OpenAI will raise $7 trillion to build cores? Jensen Huang: This money can buy all the GPUs in the world!
Following the previous foreign media revelations that OpenAI CEO Sam Altman (Sam Altman) is planning to raise billions of dollars to build a wafer factory to produce self-developed artificial intelligence (AI) chips, recently, the Wall Street Journal quoted a number of anonymous sources as saying that Altman's ambitious project will be crazier than expected, and the planned scale of funds raised is as high as $7 trillion.
However, judging by this exaggerated figure, it is clear that it is not more logical. You know, Gartner data shows that the global semiconductor market size in 2023 will only be $533 billion, while $7 trillion will reach more than 13 times. Although the global semiconductor market has encountered a trough in 2023, analysts expect that the global semiconductor market will only grow by 17% year-on-year to $624 billion this year.
And $7 trillion has accounted for 10% of global GDP last year, which is equivalent to the market value of 2.3 Microsoft, 3.8 Google, 3.8 Nvidia, 5.8 Meta or 10.4 TSMC (as of February 14). If Altman can really raise such a huge $7 trillion in funds, it will be more than enough to buy all important chip design, manufacturing and equipment manufacturers such as Nvidia, Intel, AMD, Qualcomm, Broadcom, TSMC, Samsung, and ASML. If so, then is it necessary to spend trillions of dollars to build your own fab?
Even if Altman can't think of raising $7 trillion to build a fab, then according to the current data, it may take $20 billion to $30 billion to build a cutting-edge process fab, depending on factors such as the size of the factory's capacity and geographical location, if it is estimated at $30 billion, $7 trillion is enough to build 233 cutting-edge fabs, which will be more than 10 times the total number of existing wafer fabs with the most advanced process.
The question then arises, who can build such a large-scale cutting-edge process fab, the equipment and materials required will be beyond the reach of current suppliers, and a large number of installation, maintenance and operation personnel will also be required, which is also not available in the current talent market.
For example, ASML's current lithography machine capacity is not enough to meet the needs of existing customers. At the end of 2023, ASML's undelivered order backlog remained at €39 billion. Previously, a joint report released by the Semiconductor Industry Association (SIA) and Oxford Economics also warned that the US semiconductor industry will face a shortage of 67,000 technicians, engineers and computer scientists by 2030.
So is building a fab just part of a huge $7 trillion fundraising project? Perhaps Altman plans to build his own fab while also planning to build more AI data centers to meet the global demand for AI computing power in the future, which will undoubtedly cost a huge amount of money.
If OpenAI plans to purchase 5 million NVIDIA H100 accelerator cards, it will only cost $100 billion based on the estimated price of about $20,000 for a single accelerator card. However, Nvidia's total revenue in 2023 will only be $30.3 billion, and even if its revenue doubles this year, it will only be about $60 billion. $100 billion is only about 1.4% of the rumored $7 trillion.
Obviously, the revelation that Altman plans to raise $7 trillion to build a fab seems to be an outright rumor, no matter how you look at it.
It is worth noting that on February 12, local time, Nvidia founder and CEO Jensen Huang was also asked in an interview at the World Government Summit (WGS) in Dubai, how many GPUs can be bought for $7 trillion?
Huang's answer: "All GPUs." He noted that currently, the total value of global data centers is about $1 trillion. At the same time, he also expects this figure to grow to $2 trillion in the next 4-5 years. "We are at the beginning of this new era. In the next four to five years, we will have $2 trillion worth of data centers to power software around the world. ”
Huang emphasized that we don't need much investment to build an alternative semiconductor supply chain specifically for the production of AI chips. Instead, what the industry needs is to continue to innovate GPU architectures to continue to improve performance. In fact, Nvidia has improved the performance of AI chips by a factor of 1 million over the past decade.
Huang also noted that if we were to spend $7 trillion to build a fab, assuming that chips don't go faster, "we'd need 14 planets, three galaxies and four suns to power it all." ”
Moreover, even if this amount of money is invested to build a supply chain for AI chips, it is expected to solve the problem of chip supply shortage in the next three to five years. However, this is just to create an alternative chip supply chain, and there is still a risk of oversupply of chips in the future, which is not a good approach. Instead, it should be about innovating GPU architectures to make computing faster so that companies that want to use AI locally will not have to buy a lot of equipment and build multibillion-dollar data centers.
Huang further emphasized that the fact is that Nvidia's GPUs are evolving very fast in terms of AI and HPC performance. In 2018, NVIDIA's V100 data center GPU only had 125 TFLOPS of computing performance, but NVIDIA's current H200 data center GPU provides 1,979 FP16 TFLOPS. As a result, one of the greatest contributions Nvidia has made has been a million-fold advancement in computing and artificial intelligence over the past decade. So that whatever the demand will be able to power the world, this is a fact that must be taken into account. Moreover, it is expected that the speed of transporting 10,000 will increase by another 1 million times in the next ten years.
Editor: Xinzhixun-Rogue Sword