Recently, Open AI plans to build a core buzz. According to the news, OpenAI CEO Altman has contacted Intel, TSMC want to cooperate to set up a new chip factory. In addition, when Altman visited Korea in January, he also set his sights on Samsung and SK, also for the chip.
In fact, in October last year, Altman expressed the desire to self-research chip, to the present Altman seems to want to further layout of the core. From self-research to self-made, Altman chip “ambition” can not hide.
Ultraman’s Ambition – Self-developed Chip
The matter of self-developed chips is not only Open AI, but also many big manufacturers have tried. Only, Open AI’s entry is a little later. In October last year, foreign media reported that Open AI was exploring the manufacture of its own artificial intelligence chip.
And, Open AI’s website began to recruit hardware-related talent, the official website has several hardware and software co-design positions in the recruitment, at the same time in September last year Open AI also recruited a famous bull in the field of artificial intelligence compiler Andrew Tulloch to join, which seems to confirm that Open AI self-research chip investment.
In a public conversation, Altman said OpenAI is not ruling out the option of developing its own chips.
Altman said, “We’re still in the process of evaluating whether or not to go with custom hardware [chips]. We are trying to determine how to scale up to meet the needs of the world. While it’s possible that we won’t develop a chip, we’re maintaining good collaboration with partners who have produced excellent results.”
build one’s own chip
Open AI’s real ambitions go beyond building its own chips.
Now, Open AI also wants to build its own chips. OpenAI CEO Sam Altman is reportedly planning to use billions of dollars to build a semiconductor fab of a certain size. The centerpiece of this project is to establish a “network of factories” to increase chip production.
In order to raise funds for this project, Altman is pulling together the Middle East governments, including Abu Dhabi and Saudi Arabia, as well as Silicon Valley investors. According to the Financial Times, Altman is in talks with Middle Eastern investors about chip investments to fund the chips and chip manufacturing plants needed to develop big AI models.
The negotiators include the richest investor in the Middle East and one of Abu Dhabi’s wealthiest and most influential figures – Sheikh Tahnoon bin Zayed al-Nahyan. He is the brother of Sheikh Mohammed bin Zayed al-Nahyan, President of the UAE, and the UAE’s National Security Advisor, serves as Chairman of the Abu Dhabi Investment Authority (ADIA), Abu Dhabi’s third-largest sovereign wealth fund ADQ (Abu Dhabi Development Holding Co.), is the Abu Dhabi chairman of artificial intelligence firm G42, which also has partnerships with Microsoft and OpenAI.
Altman is also pulling in top chipmakers, including TSMC and Intel. In addition, during his recent visit to South Korea in January, Altman discussed collaboration with South Korean chip giants Samsung and SK. One of the main topics discussed in their meetings was high-bandwidth memory (HBM) chips, the most modern memory chips used to process large amounts of data, which are critical to AI processors.
In fact, Sam Altman noticed the two chip giants during his visit to Korea back in June last year and expressed his willingness to invest in Korean AI startups.Econotimes cited an unnamed official as saying that SK and Samsung could be in charge of the development and production of memory chips if the cooperation is finalized.
In addition to building its own factories, Altman has invested early in three chip startups, namely Cerebras, Rain Neuromorphics and Atomic Semi.
Cerebras has launched mega chips before, and Cerebras’ second-generation AI chip, WSE-2, has 2.6 trillion transistors and 850,000 AI cores, breaking the world record in a number of indicators. This company open-sourced seven GPT models in one go last year, with the number of references reaching 111 million, 256 million, 590 million, 1.3 billion, 2.7 billion, 6.7 billion, and 13 billion respectively.
The other is Rain Neuromorphics, a company that designs chips that mimic the way the brain works, and is part of the Neuromorphic Chip startup category.Rain AI’s first chips are based on the RISC-V open-source instruction-set architecture, and are aimed at devices including cell phones and drones/robots, with the highlight being the ability to both train the algorithms and run them when deployed. While the first batch of hardware has yet to be delivered to customers, OpenAI has long placed a $51 million pre-order with Rain AI.
There’s also Atomic Semi, co-founded by chip guru Jim Keller and “garage chipmaker” Sam Zeloff, who aim to streamline chip production and processing in order to produce cheaper chips in a matter of hours.
Ultraman’s Worries
At this stage, Ultraman is perhaps the most GPU hungry person in the world.
GPUs have always been a heavy burden for Ultraman. Back in 2022, he publicly expressed his displeasure with NVIDIA’s GPU chip shortage, claiming it was putting a huge strain on the company.
He complained in several interviews that “OpenAI is currently suffering from severe constraints on GPU computational power, resulting in many short-term programs not being completed on time.”
In mid-November, Open AI suddenly announced that registration for ChatGPT Plus accounts had been suspended for no other reason than that the surge in visits had exceeded the capacity of the servers.
Altman explained this in a post on X: As the surge in ChatGPT usage after the OpenAI development day was more than we could handle, we wanted to make sure everyone had a good experience. You can still sign up for ChatGPT within the app to be notified when ChatGPT Plus reopens.
Regular ChatGPT accounts were still able to register at that time, but when you wanted to purchase the Plus service, you were prompted – Signed up for the waiting list with the reason: We’ve temporarily suspended the upgrade service due to high demand.
OpenAI has published a set of data that the growth rate of arithmetic power required for large model training keeps growing at a rate of 3 to 4 months/multiple, far exceeding Moore’s Law of 18 to 24 months/multiple. Powerful arithmetic power means faster data processing speed, more powerful big model service capability.
At the London hearing, Altman said ” most of the problems are caused by the shortage of GPUs.”
First of all OpenAI many customers are complaining about the reliability and speed of the API, without enough GPUs, there is no way this can be solved.
Second, due to the shortage of GPUs, ChatGPT’s longer 32k context capacity (about 24,000 words) can’t be pushed out to a wider range of customers for the time being.
Third, due to the shortage of computing power, ChatGPT’s model fine-tuning API is not well supported, and efficient fine-tuning methods such as LoRa cannot be used.
Fourth, due to the shortage of computing power, OpenAI can’t sell more proprietary customized models to customers, and the solution now is to require customers to pay up to $100,000 deposit in advance.
Based on the above problems, although Altman and OpenAI are the creators of the big model wave, but they have to rely on NVIDIA’s production capacity, NVIDIA in the field of AI occupies nearly 80% of the market share, NVIDIA’s production capacity is not a day, OpenAI will not be able to develop as soon as possible.
Open AI faces not only the shortage of GPUs, but also its high price.
Whether it’s buying NVIDIA’s GPUs or using GPU-based cloud servers, it’s too expensive. Last year, OpenAI recorded a revenue of $28 million last year, while the overall loss was $540 million, and the main reason behind OpenAI’s huge loss was due to arithmetic overhead.
According to an analysis by Bernstein, an American financial firm, if ChatGPT’s visits reach the level of one-tenth of Google’s searches (and this is one of OpenAI’s key goals for the future), OpenAI will initially need GPUs worth as much as $48.1 billion, and will need to spend $16 billion per year on chips to keep it running.
This kind of overhead could be a major bottleneck for further scaling of OpenAI in the future. After all, even as strong as Microsoft, can not support such a huge investment in the long term.
For OpenAI, building its own chips means security and more controllable costs in the long run.
variable
First, let’s look at the variables of self-developed chips. In fact, at present there are not a few large manufacturers of self-developed chips, such as Facebook’s parent company, Meta is currently developing a new type of chip, hoping to cover all types of AI; OpenAI’s main supporter, Microsoft, is also developing a custom AI chip, and handed over to OpenAI for testing.
However, there are three major problems with personally doing chip design and development.
First, the requirements for talent density are very high.
Second, the investment is not small. Hundreds of high-end chip teams as well as tens of millions of dollars of single flow cost, are heavy investments, and can not be expected to succeed in a flow. According to Semiengingeering data shows that the development of a 28nm node chip investment of about 51.3 million dollars, 7nm node chip is up to 297 million dollars.
Third, find a suitable foundry. 7nm foundry, can only find TSMC and Samsung these two, but how to make TSMC and Samsung actively cooperate with the project progress, but also test Open AI and the upstream Fab factory dealing with the ability of
Then if you add your due tofactory, the variables will be even greater.
First of all, it’s still a question of money. As we mentioned, even if the cost of self-developed chip flow is high, then the cost of directly building a fab will only go up.
Wafer fabs cost a lot of money to build. Semiconductor company Intel has said that a fab takes 6,000 workers three years to build and costs 30 billion.
But as a startup, Open AI is still short of money.
Last April, OpenAI completed the latest round of financing, valued at $27 billion to $29 billion. Microsoft is currently OpenAI’s largest shareholder, holding a 49% stake. Other financial investors include Tiger Global Fund, Sequoia, California-based Andreessen Horowitz, New York-based Thrive and K2 Global, Founders Fund, and others.
At the time, Altman responded to the valuation by emphasizing that OpenAI currently needs to raise a significant amount of capital to find R&D talent, as well as to drive innovative research rapidly.
“We need a lot of funding to fulfill the OpenAI mission. More interesting things must be discussed in our limited time than our future funding plans. Of course, we need more money; the path forward for AI is unclear, but innovation is expensive. On the product side, we are looking for R&D talent and fast-tracking innovation, and are working hard to drive various applications of AI and partner with customers.” Altman said.
By December, word came out that a new round of funding was in preliminary talks, and that OpenAI’s valuation could reach or exceed $100 billion due to that funding.
But in any case, more financing, spend more. open AI spends a lot of money, if you add the self-research chip and cover the fab, it is still not enough to plug the gap.
Secondly, is the problem of demand, the real capacity demand for AI chips can be how much? Definitely can not Cover a fab operating costs.
Take a look at the current wafer foundry, TSMC, Samsung, and SMIC, which foundry businesses due to being included in the semiconductor industry chain of various products, from 40nm to 3nm, even so in the semiconductor cycle downturn, the business has fluctuations. Therefore, specializing in building a fab for AI chip manufacturing is also not a realistic thing.
In retrospect, the curtain of the AI era opened, OpenAI actually does not have an absolute advantage, and today it is already a melee in which each one shows its strengths.
From GPT-5 to AI chip factory, I don’t know whether Open AI can hold up the ambition of Ultraman.
