ChatGPT is "stuck in the neck", it is reported that OpenAI plans to develop its own AI chip to challenge Nvidia

火讯财经 · 2023-10-07T02:33:02.000Z
Source of article reprint: The trend of AI This article was first published on Titanium Media App, author: Lin Zhijia According to reports on October 6, OpenAI, the American AI company behind ChatGPT, plans to develop its own AI chips. It has been conducting internal discussions since last year and has even begun to evaluate potential acquisition targets of AI chip companies to solve the shortage of AI chips on which it relies. and high costs. According to reports, the OpenAI team believes that there are three procurement options for AI large model computing power, including building its own AI chips, working more closely with chip companies such as NVIDIA, and planning to diversify chip supply outside of NVIDIA - the ultimate goal is to exceed Beyond Nvidia.
Article reprint source: AI Trends
This article was first published on Titanium Media App, author: Lin Zhijia
On October 6, it was reported that OpenAI, the American AI company behind ChatGPT, plans to develop its own AI chips. It has been conducting internal discussions since last year and has even begun evaluating potential acquisition targets of AI chip companies to solve the shortage and high cost of the AI ​​chips it relies on.
The report stated that the OpenAI team believes that there are three procurement options for AI large model computing power, including building its own AI chips, working more closely with chip companies such as Nvidia, and planning to diversify chip supplies beyond Nvidia - with the ultimate goal of surpassing Nvidia.
As early as 2022, OpenAI CEO Sam Altman publicly complained about the scarcity of Nvidia GPU chips, saying that the company was severely limited by GPUs.
Since NVIDIA dominates 95% of the global AI training market, with the scarcity of NVIDIA GPU (graphics processing unit) graphics cards and the continued rise in AI computing power costs, even a strong company like OpenAI is looking for new solutions to avoid being "choked" in the long term.
OpenAI CEO Sam Altman
As computing power costs soar, OpenAI seeks customized AI chips
Since 2023, the AI ​​big model "trend" represented by ChatGPT has swept the world, and big models are driving AI to develop in a more general direction. However, the scarcity and high cost of computing power have become the core factors restricting the development of AI.
Currently, Nvidia occupies 82% of the global data center AI acceleration market, and has monopolized the global AI training market with a 95% market share, becoming the biggest winner in this round of AI melee. As the founder and CEO of the company, Huang Renxun has made a lot of money. The company's market value has exceeded 1 trillion US dollars, and his net worth is as high as 39.9 billion US dollars.
At the same time, the surge in demand for computing power has also made Nvidia GPUs "hard to come by." The number of Nvidia A100 graphics cards a company has has become a measure of its computing power.
According to statistics, OpenAI used 10,000 to 30,000 NVIDIA GPUs to train the GPT-3.5 model. According to a report by TrendForce, if the processing power of NVIDIA A100 graphics cards is used, running ChatGPT may require the use of 30,000 NVIDIA GPU graphics cards.
In terms of price, the price of H800 available in China is as high as 200,000 yuan per card, while the price of A100/A800 has risen to around 150,000 yuan and 100,000 yuan per card. Taking the demand for 2000P computing power as an example, the computing power of a single H800 GPU card is 2P, and 1,000 cards are required, with a predicted cost of 200 million yuan; the computing power of a single A800 card is about 0.625P, and the number required is 3,200 cards, with a predicted cost of 320 million yuan.
In addition to buying GPU graphics cards, the server also needs to consider the overall machine configuration and computing, including CPU, storage, NV-Link communication connection, as well as factors such as power consumption, site rental and operation and maintenance costs. The total cost is more than 600 million yuan. OpenAI is using Microsoft's super expensive supercomputer, costing hundreds of millions of dollars, and using Nvidia's Azure cloud with tens of thousands of chips for computing training.
However, since OpenAI announced its GPT enterprise cloud business, the relationship between Microsoft and OpenAI has become increasingly distant, and there are reports that Microsoft is also developing customized AI chips.
In fact, it is not just OpenAI. In this wave of computing power demand, major technology giants such as Google, Amazon, Microsoft, and Meta have been designing AI chips into their businesses. For example, Google's latest Pixel 8 series of mobile phones use Google Tensor G3 chips and Titan M2 coprocessors, both of which are self-developed products of Google; and Amazon's latest self-developed ARM server CPU chip Graviton 3E is also used in its AWS cloud business.
However, OpenAI has no experience in developing AI chips. It takes at least two years for a chip to go from project initiation to mass production. It is not clear whether OpenAI will continue to advance its customized chip plan, and the effectiveness of the chip remains to be tested by time.
Reuters quoted industry insiders as saying that this is a huge investment, and the annual R&D cost may be as high as hundreds of millions of dollars. At the same time, OpenAI is considering acquiring some AI chip companies to speed up the process of developing its own chips.
But custom AI chips may take years, which means that OpenAI will still rely on chip suppliers such as Nvidia and AMD for a long time.
Interestingly, The Information reported on October 4 that Meta gave up on customizing VR/AR chips, the team had been disbanded, and turned to processor chips provided by Qualcomm.
Therefore, it remains to be seen whether OpenAI's customized AI chip will eventually be usable.
Other reports show that SoftBank plans to invest $1 billion to establish a new AI hardware company with Ultraman, and Apple designer Jony Ive has also discussed with Ultraman about creating a new AI hardware product.
Competition in the large model market intensifies, Google and Amazon turn to OpenAI competitors
Just a few days after Amazon announced its investment, it was reported that OpenAI's competitor, artificial intelligence (AI) unicorn company Anthropic, was raising additional funds.
According to The Information on October 4, Anthropic, which was founded only two years ago, is negotiating with Google and other investors, planning to complete a new round of financing of at least US$2 billion (approximately RMB 14.394 billion). The post-investment valuation may reach between US$20 billion and US$30 billion - more than five times the company's US$4 billion valuation in March this year.
Subsequently, competition in the AI ​​large model market intensified, and Google and Amazon also began to pay attention to and lay out OpenAI's competitors.
It is reported that Anthropic AI was founded in 2021 by Dario Amodei, former vice president of research at OpenAI, Tom Brown, the first author of the GPT-3 paper, and others. At that time, researchers led by Amodei left OpenAI after disagreements about the company's development direction. They were worried that Microsoft's investment in OpenAI would lead it to a more commercial path and deviate from the company's original idea.
In December 2022, the Anthropic team published a paper titled "Constitutional AI: Harmlessness from AI Feedback" on arxiv, describing a 52-billion-parameter model, AnthropicLM v4-s3, trained in an unsupervised manner, which directly targets OpenAI's GPT-3 model.
In January this year, Anthropic released Claude, a new AI chatbot model product based on AnthropicLM v4-s3, which is considered to be a strong competitor to ChatGPT.
On February 4 this year, Google Cloud, the cloud computing division of Google, announced a new partnership with Anthropic. It is reported that Google has invested nearly US$400 million (approximately RMB 2.03 billion) in Anthropic. The new financing will increase Anthropic's post-investment valuation to nearly US$5 billion.
Before Google's investment, Anthropic had raised more than $700 million in financing, with investors including Alameda Research, a cryptocurrency hedge fund founded by FTX founder SBF. One of the reasons why Google invested in Anthropic was that the company's CEO repeatedly expressed the threat and impact that ChatGPT brought to its search, advertising and other businesses.
According to The Information, Anthropic told investors that the company's annual revenue had been around $100 million, but due to its in-depth cooperation with Amazon, its revenue is expected to increase significantly in 2023 to $200 million, with monthly revenue approaching $17 million; by the end of 2024, Anthropic hopes that its annual revenue will reach $500 million, which is 1/200 of Anthropic's valuation and much higher than OpenAI's valuation multiples.