SemiAnalysis: Musk Breaks Through Again, How SpaceX Competes for the World's Largest AI Cloud?

TL;DR · According to SemiAnalysis tracking, SpaceX has raised its 2027 data center scale target to approximately 13GW, with about 3.6GW currently under construction and identified project capacity exceeding 9GW. · Expansion relies on the construction of the Colossus campus, reuse of rocket bases, and parallel processing of multiple procedures, with some projects expected to take only 4 to 5 months; Tesla may further provide modular electromechanical equipment. · On-site gas power generation helps shorten the waiting time for grid access, while nearly 6GWh of energy storage and public grid access will undertake subsequent scheduling tasks, but the ongoing power supply gap still needs to be addressed. · Some sold computing power enjoys high premiums, but short-term contracts, customer exit clauses, and lack of prepayment arrangements make it difficult to directly extrapolate current revenue levels to the long term. · SemiAnalysis expects that SpaceX may support expansion by lowering prices and increasing the proportion of prepayments; adding over 7GW of capacity in 2027 alone may require hundreds of billions of dollars in funding. · Tight chip supply and financing difficulties for competitors may support its pricing power, while gas turbine licensing lawsuits and regulatory changes will still affect construction plans.

Two months ago, when SpaceX proposed a plan to reach a 10GW data center scale by 2027, there were still many doubts in the market. Now, SemiAnalysis has tracked that several existing and newly built campuses have about 3.6GW of capacity under construction, and SpaceX has raised its expected scale for 2027 to approximately 13GW.

According to SemiAnalysis's comparison, if this target is achieved, it would be equivalent to building all of Google's data center capacity accumulated from 1999 to 2026 in about a year and a half. The agency previously believed that the 10GW plan could create conditions for over $300 billion in annual recurring revenue (ARR), but this calculation does not equate to locked-in annual revenue.

Supporting this judgment is the continuously expanding project reserve and the shortening delivery cycle. SemiAnalysis has identified over 9GW of data center project capacity, while also tracking over 8GW of potential self-supplied power capacity. However, there are still multiple hurdles regarding project scale, actual built capacity, and stable operational capability, including power supply, equipment, permits, and funding.

Compressing Construction Time to Months, Rocket Bases Also Begin Building Data Centers

SpaceX's most direct advantage is compressing construction work that originally required a longer cycle into a few months.

According to SemiAnalysis tracking, the first phase of Colossus 1 was delivered with a capacity of 150MW in 122 days, while Colossus 2 completed 210MW in 91 days. Based on this, the project codenamed Project SPAM, located within the Colossus 2 campus, started construction on October 5, 2026, aiming to bring the first batch of capacity online by December, which is expected to further shorten delivery times.

The construction cycle for the MiniHard project in Memphis is about 5 months, while the newly acquired Gun Range site in Mississippi is expected to take about 4 months. In comparison, SemiAnalysis states that many peers are still striving to control construction cycles within 8 to 12 months. If related projects are delivered as planned, SpaceX will shorten this cycle by about half.

The construction speed has also extended beyond the campus. Musk is attempting to acquire hundreds of residential units around Colossus, both to alleviate community conflicts caused by noise and to provide housing for construction personnel.

New projects have also exceeded the existing Colossus campus. The Texas Starbase, originally responsible for rocket manufacturing and launching tasks, is currently building two data centers. SemiAnalysis tracking shows that the site has moved from land leveling at the end of August to the above-ground main construction phase by the end of September, and is expected to transition to internal installation of the machine room soon.

Combining building volume, the construction timeline for 2027, and its assumptions regarding the deployment of NVIDIA's Vera Rubin platform, SemiAnalysis estimates that the capacity at this site could significantly exceed 1GW. The simultaneous advancement of steel structure construction and roofing also reflects its method of compressing the total construction period through parallel processing of multiple procedures.

TeraFab may provide the next batch of capacity. SemiAnalysis states that this chip manufacturing project, with a planned area of about 100 million square feet and a power scale exceeding 10GW, initially started as a data center project and later shifted to wafer fabrication planning. Local documents and official communication records from August 2026 indicate that the campus may deploy computing power, and relevant representatives have not ruled out the possibility of leasing to third parties.

However, the data center use of TeraFab has not yet been finalized. On-site engineering may serve one or more data centers, or it may support gas supply facilities, central power stations, and other wafer fabrication support. SemiAnalysis leans towards the former explanation, while also noting that the slow progress of main construction may be related to ensuring compliance with relevant agreements and permits. If the data center use is not approved, the early-leveled land may still be repurposed for other uses.

Another Starbase campus in Louisiana has also entered the observation range. Local meeting minutes do not rule out the possibility of building data centers, and the campus also involves on-site power generation planning. SemiAnalysis states that dredging in the waterfront area has begun, heavy road construction is expected to start in the first quarter of 2027, and it is speculated that this site may contribute over 1GW of capacity in 2027. This still falls under early project judgment.

Tesla Supply Chain and Self-Supplied Power Get Construction Started

Behind the rapid construction is the combined effect of engineering organization, equipment supply, and power supply arrangements.

SemiAnalysis estimates that for every 1GW of capacity built by Musk's team, personnel configuration can be controlled at about 3,000 people. The agency also assesses that Tesla may be manufacturing modular electromechanical equipment (MEP) to reduce on-site installation work and external procurement waiting time by increasing the degree of prefabrication.

This approach is similar to some data center developers manufacturing switchgear themselves to shorten delivery cycles; however, SpaceX and Tesla's collaboration may cover a larger scale. If this judgment holds, Tesla's role in SpaceX's expansion will extend from energy storage supplier to data center infrastructure supplier.

Power is adopted in a manner of implementing while constructing. The over 8GW of potential behind-the-meter power supply capacity tracked by SemiAnalysis involves suppliers like FTAI and Solaris. The so-called behind-the-meter power supply refers to configuring power generation facilities on the consumption side to directly supply power to the campus, thereby reducing reliance on new access to the public grid.

At Texas Starbase, about a month after the main construction started, APR mobile power generation equipment began to be deployed. SemiAnalysis estimates that this site may use over 1GW of APR equipment, but from the construction scale, this may still be insufficient to meet final demand. The two natural gas pipelines being laid lead the agency to judge that the subsequent power supply scale of the campus may reach several gigawatts.

Local arrangements also involve laying fiber optic channels through the Port of Brownsville and leasing waterfront sites. Based on this, SemiAnalysis speculates that part of the temporary gas turbines from Colossus 2 may be relocated to Starbase in the future to support the startup of the new campus.

Equipment orders are also progressing. According to information obtained by SemiAnalysis, the first 380MW Doosan gas turbine is being transported, with plans to deliver two units in 2026 and three more in 2027, totaling a nameplate capacity of 1.9GW. Recruitment information suggests that these devices may be used at Texas Starbase, but the specific location has not yet been confirmed.

The agency also infers from recruitment and project descriptions that SpaceX's gas turbine static and dynamic blade casting project may aim to self-supply core components with longer delivery cycles, and then collaborate with complete machine manufacturers. Relevant collaborations have not yet been announced.

A longer-term path involves photovoltaics and energy storage. SemiAnalysis's ChipBook team has tracked the import of 1,730 monocrystalline furnaces, with the first batch of photovoltaic modules expected to be installed as early as the end of 2027, with an estimated annual production capacity of about 25GWp.

Here, GWp measures the peak power of the modules, and the annual production capacity of the modules cannot be directly equated to the operational power generation capacity. The agency further estimates that based on an effective load carrying capacity (ELCC) of about 20% (which measures the contribution of power sources to reliable supply), the related deployment may contribute about 4 to 6GW of reliable capacity once fully rolled out, but this result depends on actual construction, energy storage configuration, and system operating conditions.

Nearly 6GWh of Energy Storage Cannot Replace Continuous Power Supply

As the scale of Colossus 2 expands, the power supply plan is shifting from temporary gas power generation to a collaboration of public grid, on-site power generation, and energy storage.

According to SemiAnalysis citing information from the Tennessee Valley Authority (TVA), Macrohard and MiniHard have respectively applied for 700MW and 450MW of power supply services. The agency temporarily assumes that this grid access can cover the entire demand of the two facilities.

A 500kV gas-insulated substation near MiniHard is also under construction, scheduled to be completed by December 31, 2026, and connected to the Freeport substation. This timeline is particularly critical for subsequent adjustments to gas turbine units.

SemiAnalysis states that, according to binding requirements signed on July 30, 2026, the currently operational nameplate capacity of the Southaven power station is 2.7GW, but will drop to 1.24GW by July 2027.

Even if the public grid provides about 1.2GW of power, supply and demand may not match. According to SemiAnalysis's estimates, after deducting grid power supply, the remaining IT equipment load is about 1.3GW, which is higher than the 1.24GW nameplate capacity of power generation at that time. Considering unit derating, operational losses, and power consumption of cooling and other supporting facilities, the gap may exceed 300MW.

This is also why large-scale battery energy storage systems (BESS) are receiving attention. SemiAnalysis states that Colossus 2 has 5.9GWh of energy storage being installed, approaching 6GWh, which may become one of the largest energy storage deployment projects in the world. The agency estimates that SpaceX may be the largest buyer of Tesla Megapacks, with procurement volumes potentially reaching about 8GWh under its high-end estimates, and speculates that Tesla may still hold large orders that have not yet been reflected in revenue.

SemiAnalysis envisions that the campus can prioritize power supply for gas turbines and energy storage in a coordinated manner, distinguishing between customer service and internal training, and setting different service level agreements (SLAs) for different clusters to enhance computing power utilization under existing power conditions.

If the Power Usage Effectiveness (PUE) indicator can be maintained between 1.10 and 1.15, the power consumption of supporting facilities can also be controlled. PUE refers to the ratio of total power consumption of the campus to the power consumption of IT equipment, with a lower ratio indicating a lower proportion of auxiliary power consumption.

However, energy storage can shift power supply time and buffer load fluctuations, but it does not generate electrical energy itself. If a 300MW gap exists continuously over the long term, merely expanding battery capacity cannot eliminate this gap; additional power generation, surplus power available for charging, or reducing and staggering some loads will be needed. Therefore, whether this plan can be established still depends on the actual output of the units, load curves, and scheduling design.

High-Value Orders Bring Cash Flow, Expansion Still Requires Hundreds of Billions

Whether the construction speed can translate into a profit advantage also depends on how long customers are willing to pay a premium for early access to computing power.

According to SemiAnalysis citing SpaceX CFO Bret Johnsen, the company has sold over 1GW of capacity at an annualized price close to $50 million per MW, corresponding to a payback period of less than a year. The five-year contract market price reference provided by the agency is about $12 million to $16 million per MW per year. Due to differences in contract duration, exit rights, and delivery conditions, this price difference cannot be directly viewed as a premium under the same conditions.

SemiAnalysis also states that some customers received GB300 quotes exceeding $15 per hour, but the materials did not further clarify the complete billing basis for this quote.

The high prices are based on short-term scarcity. Some of SpaceX's current transactions have a term of about 6 months, allowing customers to notify exit 90 days in advance without prepayment; agreements with Anthropic and Google are irrevocable only until January and March 2027, respectively. Such contracts can quickly monetize scarce computing power, but they also mean that revenue continuity cannot simply be extrapolated at current prices.

SemiAnalysis expects that over 50GW of new data center capacity will be added globally in 2027. Against this supply backdrop, SpaceX needs AI laboratories to continue exceeding revenue expectations and consistently underestimate their computing power needs to maintain strong additional procurement.

The consumption of funds has already begun. The agency states that SpaceX had nearly $100 billion in net cash when launching its cloud business, but this amount is rapidly decreasing. To add over 7GW of capacity in 2027 alone may require hundreds of billions of dollars.

To expand the customer base and support GPU procurement financing, SemiAnalysis expects that SpaceX may accept lower prices in exchange for higher prepayments, such as adopting 6 to 12-month contracts, over 50% prepayment, and a low-end scenario price of about $30 million per MW per year. This is the agency's prediction of future contract structures and not a publicly announced uniform quote.

Chip supply and competitive landscape may ease pricing pressure. SemiAnalysis believes that Vera Rubin supply is becoming tight, and by the end of 2027 or early 2028, AI chips may replace power and data centers as the main bottleneck for computing power expansion.

Meanwhile, some emerging cloud vendors, although they have received NVIDIA quotas, still need to find customers, data centers, and financing simultaneously. After the capital market tightens, they may rely more on long-term purchase agreements from large-scale customers, making it difficult to directly participate in the high-premium on-demand market where SpaceX operates. If this judgment holds, reduced competition and chip shortages will help SpaceX maintain pricing power for a longer period.

Federal Support Benefits Power Supply, But Lawsuits Are Not Over

Gas turbine licensing is another hurdle that the expansion plan still needs to cross.

According to SemiAnalysis, the NAACP filed a lawsuit against xAI in April 2026 under the Clean Air Act, targeting the use of unlicensed mobile gas turbines at the Southaven site supplying power to Colossus 2, demanding a shutdown until a license is obtained.

The agency states that the U.S. Department of Justice sought to intervene in June on grounds of national, economic, and energy security, including Grok's support for defense-related actions, and subsequently submitted related procedural issues to the Fifth Circuit Court of Appeals. On October 19, the judge allowed the federal government to participate in the case but did not terminate the lawsuit based on this; on October 30, the judge rejected xAI's claim that mobile turbines could be exempt from licensing as "mobile pollution sources." The case still faces preliminary injunction disputes.

Rule-making may also change licensing requirements. SemiAnalysis states that the COMET alliance, composed of gas turbine manufacturers and users, has petitioned the EPA to classify mobile turbines as "non-road engines," thereby changing the applicable licensing framework.

The agency indicates that the proposed rules were submitted to the White House for review on October 15; if ultimately approved and applied as expected, it may benefit xAI, but its legal effects still depend on the final rules and subsequent judicial processes.

SemiAnalysis also cites an executive order from July 2025 to expedite data center power supply licensing, as well as Trump's statements about protecting Musk's enterprises, judging that there is political momentum at the federal level to support the continued operation of the project. This does not mean that gas turbines can ensure continuous operation, as community disputes, emission permits, and court rulings may still affect power supply arrangements.

The logic of SpaceX competing for the AI cloud market is thus gradually becoming clear: by delivering data centers and power faster, it can seize customers' urgent additional computing power needs, and then use high-value orders and customer prepayments to support subsequent construction.

However, 13GW remains a target. Whether it can become a stable, revenue-generating capacity will depend on whether power supply continuity, chip delivery, contract renewals, and financing progress can keep pace simultaneously.