From GPT-5 to AI Chip Factories: How Big Is Sam Altman's Game?
-
This time OpenAI's ambitions are truly enormous.
Massive investment attraction to build global AI chip factories
As OpenAI's valuation approaches the $100 billion mark, Sam Altman's ambitions can no longer be concealed. He is no longer content with competing against Google and Apple, but is directly challenging the computing power giant Nvidia. According to informed sources, Sam Altman is actively seeking massive financial support from global investors to build an AI chip manufacturing plant. Among these investors are Abu Dhabi's G42 and Japan's SoftBank Group, with negotiations involving G42 alone reaching approximately $8 to $10 billion.
In fact, as early as October last year, G42 signed a cooperation agreement with OpenAI to provide artificial intelligence services for local and regional markets. Subsequently, in December, reports emerged that G42 was attempting to raise $100 billion, though at the time there was no information on the purpose of these funds. It now appears highly likely that this was in preparation for building an AI chip factory.
After all, realizing this plan requires facing enormous financial demands. From a cost perspective, advanced 2-nanometer or 3-nanometer process technologies alone require billions of dollars in development, with costs increasing as the node size decreases. Establishing a modern wafer fabrication plant capable of mass-producing chips at 2-nanometer or 3-nanometer nodes would require an investment of up to $30 billion. Additionally, the cost of a single low-numerical-aperture EUV lithography tool is approximately $200 million, while high-numerical-aperture lithography machines are expected to cost between $300 and $400 million. This means that Sam Altman would need to raise at least tens or even hundreds of billions of dollars to build an AI chip factory.
Although the complete list of partners and investors involved in the project has not yet been finalized, sources reveal that Sam Altman is attempting to collaborate with top-tier chip manufacturers, which may include TSMC, Samsung, and Intel, to secure the necessary capital.
Once the new chip factory is established, it will supply AI chips to the global market. Training GPT-5 Without Chips? A Full-Scale Battle Against NVIDIA
In fact, Sam Altman's rush to find partners to establish an AI chip factory is more out of necessity than choice. According to related media reports, OpenAI may already be facing a critical shortage of chips. This can be seen from Sam Altman's public statements over the past year, where he has consistently complained about the shortage of Nvidia GPUs, severely limiting OpenAI's GPU supply.
Due to this issue, OpenAI has been forced to postpone many short-term plans, such as fine-tuning, dedicated capacity, 32k context windows, multimodal capabilities, and more. It has even affected the reliability and speed of the API.
Additionally, the operational costs of ChatGPT are already enormous. According to statistics, each ChatGPT query costs approximately $0.04. If ChatGPT's query volume grows to one-tenth the scale of Google Search, it would initially require deploying AI chips worth about $48.1 billion for computation, with an annual need for chips worth around $16 billion to sustain operations. Amid intertwined factors, OpenAI has to seek new approaches to prevent being perpetually constrained by chip supply. According to Sam Altman, computational power limitations will become the primary obstacle for running artificial intelligence models.
He anticipates that the demand for AI applications and related computing power will surge dramatically in the coming years. OpenAI requires high-end chips, but there are already few global factories capable of producing them, which will severely restrict the training of next-generation AI models.
To address this issue, Sam Altman has taken the lead in proposing an initiative to establish a global network of factories to increase chip production and ensure sufficient computational resources. It is reported that Sam Altman has discussed chip supply issues with U.S. Congress members, including where and how to build new factories.
Considering the White House's recent series of subsidy policies for the chip industry, Sam Altman is likely to receive support. In August 2022, the U.S. CHIPS Act, totaling $280 billion, was officially signed into law. Of this, $52.7 billion will be allocated over the next five years to subsidize the construction and upgrading of chip factories, aiming to bring semiconductor manufacturing back to the U.S. and encourage companies to research and produce chips domestically.
This indicates that the U.S. government's plan to invest heavily in increasing chip production aligns with Sam Altman's current approach. Can OpenAI Complete This Grand Strategy?
However, Sam Altman's plan to rapidly build AI chip manufacturing plants is far from simple.
Globally, tech giants like Google and Microsoft are vigorously developing their own chips while still eagerly seeking business collaborations with Nvidia. This phenomenon stems not only from Nvidia's outstanding GPU performance but also from the comprehensive advantages of its long-cultivated CUDA ecosystem. This is precisely the constraint OpenAI is trying to break, but from a developer's perspective, OpenAI seems to be merely repeating similar patterns, changing the form but not the substance.
Moreover, OpenAI appears to carry an invisible "bomb." Due to significant risks in OpenAI's governance structure, developers' confidence in building services based on GPT will be significantly reduced, providing opportunities for competitors like Google and Meta.
Even if OpenAI one day overcomes all difficulties and builds an AI chip factory, the ensuing problems will not diminish. Taking CoWoS packaging capacity as an example, the shortage of NVIDIA's H100 chips is mainly attributed to insufficient CoWoS packaging capacity. TSMC recently stated during its earnings call that it would continue to expand advanced packaging, including CoWoS capacity. According to estimates from equipment manufacturers, TSMC's total CoWoS capacity in 2023 will exceed 120,000 units, and it is expected to reach 240,000 units by 2024.
This means that doubling production will take nearly a year, likely due to the complexity of the manufacturing process. Additionally, the HBM3 memory used in the H100 poses another significant production challenge, as this critical component is controlled by companies like Micron, SK Hynix, or Samsung.
Currently, almost all AI training and inference rely on the same type of generative AI chips. However, over time, more advanced GPUs, CPUs, or other new processors are likely to emerge, potentially leading to an oversupply of the currently used AI chips. These are issues that OpenAI will inevitably have to address. From GPT-5 to AI chip factories, OpenAI still has a long way to go.