GPT-5 Bottlenecked by 50,000 H100 GPUs! Altman Rushes to Raise Billions
-
Recent reports reveal Sam Altman is planning to raise billions to build a global semiconductor wafer fab network for OpenAI. GPT-5 training faces severe chip shortages. Is OpenAI determined to break Nvidia's dominance by becoming the new AI chip powerhouse itself?
Sam Altman is raising billions to establish a global semiconductor wafer fab network!
The underlying reason for this move is likely that OpenAI currently lacks sufficient chips to train GPT-5. Previously, according to the Financial Times, OpenAI is developing a new AI model that will be a "significant upgrade" to GPT-4, expected to be released later this year.
Training GPT-4 used approximately 25,000 A100 GPUs. For GPT-5, an additional 50,000 H100 GPUs will be required. Currently, Nvidia's H100 is priced between $25,000 and $30,000.
Nvidia's AI chips have essentially monopolized the market. How can such a critical lifeline be left in others' hands?
Indeed, recent foreign media reports have revealed that Sam Altman is negotiating with Middle Eastern investors and TSMC to establish partnerships for chips to train and run AI models. Computing power currency determines OpenAI's future, and it must not be controlled by NVIDIA! Altman is determined this time.
Training GPT-5 will only increase OpenAI's demand for chips.
Altman's ambition is to build a self-sufficient semiconductor supply chain empire in the coming years as AI technology becomes more widely applied.
So, is OpenAI also planning to compete with industry giants like Intel, TSMC, and Samsung? Altman stated at the Davos Economic Forum that the two major currencies of the future world will be computing power and energy.
Compared to tech giants like Amazon, Google, and Microsoft, Sam Altman clearly has a more ambitious plan: to build a network of AI chip factories.
Altman seems quite certain that current manufacturers like TSMC, Samsung, and Intel will be unable to meet the demand for AI chips in the coming years.
Now, Altman is raising billions of dollars with the goal of establishing a global network of AI chip factories. He is currently in negotiations with several potential large investors, including Abu Dhabi-based G42 and SoftBank Group.
Clearly, in the era of AGI, chips will be in short supply. Altman is now very concerned that as AI technology becomes increasingly widespread, the existing chip supply will not be able to meet the demand for large-scale deployment.
The current production of AI chips is far behind the expected demand. Only immediate action can ensure an adequate supply of chips within the next decade.
However, building a global network of chip factories requires massive investment and will take many years. Moreover, unlike other companies in the industry, the cost of building and maintaining semiconductor fabrication plants is much higher. The construction cost of an advanced facility can reach tens of billions of dollars.
Companies like Amazon, Google, and Microsoft tend to design their own custom chips and outsource manufacturing because the cost of building and maintaining semiconductor wafer fabs is simply too high!
After all, constructing a state-of-the-art wafer fab may require an investment of tens of billions of dollars, and establishing such a network of facilities can take years. According to Bloomberg, the negotiations between OpenAI and G42 alone involve an amount close to $8 to $10 billion.
OpenAI's previous major investor was Microsoft. Now, to raise funds, Altman has reached out to wealthy Middle Eastern investors, including some from the UAE.
One of these investors is Sheikh Tahnoon, one of the wealthiest and most influential figures in Abu Dhabi. Sheikh Tahnoon is among the most powerful individuals in the UAE, a brother of President Sheikh Mohammed, and also serves as the UAE's National Security Advisor. He is responsible for overseeing a rapidly expanding business empire and serves as chairman of several of Abu Dhabi's strongest sovereign wealth funds. These include the $800 billion Abu Dhabi Investment Authority and another state-owned investment entity ADQ.
Additionally, he chairs International Holding Company and G42. The former is a massive conglomerate that has rapidly become the UAE's largest listed company, while G42 is an ambitious AI firm that has established partnerships with Microsoft and OpenAI.
It remains unclear about the exact fundraising amount for Altman, but to compete with Nvidia whose market value approaches $1.5 trillion, it would conservatively cost at least billions of dollars. G42 Group CEO Peng Xiao and Sam Altman signed an agreement.
Putting everything else aside, Sam Altman's social skills have truly reached the pinnacle of human capability.
Building factories requires astronomical sums of money Let's break down how much Sam Altman would need to spend to build a semiconductor fab.
The development cost for 2nm or 3nm process technology runs into tens of billions of dollars, and this figure continues to rise as process nodes shrink.
Meanwhile, a modern fab capable of mass-producing 3nm or 2nm chips can now cost up to $30 billion.
Moreover, fab costs are rapidly increasing. For instance, a low numerical aperture (Low-NA) extreme ultraviolet (EUV) lithography machine costs approximately $200 million, while high numerical aperture (High-NA) lithography equipment is expected to be priced between $300 million and $400 million. To produce the most advanced AI and high-performance computing (HPC) chips, a leading fab would need several of these machines. In the AI wave, NVIDIA is reaping the benefits.
Currently, 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 H100 chips in its servers by the end of the year.
It can be said that NVIDIA monopolizes the current AI computing market and holds pricing power, leading to its rapid revenue growth.
Chip startups like Graphcore find it increasingly difficult to compete with NVIDIA's dominance, a situation driven by the interplay between software and hardware. Simply making a faster chip is far from sufficient, especially since this is already quite challenging. Clearly, major tech companies started their strategic layouts early on.
Tech giants like Amazon, Google, and Microsoft have adopted this approach—designing their own custom semiconductor products while outsourcing the manufacturing process to other companies.
Now, each company is reaping the rewards. By late November last year, Microsoft had launched its first AI chip while intensifying collaboration with AMD. Before spring 2023, Meta also introduced its own chip. Meanwhile, Google and Amazon have been developing TPU and Trainium chips respectively for many years. OpenAI Once Spent $51 Million to Buy AI Chips
In December last year, Altman was exposed to be engaged in "chip deals," which is suspected to have become the trigger for the OpenAI power struggle.
At that time, foreign media reported that OpenAI signed a $51 million letter of intent with AI chip startup Rain AI in 2019, agreeing to purchase chips once 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. The chip is expected to tape out in December and begin shipments in October 2024. Notably, Sam Altman, a shareholder in Rain AI, personally invested $1 million in the project.
According to an anonymous source, Sam Altman's previous dismissal from OpenAI's board was partly due to the entanglement between his other investments and OpenAI. It is reported that Altman had raised funds in the Middle East for this project, codenamed Tigris.
The neuromorphic processing unit (NPU) developed by Rain AI mimics the functionality of the human brain and promises higher processing power and energy efficiency compared to today's GPUs.
This 'brain-like' NPU chip claims to offer 100 times the computing power of GPUs and up to 10,000 times the energy efficiency in training tasks. Rain's goal is to provide a chip that can be used for both model and algorithm training and subsequent inference operations.
According to Rain, this chip will allow AI models to customize or fine-tune in real-time based on their surrounding environment.
From this perspective, it is not a direct competitor to current GPUs like NVIDIA's H100.
Relevant sources indicate that these features are highly attractive to OpenAI, which hopes to use these chips to reduce data center costs and deploy its models in devices such as smartphones and watches. If successful, OpenAI would no longer be constrained by Nvidia.
In fact, Altman's planning began very early.
As far back as 2018, he led the seed round financing for Rain, and a year later, OpenAI approved this $51 million letter of intent to purchase chips.
Of course, Altman's plan hasn't been without its challenges. Rain had previously undergone leadership reshuffles and changes in investors. A cross-departmental government agency overseeing national security risk investments required Saudi-affiliated fund Prosperity7Ventures to divest its stake in Rain, after which Silicon Valley's Grep VC acquired these shares.
These changes may increase the difficulty for Rain to bring its new chip technology to market, while also making the delivery date for OpenAI's $51 million order less certain.
Overall, this deal with Rain demonstrates OpenAI's willingness to spend significant funds to secure chip supplies for its AI projects.
Currently, the $51 million AI chips purchased from Rain AI represent just a fraction of OpenAI's massive investments in AI chips. OpenAI's ambitions are quietly being deployed.
Years ago, OpenAI appointed the former head of Google's TPU as its hardware lead and is currently recruiting 'data center facility design experts'. Reports suggest that Richard Ho, a chip engineer, is leading a new division at the generative AI startup to optimize partners' data center networks, racks, and architecture. 'He played a significant role in the creation of the TPU. He has also worked at DE Shaw and designed ASICs. So far, he has been more involved in software-hardware integration, data center design, and accelerator chip selection. But OpenAI has recently hired many experts in compilers and kernels.'
Additionally, Altman has held discussions with semiconductor executives, including those from chip design company Arm, on how to design new chips early to reduce costs for OpenAI.
How severe is OpenAI's GPU shortage? Altman: Everyone, stop using ChatGPT for now. Throughout last year, Sam Altman has been complaining that OpenAI is suffering from a severe GPU shortage.
The explosive growth of generative AI shows no signs of slowing down, placing even higher demands on computing power. Many companies are using NVIDIA's high-performance H100 GPUs to train models, but the H100 is extremely expensive.
Elon Musk has even remarked that GPUs are now harder to come by than drugs.
Sam Altman stated that OpenAI is severely constrained by GPU limitations, forcing the company to postpone numerous short-term plans (fine-tuning, dedicated capacity, 32k context windows, and multimodal capabilities). In fact, due to the GPU shortage, Altman doesn't want too many people using ChatGPT.
We're extremely short on GPUs. The fewer people using our product, the better.
We'd actually be happy if fewer people used it, because we simply don't have enough GPUs.
Many OpenAI users have complained about API reliability and speed issues. Sam Altman explains that this is also due to the critical GPU shortage. Training GPT-5 requires 50,000 H100 GPUs, but previous reports indicated that Nvidia's best chip, the H100, may sell out before 2024.
If the GPU supply cannot keep up, it will hinder OpenAI's ability to improve and train new models.
Now, the question is: Can Altman, who is actively seeking funding, raise tens of billions or even trillions of dollars to build semiconductor fabs? And can he sustain their operations?
For now, we only know that his actions could potentially reshape the entire foundry market landscape.