Five months ago, the outlook for xAI looked considerably less convincing. By late March, the last two of the original 11 researchers who co-founded the company with Elon Musk had reportedly departed. Musk acknowledged the problem himself, writing that xAI “was not built right first time around” and was being rebuilt “from the foundations up”.
On 12 August, the organisation, now part of SpaceX, released Grok 4.6. Artificial Analysis gives the model a score of 61 on its Intelligence Index, level with OpenAI's GPT-5.6 Sol at maximum reasoning effort and behind only Anthropic's leading models. Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens below 200,000 prompt tokens. Artificial Analysis describes that headline pricing as more than 60% below GPT-5.6 Sol and Claude Opus 5. Two days after launch, Grok 4.6 was available inside GitHub Copilot.
The same week, SpaceX completed its acquisition of Cursor, following an agreement that values the AI coding company at $60bn. Axios has described it as the largest acquisition of a venture-backed startup on record. Cursor says access to SpaceX's compute infrastructure will allow it to build more capable models at lower cost. Look at those events purely as an AI model story and the turnaround is striking. Look at them as an infrastructure story and they begin to make considerably more sense.
The important question is no longer whether Grok can beat GPT or Claude on the next leaderboard. It is whether SpaceX is building a structural advantage by controlling progressively more of the industrial system required to manufacture intelligence.
A frontier model is only the visible layer
AI competition is usually narrated through models. OpenAI releases one, Anthropic responds, Google publishes a new Gemini generation and SpaceXAI releases Grok. Benchmarks move, token prices change and attention switches to whoever occupies the frontier that week. That increasingly describes only the visible part of the competition.
Nvidia chief executive Jensen Huang offers a more useful way to view the market. He describes AI as a five-layer stack: energy → chips → infrastructure → models → applications. Every application ultimately depends on the physical layers underneath it, from the electricity supplying a data centre to the processors converting that energy into computation.
Viewed through that framework, SpaceX looks unusual. It does not yet literally own every layer. Its semiconductor ambitions remain under construction and its orbital compute plans remain plans rather than operating infrastructure. But it is attempting to establish meaningful control across all five.
That distinction matters because most frontier AI companies optimise primarily around models and applications while contracting extensively for the physical infrastructure beneath them. SpaceX is attempting something closer to industrial vertical integration. Its real AI strategy may therefore be less about building the best model than about controlling more of the system that determines how cheaply, quickly and repeatedly frontier models can be built.
The first layer is increasingly energy
AI's physical constraints are becoming difficult to hide behind cloud abstractions. Training and serving frontier systems requires enormous quantities of electricity, and SpaceX has been unusually aggressive about securing it for Colossus. SemiAnalysis reports that the Southaven power plant supporting SpaceX's AI infrastructure expanded from 27 gas turbines providing roughly 495MW in February to 69 turbines providing around 1.7GW by July. SpaceX has separately said it intends to transition towards permanent power infrastructure.
The rapid expansion has also generated significant environmental and permitting controversy, including litigation over the use of temporary turbines. Those environmental and regulatory issues are material. Vertical integration does not eliminate externalities or regulatory constraints simply because infrastructure can be built quickly.
Strategically, however, the development shows where SpaceX believes the bottleneck sits. Power is not treated as somebody else's utility problem. It is part of the AI system. That is a recurring pattern across Musk's companies: move a dependency inside the operational boundary when the external market cannot supply it at the required speed, cost or scale. For AI, that logic now starts with electrons.
Chips are the next dependency SpaceX wants to internalise
The semiconductor layer is less mature, but potentially more consequential. SpaceX and Tesla announced in August that they intend to invest an initial $16.8bn in Terafab, a vast semiconductor manufacturing project in Grimes County, Texas. Intel is participating in the initiative, although the precise nature of its contribution remains less clearly defined. SpaceX's SEC filings describe a long-term ambition to create an integrated facility spanning logic and memory fabrication, packaging and associated chip design, with an eventual target of producing one terawatt of compute hardware annually.
This should not be confused with current self-sufficiency. SpaceX explicitly says it expects to continue sourcing a significant proportion of its compute hardware from third-party suppliers. Terafab is an attempt to augment those relationships and reduce future constraints, not evidence that SpaceX has already displaced Nvidia, TSMC or the wider semiconductor supply chain.
The strategic direction is still unmistakable. SpaceX's own prospectus describes the constraints on AI growth as increasingly physical: chip manufacturing, data-centre infrastructure and power generation. Its response is to move further down the stack. That changes the unit of competition from model development to industrial capacity.
Colossus is becoming a business, not merely an internal asset
The infrastructure layer is where that strategy is already producing remarkable economics. SpaceXAI says Colossus contains more than 200,000 Nvidia GPUs. More importantly, SpaceX has started monetising that capacity externally rather than reserving it exclusively for Grok.
Its SEC prospectus discloses agreements under which Anthropic has access to approximately 325,000 Nvidia GPUs across Colossus and Colossus II, paying $1.25bn per month once the capacity is fully ramped. Google subsequently agreed to pay approximately $920m per month for access to around 110,000 GPUs and associated compute infrastructure. Annualised, those two agreements alone represent approximately $26bn of revenue.
For context, SpaceX reported total 2025 revenue of approximately $18.7bn across Space, Connectivity and AI. In other words, the annualised value of two major compute contracts now exceeds the company's entire revenue base from the previous year. Anthropic is a direct competitor to Grok at the model layer. Google operates Gemini and is another of the world's most important AI developers. Yet both are becoming customers of infrastructure owned by SpaceX.
Competition at one layer is becoming revenue at another.
Your competitor can also be your customer
This is where the architecture becomes strategically interesting. If Grok wins, SpaceX benefits. If Cursor grows, SpaceX benefits. If third-party developers use Grok through GitHub Copilot, SpaceX benefits. But if Anthropic needs more compute to grow Claude, SpaceX can benefit from that too. If Google's AI demand exceeds the capacity it wants to provide internally, SpaceX can monetise the shortfall.
That produces a different kind of competitive position. Most AI companies need their proprietary product to win in order to capture the majority of the upside. An infrastructure owner can potentially benefit from demand regardless of which application captures the end customer. AWS demonstrated the power of that position in cloud computing. Nvidia demonstrated it again in accelerated computing.
SpaceX appears to be exploring whether the same principle can apply to the production of intelligence itself. The stronger the AI market becomes, the more demand there is for power, GPUs, networking, storage and data-centre capacity. A vertically integrated supplier does not necessarily need to predict the eventual model winner if it owns scarce infrastructure every winner needs.
Grok 4.6 changes what the infrastructure story means
Infrastructure alone would not be enough. If SpaceX became an enormous neocloud provider but consistently failed to build competitive AI systems, the strategy would look less like vertical integration and more like diversification. That is why Grok 4.6 matters.
Artificial Analysis places it back at the frontier, scoring 61 alongside GPT-5.6 Sol at maximum reasoning effort. It remains behind Anthropic's strongest models on the same index, so claims that SpaceXAI has suddenly become the leading AI laboratory would be premature. But Grok 4.6 substantially closes the capability gap while competing aggressively on price.
The timeline is equally important. In March, Musk was publicly acknowledging that xAI needed rebuilding after the departure of its founding research team. By August, the organisation had produced a model operating at the frontier of a major independent benchmark. That does not prove that infrastructure caused the improvement. AI performance depends on talent, data, training methods, model architecture, evaluation, post-training and many other factors.
But it does challenge a common assumption about where durable AI advantage resides. Research talent matters enormously. So do algorithms. Yet when large parts of the frontier converge on broadly similar architectures and compete for the same specialised labour, the ability to run more experiments, train at greater scale, absorb failures, provision capacity quickly and reduce marginal inference cost becomes increasingly important.
The model is produced by the factory around it.
Cursor extends the stack into applications
The Cursor acquisition adds the fifth layer. Cursor has developed from an AI-assisted code editor into an increasingly agentic software-development environment. Its acquisition gives SpaceX a significant application through which frontier models and compute can reach professional users.
Cursor itself makes the infrastructure rationale explicit. In announcing completion of the acquisition, it said access to SpaceX's GPU capacity would allow it to develop stronger models that are more economical to run. It described Grok 4.6 as an early example of what the companies could build together. SpaceX now has several routes into the application layer: Grok, X, Grok Bot and Cursor, alongside distribution through external platforms such as GitHub Copilot.
That creates a potential feedback loop: applications create demand → demand generates data and revenue → revenue funds infrastructure → infrastructure enables larger and cheaper model development → better models strengthen applications. Vertical integration becomes valuable when each layer reinforces the next. This is considerably harder to reproduce than a benchmark lead.
SpaceX is also trying to move the data centre off the planet
The most ambitious part of the strategy remains speculative. In January, SpaceX filed an application with the US Federal Communications Commission seeking permission for a non-geostationary system of up to one million satellites forming what it calls the SpaceX Orbital Data Center system. The FCC accepted the filing for consideration in February.
The proposal should not be described as one million operational data centres. It is an application for an enormous future constellation, and substantial engineering, economic, regulatory and environmental challenges remain. Reuters reports that analysts expect SpaceX's near-term AI economics to remain firmly terrestrial, with Colossus and Colossus II providing the more credible source of growth.
Even so, the filing reveals the direction of travel. SpaceX is asking whether advantages it already possesses in reusable launch, satellite manufacturing, orbital communications and solar exposure can eventually be translated into compute infrastructure. For another AI laboratory, orbital data centres would require an entirely new industrial capability. For SpaceX, they are an extension of capabilities it already possesses.
That is the strategic logic of vertical integration: an asset developed for one market becomes an advantage in another.
The physical stack may become the real moat
For the past three years, the AI industry's most visible source of differentiation has been the model. But model advantages are proving transient. A frontier release leads for several months. Competitors catch up. Techniques diffuse. Researchers move. Prices fall. Distillation, open models and improved training methods continually reduce the time during which any single capability remains unique.
Physical infrastructure behaves differently. Power plants take years to permit and construct. Advanced semiconductor fabrication requires enormous capital and scarce expertise. Hyperscale data centres require land, electricity, networking and equipment. GPU fleets represent tens of billions of dollars of deployed capital. Launch infrastructure, satellite manufacturing and orbital operations require capabilities that cannot be reproduced by hiring 20 researchers from a competitor.
The deeper SpaceX moves into these layers, the more its potential advantage shifts from intellectual property alone towards industrial capacity. That may prove more durable.
Vertical integration also concentrates risk
There is another side to the argument. Owning more layers means funding more layers. SpaceX is pursuing capital-intensive projects spanning power generation, semiconductor manufacturing, terrestrial data centres, orbital compute, frontier model development and applications. Each exposes the company to different regulatory, technological and execution risks.
The environmental controversy surrounding its Memphis infrastructure illustrates one problem. Semiconductor manufacturing introduces another level of technical difficulty entirely. Orbital compute remains highly uncertain. Frontier models themselves are economically volatile and face powerful competitors.
There is also a governance question. A company that simultaneously supplies infrastructure to competitors, builds competing models and controls downstream applications occupies a complicated position in the value chain. Questions around customer confidentiality, resource allocation, competitive neutrality and concentration of infrastructure power become increasingly important as those businesses grow.
Vertical integration removes external dependencies by creating internal ones. The architecture is powerful precisely because so much depends on the organisation being able to execute across every layer.
The market is still deciding what SpaceX has become
Investors have not treated the strategy as an uncomplicated success. SpaceX floated in June at $135 per share. Its stock subsequently fell below $105 before recovering sharply following its second-quarter results and the launch of Grok 4.6. Recent trading around $146 puts the shares back above the IPO price, but the volatility reflects a wider uncertainty over how the company should be valued: as a rocket company, a communications platform, an AI infrastructure provider or some combination of all three.
That uncertainty is reasonable. Terafab is not yet producing chips. Orbital compute is not yet operating. Grok has returned to the frontier but has not established sustained leadership. Cursor's $60bn price embeds considerable expectations about future application value. The investment thesis therefore remains much more ambitious than the current operating reality.
But the direction of the company is becoming difficult to miss.
The model may be the least important part of the strategy
The conventional AI race asks who will build the most capable model. SpaceX suggests a different question: Who controls the system required to keep producing increasingly capable intelligence?
Jensen Huang's five-layer cake is useful because it reveals why those questions are different. At the energy layer, SpaceX is expanding dedicated generation. At the chip layer, it is investing with Tesla and Intel in Terafab. At the infrastructure layer, Colossus has become large enough that competing frontier labs are paying billions to use it. At the model layer, Grok 4.6 has returned SpaceXAI to the frontier. At the application layer, SpaceX now owns Grok, X and Cursor while expanding into agentic products such as Grok Bot.
The important qualification is that this stack is still being assembled. SpaceX does not yet completely own it. But it may already have demonstrated the more important principle: the next phase of AI competition will not be won at the model layer alone.
If intelligence becomes an industrial product manufactured continuously from energy, chips and compute, then controlling the factory may prove as strategically important as designing what comes out of it. And that creates perhaps the most unusual part of the SpaceX thesis.
OpenAI can beat Grok. Anthropic can beat Grok. Google can beat Grok. SpaceX can still get paid.
