SpaceX's compute pays for itself within a year; Meta's Muse is hard to launch globally due to compute costs

This week, SpaceX, Tesla, and Meta successively released their third-quarter financial reports.

On October 20, SpaceX disclosed on its third-quarter earnings call that, at current prices, newly deployed compute capacity can pay for itself in less than a year. A day later, Zuckerberg told investors that Meta's personal agent Muse cannot yet be launched globally, one reason being that running it is too compute-intensive.

The same compute carries two different prices at the two companies. SpaceX can sell it to external customers, while Meta must first bear the cost and wait for free users to generate revenue. Tesla's financial report provides a third growth sample: vehicle deliveries fell year over year, yet paying FSD users are increasing.

Musk's SpaceX has given big companies and the capital markets a more intuitive pricing anchor. As long as compute remains scarce and can be allocated across different businesses, the revenue from renting it out externally will become the benchmark for internal projects. Free products need budgets; proving they can acquire users is no longer enough—they must also prove that these users are worth the company giving up the rent it could collect right now.

Once compute has a price, winning or losing on models no longer determines who makes money

SpaceX's third-quarter Superintelligence business SI revenue was $7.56 billion, accounting for about 52% of total revenue. The company did not attribute all the growth to Grok, but explicitly mentioned that the start of execution of new cloud service agreements was one of the main drivers.

In May this year, Anthropic announced it would rent all the compute capacity of SpaceX's Colossus 1 data center, with more than 300 megawatts of new capacity, to raise usage quotas for Claude paying users. Another agreement disclosed in June stipulated that Google would pay $920 million per month starting in October to purchase computing power.

Musk's investment model therefore has a path to returns that does not require waiting for Grok to win. The stronger Claude's user demand, the more likely it is to bring compute revenue to SpaceX. Competition among models can no longer directly determine which of the two companies makes money and which bleeds.

This also creates new pressure on cloud providers. Google has its own cloud business yet still buys computing resources from SpaceX. At least in a phase of tight supply, whoever can deliver usable compute on time has a chance to take the revenue. Long-term customer relationships and advantages in software services cannot solve the immediate capacity gap.

SpaceX's management also made clear that it will compare the returns of training its own models, inference for its own products, and external cloud services, and ultimately decide who gets the compute. This choice matters: SpaceX can continue betting on its own models while also earning returns from the growth of other models.

However, this price has a shelf life. The agreement between Google and SpaceX allows either party to terminate with 90 days' notice after December 31 this year, and "paying back in less than a year" is also explicitly based on current prices. But as long as compute remains tight, this price will stay on the table at every big company's internal meetings.

Free agents must find revenue thicker than advertising

Roughly converting SpaceX's payback period, it is equivalent to recovering more than double the investment each year. This is the benchmark set by external customers, against which a free product like Muse must be measured.

Meta's situation is therefore somewhat awkward.

Muse was first launched in the United States and Canada, and management acknowledged that compute limits the pace of expansion. The company hopes to improve computing cost efficiency more than tenfold over the next year, yet the finance chief simultaneously expects that total serving costs will still rise as users increase. Unit costs falling while total spending rises means cheaper inference will bring more and more complex tasks, eating up the savings again.

Meta's revenue structure is also not easy. Meta's third-quarter revenue was $64.6 billion, up 26% year over year, of which ad impressions grew about 10% and average prices rose about 14%; multiplied together, the two roughly match the revenue growth rate. Growth still comes from advertising, and AI has not yet visibly appeared on the revenue line. At the same time, the company raised its capital expenditure forecast for this year to $135 billion to $145 billion.

The way out Zuckerberg offered is charging for transactions. Muse is free for most people, with merchants paying for transactions facilitated by Muse, and higher-tier models, extra usage, and enterprise features available by separate subscription. He revealed that shopping is the most common scenario for Muse, with more than a quarter of users participating, and an average purchase amount of more than $200.

Whether merchants are willing to pay is the biggest variable in this model. Recommended ads help merchants win attention, while agents promise to compare prices for consumers and help them take fewer detours. The more effective it is, the harder it becomes to charge merchants an extra layer simply for "bringing users."

If an order would have been completed through Meta ads anyway, and is instead completed by Muse with an added service fee, merchants will most likely deduct it from their advertising budget. What Meta needs to prove is that Muse creates incremental transactions, not just renames advertising revenue.

An agent with only traffic and no payment and fulfillment data cannot stand firm here.

Therefore, big companies with massive user bases cannot simply move users of old products unchanged into high-cost agents. Which regions to cover first, which tasks get stronger models, and how much usage to reserve for paying customers will directly affect the pace of expansion. Traffic advantages are still useful, but the return on serving that traffic must be recalculated.

The same logic applies to agent products such as Qwen GO and Doubao Shopping.

Tesla's comparison is more direct. Third-quarter vehicle deliveries were 486,500, down 2% year over year; the active FSD subscription metric cited in the earnings report reached 1.74 million, up 67% year over year. This metric includes one-time purchases and monthly subscriptions, excludes free trials, and FSD still requires the driver to actively supervise.

Fewer cars sold, yet more software sold—the two growth curves have already separated. The reason is not big words like "the car is the AI entry point," but that FSD corresponds to the clear task of driving; users can feel the effect after driving a few times themselves, and whether to pay is also up to the user.

From this, Muse's real situation can also be seen. Its most frequently used scenario is shopping, and shopping happens to be a concrete task with an amount and an outcome, closer to a form that can charge than general chat. But FSD users' willingness to pay for assisted driving does not mean they are willing to pay for a general assistant in the car, and it certainly does not imply Robotaxi profits. In the third quarter, Tesla Robotaxi's cumulative unsupervised driving mileage approached 2 million miles, about 8 times the level at the end of the second quarter. Mileage growth shows the service is expanding, but whether the fleet makes money still depends on factors such as fleet operating efficiency, vehicle utilization, and cost per mile.

Another move by Tesla this quarter is equally key. It lowered its full-year capital expenditure forecast from more than $25 billion to about $20 billion, and free cash flow turned positive at $1.531 billion for the quarter. Better cash performance comes partly from new businesses and partly from a changed pace of spending. Therefore, the maturity of the software business should not be overestimated either.

Only features that can already charge are qualified to get long-term budgets. Projects further from charging depend on the company being willing to wait longer. Alongside its earnings release, Meta announced the creation of non-voting Class C shares, adding tools for future financing, employee compensation, and potential acquisitions. This allocation itself brings no new cash, but it gives Muse a longer waiting period.

The compute that cloud vendors sell to model companies is a cost for the model companies. The commissions agents charge merchants may come from merchants' advertising budgets that have been squeezed out. Multiple companies reporting growth at the same time does not mean end customers have paid out the same amount more.

Big companies' own accounting standards are also affecting the readings.

On its July conference call, Microsoft announced that starting in fiscal year 2027, the depreciation period for data centers and office buildings would be extended from 15 years to 25 years, and more data center leases would shift from finance leases to operating leases. Finance leases count toward capital expenditure, while operating leases do not; after the standard changed, Microsoft's calendar-year 2026 capital expenditure forecast displayed as about $175 billion instead of about $190 billion. Management stressed that the investment plan itself had not changed. Based on a rough calculation of its disclosed building costs, the depreciation period alone could reduce annual depreciation charged to the income statement by several billion dollars.

The money was not spent less; it just looks less painful.

Cash flow is another side. According to calculations from Google's second-quarter earnings report, its operating cash flow minus capital expenditure was negative, the first time since going public. Meta also faces the coexistence of raised capital expenditure and cash pressure in the third quarter. While raising compute investment on one hand and seeking returns for free products on the other, big companies' cash flow bears pressure from both ends.

Meta is still choosing to increase its options. On September 28, it announced the establishment of Meta Enterprise Platform, pushing Muse, enterprise agents, model APIs, and coding tools to enterprises, with head CJ Desai reporting directly to Zuckerberg. This brings paid demand beyond consumers into Meta's resource allocation.

It is also the consensus among big companies: reduce revenue dependence on the success of any single proprietary app and direct resources toward the B2B side, which is closer to monetization. For example, domestic AI office competition became white-hot this quarter.

So the earnings reports of other major tech companies released next can likewise be stripped of inflated numbers using these questions. How much of AI revenue comes from external customers, and how much is internal transfer? Has new AI revenue cannibalized existing business revenue? After adjustments to capital expenditure standards, how much did real compute investment grow?

SpaceX has already demonstrated the return rate of selling compute to external customers, Tesla has demonstrated paid demand for a specific task, and Meta is trying to turn shopping into a third kind of revenue. AI products still pulling in free users still owe the earnings report one key number: after these users are done using them, how much incremental profit have they actually left the company.

Note: The data in this article comes from public sources and does not constitute investment advice.

This article comes from the WeChat public account "Emphasis Next" (ID: leo89203898), author: Hua Wen, editor: Xiaobai, authorized for publication by 36Kr.