SpaceX is poised to officially commence its share offering on June 11 and go public on June 12, aiming to raise a minimum of $75 billion with a valuation of $1.77 trillion. A plethora of information regarding the specifics of this, the largest IPO in human history, including the prospectus and roadshow PPT, is readily available online, so I won't delve into the details here. SpaceX categorizes its business into three primary segments:

1. Space: This segment primarily centers on space launch services, encompassing the reusable spacecraft Starship, which is slated to commence commercial operations in the latter half of this year (a move that was previously deemed 'unrealistic' by numerous experts).

2. Connectivity: Mainly represented by Starlink's satellite communication business, which was once cited by communications industry expert Lv Tingjie (a frequent speaker at brokerage strategy meetings) as evidence that 'Musk doesn't understand communications'.

3. AI: This encompasses the former xAI business, including data centers, Grok, and X (formerly known as Twitter), as well as a chip factory co-built with Tesla, among other ventures.

From a capital operations perspective, Musk's strategy of initially merging X Platforms into xAI and subsequently integrating xAI into SpaceX is nothing short of brilliant. The first merger enabled the user bases of X and Grok to be combined, leading to the claim that 'xAI has 550 million monthly active users'. The second merger not only facilitated xAI's early entry into the capital markets but also provided SpaceX with a more compelling narrative. In the roadshow PPT, management emphasized that the market size for the space and connectivity businesses is $2 trillion, while the potential market size for the AI business is a staggering $26.5 trillion, accounting for 93% of the total potential market (TAM). Whether Wall Street believes this remains to be seen, but Musk's ability to articulate this story coherently and confidently underscores his exceptional storytelling skills and psychological resilience.

PS: The well-worn cliché that 'Musk has a high IQ but low EQ' comes to mind. I believe the inventors and propagators of this cliché are destined to be nailed to the pillar of historical shame, but this might be their only chance to be remembered in history—after all, infamy is still fame.

Now, let's get down to business. I currently subscribe to three foundational models: GPT, Claude, and Grok. I also subscribed to Gemini for a while but have since discontinued. There's not much to say about GPT—I've been a user since 2023 and have an uninterrupted payment history of 25 months, so I'm accustomed to discussing everything with it. Claude is also self-explanatory; in the wave of Agentic Workflow, not using it would be abnormal (though I rarely use Agent in my daily work). That leaves Grok: its fees are significantly higher, with a basic membership costing $30 per month (compared to $20 for the other two), and its model capabilities are noticeably weaker. If I were to cut one more paid model, it would undoubtedly be Grok, not the other two.

In the United States and even globally, Grok is a foundational model with a relatively awkward positioning:

1. On the consumer side (C-end): GPT and Gemini boast a much stronger user base, with the former having a first-mover advantage and the latter supported by Google's ecosystem. Grok's only strength lies in its integration with X (Twitter), but how can the latter compare to Google?

2. On the business side (B-end): Claude and GPT can outperform Grok by a wide margin, with capabilities that are not even in the same league. If we talk about cost-effectiveness, Grok is clearly no match for several Chinese open-source models, including but not limited to DeepSeek.

3. In terms of multimodality: Grok is obviously no match for Google's series of models, which have reached a high level of sophistication, and it doesn't even rank in the video space. X's data is primarily text-based, which probably won't be of much help in training multimodal models.

For a while, Grok pursued an open-source route and even briefly took over the torch of 'North America's strongest open-source model' from LLaMA. However, since Grok 2.5, the open-source route has been abandoned, while Google has returned to the open-source space with its Gemma series. Unless Musk has a change of heart, it's unlikely that Grok will become the flagbearer of open-source models.

So, where does Grok's advantage lie? Why do people still pay for it? There's an open secret: writers in the fan community and those who create borderline content (and their readers) particularly love Grok because its content moderation is very lax, almost non-existent. Claude is the most fearful of users generating 'unhealthy content', constantly worried about you crossing any lines; GPT has further tightened its restrictions since entering the 5.0 era. Gemini is much more lenient but still not entirely so, refusing to budge on principled issues. Only Grok is literally 'without any scruples'—if you dare to ask, it dares to answer.

Additionally, Gemini, which is also relatively lenient, has a major issue: its contextual hallucinations are too severe, making long-term conversations less valuable. Anyone who has used Gemini to analyze long-text content knows that it's best to feed it the entire text at once rather than piece by piece; otherwise, it will generate many bizarre cross-references, making it highly valuable for brainstorming but worthless for professional purposes. This once again highlights Grok's significance—at least it's usable in the long-text domain. Grok is also unique in its 'anti-political correctness': each version of GPT becomes more left-leaning, and Claude was already left-leaning to begin with; compared to them, Grok can be considered the only right-leaning frontier large model. Especially on topics involving the conservative-liberal binary opposition, the responses from other large models are basically unreadable, while Grok at least maintains a nominal neutrality. Whether this reflects Musk's own ideology, I have no idea.

I'm also somewhat surprised by Grok's rich knowledge in areas like anime and gaming. Whenever I play a game or watch an anime, I definitely discuss it with Grok first, but not for other topics. I think Grok seems to have formed a deep cultural imprint of anime, not just something created by stacking corpora. As we all know, Musk is an anime enthusiast and a gamer—whether Grok's characteristics are related to this, I have no idea.

Besides that, there's another technical characteristic: Grok's search capabilities seem particularly strong, especially its ability to directly read large amounts of web information, even some behind 'paywalls'. Try asking GPT, Claude, Gemini, and Grok simultaneously about the latest report from a European and American paid media outlet like WSJ—Grok seems to be the only large model capable of breaking through the 'paywall' and directly quoting the original text. I believe many users use this method to 'freeload' a large amount of information that would otherwise require payment, and perhaps this is one of Grok's core competitive advantages?

However, this doesn't mean Grok has mastered some 'black technology' that others don't have. I think the most reasonable explanation is that Grok's goals are not ambitious enough, so fewer people are paying attention to these 'gray area' practices. If the same thing happened to GPT or Claude, it would probably make global headlines tomorrow and lead to lawsuits demanding billions in damages. If Grok were to develop hundreds of millions of paying users, these 'abilities' would certainly be changed or at least significantly restricted.

In fact, I think Grok's other 'advantages' can probably be explained in the same way—GPT and Gemini have become stricter in content moderation because their targets are too big, making it necessary to be highly vigilant. (Note: Claude is an exception—its content moderation has been strict from the beginning, purely because its management is perverse and takes pleasure in torturing users, commonly known as 'arrogance due to talent', a term used to describe Yang Xiu in the famous Chinese classical novel Romance of the Three Kingdoms.) In other words, all of Grok's advantages are based on the fact that it's not big enough yet... although given its current trajectory, its chances of growing larger are not high.

During the IPO roadshow, SpaceX explicitly stated its plan to 'begin deploying space-based computing power in 2028' and believes that 'orbital AI computing power is technically ready, cheaper, and faster' (Readily available, cheaper, and faster). In the long term, SpaceX positions itself as a 'space technology infrastructure supplier', a role that doesn't necessarily require developing its own large models. As we all know, the currently most advanced AI cloud service providers globally are Microsoft and Amazon, both of which have relatively weak self-developed large models; Oracle, which has been rising in prominence over the past two years, has barely wasted resources on large models.

In the near term, xAI (which no longer exists and has been fully integrated into SpaceX's departments) also shows clear signs of shifting toward a computing power supplier model: it has signed computing power leasing contracts with Anthropic and Google worth $1.25 billion and $920 million per month, respectively. Musk once proposed a formula for making money: 'Buy a bunch of GPUs and then make money'—without needing to go through the step of 'via your own large model/application layer'. After all, Anthropic and Google have already established working AI business models, so why not make money from them?

There's only one uncertainty here: SpaceX holds an option to acquire Cursor for $60 billion, and Cursor is the most popular AI programming tool. Coding & Agent is currently the AI space with the highest token consumption and the most certain business model, and Musk clearly wants to retain a ticket to this space, and the best ticket available at this stage, since Claude Code and GPT Codex are not for sale.

But again, this doesn't require maintaining a strong self-developed large model. Cursor is inherently 'model-neutral'—many people use it with Claude, and some with GPT. The Cursor team's model development capabilities are negligible; they recently came up with a 'self-developed model' that later turned out to be a modified version of Kimi. This also illustrates that having a first-party foundational model is not important for a tool like Cursor. We can even consider this a symbol of SpaceX's AI business fully shifting toward 'infrastructure': computing power is provided as an infrastructure to all users, and tools are the same, as long as the price is right.

Regardless, I hope Grok can continue to operate—it doesn't need to become one of the world's most powerful models, but it must exist. I thoroughly enjoy discussing anime works like Zombie Land Saga, Dandadan, and The Girl I Like Forgot Her Glasses with Grok—its insights have been incredibly beneficial to me. It would be even better if the monthly price could be halved. I hope Musk never considers shutting down Grok!