To The Moon

How to Rebuild SpaceX’s Valuation Framework?

An Earnings Report That Puts Launch Services and Starlink in the Background:

For years, investors have associated Spacex with two businesses: rocket launches and Starlink. The former offers formidable technical barriers and Elon Musk’s personal appeal; the latter offers a growing subscriber base and cash flow. Many investors still place AI at the bottom of their valuation models, treating it as an option that might pay off someday.

But the third-quarter results suggest that this view is falling behind reality. SpaceX reported quarterly revenue of $14.58 billion, up 161% year over year and roughly 15% above the consensus estimate of about $12.7 billion. The AI segment contributed $7.56 billion, exceeding the combined revenue of the Connectivity and Space segments. While the market is still debating how much to pay for rockets and satellites, AI already accounts for more than half of the company’s revenue.

Several other figures deserve to be considered together. The Connectivity segment, which includes Starlink, generated $5.25 billion in revenue and reached 13.6 million subscribers. Space segment revenue was $1.77 billion, compared with $1.67 billion a year earlier. Net income reached $4.4 billion, versus a $1.05 billion loss in the same quarter last year, while adjusted EBITDA (a non-GAAP measure) rose 306% to $8.8 billion. The main source of incremental growth is already clear.

This transformation came little more than six months after SpaceX acquired xAI in February 2026. In May, the company signed a compute contract with Anthropic, providing access to compute capacity backed by approximately 325,000 Nvidia GPUs for $1.25 billion a month. In June, Google agreed to rent compute capacity backed by roughly 110,000 GPUs for $920 million a month under a contract running through June 2029. Another contract disclosed in September is set to generate $1.11 billion a month starting in December. Compute rental has become a business with revenue that can be calculated month by month.

For comparison, SpaceX generated $18.7 billion in revenue for the whole of 2025. Starlink contributed $11.4 billion, while the AI segment generated just $3.2 billion and lost approximately $6.4 billion. The revenue mix has changed dramatically in less than a year. The AI segment also includes Grok, X, and coding tools, however, so the entire $7.56 billion cannot be counted as GPU rental revenue. What needs reassessing is the importance of compute rental within the segment and how long the underlying demand can last.

Core Argument: Cloud Compute Demand Is More Resilient Than the Market Assumes:

Some investors may ask whether these large contracts are simply a temporary response to a compute shortage. Once models become more efficient, the capabilities of open-weight models—models whose weights are available for download—catch up, and companies move models back onto their own servers, should cloud demand not cool down?

This argument underestimates the practical barriers ordinary businesses face when adopting AI. Being able to download a model is very different from being able to run it reliably and economically over the long term. Companies ultimately compare how much they must spend, how many problems a system can solve, and who takes responsibility when something goes wrong. Once all of these factors enter the calculation, frontier models delivered through the cloud often look more attractive than expected.

Dimension One: High Hardware Costs Keep Raising the Barrier to Reliable On-Premises Model Deployment:

Start with hardware costs, which are easy to underestimate. Many demonstrations of on-premises solutions make the process look simple: download an open-weight model, add a server, connect a few business systems, and the company has its own AI. Once procurement begins, however, GPUs, memory, servers, and ongoing operations all cost money.

Recent price movements have hardly favored the do-it-yourself approach. Prices of hardware such as DRAM and NAND have surged. The shortage of graphics cards on the market is becoming increasingly severe. Even AWS raised its H200 Capacity Block prices by approximately 15% in January. Businesses cannot build their budgets around the assumption that hardware will soon become cheap.

Take a 70-billion-parameter model. Even with 4-bit quantization, its weights alone require around 40 GB of GPU memory, before accounting for the key-value (KV) cache and other overhead. Higher throughput, longer context windows, or serving multiple employees at once require more hardware. What initially looks like a software purchase can easily turn into a server investment.

Many of the servers already available to corporate IT departments are simply virtual machines without GPUs. CPUs can certainly run some smaller mixture-of-experts (MoE) models with fewer active parameters, but speed, concurrency, and output quality still need to be tested. Generating a paragraph in a demonstration does not mean a system can support employees’ daily work, much less replace frontier models in complex analysis.

Cloud providers can negotiate long-term GPU contracts, use their own chips, and pool demand from different customers to improve utilization. A company buying GPUs for itself may find that it lacks capacity during busy periods but cannot easily rent out idle equipment to spread the cost. Saving on token fees does not make depreciation, electricity, or maintenance disappear.

This is particularly evident among industrial customers in developing countries with limited budgets. These customers often hesitate to send their data outside the company while also lacking the budget for substantial hardware purchases. Data concerns can be addressed progressively through contracts, deployment choices, and technical safeguards; the procurement budget remains a hard constraint. The solution that ultimately gets deployed often processes sensitive data locally and uses the cloud for tasks requiring stronger model capabilities.

Dimension Two: Shadow AI Raises User Expectations and Weakens the Appeal of Assembled On-Premises Open-Weight LLM + Agent Solutions:

When companies procure AI solutions, the people actually making the judgments are individual employees. Many of these employees have long been using frontier models privately through personal accounts. The barrier is extremely low: Anthropic’s Sonnet 5.5, for example, is available to free users on the web and in mobile apps. Employees can export data, paste it into a chat window, and complete a great deal of analysis, writing, and coding; management often does not know, or simply looks the other way.

This is “shadow AI,” and it has an underestimated consequence: employees have already seen the reasoning capabilities of frontier models firsthand. When an AI vendor bids with “a small or midsize open-weight model deployed on premises, combined with its own assembled agent workflows,” these employees often see no technical advantage in the solution:

• Capability: it is clearly weaker than the frontier models employees use every day.

• Approach: the so-called agent merely connects model calls, tools, and workflows—employees can do similar things in their personal accounts.

• Price: it also requires the customer to buy additional hardware.

The deeper problem lies on the supply side. Many AI vendors’ so-called “AI engineers” actually come from software development. Their skills center on agents themselves: their main work is assembling model calls, prompts, tools, and validation workflows. Yet they are unfamiliar with the underlying model architectures and training algorithms and have rarely worked on complex machine learning for scientific and industrial processes: prediction, optimization, and control. As a result, “agents” are packaged as the entirety of “AI.” The solution remains a flowchart in a PowerPoint deck, fails to solve real business problems, and performs far worse than assembling the workflow directly on GPT or Claude.

This contrasts with the strategy of model developers. They are training agent capabilities directly into their models and making them available through official SDKs and tools. Model developers themselves are absorbing the generic agent orchestration layer, leaving less and less room for intermediaries.

The harder work often lies elsewhere: understanding what a factory’s data actually mean, defining production constraints, validating a model’s recommendations, and maintaining the system when something goes wrong. These problems require industry experience and can create real barriers to entry. Solving these problems does not inherently require a company to run a large language model on its own premises. Calculations and analyses grounded in physical and chemical processes often call for machine learning of the kind used in physical AI, rather than small or midsize LLMs. It is often easier to process sensitive data locally with deterministic code and let a cloud model work with de-identified summaries.

Dimension Three: Improving Data Sovereignty and Compliance Options Make Cloud AI More Accessible:

Some investors may still ask whether restrictions on sending corporate data across borders will ultimately force companies back to on-premises deployment. This has been one of the strongest arguments for local deployment in recent years. Cloud providers have invested heavily in addressing it, however, and customers now have far more deployment options than they did a few years ago.

In 2025, AWS opened new regions in locations including Thailand, Taiwan, Mexico, and New Zealand. In January 2026, the AWS European Sovereign Cloud launched in Germany, emphasizing physical and logical separation from other regions, with a planned investment of €7.8 billion. For local businesses, data centers are closer, and it is easier to design data management arrangements and service contracts around a local region.

Microsoft’s sovereign private cloud supports hybrid or fully disconnected deployment, while Google offers Google Distributed Cloud air-gapped for sensitive workloads. Mistral markets sovereign AI around data boundaries, model control, and system auditability, and sees Southeast Asia as an expansion opportunity. These offerings share a common feature: they turn capabilities that companies once had to build and maintain themselves into products they can purchase.

Data protection is also shifting from whether companies can use the cloud to how they can control their data. Providers including OpenAI, Anthropic, and AWS now offer conversation deletion, retention controls, and zero data retention for eligible models and APIs. Subject to the applicable contracts and product terms, businesses can reduce the long-term retention of sensitive data on providers’ systems. There are now concrete ways to address the concern that using a cloud model means leaving conversation history in someone else’s hands indefinitely.

An air-gapped private cloud and a public cloud API are, of course, different services. The availability of the former does not mean a company can send all its data to public models. The appropriate solution still depends on the type of data, local rules, contract terms, and confidentiality requirements imposed by technology licensors. The more accurate conclusion is that businesses have more workable compromises. The choice is increasingly broader than either exposing their data or buying their own GPUs.

For customers able to adopt these solutions, large providers may still build and operate the compute infrastructure; what changes is the service location, network boundary, and payment arrangement. Sovereignty requirements reshape cloud services, but they do not necessarily return demand to small data centers built by individual companies. This is another development that matters when assessing cloud compute demand.

Dimension Four: The Gap Between Open-Weight and Closed-Weight Models Is Widening as Compute Concentrates Among Large Cloud Providers

Many investors are familiar with a particular open-source narrative: closed-weight models move ahead, the open-source community catches up, and model capabilities become cheaper. That process does occur across many tasks. Assuming it will naturally eliminate the gap at the frontier, however, may be too optimistic.

Epoch AI estimates that since January 2026, the best open-weight models have trailed the closed-weight frontier by an average of about four months. From 2023 through October 2025, the gap was approximately three months. Another dataset based on Arena evaluations puts the gap between the strongest closed-weight and open-weight models at 29 Elo points as of September 2026, after the two had briefly been close to parity in January 2025. These measures do not capture every industry task, but they do not support the claim that open-weight models have fully caught up with the frontier.

Why has catching up proved harder than expected? Ultimately, competing at the frontier increasingly requires capital, compute, and fully staffed research teams. Anthropic’s prospectus disclosed $7.33 billion of spending on compute and infrastructure in 2025 and at least $518 billion in infrastructure commitments over roughly the next decade. A small team backed by a few conventional financing rounds can hardly replicate that level of investment.

Community collaboration played an essential role in the growth of open-source software. A project can be divided into many small tasks, with developers contributing code that is gradually integrated through reviews and testing. Linux is a prominent example. Frontier model training, by contrast, requires centralized coordination across many stages, from pretraining and data engineering to reinforcement learning and evaluation. A single large experiment can consume enormous sums, making it much harder to assemble progress from scattered contributions.

Much of the critical know-how never appears in full in research papers. How to organize data, keep training stable, and decide whether a change merits further investment all depend on sustained experimentation and teamwork. As talent and compute concentrate, so do the players capable of running those experiments continuously. Open-weight models will keep improving, but the frontier they are chasing is also moving forward. Leading closed-weight model providers are also continually refining their defenses against model distillation.

The distribution of compute already reflects this concentration. According to an Epoch AI analysis published in April 2026, Amazon, Google, Meta, Microsoft, and Oracle together controlled approximately 71% of the world’s installed AI compute capacity as of the end of the fourth quarter of 2025, measured in H100-equivalent compute. The figure was around 63% in the first quarter of 2024. This 71% represents the share of installed AI compute capacity, not cloud revenue market share. At the time, Epoch also projected hyperscaler capital expenditure of approximately $770 billion in 2026 and more than $1 trillion in 2027.

Consider it from another angle. Even if an ordinary company can run a “good enough” model locally, that does not eliminate its need for cloud capabilities. Tasks involving process technology are handled locally, while general-purpose and programming tasks may still call frontier models. Cheaper individual calls may also encourage AI adoption in work that was previously uneconomical to automate. Any reduction in compute demand from model efficiency must be assessed alongside task volumes and usage frequency, rather than the cost of each inference alone.

Dimension Five: Musk’s Preferential Access to Nvidia GPUs and Strong Financing Capacity Help SpaceX Join the Cloud Compute Leaders:

If compute demand continues to concentrate among large suppliers, the next question is why SpaceX can enter that group. Cloud computing already has Amazon, Microsoft, and Google. Having enough money to buy GPUs alone is not sufficient to compete with them.

What distinguishes Musk is the combination of execution capability in large construction projects, demand from his own models, and a working relationship with Nvidia. When supply is tight, these factors directly affect how quickly GPUs can be obtained, how quickly data centers can be built, and how soon the equipment can be put to work.

The first Colossus cluster, equipped with 100,000 H100 GPUs, went from the start of construction to operation in just 122 days. Jensen Huang publicly praised that pace and said projects of a similar size would ordinarily take years. Bringing a data center online several months late means losing several months of revenue and potentially missing the window when customers need capacity most. Construction capability ultimately shows up in contracts and cash flow.

Nvidia participated in xAI’s financing and consequently holds an equity stake in the combined SpaceX, helping SpaceX secure an advantage in procuring Nvidia equipment.

At a larger scale, Musk has disclosed that Colossus 2 has deployed 110,000 GB200 and 440,000 GB300 GPUs and continues to expand. SpaceX says its data centers collectively contain approximately one million GPUs, with 1.4 GW of installed power capacity as of the end of June 2026. Those headline figures must be considered alongside actual commissioning and utilization. Orders from Google and leading model developers at least provide evidence that external customers are willing to pay for the capacity.

This is where comparisons with conventional GPU cloud providers, or neoclouds, become useful. Companies such as CoreWeave depend heavily on compute rental, making customer concentration, equipment depreciation, and financing costs important to profitability. Beyond compute, SpaceX has operating cash flow from Starlink and demand from its own model and internet businesses, allowing it to allocate capital and compute within the group.

Financing capacity and access to resources are additional SpaceX advantages. Musk’s business and investment network has substantial fundraising capacity. Backed by Starlink’s cash flow, group assets, and industry partnerships, it can combine external financing with internal funding to expand. Ordinary neoclouds struggle to match that financial strength and ability to mobilize resources.

For example, when external rental demand is strong, compute can be prioritized for revenue-generating contracts. When contracts expire and capacity becomes available, it can be used to train and run Grok or support X and coding tools. For a conventional rental provider, idle equipment simply continues to depreciate. SpaceX at least has internal uses for it. Internal consumption must still generate a business return, however; resources spent on training cannot simply be treated as rental cash already received.

Rebuilding the Valuation Framework

Returning to valuation, the AI segment now merits a separate assessment. In the past, it could be assigned a rough estimate of its option value. Today, contracts and costs need to be examined individually. Launch services, Starlink, and AI grow in different ways. Applying the same revenue multiple across all three risks obscuring the differences that matter most.

The profit mix provides a good example. In the third quarter, the Connectivity segment, which includes Starlink, generated $2.37 billion in operating income, compared with $431 million for AI. The AI segment had lost approximately $1.3 billion in the previous quarter. On an operating-income basis, Starlink remains the main pillar. On an adjusted EBITDA basis, however, AI generated about $5 billion, already exceeding Connectivity’s approximately $3.5 billion. If the current growth momentum continues, the AI segment’s reported operating income could also surpass Starlink’s in roughly two quarters.

Why do the two measures differ so much for the same business? The main reasons are depreciation, amortization, and share-based compensation. Approximately $1.2 billion represents Cursor acquisition-related share-based compensation recognized in the quarter. The awards carry future service and vesting conditions, and expenses will continue to be recognized in future periods. This cost therefore cannot simply be labeled “one-off” and removed entirely from future earnings estimates.

For compute rental, contract duration and fee levels clearly matter. According to reports, the disclosed contracts imply an annualized compute rental revenue run rate of approximately $41.1 billion by December. That figure helps indicate the revenue scale at that point, but it is an annualized run rate calculated from monthly fees. It is neither revenue already earned for the whole of 2026 nor unconditional cash flow locked in for future years. Delivery, utilization, and early termination provisions all affect what the contracts are ultimately worth.

Capital expenditure cannot be ignored either. SpaceX spent $19.59 billion in the third quarter, versus $6.88 billion a year earlier, exceeding its revenue for the quarter. Adjusted EBITDA is useful for assessing operating scale, but the returns available to investors must also account for the ongoing cost of buying GPUs, building data centers, and maintaining infrastructure.

GPU depreciation requires the same scrutiny. Depreciating equipment over five to six years for accounting purposes does not mean it can command the same rental price throughout that period. New chip generations change the rental rates, utilization, and maintenance costs of older equipment. Shortening the assumed useful life increases annual depreciation expense and reduces estimated earnings, all else being equal. This variable deserves repeated analysis rather than being treated as a minor footnote to the accounts.

Quarterly net income of $4.4 billion also exceeded operating income of approximately $2.8 billion, indicating that non-operating items affected the final result. Applying a price-to-earnings (P/E) multiple without normalizing earnings for non-operating items risks carrying those effects into the equity valuation. A more useful approach is to assess growth in the Space business, Starlink’s subscriber base and average revenue per user (ARPU), and the cash AI contracts can retain after equipment costs and reinvestment.

The case for rebuilding the valuation framework has therefore emerged, but there is still a calculation to be made between “deserves a separate valuation” and “deserves a higher valuation.” Starlink can remain the cash-flow pillar. The Space business requires monitoring launch prices, market share, and Starship progress. AI requires attention to renewals, customer concentration, revenue per GW, equipment utilization, and ongoing investment. Separating these variables makes it possible to see how much value growth actually creates.

Over a longer industry horizon, hardware investment, frontier model research, and enterprise service capabilities continue to concentrate AI competition among a small number of large players. Ordinary businesses will look for cheaper, more compliant ways to use AI, but that does not necessarily mean maintaining their own model teams and building their own data centers. The more complete the services offered by cloud providers and model developers become, the stronger the case for businesses to purchase those capabilities directly.

SpaceX has already secured large orders from this demand. That is the development in its third-quarter results that most deserves reassessment. Investors should now track whether customers renew, whether compute can continue to be rented at attractive prices, and how much cash remains after expansion. Rockets and Starlink still matter. But any valuation of SpaceX today must take AI compute seriously as a business that needs to be modeled in its own right.

SpaceX also faces risks beyond its business fundamentals. Musk’s high level of political involvement and outspoken positions mean that shifts in U.S. politics, policy, and regulation could affect the company’s operations, valuation, and share price through government contracts, regulatory approvals, and market sentiment. Its technological advantages and compute orders do not eliminate this political risk, which also needs to be reflected in the valuation.

This article is an analysis of the industry and the company and does not constitute investment advice.

    $SpaceX(SPCX)$  $英伟达(NVDA)$  $微软(MSFT)$  $亚马逊(AMZN)$  $谷歌(GOOG)$  

@Tiger_comments  @TigerStars