AI Value Moves to Land and Energy Infrastructure Layer
- Hurratul Maleka Taj
- 6 hours ago
- 4 min read
Value in AI is moving to the layer that owns the land and energy infrastructure
SpaceX's Q2 earnings, its first full quarter with xAI consolidated, look like a company losing the model race and spending to stay in it. The structure of the market points the other way. The short version is that capable models are multiplying while land and energy infrastructure is scarce, so value is migrating from the model at the top of the stack to the physical plant at the bottom.
The data
· AI capital expenditure rose 105% quarter over quarter, from 7.7 billion to 15.8 billion.
· Q2 AI spend alone exceeded the company's entire 2025 AI spend.
· Across 2026, AI is 82.7% of SpaceX capex, or 23.5 billion of 28.4 billion.
· xAI adjusted EBITDA moved from negative 609 million in Q1 to positive 1.1 billion in Q2.
· Grok adoption is stalling in consumer and enterprise, which frames the shift toward renting out infrastructure as a neocloud.
Figure 1: AI dominance over SpaceX Capex accelerates
The market structure
The AI stack has four layers. They are different businesses with different economics, and value is migrating from the top to the bottom.
Models. OpenAI, Anthropic, and xAI build the models. This was assumed to be where value concentrates. It is now the most exposed layer, because capable models are multiplying and converging, and a capability that many firms can supply loses pricing power. The model is commoditizing.
Silicon. Nvidia designs and sells the GPUs that run the models, with AMD and in-house accelerators a smaller share. This is not where operators differentiate. Every operator buys from the same short list on similar terms, so access to chips is a common input, not an edge.
Components. Vertiv, Eaton, and Schneider supply the power and thermal systems installed inside the facility. Demand here is the clearest read on where capital is flowing. Vertiv's backlog has more than doubled to above 15 billion, covering 12 to 18 months of forward revenue, and 2026 guidance implies organic growth near 30%. But these firms are suppliers. They sell equipment into a plant that someone else owns and runs.
Land and energy infrastructure. Someone has to secure the electricity, acquire the site, build the facility, and engineer power and cooling as one system so the chips run at high density without failing. This is now the hardest and scarcest layer, because power availability, not GPU supply or capital, is the binding constraint on buildout.
The logic follows directly. Because silicon is a common input available to every operator on equal terms, durable advantage cannot come from the chip, or from the model that runs on it. It moves to the inputs that are genuinely scarce and not equally available, which are low-cost energy, buildable land, and the integrated plant that ties them together.
The owner-operators
Most operators do one job and buy the rest. A merchant cloud rents a building, buys Nvidia chips, buys Vertiv components, buys grid power, and integrates them, paying a vendor margin on every input. SpaceX and Crusoe own the bottom of the stack instead of renting it. They secure the energy and engineer the physical plant directly, which removes the stacked vendor margins and sets a lower cost floor.
They are not identical, and the difference matters.
SpaceX buys its chips from Nvidia like everyone else and still sources some components from vendors. Its edge is integration and operation, not manufacturing. Michael Mealling, GP at Starbridge Venture Capital, notes that the hard problems of space hardware transfer directly to dense compute. Heat dispersal, power distribution, and fire management are shared constraints between launch systems and data centers. SpaceX engineers the building-level power and cooling as a single system, which is hard to match by assembling third-party parts.
Crusoe is the more fully integrated case today. It owns power generation and manufactures its own power and data modules in-house, running the full value chain from energy to compute. It has contracted 4.9 gigawatts against a pipeline above 40 gigawatts, anchored by Oracle and Microsoft. Crusoe reached this position by starting as an energy company. SpaceX is reaching it by extending hardware competency out of launch systems. Two paths, one category.
Thesis
In this cycle the binding constraint is shifting from model quality and GPU supply to ownership of land and energy infrastructure. Pricing power follows the constraint. Model developers and merchant GPU operators both become dependent on the owner-operator layer beneath them.
If that thesis holds, three specific claims follow, each testable against evidence and each carrying a condition that would disconfirm it.
Hypotheses
H1. The neocloud, not Grok, is the asset under construction. Grok is captive demand that underwrites the buildout. The Q2 EBITDA reversal reflects the infrastructure segment reaching positive contribution, not the model gaining share.
Falsification: if the swing traces to internal cost allocation or one-time timing rather than infrastructure monetization, H1 fails.
H2. Margin accrues to ownership of land and energy infrastructure, not to chips or models, which are common inputs. Owner-operators set the cost floor. Merchant operators renting GPUs on third-party power, such as CoreWeave and Lambda, face margin compression as capacity scales through 2027.
Falsification: if merchant and owner-operator margins converge through the capacity ramp, H2 fails.
H3. This is a category, not a single company. SpaceX and Crusoe define an owner-operator class distinct from model labs and merchant clouds, and external tenant revenue is its proof. Anthropic, a direct competitor, already pays 1.25 billion per month for compute on Colossus, on a contract either party can terminate on 90 days notice.
Falsification: if owner-operators cannot hold external tenants beyond a single anchor, the category reads as one customer, not a market, and H3 fails.
References
1. Jacob Robbins, "xAI's neocloud pivot eats up 82.7% of SpaceX capex," PitchBook News. https://pitchbook.com/news/articles/xais-neocloud-pivot-eats-up-82-7-of-spacex-capex
2. SpaceX, Form S-1 Registration Statement, SEC EDGAR (CIK 1181412), filed May 20, 2026. https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=1181412
3. "SpaceX skeptics have added reason for concern after Musk comments diverge from IPO filing," CNBC, May 29, 2026. https://www.cnbc.com/2026/05/29/spacex-skeptics-concerned-as-musk-comments-diverge-from-ipo-filing.html
4. "Crusoe's Contracted AI Infrastructure Capacity Approaches 5 Gigawatts Across Data Centers and Cloud," Crusoe newsroom, June 9, 2026. https://www.crusoe.ai/resources/newsroom/crusoes-contracted-ai-infrastructure-capacity-approaches-5-gigawatts-across-data-centers-and-cloud
5. "Welcome to the era of BYO power," Crusoe blog, March 2026. https://www.crusoe.ai/resources/blog/welcome-to-the-era-of-byo-power
6. Vertiv Holdings Co, "Vertiv Reports Strong First Quarter with Diluted EPS Growth of 136%; Raises Full-Year Guidance," April 22, 2026. https://investors.vertiv.com/news/news-details/2026/Vertiv-Reports-Strong-First-Quarter-with-Diluted-EPS-Growth-of-136-Adjusted-Diluted-EPS-Growth-of-83-Raises-Full-Year-Guidance/default.aspx
Comments