SpaceX’s ambitions in artificial intelligence are undergoing a significant evolution in southeastern Arizona, where an affiliated entity has proposed what may become one of the world’s largest dedicated supercomputing and AI infrastructure campuses. This 747-acre development in Cochise County features a substantial projected electrical demand of up to 6 gigawatts, to be fulfilled across four construction phases.

If successfully executed, the project would represent an unprecedented concentration of computing infrastructure, necessitating millions of advanced accelerators, extensive electrical grid upgrades, large-scale energy storage, and a robust supply chain capable of hardware procurement at a scale rarely attempted.

However, there is a fundamental distinction between the proposal of a 6 GW supercomputing facility and the practical capability to construct, power, and equip such an installation. This reality is central to the application submitted by Tombstone Resources LLC, a SpaceX-affiliated entity operating as SpaceXSI. The proposal outlines a development near Cochise, Arizona, comprising up to nine buildings of approximately 493,000 square feet each, phased into four 1.5 GW stages.

This project raises questions that extend beyond a single technology firm's objectives, underscoring how the global race for AI infrastructure is increasingly testing the limits of regional power systems, hardware manufacturing capacity, transmission infrastructure, and environmental resources.

Six gigawatts is an extraordinary computing commitment

To put the proposed scale into perspective, 6 GW represents approximately 19 percent of Arizona's stated 32 GW electrical grid capacity. The project's projected demand would also be roughly 2.5 times the combined demand attributed to data centers across Maricopa County in the reporting cited by the vetting brief.

The development would be positioned adjacent to the Apache Generating Station, an existing power complex with approximately 626 MW of coal, natural gas, and solar generation capacity. The proposal envisions expanding the surrounding energy infrastructure to support a campus whose ultimate demand could exceed 6 GW.

The difference in scale is remarkable.

This is not simply another hyperscale data center expansion. It is a proposal to establish an enormous computing destination whose electrical requirements would rival those of major industrial regions.

The application identifies approximately $3.6 billion in grid infrastructure upgrades, including an initial phase estimated at $500 million. Construction could begin in March 2027 if the required zoning approvals are obtained.

However, the proposed electrical infrastructure is not the same as secured, operational computing capacity. The application describes an ambitious development schedule, but the availability of the necessary generation, transmission equipment, construction resources, and computing hardware remains a central question.

For the supercomputing industry, that distinction matters.

A building capable of housing millions of processors is not a supercomputer. The actual computing system depends on accelerator availability, memory capacity, high-speed networking, storage, power distribution, cooling, and the software infrastructure required to operate the machines efficiently.

At 6 GW, every one of those requirements becomes a major engineering and procurement challenge.

The GPU question: Where do millions of accelerators come from?

One of the most revealing aspects of the proposal is what it does not disclose.

The application does not identify the intended computing workloads, a specific customer, an accelerator supplier, or the hardware architecture that would populate the campus.

That omission makes it difficult to independently evaluate the project's ultimate computing capacity.

A useful way to understand the scale is to perform an illustrative hardware calculation.

Using the approximately 1.9 kW per GPU figure cited in the vetting brief for NVIDIA's GB300-class hardware, a 6,000 MW IT load would correspond to approximately 3.16 million GPUs.

That is an order-of-magnitude illustration, not a procurement forecast. Future accelerator generations will have different power characteristics, and the actual facility electrical demand would also depend on cooling, networking, storage, and other infrastructure loads.

Nevertheless, the calculation exposes the extraordinary scale of the proposed undertaking.

Millions of accelerators would require an equally substantial supporting infrastructure. High-bandwidth memory, advanced packaging, high-speed interconnects, network switches, optical connectivity, storage systems and power-delivery equipment would all have to arrive on a coordinated schedule.

The challenge is not merely purchasing GPUs.

It is assembling a complete, operational computing platform at a scale that could strain multiple segments of the semiconductor and data-center supply chains.

SpaceXSI has not publicly identified the accelerator procurement strategy for the Cochise development. Nor has it disclosed a workload roadmap explaining whether the campus would support large-scale model training, inference, AI cloud services, or a combination of computing operations.

Without those details, the 6 GW figure describes an infrastructure ambition rather than a verified measure of future computing performance.

A 24 GWh battery cannot manufacture electricity

The proposal also includes a 24 GWh battery energy storage system intended to stabilize power delivery and reduce the impact of the campus on the surrounding electrical grid.

That is a substantial installation, but its role needs to be understood correctly.

At a continuous 6 GW draw, a fully charged 24 GWh battery would theoretically provide approximately four hours of electricity before depletion, assuming the entire rated capacity were available and ignoring conversion losses, operating reserves, and other limitations.

That makes the battery a potentially valuable grid-management component. It does not, however, represent an independent source of sustained electricity.

The distinction is important because a computing campus operating at this scale requires dependable power around the clock.

Battery storage can help manage fluctuations, provide short-duration backup, and smooth the interaction between generation and demand. It cannot replace the generation and transmission infrastructure needed to supply a continuous multi-gigawatt computing load.

The project's ultimate success will therefore depend on whether its proposed electrical upgrades and associated generation resources can be delivered on schedule.

There is also a financial question: who bears the cost if the campus is delayed, substantially reduced in size, or never reaches its proposed capacity?

The application identifies potential electricity-bill reductions of 10 to 40 percent for members of Sulphur Springs Valley Electric Cooperative. Those figures are applicant projections, not independently established outcomes, and are subject to regulatory approval.

The Arizona Corporation Commission's review of the proposed Energy Services Agreement will be an important test of the project's financial structure and protections for existing electricity customers.

The water issue is bigger than the cooling system

SpaceXSI's application emphasizes a dry-cooling design that would avoid conventional cooling towers and evaporative spray during normal operation.

The company estimates annual direct water consumption of approximately 7 to 10 acre-feet at full build-out, primarily for domestic use by facility personnel. It also states that no new groundwater well would be required for industrial cooling.

On the surface, those figures appear relatively modest for a development of this size.

But direct facility water consumption is only part of the equation.

The proposed campus would require enormous quantities of electricity. Generating that electricity can also consume water, depending on the generation technologies, cooling systems, and fuel sources involved.

Reports cite estimates that power generation can increase the overall water footprint associated with data centers by 60 to 90 percent. Yet the application does not quantify the indirect water consumption associated with supplying the full 6 GW campus.

That is a significant information gap.

The proposed development is located in the Willcox groundwater basin, one of Arizona's most severely depleted groundwater regions.

According to the Arizona Department of Water Resources, groundwater declines have reached approximately 400 feet in parts of the basin between 1940 and 2015. An estimated 5.7 million acre-feet of groundwater has been removed from storage.

The state's hydrologists have warned that even if pumping stopped entirely, recovery could take more than 280 years.

The issue is therefore not simply whether the data center itself uses a relatively small amount of water. It is whether the full energy infrastructure required to operate the campus can be supported without creating additional pressure on an already stressed regional resource.

Dry cooling addresses one part of that challenge. It does not answer the larger question.

The Apache power connection deserves scrutiny

The proposed campus would be located beside the Apache Generating Station, an energy complex whose existing generation includes coal, natural gas and solar resources.

The station had previously been expected to retire by 2028, but its operating outlook changed following federal support for upgrades.

Reports identify approximately $21 million in federal funding intended to support continued operation.

That creates an interesting contrast with the project's presentation as a modern, battery-supported, dry-cooled AI infrastructure development.

The proposed computing campus would be built alongside an energy complex whose continued operation includes fossil-fuel generation.

This does not establish the project's eventual emissions profile. That will depend on the actual generation mix, power contracts, grid arrangements, and any additional infrastructure ultimately constructed.

But it does raise an important question for an industry increasingly making claims about efficient and sustainable AI infrastructure: What is the real energy and environmental footprint of the computing systems once the electricity required to operate them is accounted for?

The answer cannot be determined from the campus's cooling design alone.

Is this the next stage of SpaceX's AI infrastructure ambitions?

The Cochise proposal arrives amid SpaceXSI's broader expansion of AI computing infrastructure associated with xAI and Grok.

The company's existing Memphis-area computing operations have been described as exceeding 1 GW of capacity, although the precise scale has been disputed in outside reporting.

The Arizona proposal would represent a substantial escalation in ambition.

However, there is an important distinction between the Memphis operations and the Cochise application. The Arizona filing does not identify the campus as a new Colossus facility, does not establish a direct connection to the Colossus 2 expansion, and does not disclose a specific customer or workload.

It should therefore be treated as a separate proposed development rather than automatically labeled Colossus 3.

The broader strategic direction is nevertheless apparent: access to large-scale computing capacity is becoming increasingly dependent on the ability to secure dedicated energy infrastructure.

For AI developers, owning or controlling the electrical supply can become just as strategically important as securing advanced accelerators.

But electricity alone does not produce competitive computing performance. The quality of the system will ultimately depend on how efficiently the power is converted into useful computational work.

That means accelerator utilization, memory bandwidth, network performance, storage throughput, thermal management, and software efficiency will determine whether an enormous facility delivers corresponding computing value.

A 6 GW electrical commitment is not a benchmark.

The December hearing is the first major test

The Cochise County Planning and Zoning Commission is scheduled to consider the proposal on December 9, 2026, at 2 p.m. in Bisbee.

Written comments intended for inclusion in the hearing packet are due November 27 at noon.

The county has explicitly stated that the project figures are based on information supplied by the applicant and have not been independently verified or endorsed by the county.

That qualification is essential.

The proposed 1,000 jobs, projected tax benefits, electricity-bill reductions, infrastructure costs, and construction schedule should all be treated as applicant claims until independently substantiated.

The zoning hearing is also not the only regulatory consideration. The proposed energy agreement will face a separate review by the Arizona Corporation Commission.

The project therefore has multiple milestones ahead before its proposed computing capacity can become an operational reality.

The real benchmark is delivery, not gigawatts

The SpaceXSI proposal exemplifies the extraordinary scale of the infrastructure competition currently emerging within the artificial intelligence trade. Establishing a 6 GW supercomputing campus necessitates a level of coordination across power generation, grid transmission, semiconductor manufacturing, and hardware integration that remains largely unprecedented.

While this project serves as a compelling case study regarding the integration of large-scale computing and energy infrastructure, the industry's tendency to quantify ambition in gigawatts should not overshadow critical engineering and operational uncertainties. Fundamental questions remain regarding the procurement of millions of accelerators, the definition of target workloads, the actual conversion efficiency of electrical capacity into computing power, and the long-term financial accountability for grid upgrades. Furthermore, the project must address substantial risks concerning construction timelines and the indirect water consumption necessitated by such a massive energy demand.

Ultimately, these are not peripheral concerns but are essential to determining the project's viability. Cochise County currently holds a development proposal, not a functioning supercomputing installation. The next generation of AI infrastructure will not be defined by the magnitude of power commitments, but rather by the capacity to transform energy into reliable, efficient, and economically sustainable computing performance. For SpaceXSI, the 6 GW ambition represents only the initial phase of this rigorous challenge.

Microsoft's decision not to exercise a massive expansion option exposes the financial, power-delivery, and deployment risks facing the AI supercomputing industry.

Microsoft’s decision to forgo a potential multibillion-dollar expansion of its AI computing agreement with British infrastructure provider Nscale has prompted critical questions regarding the economic and operational viability of the rapidly expanding supercomputing trade. The decision concerns additional capacity at Nscale’s Cedarvale campus in Ward County, Texas, where Microsoft currently has an agreement for about 234 megawatts of infrastructure using roughly 104,000 NVIDIA GB300 GPUs. 

Following a report by the Financial Times on November 5, it was confirmed that Microsoft declined an option to more than double its existing commitment, a move that potentially precludes tens of billions of dollars in future revenue for Nscale. While the core agreement remains in place, the abandonment of this expansion highlights a complex trend in the AI infrastructure market, where massive contractual commitments are increasingly scrutinized against the practical, high-stakes challenges of power delivery, facility construction, hardware installation, and the timely commissioning of computing capacity. In an industry defined by hundreds of billions of dollars in capital expenditure, these operational distinctions are of paramount importance.

A 234 MW AI supercomputing deployment, with a much larger question mark

Nscale's existing Texas commitment represents a substantial deployment of GPU-based computing infrastructure.

The approximately 234 MW facility is designed to accommodate roughly 104,000 NVIDIA GB300 GPUs, with phased delivery beginning in the third quarter of 2026.

This is not simply a conventional colocation facility. It is part of the infrastructure required to support large-scale AI training and inference workloads, where accelerator density, memory bandwidth, high-speed networking, storage throughput, and power availability determine how effectively computing resources can be utilized.

At this scale, the challenge extends well beyond purchasing GPUs.

A deployment involving more than 100,000 accelerators requires extensive electrical distribution infrastructure, high-density cooling systems, networking fabrics capable of supporting large distributed workloads, and storage architectures that can continuously feed data to thousands of interconnected compute nodes.

The facility must also deliver predictable performance under sustained workloads. A large GPU installation that cannot maintain adequate power delivery, cooling capacity, or network performance does not translate automatically into productive AI computing capacity.

Microsoft's decision not to expand its commitment therefore raises an important question: What changed between the original infrastructure plans and the additional capacity Microsoft ultimately declined?

The exact reason has not been publicly established.

Nscale's $103.4 billion backlog deserves closer examination

The more consequential issue may be Nscale's reported contractual backlog.

In its September IPO filing, Nscale disclosed approximately $103.4 billion in total contract value. Its agreements with Microsoft represented up to approximately $43.8 billion through December 2033, excluding optional extensions. Anthropic represented another approximately $44.6 billion.

Those numbers suggest extraordinary demand for Nscale's computing infrastructure.

But the operational picture is considerably less impressive.

As of August 31, 2026, approximately $2.6 billion of the company's reported $103.4 billion contract value was operational, representing roughly 2.5% of the total. The filing also indicated that approximately 25,000 of 461,000 contracted GPUs were active.

This distinction is critical.

A contractual commitment represents an expectation of future computing services and associated revenue. It does not establish that the physical infrastructure supporting those services has been completed, energized, commissioned, and made available to customers.

The gap between contracted value and operational capacity exposes Nscale to a difficult execution challenge.

The company must transform an enormous pipeline of agreements into functioning infrastructure while managing construction schedules, power availability, financing obligations, hardware procurement, and customer deployment requirements.

Microsoft's decision does not reduce the existing $103.4 billion backlog because the declined expansion option was not included in that figure.

Nevertheless, Nscale loses the potential future business associated with the additional Texas capacity.

That is an important distinction for prospective investors evaluating the company's proposed public-market valuation.

The 500 MW versus 700 MW discrepancy raises additional questions

The size of the abandoned expansion remains another unresolved issue.

The Financial Times described an option that would have more than doubled Microsoft's existing commitment. One account placed the additional capacity at approximately 500 MW.

However, Nscale's original October 2025 announcement referred to an option for a second 700 MW phase beginning in late 2027.

On November 5, the company corrected that announcement, removing the statement connecting Microsoft to the 700 MW second phase and acknowledging that the original wording was incorrect.

The corrected announcement still describes the Texas campus as having the potential to scale to 700 MW, with Nscale pursuing a broader expansion of its Texas footprint.

These are materially different propositions.

An infrastructure site having the technical potential to reach 700 MW does not mean Microsoft contracted for that capacity.

The distinction matters because the difference between 500 MW and 700 MW represents a substantial amount of additional electrical infrastructure, GPU deployment, cooling capacity, and potential computing revenue.

Neither figure should be treated as definitively established for the declined option.

The timing of the correction also deserves scrutiny. The company revised its previous announcement on the same day the Financial Times report became public.

Nscale has not publicly explained the discrepancy in sufficient detail to resolve precisely what the original option covered.

For a company seeking a major public offering, clarity surrounding customer commitments and future capacity should be fundamental.

Figure could become an important customer, but replacement is not guaranteed

Nscale is not without potential alternatives.

In September 2026, the company announced a multiyear partnership with Figure, the humanoid robotics company, involving the potential deployment of up to 100,000 NVIDIA Vera Rubin GPUs at its Barstow, Texas, facility.

The initial deployment is targeted for the second half of 2027.

Figure's initial computing commitment is approximately $3.5 billion, with an expressed intention to scale beyond $6 billion. Nscale will also become a shareholder in Figure and its preferred computing provider.

The agreement is significant because it connects advanced robotics with next-generation AI supercomputing infrastructure.

Humanoid robotics requires substantial computational resources for model training, simulation, perception, reasoning, and the development of increasingly sophisticated control systems.

Large-scale GPU infrastructure could support the computational workloads required to train and improve these systems.

However, Figure's agreement should not automatically be interpreted as a replacement for Microsoft's declined Texas expansion.

The financial commitment is substantially smaller than the reported value of the abandoned option. The deployment involves NVIDIA Vera Rubin rather than the existing GB300 infrastructure, and the initial deployment is still targeted for a future date.

There is also an important distinction between a customer's intention to scale and a fully contracted, financed, and operational computing deployment.

Nscale has reportedly been discussing the remaining Texas capacity with several major potential customers.

Whether those discussions translate into commitments of comparable scale remains to be seen.

The lease and financing obligations create additional exposure

One of the more important details surrounding the Texas project is that Nscale does not own the underlying Cedarvale facility.

The company leases approximately 234 MW of capacity from Ionic Digital under a reported 126-month triple-net lease.

Under this arrangement, Nscale carries rent and operating-cost obligations over the lease term, creating financial exposure even if customer demand or infrastructure utilization falls short of expectations.

Secondary summaries of Ionic Digital's filings put the lease value at approximately $1.95 billion through January 2037.

The financing structure also includes a reported NVIDIA guarantee covering up to $860.3 million in rent during the first five years, with Nscale's parent company guaranteeing subsequent payments.

Nscale also closed approximately $3.05 billion in senior secured delayed-draw term loans on August 31, 2026. Of that amount, approximately $1.85 billion was allocated to Ward County, with the remainder supporting a separate project in Madison County, North Carolina.

These financing arrangements predate the November announcement and should not be interpreted as a response to Microsoft's decision.

Nevertheless, they demonstrate the scale of capital required to develop and equip modern AI infrastructure.

The fundamental problem is that infrastructure financing and customer demand do not always move at the same speed.

A developer can secure financing, negotiate GPU procurement, and sign long-term lease agreements while still facing uncertainty about when the electrical infrastructure will be available and when customers will begin consuming computing services.

For infrastructure providers operating at this scale, utilization and delivery schedules are central to the economics of the business.

Microsoft's decision does not signal an AI demand collapse

There is an important reason not to interpret Microsoft's decision as evidence that the AI computing market is weakening.

Microsoft's October 28 fiscal first-quarter 2027 results showed Azure revenue growth of 48%, reaching approximately $33.2 billion.

Capital expenditure, including finance leases, reached approximately $45.2 billion. That figure was below the company's earlier guidance of more than $50 billion, with third-party infrastructure delivery delays cited as a contributing factor.

Microsoft also indicated that demand for computing capacity continued to exceed supply.

Those figures suggest that the company's infrastructure requirements remain substantial.

The more plausible questions concern the timing, location, economics, and reliability of individual infrastructure projects.

Microsoft may be reallocating capacity commitments toward projects that can deliver power and computing resources on a more predictable schedule. It may also be reassessing the economics of individual leases or the timing of hardware deployments.

The company's earlier withdrawal from Nscale's planned Loughton project in Essex, where grid-connection delays pushed expected power availability into the 2030s, reinforces the importance of infrastructure readiness.

These decisions are not necessarily evidence of reduced AI demand.

They may instead indicate that hyperscale customers are becoming more selective about where they commit capital and how much risk they are willing to assume for infrastructure that has not yet been delivered.

That distinction is increasingly important as AI infrastructure spending accelerates.

Texas introduces another layer of infrastructure uncertainty

The Texas expansion also faces a changing regulatory environment.

Governor-elect Gina Hinojosa has pledged to halt new data-center approvals until legislation addresses concerns surrounding electricity costs, water consumption, environmental impacts, and local decision-making.

Texas has become one of the largest destinations for data-center investment because of its land availability, energy resources, and historically business-friendly environment.

But the growing electricity requirements of AI infrastructure are placing additional pressure on the state's power system.

For a campus with ambitions to scale toward hundreds of additional megawatts, access to electricity cannot be treated as a secondary consideration.

Power availability, transmission infrastructure, regulatory approvals, and community acceptance are increasingly becoming determining factors in whether proposed AI supercomputing projects can proceed on schedule.

The Texas situation illustrates a broader industry challenge: the availability of GPUs does not guarantee the availability of the infrastructure needed to operate them.

The real test is operational computing capacity

Nscale’s current situation underscores a fundamental challenge facing the AI supercomputing industry. While the market has proven highly adept at announcing ambitious infrastructure projects, securing multibillion-dollar computing agreements, and projecting extensive GPU deployments, the execution of these initiatives remains significantly more complex.

Success in this sector requires the precise synchronization of semiconductor supply chains, advanced packaging, high-bandwidth memory, high-speed networking, power generation, grid integration, liquid cooling, facility construction, and long-term financing. Any failure in this orchestration can have cascading consequences: electrical infrastructure delays can leave completed facilities dormant, while bottlenecks in cooling or networking can render installed accelerators inefficient. Furthermore, shifts in customer deployment schedules impose substantial financial pressure on developers who bear significant lease and financing obligations.

For Nscale, the immediate imperative is to demonstrate that its extensive contractual backlog can be effectively converted into operational computing infrastructure and consistent revenue. Conversely, for Microsoft, the objective is to secure reliable computing capacity without committing to projects that fail to meet stringent operational and economic requirements.

For the broader industry, the lesson is becoming increasingly clear: the next phase of the AI infrastructure race will not be defined solely by the volume of GPUs announced or the total value of computing contracts signed. Rather, it will be determined by the capacity to deliver the power, cooling, networking, storage, and operational stability required to ensure system productivity. Nscale's recent decision regarding its Texas expansion serves as a critical reminder that a multibillion-dollar computing agreement is not synonymous with operational supercomputing capacity. As the industry advances toward increasingly massive AI facilities, this distinction will likely become one of the most reliable metrics for evaluating the long-term sustainability of the current infrastructure expansion.