GPT-5 was stuck in the neck by 50,000 H100s! Ultraman raised billions of dollars to replace Nvidia to build an AI chip empire
Recently, foreign media revealed that Sam Altman is planning to raise billions of dollars to build a global semiconductor fab network for OpenAI. GPT-5 training is seriously lacking in cores, and OpenAI vows not to let Nvidia take the initiative, so why don't you become the new AI chip overlord yourself?
Sam Altman is raising billions of dollars to build a global network of semiconductor fabs!
The reason behind this incident is likely to be that OpenAI no longer has a "core" to train "GPT-5".
Previously, according to the Financial Times, OpenAI is developing a new AI model that will be a "major upgrade" to GPT-4 and is expected to be released later this year.
GPT-4 was trained with about 25,000 A100 GPUs. And to train GPT-5, you need 50,000 H100s. Today, Nvidia's H100 sells for $25,000 to $30,000.
NVIDIA's AI chips basically monopolize the market, how can this lifeblood be in the hands of others?
Sure enough, foreign media have recently revealed that Sam Altman is negotiating with Middle Eastern investors and TSMC to start a partnership to train and run AI model chips.
Computing currency determines the future of OpenAI, and it must not be in the hands of NVIDIA! Altman has made up his mind this time.
Training GPT-5, OpenAI's demand for chips will only increase
Altman's ambition is to build a self-sufficient semiconductor supply chain empire in the coming years with the widespread adoption of AI technology.
So, OpenAI is also interested in competing with industry giants such as Intel, TSMC and Samsung?
Altman said at the Davos Economic Forum that the world's two major currencies in the future will be computing power and energy
Compared with Amazon, Google, Microsoft and other big manufacturers, Sam Altman obviously has a bigger plan: to build a network of AI chip factories.
Altman should be very sure that the current foundries such as TSMC, Samsung, and Intel will definitely not be able to meet the demand for AI chips in the next few years.
Now, Altman is raising billions of dollars with the goal of building a global network of AI chip factories.
Now, he is in talks with a number of potentially large investors, including Abu Dhabi-based G42 and SoftBank Group.
Obviously, in the era of AGI, chips will be in short supply. Altman is now concerned that as AI technology becomes more widely available, the existing chip supply will not be able to meet the demand for large-scale deployments.
The AI chips currently being produced are far behind the expected demand. Only by acting now can we ensure that there will be an adequate supply of chips within a decade.
However, building a global network of chip factories requires a huge investment and takes many years.
And, unlike other companies in the industry, the cost of building and maintaining a semiconductor factory is much higher. The cost of building a state-of-the-art plant can be tens of billions of dollars.
Amazon, Google, and Microsoft all tend to design their own custom chips and outsource manufacturing, which is also because the cost of building and maintaining semiconductor fabs is too high!
After all, building a state-of-the-art fab can cost tens of billions of dollars in investment, and building such a network of facilities can take years.
According to a Bloomberg report, the amount involved in OpenAI's negotiations with the G42 alone is close to $8 billion to $10 billion.
OpenAI新金主:中东土豪
OpenAI's former financier was Microsoft. And now, in order to raise funds, Altman has reached out to Middle Eastern tycoons, such as some investors in the United Arab Emirates.
One of them is Sheikh Takhnoon, one of the richest and most influential people in Abu Dhabi.
Sheikh Takhnoun is one of the most powerful people in the UAE, the brother of President Sheikh Mohammed, and the UAE's national security adviser.
He also oversaw a rapidly expanding business empire and chaired some of Abu Dhabi's strongest national investment funds. These include the $800 billion Abu Dhabi Investment Authority and another state-owned investment entity, ADQ.
In addition, he is the Chairman of the International Holding Company and the G42. The former is a huge conglomerate that has quickly become the largest publicly traded company in the UAE, while G42 is an ambitious AI company that has established partnerships with Microsoft and OpenAI.
It's unclear exactly how much Altman will raise, but it would cost at least billions of dollars to compete with Nvidia, which has a market capitalization of nearly $1.5 trillion.
G42集团CEO Peng Xiao和Sam Altman签订协议
Not to mention, Sam Altman's social skills have really reached the level of the human ceiling.
The cost of building a factory is astronomical
Let's take a look at how much it will cost Sam Altman to build a fab.
The cost of developing 2nm or 3nm process technology is in the billions of dollars, and this cost is rising as process nodes shrink.
At the same time, a modern fab capable of mass-producing 3nm or 2nm chips can now cost up to $30 billion.
In addition, the cost of fabs is rising rapidly, for example, a low-NA extreme ultraviolet (EUV) lithography equipment costs about $200 million, while a high-NA lithography machine is expected to sell for between $300 million and $400 million.
For a leading fab to produce state-of-the-art AI and high-performance computing (HPC) chips, it would have to have several of them.
In the wave of AI, Nvidia is reaping the benefits of fishermen
Now, Google, Amazon, Meta, OpenAI, and Microsoft are all using Nvidia's GPUs to train AI and deploy models to customers. Meta alone plans to install 340,000 H100s for servers by the end of the year.
It can be said that Nvidia has monopolized the current AI computing power market and holds pricing power, which has led to the rapid growth of its revenue.
Chip startups like Graphcore are already struggling to compete with Nvidia's dominance, which is caused by the interplay between software and hardware. It's not enough to just make a faster chip, not to mention that it's quite difficult in itself.
Obviously, the major technology companies have started to lay out early.
Big tech companies like Amazon, Google, and Microsoft do just that – designing their own custom semiconductors and outsourcing the production process to other companies.
Now, it's harvest time for every family. At the end of November last year, Microsoft launched its first AI chip, and at the same time, Microsoft is also stepping up cooperation with AMD; before the spring of 23, Meta also launched its own chip; Google and Amazon, respectively, have developed TPU and Trainium chips for many years.
Microsoft's latest Azure Maia 100 chip
OpenAI曾怒甩5100万刀狂买AI芯片
In December last year, Altman was revealed to be conducting a "chip transaction", which is suspected to have become the fuse of OpenAI's palace fight.
At that time, foreign media revealed that OpenAI signed a letter of intent worth $51 million with AI chip startup Rain AI in 2019 to buy chips after Rain AI's chips were launched.
Rain AI is developing a "brain-like" NPU chip that can significantly reduce the cost of AI computing power, and is expected to be taped out in December and available in October 2024.
And it is worth noting that Sam Altman, as a shareholder of Rain AI, also personally invested $1 million.
According to a person who did not want to be identified, Sam Altman was previously fired by OpenAI's former board of directors, in part because of his other investments tangled with OpenAI.
It is reported that Altman had raised funds in the Middle East for the project, codenamed Tigris
The neuromorphic processing unit (NPU) developed by RainAI mimics the functions of the human brain and has the potential to provide higher processing power and energy efficiency than today's GPUs.
This NPU "brain-like" chip claims to have 100 times more computing power than GPUs, and the energy efficiency in training is even 10,000 times higher than that of GPUs.
Rain's goal is to provide a chip that can be used both for model and algorithm training and for later inference runs.
According to Rain, such a chip will allow AI models to be customized or fine-tuned in real-time based on their surroundings.
- From this point of view, it is not a direct competitor to GPUs such as the Nvidia H100.
Officials say these features are a strong attraction for OpenAI, which hopes to use the chips to reduce costs in data centers and deploy its own models in devices such as phones and watches.
If it can be successfully developed, OpenAI will naturally no longer be subject to NVIDIA.
Actually, Altman's layout was very early.
Back in 2018, he led Rain's seed round of funding, and a year later, OpenAI approved the $51 million letter of intent to purchase the chip.
Of course, Altman's plan wasn't all smooth sailing.
Previously, Rain's leadership had been reshuffled, and the company's investors had changed. A multiagency government agency that oversees national security ventures asked Saudi Arabia-affiliated Prosperity7 Ventures to sell its stakes in Rain, which Silicon Valley's Grep VC acquired.
-- All these changes may make it more difficult for Rain to bring the new chip technology to market, and also make the fulfillment date of OpenAI's $51 million order less clear.
Overall, the deal with Rain also shows that OpenAI is willing to spend a lot of money to secure the supply of chips needed for its own AI projects.
Now, the $51 million worth of AI chips purchased from Rain AI is just a small part of OpenAI's huge investment in AI chips.
OpenAI's ambitions are being quietly deployed
A few years ago, OpenAI appointed the former head of Google's TPU as its head of hardware, and is hiring "data center facility design experts".
Silicon engineer Richard Ho is leading a new division of generative AI startups to help optimize partners' data center networks, racks, and architectures, according to the news.
"He also played an important role in the creation of TPU. He also worked at DE Shaw and designed ASICs. So far, he has been more responsible for software and hardware integration, DC design, and accelerator chip selection. But OpenAI has also recently hired a number of compiler and kernel experts. 」
In addition, Altman has also held discussions with semiconductor executives, including chip design company Arm, on how to design new chips as soon as possible to reduce costs for OpenAI.
How serious is OpenAI's GPU shortage?Altman: Don't use ChatGPT yet
Throughout last year, Sam Altman complained that OpenAI had fallen into a severe GPU shortage.
The explosion of generative AI has not slowed down, putting forward higher requirements for computing power. Many companies are using NVIDIA's extremely high-performance H100 to train their models, but the H100 is very expensive.
Musk said that GPUs are now more in demand than drugs.
According to Sam Altman, OpenAI has been severely constrained by GPUs and has had to postpone numerous short-term plans (fine-tuning, dedicated capacity, 32k context windows, multimodality).
Even, because of the shortage of GPUs, Altman doesn't want a lot of people to use ChatGPT.
We have a huge shortage of GPUs, and the fewer people who use our products, the better.
If people use less, we'll be happy because we don't have enough GPUs.
Many OpenAI users have complained about the reliability and speed of the API, and Sam Altman explained that the reason is also that there is a shortage of GPUs.
To train GPT-5, you need 50,000 H100s, but it has been previously reported that Nvidia's best chip, H100, will be sold out by 2024.
If GPUs on the market can't keep up, it will hinder OpenAI's ability to improve and train new models.
Will Altman, now on the go, be able to raise billions or even hundreds of billions of dollars to build a fab, and will it be able to keep the fab afloat afloat?
At present, we only know that his move has the potential to change the pattern of the entire foundry market.
Resources:
https://www.tomshardware.com/tech-industry/artificial-intelligence/openais-sam-altman-raises-billions-to-build-chip-empire-report
https://the-decoder.com/gpt-4-successor-reportedly-a-major-upgrade-openais-altman-in-talks-with-tsmc-for-ai-chips/
https://www.ft.com/content/1cdaadc3-b384-4f50-88ff-291c062c8376