Why are institutions going all-in on SpaceX?

👆Click the blue text to follow us

Conspiracy Theory

Jason heard a conspiracy theory today about why some institutions have recently started aggressively bullish on SpaceX. It may be because, by a certain week before September 30, the company must complete the acquisition of Cursor at the average price of that week. Therefore, the higher the stock price, the lower the cost for SpaceX, the less dilution for class A shareholders, and the more beneficial it is. While Musk is fighting off short sellers, he can also reduce costs on the Cursor acquisition, and in the future, can subsequently move on to acquiring Tesla, achieving three goals at once. What do you think? Let's hear the latest research from SemiAnalysis.

SemiAnalysis has evaluated the possible sites SpaceX may use one by one and has tracked over 30 gas turbine, engine, and fuel cell suppliers on a quarterly basis. The conclusion is that, although land and electricity are currently tight, as long as there is massive adoption of on-site gas power generation and higher equipment procurement prices are acceptable, there is indeed a path to achieving 10 GW. If half of this is used for commercial inference, the company's ARR could reach $300 billion by the end of 2027.

The team believes that Musk's core strategy is speed first. Large transformers have more than a two-year wait, mainstream gas turbines more than five years for delivery, so the company uses Chinese-made power modules and procures used equipment through the secondary market, while keeping construction phases as parallel and pre-assembled as possible, thereby compressing commissioning cycles. If everything goes smoothly, adding ten gigawatts by 2027 is not impossible.

In terms of revenue and price, the team said that two major OA model companies may generate over $10 billion in annual revenue per gigawatt, with gross profit margins typically above 60%. This also means that, as long as cluster utilization rates and Token prices can be maintained, procuring one gigawatt of computing power per year for $4 billion still leaves profit space for AI model companies.

The team added that Microsoft currently faces a demand gap. Microsoft had slowed datacenter leasing by the end of 2024, but by 2025 and 2026, expansion has already started.

It is estimated that Microsoft has already signed over 10 GW of binding commitments this year through leasing, self-building, power purchase agreements, and energy service agreements, with total contract value exceeding $300 billion. However, most of these capacities will not be online until the end of 2027 or 2028. So, until then, Microsoft's computing power gap still exists.

This gap leaves a window for SpaceX. SemiAnalysis speculates that Microsoft may purchase 3 GW of computing power from SpaceX, priced at $5 billion per GW per year, making annualized contract value as high as $150 billion. But, just like the Anthropic-Google contract, there’s a 90-day termination clause, which reduces Microsoft's long-term commitment risk and makes internal approvals easier.

So, whether to meet Microsoft’s needs or their own, if SpaceX wants to add 10 GW of capacity, where does the funding come from? SemiAnalysis suggests two possible paths.

The first is for Nvidia to provide vendor financing to reduce SpaceX’s upfront cash outflow. In the latest financial report, Musk already announced that their future computing infrastructure will exclusively use Nvidia chips, and financing conditions may be one consideration.

The second is relying on high prices brought by quick delivery, selling computing power at $30 to $50 million per megawatt per year, allowing capital recovery periods to be pushed below one year, and then using operational cash flows to support the next round of construction.

In the end, SemiAnalysis assumes SpaceX will add 10 GW of computing power by 2027, with 5 GW for commercial inference, calculated at $5 billion per GW per year, leading to an ARR of $300 billion by the end of 2027. If this model materializes, SpaceX will quickly transition from a space and satellite internet company to a giant-scale AI computing provider.

Jason believes we can say SpaceX is extremely capable in launching projects, also very strong in supplying electricity in various ways, Musk and Jensen Huang have a great relationship, SpaceX could be allocated more Nvidia GPUs and get guarantees and leasing support from Nvidia, but I think this article has some bugs that must be pointed out, and the returns are not calculated this way.

Let's set aside the GW scale increments, and first look at the price per unit and estimates of return on investment.

The article says SpaceX can price at $30 to $50 million per MW; using the $50 million upper limit and baseline cost of $5 billion per GW, it's indeed possible to break even in a year.

But here's the problem—is the $30 to $50 million price per unit reasonable? Is it sustainable?

To judge this, let's look back at the Anthropic-Google and SpaceX contracts: Anthropic pays $26.8 million per MW, Google pays $50 million. I think SemiAnalysis rounded these two contracts to the $30-50 million range, and extrapolated linearly to Microsoft or other customers at that price.

Keep in mind, these high prices were an emergency response to the extremely tight demand at that time for Google and Anthropic. Even without counting the 90-day cancel clause or the time lag between signing and going online, one should at least apply a 20% discount for caution—so, $24-40 million per MW, with a revenue recovery period close to two years.

If you factor in depreciation, electricity, maintenance, and financing costs, the comprehensive recovery period will be about three years, not the one year SemiAnalysis claims. And this doesn’t take into account the research team's point that the company needs to pay for more expensive gas turbines to speed up construction.

Next is Microsoft. The article first says Microsoft has $300 billion worth of computing contracts for 10 GW capacity, which means a unit cost of about $30 billion. But for SpaceX, all computing power is calculated at the scarce peak rate of $50 billion return, and a one-year investment payback is concluded. In my opinion, this is an inconsistent comparison.

Even if the final scale is $50 billion a GW, Microsoft, like Google and Anthropic, will most likely include a 90-day early termination clause. This means that contracts can only be taken one step at a time—you shouldn’t project overall asset returns linearly based on current ARR and peak prices, and there should be a heavy discount on the corresponding asset valuations.

So, SemiAnalysis bases its one-year capital recovery argument on a 10 GW increment, peak prices, and an ARR lens. All I can say is, my friend, Musk might not be happy with this portrayal.

Alright, next up are Bloomberg’s Top 5 Tech News:

Fourth: Meta has released an open-source AI model, Muse Glimmer, with 3 billion parameters, capable of running on a personal computer and is focused on scheduling and other agent tasks. Meanwhile, Zuckerberg published a lengthy article opposing centralization of AI power, insisting on an open-source path, claiming AI will drive new jobs, establishing a $1 billion fund to give back to data center communities, and emphasizing continued curbs on chip exports to China. (What a show)

Third: Intel announced a $15 billion new stock issuance, the first issuance since it went public in 1971, to raise funds for expanding AI and custom chip businesses. The stock price fell 4.1% that day, but is still up over 160% year to date.

Second: Microsoft plans to massively ramp up next-generation AI chip production, already negotiating with TSMC to deliver over 300,000 chips by 2027. The new Maia 300 chip will launch this fall, aiming to reduce reliance on Nvidia.

First: OpenAI completed a tender offer, helping employees sell about $7 billion in equity. The funds came from within the company. The valuation remains at $852 billion, seen as a routine operation before an IPO, highlighting fierce competition with Anthropic.