The chip war is here! 1 million cards vs 50 million cards, Ultraman and Musk "gods fighting"

OpenAI will launch 1 million GPUs by the end of the year, which intuitively signals the start of a new round of chip wars.

On July 21, Ultraman announced on Twitter that OpenAI will launch over 1 million GPUs by the end of the year. He added, "Proud of the team, but now they better think about how to scale this number by 100 times."

Ultraman officially announces OpenAI's "million GPU" goal

This seemingly simple announcement actually hides multiple signals:

First, scale crushes competitors. OpenAI's goal is no longer 100,000 cards or 200,000 cards; if they are going to do it, they will do 1 million cards, which will reach 5 times the 200,000 GPUs used by Musk's xAI to train Grok 4;

Secondly, strategic autonomy. The leap in computing power infrastructure means OpenAI is gradually breaking free from its dependence on Microsoft Azure—previously, its computing power was highly bound to Microsoft, but now it is gaining initiative through self-built data centers (such as the Stargate project);

Finally, OpenAI's industry ambitions are also laid bare. "Scaling by 100 times" directly points to the ultimate computing power goal required for AGI, and an AI arms race based on computing power has entered a fever pitch.

There is also a small subplot—just after Ultraman set the goal of reaching 1 million cards, The Wall Street Journal came out to undermine it, stating that the Stargate project was not progressing smoothly and SoftBank was slow to provide funding. But OpenAI quickly "put out the fire": not only did they officially announce increased investment with Oracle, expanding the Stargate project to 4.5 gigawatts, but they also emphasized that the first phase of the project has partially gone into operation and that multi-party cooperation is progressing smoothly.

Subsequently, Musk even directly "launched a satellite," stating that xAI aims to deploy the equivalent of 50 million H100 GPUs within five years.

Musk proposes xAI's goal of 50 million cards in five years

Based on a rough estimate of an average of $40,000 per card, the scale of 1 million cards would mean that the GPU portion alone is worth as much as $40 billion. This level and speed of spending is unprecedented in the tech industry, basically approaching the annual capital expenditures of leading giants.

NVIDIA is naturally pleased with this, but the question is, what height will the demand for tens of thousands of computing cards push NVIDIA's market value to?

Let's turn our attention back to computing power; the most significant recent case where OpenAI was affected by computing power was the launch of the "Ghibli-style" image generation feature in March, where the official product was temporarily limited, including restrictions on the rate of image generation, limiting free users to 3 generations per day Ultraman is still shouting on Twitter, "Our GPUs are about to melt," which on the surface is product promotion, but can also be seen as giving hesitant investors a "wake-up call."

Clearly, the Stargate project is still in the fundraising stage. Although OpenAI, SoftBank, and Oracle can gather over $50 billion, there is still a half gap that needs to be filled through debt financing. To stimulate investors to open their wallets, some signals of the Stargate project's rationality must be released.

Ultraman "is straddling three boats"

OpenAI has long pursued computing power, with its sources including self-developed chips, the Stargate project, and Microsoft as the three main channels.

Regarding the self-developed chip, there were rumors that Ultraman was looking to raise $7 trillion to enter the chip manufacturing business. However, in February last year, Ultraman subtly denied this, stating in a fireside chat with former Intel CEO Pat Gelsinger, "We do believe that the world needs to invest heavily in AI computing (chips)."

According to industry tracking and research data, OpenAI's self-developed chips have been progressing in an orderly manner, with its first product expected to be launched as early as 2026.

Technology companies' self-developed chip roadmap, with red indicating released products, \* representing pending confirmation, source HSBC

HSBC's research team disclosed a progress report on technology companies' self-developed ASICs in late June, including Silicon Valley companies such as Google, Meta, Amazon, Microsoft, and xAI, all of which are developing AI chips in-house.

The research report shows that OpenAI's first 3-nanometer self-developed chip, designed by Broadcom, is codenamed Titan V.1 and will be released in 2026. A more advanced Titan V.2 chip is expected to be launched in 2028, but it is uncertain whether it will be based on 2nm technology or A16 (1.6nm) technology.

A long-time semiconductor industry analyst, Paul, disclosed detailed specifications and release dates for OpenAI's self-developed chips on Twitter (as shown in the image above), emphasizing that Titan V.1 will be launched in the third quarter of 2026, with core configurations including N3 technology, 144GB HBM3e memory, two computing chips, and CoWoS-S packaging. However, he believes that Titan V.2 will be launched in the third quarter of 2027, slightly ahead of HSBC's analyst team's prediction of 2028.

Self-development is a long-term plan. Before this path is fully realized, OpenAI's foot is starting to extend to "another boat" beyond Microsoft, leading the construction of computing power infrastructure. In January of this year, OpenAI, in collaboration with SoftBank and Oracle, launched the Stargate project, planning to invest $500 billion in the United States over four years to build computing infrastructure, with an initial investment of $100 billion, where SoftBank assumes financial responsibility, and OpenAI takes on operational responsibility.

The key point here is that operational control is what Ultraman wants to obtain in the Stargate project—how to allocate resources is entirely up to them, and whether to engage in price wars is also their decision.

Aerial view of the Stargate project site in Abilene, Texas, USA, source: OpenAI

Four months later, OpenAI assembled a "UAE version" of the Stargate project, planning to collaborate with partners such as G42, Oracle, NVIDIA, and SoftBank to build a 1-gigawatt data center locally, expected to be operational by 2026.

Before these grand infrastructure projects come to fruition, OpenAI's computing power supply still relies on Microsoft—the two companies have been collaborating since 2019, with Microsoft providing over $13 billion in direct investment while becoming OpenAI's exclusive computing power provider. Microsoft gains priority in areas such as OpenAI's models and revenue sharing, for example, obtaining a 49% profit-sharing right from OpenAI, potentially reaching $120 billion.

Relying on Microsoft also means avoiding being "choked" by Microsoft. From the $7 trillion chip-making rumors to the $500 billion Stargate project, and then to the UAE version of Stargate, OpenAI's core logic is to build a grand narrative of computing power led by itself, continuously stacking up.

Without scaled computing power, it could be crushed at any moment by Google's price wars, and scaling is Google's inherent advantage. On the product level, lacking computing power is like a "good cook unable to make a meal without rice," leading to more issues like limited "Ghibli-style" image generation capabilities. This is why there were previous rumors that Ilya left in anger due to computing power demands, and that GPT-5, DALL-E, etc., were forced to delay their releases due to computing power shortages.

Coincidentally, while OpenAI was "stepping on the gas," Microsoft lightly tapped the brakes.

In April of this year, TD Cowen analysts stated that Microsoft abandoned its 2-gigawatt new data center projects in the U.S. and Europe. Microsoft's official response indicated that data center capacities were planned years ago, and that their layouts are now complete, leading to some strategic adjustments for flexibility.

Microsoft's strategic contraction can actually be traced back to the end of last year when Nadella, in an interview on the BG2 podcast, openly emphasized the differences with Ultraman, stating, "We need to think rigorously about how to effectively utilize existing equipment. We also need to consider the lifespan of the equipment and cannot simply purchase new devices. Unless the performance and cost of GPUs can bring significant improvements, allowing profit margins to reach or exceed those of large cloud service providers, we will not act lightly." Everyone is chasing the sense of security in computing power. Nadella believes that the existing computing power is sufficient and needs refined operations, while Altman is concerned that insufficient computing power will become a bottleneck for new models and products.

As a result, the two sides are drifting further apart.

In January this year, Microsoft chose to let go and revised its cooperation terms with OpenAI, allowing it to use computing resources from third-party suppliers. Soon, cloud providers like Oracle and CoreWeave signed leasing agreements with OpenAI one after another. Of course, to maintain decorum, Microsoft still retains the priority cooperation rights for providing computing power.

The Information cited news from an investor meeting stating that OpenAI plans to shift 75% of its computing power sources to the Stargate project by 2030.

The Computing Power War "Burns" 25 Trillion

OpenAI is pursuing computing power, aiming for "self-controlled computing power" internally, while externally responding to the "computing power war" among Silicon Valley giants.

On July 16, The Information published an exclusive interview with Meta CEO Mark Zuckerberg, who stated that Meta is building multiple data center clusters.

"Our team is working day and night on the Prometheus and Hyperion projects, which are our first two Titan clusters, both exceeding 1 gigawatt. Hyperion will expand to 5 gigawatts in the coming years. I've shared pictures of it; in terms of land area, this data center occupies a significant portion of Manhattan. It's enormous," Zuckerberg said.

Meta's Hyperion data center project in Manhattan, Source: Zuckerberg

What does a 1 gigawatt data center mean?

Assuming that the 1 gigawatt Hyperion data center being built by Meta is fully equipped with GB200 NVL72 racks, with a power consumption of 140KW per rack, it can accommodate over 7,100 racks in total. Since each rack is equipped with 72 GPUs, that totals approximately 510,000 GPUs. Calculating at 3 million dollars per rack, the total cost for over 7,100 racks exceeds 21 billion dollars.

If OpenAI and Oracle's newly expanded 4.5 gigawatt project materializes, then in the future, OpenAI could potentially control nearly 2.5 million GPUs through the Stargate project.

To the ultra-large-scale training clusters by 2026, source: SemiAnalysis

On July 21, the well-known research institution SemiAnalysis disclosed the training cluster data of Anthropic, OpenAI, and Meta by the end of 2026 based on its data center and accelerator models. SemiAnalysis listed another 1-gigawatt capacity Prometheus data center project from Meta, which uses a mix of GB200/300, with a total of 500,000 GPUs, consistent with our estimates for the Hyperion data center.

In terms of energy consumption, a 1-gigawatt GB200 NVL72 data center running at full capacity 24 hours a day for 365 days a year is expected to require 8.76 billion kilowatt-hours of electricity. In comparison, Tokyo, Japan's total electricity consumption for the year 2023 is only about 130 billion kilowatt-hours.

xAI is not included in SemiAnalysis's tracking data, but as the number one competitor to OpenAI, xAI is also "madly" investing in infrastructure.

On July 10, xAI announced its Grok 4 model, and Musk revealed during a live broadcast that this model is on a supercomputer cluster with over 200,000 H100 GPUs. The emphasis is not just on this 200,000-card cluster but also on the speed of xAI's data cluster construction—only 9 months have passed since the completion of the previous "100,000-card" cluster.

Even more exaggerated, xAI's first 100,000-card level Colossus AI supercomputer cluster took only 122 days from construction to operation, showcasing an astonishing construction efficiency.

Regarding why to push so hard on infrastructure construction, Musk revealed his logic during the live broadcast, emphasizing that relying on cloud providers' computing power to coordinate a 100,000-card cluster would take an estimated 18 to 24 months. "We thought, 18 to 24 months means failure is inevitable," Musk said.

Building in 122 days, while coordinating cloud providers' computing power would take at least 18 months, also partly explains why OpenAI does not intend to collaborate with Microsoft anymore—relying on external partners for computing power coordination is too inefficient, and leasing computing power can only serve as a short-term transitional solution; only self-leadership can ensure control.

One can imagine a scene: when OpenAI launches Ghibli-style image generation, Ultraman says, "Our GPUs are about to melt," and turns to Microsoft for computing power support only to hit a wall—receiving the reply, "Just wait a bit longer." At this moment, Ultraman probably can only let out a helpless sigh.

From 2023 to 2025, the trend of capital expenditure in AI infrastructure construction Source: The Business Engineer

Returning to the 200,000-card cluster of xAI, with the price of H100 single cards at $25,000-$30,000, the cost can be roughly estimated, and the total cost of the GPU part will require $5 billion-$6 billion, not including the costs of infrastructure, operation, and maintenance.

The investments of OpenAI, xAI, and Meta in data centers are a microcosm of the industry's AI capital expenditure expansion.

The Business Engineer analyst Gennaro Cuofano released a research report in May this year, citing the performance of Silicon Valley companies and industry forecast data, outlining the capital expenditures of major Silicon Valley companies on AI for the years 2023, 2024, and 2025, with corresponding values of $170 billion, $256 billion, and $360 billion.

For the whole year, $360 billion converts to over 2.5 trillion yuan, representing a growth of over 110% compared to 2023. More importantly, the AI expenditures of large companies account for over 85% of the entire industry, which also means that the "Matthew effect" of AI infrastructure construction is continuously strengthening—future leading cloud providers will hold the core resources of the industry.

The giants are all involved in this $2.5 trillion computing power war, and there is another noteworthy background—the signing and passing of the OBBB (Big Beautiful Act).

According to the act, large data center infrastructure construction and R&D by tech giants can receive tax credits. For example, in the case of full depreciation of equipment, if a company purchases $100 million worth of servers and other data center hardware, under traditional depreciation rules, it needs to be spread over 5 years, allowing only $20 million to be deducted each year. According to the act, companies can deduct $100 million of taxable income in the year of purchase.

With business demand, competitors are all in the game, and policies have indirectly acted as a catalyst, all stimulating the likes of Altman, Zuckerberg, and Musk to eagerly engage in another Silicon Valley chip war.

If one must ask a question, with millions of GPUs, can humanity open the door to the AGI era?

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this is at one's own risk