Original source: Lei Technology

Image source: Generated by Unbounded AI

NVIDIA is living a very prosperous life now. From the virtual currency craze to the era of large AI models, NVIDIA's development speed in the past few years has exceeded any previous period, which has also helped the chip company's market value successfully exceed 10,000. billion dollar mark.

However, compared to the castle-like virtual economy of virtual currency, the "real demand" brought by large AI models is the core driving force for NVIDIA to break through the trillion-dollar market value mark. NVIDIA's H100 is reported to be from ordering to delivery The time required has been as long as several months, and the premium of the spot was once close to 100%.

However, NVIDIA's good days may not last long. As large AI models are recognized as a "broad road", major companies are stepping up their efforts to purchase NVIDIA graphics cards and build their own training servers. They are also seeing funds flowing out like a flood. , also made his own little calculation.

Recently, **OpenAI announced that it will begin to develop its own AI chips to reduce its dependence on Nvidia. Coincidentally, Microsoft, which is building a large-scale AI server, also announced its own AI chip plan. **Interestingly, although OpenAI now nominally belongs to the Microsoft camp (Microsoft has previously completed the acquisition of OpenAI), OpenAI and Microsoft do not seem to have any plans to share chip plans.

In addition to OpenAI and Microsoft, there are many manufacturers who are also ready to make a move.

Embattled on all sides

The cost of supporting a large-scale data center is not low. The initial hardware investment alone is measured in "hundred million". The European data center plan announced by Microsoft some time ago has an initial investment of up to US$500 million, not including subsequent maintenance. Waiting fees. Among the US$500 million, in addition to infrastructure construction and other expenses, the largest expenditure is the purchase of professional computing cards produced by Nvidia.

According to analysis some time ago, the difference between the cost and selling price of Nvidia's chips may be more than 10 times. Taking the H100, which is most popular among large enterprises, as an example, the cost of the computing card is about 2,000-2,500 US dollars, while the official selling price is more than 25,000 US dollars.

Whether it is to save money or to take advantage of this emerging market, the implementation of its own AI chip research and development plan is imminent. **Judging from the currently known information, semiconductor giants such as Intel and AMD have announced a new round of AI chip research and development plans. Intel uses the CPU as a breakthrough to create another AI chip in a different way, and has even released the first generation of AI chips. products, AMD is trying to challenge Nvidia's position in the GPU field. **

It’s not surprising that traditional semiconductor giants are trying to get a piece of the pie. What caught Nvidia’s attention even more is that OpenAI and Microsoft announced that they will launch AI chip research and development plans. As the two core users, if they abandon Nvidia, they will obviously have negative consequences for Nvidia. Ecological status and revenue have serious consequences.

OpenAI's chip plan has only been exposed for the first time recently. For an AI company, I have doubts about OpenAI's chip research and development capabilities. Moreover, judging from the recent recruitment information released by OpenAI, they are building a research and development team from scratch. It may take at least a year before they can produce preliminary results, and there is a high probability that they will not be able to compete with Nvidia's flagship chips.

Relatively speaking, Microsoft's chip plan is more concerning. Microsoft's investment in the chip field has actually been quite high, and it has produced a lot of products in recent years. ** And the chip codenamed "Athena" that was recently exposed, According to internal sources, research and development began as early as 2019 and has now entered the trial production stage. **

It is reported that OpenAI has secretly tested the Athena chip. As a chip designed for training and running large models, its performance is very good in terms of performance, at least comparable to mainstream chips from Amazon, Google and other companies.

Of course, Athena's performance is definitely not comparable to Nvidia's flagship chips, but it can give Microsoft greater initiative and allow Nvidia to be slightly more restrained in supplying chip quotations. Moreover, Athena is only Microsoft's first professional AI chip, and its R&D investment of more than 2 billion US dollars will obviously not produce just one result.

As the largest sponsor of OpenAI, Microsoft will most likely require OpenAI to provide a testing and deployment environment for the Athena chip. After all, Amazon and Google have done so. Long before Microsoft, Amazon and Google had invested in many AI companies. While Amazon provided $4 billion in financial support to Anthropic, it also required the other party to use two AI chips developed by Amazon. **

When leading AI companies begin to switch to other chips or self-developed chips, it will inevitably have a significant impact on the hardware selection of the entire AI industry. This is exactly what NVIDIA does not want to see. How will NVIDIA respond?

Nvidia’s Countermeasures

The charm of large AI models has immersed many technology companies in it, and some even believe that this is the beginning of the next industrial revolution. Of course, let’s not discuss how many new technologies have been dubbed the “beginning of the industrial revolution.” At least judging from the current development route, the large AI model should be the most closely related to ordinary people in recent years. skill improved.

The close relationship with ordinary people means that this technology has a very broad application market and can be quickly promoted and commercialized to bring profits. From the birth of the technology to its commercial use, few technologies have progressed as fast as large AI models. From ChatGPT being announced and open for use, to various AI large models springing up and being open to the public, the whole process took just one day. It will be completed in less than a year.

From productivity to entertainment, consumption, travel, and education, large AI models have been implemented in many applications. Because of this, some powerful companies are also stepping up their efforts to build their own data centers and computing centers to deploy and train larger-scale models. AI models give you an advantage over the competition.

**As the AI market enters a competitive stage, companies are also seeking more efficient training methods and more powerful models. In addition to optimizing algorithms and other aspects, professional computing cards with stronger computing capabilities are also a must. **So, NVIDIA’s countermeasures are actually very simple. Stabilize the R&D team and launch AI chips that are far ahead of other manufacturers.

Hardware performance is NVIDIA's biggest advantage. Whether it is Amazon or Microsoft, as long as they want to find the best balance between performance and energy consumption, NVIDIA is their first choice. There are only two reasons that stimulate manufacturers to use self-developed chips. One is that Nvidia's chips are too expensive, and the other is that supply is limited and they need to wait for stocking, which has an impact on manufacturers' expansion plans.

At present, Nvidia's production capacity is gradually increasing, and purchase volume is gradually declining, and it should soon reach a stage of balance between supply and demand. Then the only problem is the price. Considering that Nvidia's cost and selling price are nearly 10 times different, there should be ample room for price reduction.

**Personally, I believe that as long as NVIDIA is willing to lower the price, it will still be a cost-effective deal for many companies to purchase NVIDIA's professional computing cards to build high-performance data centers. **As for self-developed chips? In fact, data centers require different types of chips depending on their size and purpose. Some data centers with lower performance requirements are suitable to be built with self-developed chips.

To put it simply, training and development centers use Nvidia's professional computing cards to improve training efficiency, while data centers for ordinary users use self-developed or other chips to reduce construction costs and subsequent maintenance costs. With the application scope of AI models, To expand, companies obviously need to build more data centers around the world to respond to user needs nearby.

Therefore, the advantages that NVIDIA has accumulated in the past will not be easily lost even in the future. However, as other companies enter the game, NVIDIA's voice will be reduced. In terms of product pricing and other aspects, NVIDIA may give away part of its profits to maintain the market. share.

However, compared to the previous battles between gods and mortals, this time many AI companies besieged "Guangmingding", which can allow small and medium-sized AI companies to obtain cheaper data center deployment solutions.