Ultraman reveals the first phase of the "Stargate" factory, with 20 supercomputers being built across the U.S.! 700,000 GPUs make Jensen Huang a profit of 37.5 billion

Early in the morning, Ultraman just finished a live broadcast and showcased the first phase of the "Interstellar Gateway" factory. What we see above is just Base 1 of the "Interstellar Gateway." It is reported that Abilene, Texas has been approved for 875 acres of land (approximately 3.54 million square meters), an area comparable to New York's Central Park, with plans to build 10 data centers. In the future, a total of $500 billion will be spent over 4 years to construct a total of 20 supercomputing centers across the United States. Ultraman's latest post is a strong response to Musk's mockery yesterday. Ultraman and Musk are directly confronting each other, tearing each other apart. Could it be that everything Musk said is just a gimmick? Since the official announcement of the "Interstellar Gateway" on Tuesday, it has attracted widespread attention, especially the $500 billion investment which has sparked heated discussions throughout the tech community. It is said that this amount is equivalent to the total of the "Manhattan Project + Apollo Program." ![](https://mmbiz-qpic.wscn.net/sz_mmbiz_png/UicQ7HgWiaUb1xNzGq6qEaIWnJyewa9icAvP0QQSRWibbeGGwC8cV8BiboJbD3HhACWSwz8xU4UF2npFjb62phicE3icg/640? This astonishing figure has inevitably sparked speculation from all sectors. As Musk mentioned in yesterday's battle, the actual financing obtained by SoftBank is far less than $10 billion. So, what is the real situation? SemiAnalysis exclusively revealed that, in fact, the first phase of the "Stargate" has been confirmed and is underway at the data center base in Abilene, Texas. However, they confirmed Musk's statement—SoftBank indeed does not have enough funds to build such a large-scale computing cluster. In the following text, they will analyze in detail the actual situation of the "Stargate," the current crisis facing Microsoft, the project structure and costs, as well as the true state of this computing cluster. $100 Billion Cluster Locked In In addition to the real scene released by Ultraman, SemiAnalysis also created a macro map of the cluster construction in Texas. The confirmed first phase is only a part of the promised $100 billion investment. At the very least, it can be said that the calculation method for the final figure (which is $500 billion) is indeed intriguing. Four months ago, a well-known analyst publicly discussed the planning of the first phase at the data center base in Abilene, Texas, in the "Multi-Datacenter Training" report. The Real Data Behind "Stargate" They believe that this project is measured by Total Cost of Ownership (TCO), rather than capital expenditure. In other words, the $100 billion and $500 billion refer to TCO, which includes capital expenditures for data centers, servers, network facilities, power supply, and various operating costs including debt financing. Moreover, currently, only the initial $100 billion project has confirmed the data center site and power supply. The overall funding for the "Stargate" project has not yet been fully secured Another detail that can easily lead to misunderstanding is that the first computing power cluster of this large project is actually the collaboration project between Oracle and OpenAI announced last year. The project is entirely constructed by Lancium, Crusoe, and Oracle. According to analysts, SoftBank and other parties like MGX are not involved. 20 Supercomputers Across the U.S., 10 Already Under Construction In fact, OpenAI clearly stated in a blog post on Tuesday that "Stargate" is a newly established joint venture company. In this project, they plan to invest $500 billion in AI infrastructure over the next four years, with $100 billion to be put into use immediately. This is the first major AI plan announced by the second Trump administration, and due to significant relaxation of regulatory policies, more similar plans may be launched in the future. The equity founding parties of the Stargate plan include SoftBank, Oracle, MGX, and OpenAI, with the first three companies responsible for the initial funding. SoftBank and OpenAI are core partners, with SoftBank handling financial management while OpenAI leads operational work, including almost all computing power cluster management software. As mentioned earlier, the project plans to build a total of 20 data centers, with 10 already under construction in Abilene, Texas, each covering an area of approximately 46,500 square meters (500,000 square feet). Details of the Initial $100 Billion Expenditure The initial Oracle/Crusoe project, as one of the data centers, will interconnect with other super parks to achieve distributed training at a scale several orders of magnitude larger than any existing training operations. Previously, the "under construction" Stargate project park mentioned by Ultraman refers to the Oracle/Crusoe Abilene park. This park will accommodate 100,000 GB200 computing power clusters for OpenAI. This development project and the GB200 servers scheduled for delivery in the second and third quarters will be deployed within the Oracle/Crusoe data center. Below are high-definition photos of the park taken on December 29, 2024. Four modules of approximately 11,150 square meters make up a building of 44,600 square meters, with only two buildings currently under construction, and the key IT load capacity is approximately 180 megawatts To help everyone understand the scale, the entire park, consisting of 20 buildings, will contain 80 such "modules"! According to the "Accelerator Model and Data Center Model," analysts have planned a construction plan for OpenAI/Microsoft/Oracle, indicating that OpenAI can deploy the entire $100 billion total cost of ownership (TCO) without building a new park. Therefore, they believe that the initial $100 billion of the "Stargate" joint venture plan will be fully invested in this park. On Lancium's official website, this park is referred to as the "Abilene Clean Campus," claiming to achieve a 1GW power supply capacity by 2025. However, through multiple industry sources and documents, it has been revealed that the park has additionally obtained 1GW capacity, bringing the total power supply capacity to 2.2GW. Based on this, it can be estimated that the key IT load capacity of the park is 1.8GW, with a power usage effectiveness (PUE) of approximately 1.25. Although they may optimize to bring the PUE close to 2GW IT capacity, considering the power density improvements of GB300, especially VR200, this goal will be difficult to achieve. 700,000 GPUs, Jensen Huang's initial profit of $37.5 billion In addition to the 100,000 GB200 computing clusters currently being deployed, analysts estimate that they will also deploy another 200,000 GB300 and about 400,000 VR200 (the supercomputer that NVIDIA plans to launch in 2026, corresponding to Vera CPU and Rubin GPU) throughout the park. They calculated the implied number of Vera Rubin processors based on the park's $100 billion total cost of ownership (TCO) investment and the unit cost in the AI TCO model. The total IT power of approximately 700,000 GPUs is 1.8GW, which aligns with the 1,800 megawatt capacity of the Oracle/Crusoe Abilene park In short, OpenAI expects to pay approximately $2.8 per GPU per hour for the first GB200 cluster, with a total cost of ownership (TCO) for the full-scale cluster at $2.38 per hour. These figures will be adjusted accordingly with the future deployment of NVIDIA GPUs. OpenAI's spending will exceed $100 billion, but this is not an upfront capital investment. Analysts also estimate that NVIDIA's revenue from this will be relatively moderate, with IT equipment capital expenditures at approximately $37.5 billion and $49 billion, respectively. This will occupy the capacity of the entire 1.8GW Crusoe Lancium site. As it evolves to GB300 and VR200, the number of GPUs in each generation will double, and the power requirements for individual GPUs are also increasing. While 700,000 GPUs may seem small for a $100 billion investment, it is important to note that this equates to the computing power of millions of Hopper (NVIDIA's high-performance computing architecture) GPUs. SoftBank's Presence Oracle and SoftBank are well-known equity providers. MGX is a newly established technology investment platform in Abu Dhabi, co-founded by Mubadala and G42. The fund is chaired by Sheikh Tahnoon bin Zayed al-Nahyan, who reportedly oversees investment funds exceeding $15 trillion. Given Musk's claim that SoftBank raised less than $10 billion, it is noteworthy that this joint venture will also raise "debt capital" through project financing. The specific ratio of "equity financing" to "debt financing" has not yet been determined, but Oracle and MGX, with their strong balance sheets, can provide substantial financial support for the project. That is to say, even if SoftBank lacks funds, Oracle and MGX are available. It is reported that OpenAI and SoftBank each have committed to invest $19 billion on paper. Although SoftBank currently does not have that much liquidity, it is expected that it may raise the required equity investment by selling part of its ARM shares. OpenAI also lacks this funding, but they can finance after the launch of the new product o3, and they do not need to raise the full amount directly; they can achieve this by issuing "hybrid capital instruments" to other members of the consortium. It must be acknowledged that Musk's judgment is correct; they indeed have not raised this funding yet The key issue is whether the project financing risk can be separated from the parent organization. This risk isolation is crucial because OpenAI needs to raise funds to cover related expenses. For OpenAI, a favorable factor is that this funding does not need to be in place immediately. OpenAI currently has enough funds to cover the rent for 100,000 GB200 clusters this year, but they need to ensure they can afford the significantly increased rent in 2026 and beyond. While OpenAI may not be able to raise the required funds directly, senior analysts believe that their technology and product development will be able to support expenditures of this scale. Who are the biggest winners? We see many tech giants' names, such as NVIDIA, Arm, Microsoft, Oracle, and the previously mentioned investors, in the announcement of the "Stargate" joint venture project. These are impressive well-known companies, but the reality is that there are both losers and winners. If we were to say who the biggest loser is, it would probably be Microsoft. Microsoft, the biggest loser When OpenAI chose another infrastructure partner, Microsoft was effectively pushed to the sidelines. At the same time, Nadella's stance has been very clear: just hold on to the $80 billion investment! On Tuesday, Microsoft's blog stated the facts of the partnership and that Microsoft has a right of first refusal. But the reality is that this will have a gradually apparent negative impact on Microsoft's long-term development, as they are unwilling to take on the risk of this investment. In addition, it is even more shocking that Chris Young, who is responsible for OpenAI partnerships, suddenly left after the announcement was made. Optimistic analysts might say that Microsoft is shifting some of its capital-intensive AI operations to Oracle, but SemiAnalysis believes that this short-term gain cannot offset the long-term potential damage from losing OpenAI's revenue and control. While both Musk and Nadella raised the same question—"Where are the actual gains?"—Musk has devised a plan to ensure that their Memphis base in Tennessee grows at a rate no lower than OpenAI's; Nadella, on the other hand, has no plans for building cutting-edge AI models internally. Arm: Technology Partner or Fake News Arm's stock price rose by about 16% due to being listed as a technology partner, but this is merely because its Grace and Vera CPUs are used in conjunction with NVIDIA's Blackwell and Rubin GPUs. SoftBank may be pushing for Arm to join the PR camp, which superficially seems effective. However, in reality, Arm has not played a substantive role. For Arm shareholders, the reality is as described above; SoftBank may have to sell part of its stake in the company to provide some funding for the "Stargate" equity investment. We believe that most investors overlook this point, mistakenly viewing this announcement as a substantial positive news. Oracle is tasked with managing data centers and supply chain logistics; managing a 1.8GW cluster and procuring all necessary equipment as per another company's requirements is a highly challenging task. NVIDIA is clearly deeply involved, as their hardware is primarily relied upon. Meanwhile, OpenAI will maintain operational control and run almost all of its own cluster management software. All for One Training The scale of this project is impressive, but as analysts recently pointed out in Fabricated Knowledge, this seems to be another significant milestone in the AI development boom. We know that the first $100 billion will be invested in the Abilene base, while the remaining $400 billion will be used to develop more data center bases. Notably, all these data center bases will continue to be built with the goal of achieving distributed AI training. It is worth noting that simply replicating four bases equivalent to the scale of Abilene could generate 9GW of core IT computing power, all of which will be used for a single AI training According to the analysis by SemiAnalysis, with the continuous advancement of data center interconnect (DCI) construction, Marvell and Ciena will be the biggest winners, Fabrinet will gain considerable profits, while Coherent and Lumentum will see relatively smaller gains. New Intelligence, original title: "Ultraman showcases the first phase of the 'Interstellar Gate' factory, with 20 supercomputers being built across the U.S.! 700,000 GPUs make Huang explode with a profit of 37.5 billion." Risk Warning and Disclaimer The market has risks, and investment should be cautious. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk