to charges balancing, which is how the processes are distributed on his GPU network.
The result of these breakthroughs and other breakthroughs is not only an AI model which is faster to train and costs less. The longer term impact of Deepseek innovations is that it is cheaper to operate, and it can work on less capable equipment. In other words, AI inference has become much more accessible.
In a world that has the potential to execute AI systems on equipment that is holding in your pocket and for a tiny fraction of penny, there are two very large winners: Apple (AAPPL -0.67%)) And Meta-platforms (Meta 0.32%)). Here is why.
Make reality on devices
When Apple began to develop its artificial intelligence features for the iPhone and other devices, this has put data confidentiality at the forefront of its efforts. Apple Intelligence is designed to operate as much as possible on the iPhone. When it is necessary to make a call to the cloud, it takes each step it can to encrypt user data in the process.
There is a reason why the new characteristics of AI Apple last year is only available on iPhones released in the past 15 months. Since Apple tries to keep everything on the device, he needs enough treatment and memory power to execute his AI. The new iPhone chip, the A18 PRO, stimulated the memory bandwidth to support the faster AI treatment.
Apple could adopt many DEEPSEEK methods to make the iPhone plus capable of managing AI inference. This opens the door to features such as a conversational SIRI and the context, a faster translation without necessary internet connection, intelligent camera features and better productivity tools. More advanced AI features could increase the revenues of Apple sales and iPhone services.
Apple’s actions are currently negotiated for a relatively high multiple of 32.5 times the profits. But with its massive cash flows, which it uses to buy actions and improve the profitability of services from services, it can justify this high multiple, in particular given the coherence that Apple has presented in recent years. The potential stimulation of major IA improvements to disk AI could be a growth catalyst in the coming years.
AI failure at 3 billion people
Meta’s IA spending develops quickly while working to evolve its capacities and extend the features of AI to more parts of its activities. Capital spending increased by around 40% in 2024, and management said it expected an increase of 60% in 2025. These investments in AI paid well for Meta, which led to a commitment Stronger, better advertising tools and new features like Meta AI, which have the potential to become significant sources of income on the road.
An important decision taken by the meta with regard to AI was to open its LLAMA Model IA. One of the pruning behind this decision was to help make the model more effective. In fact, Deepseek used Llama as a basis for the development of R1, that’s exactly what Meta hoped.
The reduction in the cost of AI inference could unlock enormous benefits for Meta. This is a problem on which Meta has been working for a long time. “Many things are expensive, on the right, to generate an image or a video or a cat interaction,” said Zuckerberg during a call for results in February 2023. “So, one of the great interesting challenges here goes Also be how to adapt this and make it more effective for this way, we can bring it to a much larger user base.
Deepseek responds to this challenge and gives Meta the tools he needs to scale AI to its 3 billion users. Although Meta cannot slow down her expenses in AI as soon as he is able to earn much more money on the expenses he is engaged.
Meta Stock zoomed out above Deepseek News, reaching a new top of all time. However, the actions are negotiated 26.8 times the estimates of term profits to date. Meta is also a milk cow, using excess available cash flows to buy actions and support strong growth in action by action. If it can make AI more profitable, it is to see the benefits climb considerably in the coming years, which is well worth the price.
