--- title: AI Data Center Economics in Gigawatts slug: ai-data-center-economics-in-gigawatts date: "2026-07-04" summary: A practical way to compare CoreWeave, SpaceX, and broader AI infrastructure is to normalize the economics by gigawatt: revenue, capex, deployment time, and contract quality. categories: "[[Writing]]" tags: - finance - ai - infrastructure featured: false published: false --- One useful way to analyze AI data center economics is to normalize everything to a gigawatt. The broader capex debate is usually framed in aggregate terms. Can the industry spend more than a trillion dollars a year on AI infrastructure? Can inference revenue grow fast enough to support that investment? Are these data centers durable productive assets, or is the market overbuilding ahead of demand? Those questions matter, but they are difficult to answer at the headline level. A dollar of capex spent against a signed five-year customer contract is different from a dollar spent speculatively. A cluster that can be energized in a few months is different from one waiting years for power. A gigawatt earning $12 billion a year is not economically equivalent to a gigawatt earning $50 billion a year. So the cleaner unit of analysis is: - revenue per gigawatt; - capex per gigawatt; - operating cost per gigawatt; - time from capital deployment to revenue; - contract duration and customer quality. This framing does not settle the AI infrastructure debate by itself. It gives us a better way to compare the cases. CoreWeave provides a grounded baseline for contracted GPU cloud economics. The BG2 SpaceX discussion provides a more aggressive version of the same model, with higher monetization assumptions and a possible long-term cost advantage from orbital compute. ## CoreWeave as the baseline The CoreWeave article is useful because it separates the data center stack into layers. In its 100MW model, the co-location provider funds the physical data center: shell, electrical equipment, cooling, and related infrastructure. CoreWeave leases that capacity, signs a long-term cloud customer, buys the GPUs, and finances most of the GPU capex with debt. The sequencing matters. The attractive version of the business is not "buy GPUs and hope demand arrives." It is closer to: secure capacity, sign a customer, buy the GPUs, draw debt, then service the contract. The financing is more defensible because it is tied to contracted cash flows and because the debt amortizes within the customer contract period. The model's main assumptions look like this: | Item | 100MW model | Per gigawatt | | --- | ---: | ---: | | Annual customer revenue | $1.2B | $12B per GW per year | | Five-year total contract value | $6B | $60B per GW | | GPU capex | $2.7B | $27B per GW | | Physical data center capex, funded by colo provider | $1.0B-$1.3B | $10B-$13B per GW | | Year-one rent | $165M | $1.65B per GW per year | | Tenant fit-out | $75M | $0.75B per GW | | Electricity, using the article assumptions | about $78M per year | about $0.78B per GW per year | | Contract return | about 20% unlevered IRR | same project return | This is a reasonable base case for thinking about the neocloud model. At $12 billion of annual revenue per gigawatt, a $27 billion GPU capex bill can work if utilization is high, the contract is long enough, power and rent are known, and the customer pays. The article's key conclusion is that this structure can produce roughly a 20% unlevered IRR without assuming residual GPU value after the initial contract expires. That is important because residual value is one of the most uncertain parts of the model. If older GPUs can be re-leased, sold, or used for lower-tier workloads, that is upside rather than a requirement for the base case. This also explains why the debt question is more nuanced than "leverage is bad." Leverage is dangerous when the asset and liability do not match. But if the GPUs are purchased against a signed customer contract, and the debt amortizes inside that contract, the structure looks more like asset-backed project finance than speculative borrowing. The risk is still real. It just sits in specific underwriting questions: customer credit, delivery timing, renewal risk, useful life of old GPUs, and future pricing. If those assumptions weaken, the same leverage that helps CoreWeave scale also makes the downside sharper. ## What IRR does and does not tell us IRR is useful, but it is easy to misuse. A project IRR is not the same as the investor's realized return. It is the discount rate that makes the project's cash flows work. If a project returns cash early, the investor still has to find another place to put that cash. A 20% project IRR only compounds into something like a 20% investor return if similar opportunities are available repeatedly. That distinction matters here because the AI infrastructure thesis depends on repetition. A single 20% IRR cluster is interesting. A repeatable pipeline of 20% IRR clusters is a business. A market where many operators can repeatedly source power, chips, customers, financing, and execution at attractive spreads is the broader infrastructure thesis. CoreWeave's model is therefore not just about whether one cluster works. It is about whether the company can keep finding contracted demand and financing GPU purchases against that demand. The 75% debt-funded assumption is powerful because each dollar of equity can support roughly four dollars of GPU assets. But that only remains attractive if the contracts, delivery schedule, and refinancing environment remain supportive. ## The SpaceX/xAI upside case The BG2 SpaceX discussion uses a similar per-gigawatt framing, but with more aggressive assumptions. The main claim is that SpaceX or xAI may be able to monetize AI compute at a higher rate per gigawatt than the CoreWeave-style baseline. The podcast discusses implied monetization of roughly $14 billion per gigawatt per year for the AI business, with claimed deal economics around $22 billion to $23 billion for Anthropic and around $50 billion for Google. Put next to the CoreWeave baseline, the spread is significant: | Case | Revenue per gigawatt per year | | --- | ---: | | CoreWeave modeled contract | about $12B | | SpaceX implied AI business estimate | about $14B | | Claimed Anthropic deal level | about $22B-$23B | | Claimed Google deal level | about $50B | This helps explain the 55% IRR reference in the podcast. If the capex base is broadly comparable, moving from $12 billion of annual revenue per gigawatt to $22 billion, $23 billion, or $50 billion changes the project economics dramatically. The other important variable is speed. The podcast emphasizes how quickly Elon-led teams can bring large clusters online. That matters because a delayed cluster is not just an operational inconvenience. It is a financial cost. At $12 billion per gigawatt per year, one month of delay is roughly $1 billion of annualized revenue capacity not yet producing. At $50 billion per gigawatt per year, the monthly revenue opportunity is more than $4 billion on an annualized basis. The higher the monetization rate, the more valuable fast deployment becomes. This is the central difference between the base case and the upside case. CoreWeave shows that contracted GPU cloud economics can work at roughly $12 billion per gigawatt per year. The SpaceX/xAI version asks what happens if a company can both deploy faster and charge materially more for scarce capacity. ## Orbital compute as an option The most speculative part of the BG2 discussion is orbital compute. The terrestrial data center cost stack has two major pieces. The first is the silicon: GPUs, networking, and related equipment. The second is everything around the silicon: land, shell, power, cooling, generators, transformers, switchgear, and interconnection. BG2 frames terrestrial all-in cost at around $60 billion per gigawatt, with roughly $35 billion for silicon and $20 billion to $25 billion for the surrounding infrastructure. The orbital compute argument is that the non-silicon bucket could shrink if Starship launch costs fall far enough and if power, cooling, and space constraints are structurally different in orbit. The podcast suggests a rough orbital non-silicon cost of about $5 billion per gigawatt and an all-in cost closer to $30 billion per gigawatt. Those numbers should be treated as directional rather than fully underwritten. The unresolved questions are substantial. GPUs fail. Satellites fail. Radiation, maintenance, thermal management, networking, latency, and replacement cycles all matter. The economic advantage only matters if the reliability-adjusted cost is actually lower and if customers value the orbital capacity similarly to terrestrial capacity. So orbital compute should not be treated as the current base case. It is better understood as an option on the cost curve. If revenue per gigawatt rises while non-silicon deployment cost falls, the economics can improve from both sides. But the reliability and maintenance assumptions are doing a lot of work. ## Why the aggregate capex debate is hard This is why the $1.5 trillion capex versus $300 billion inference revenue debate is difficult to answer directly. At a high level, the comparison sounds concerning: a very large capex number against a much smaller current revenue number. But the comparison mixes different categories: - training capex versus inference capex; - capex incurred today versus revenue earned over several years; - infrastructure provider revenue versus AI lab product revenue; - contracted capacity versus speculative supply; - high-utilization clusters versus delayed or underutilized projects; - baseline pricing versus scarcity pricing. The better approach is to underwrite the unit. For each gigawatt, the relevant questions are: What does it cost? What does it earn? How long is the contract? How much is debt-financed? Does the debt amortize within the contract? Who is the customer? What happens after the initial term? How long before the asset is energized? If the market is full of $12 billion per gigawatt assets with $27 billion of GPU capex, long-term contracts, controlled rent and power costs, and three-year cash paybacks, it can support a lot of investment. If the market shifts toward speculative builds, weak customers, delayed power, falling utilization, and short contracts, the same headline capex number becomes much more fragile. The size of the capex program matters. But the quality of the spread matters more. ## The underwriting checklist The AI data center business sounds new, but the underwriting checklist is familiar. Customer quality comes first. A five-year contract is only as good as the counterparty. Investment-grade hyperscalers and cash-burning AI labs are different credit risks, even when both are strategically important. Duration matching is next. If the debt amortizes within the customer contract, the financing is more defensible. If the data center lease runs for 10 to 15 years and the customer contract runs for five, renewal risk becomes a major liability. Deployment execution matters because delay directly affects returns. The higher the revenue per gigawatt, the more expensive each month of delay becomes. Chip access is also central. These businesses are not simply buying commodity servers. Early access to Nvidia generations, networking, discounts, and integration support can determine who wins the customer contract. Power and land remain binding constraints. For terrestrial data centers, the bottleneck is not only GPUs. Grid access, transformers, generators, cooling, permitting, and physical location all shape the return profile. Residual life is a source of potential upside. The CoreWeave model does not require residual GPU value to reach its return target. But if old GPUs can be re-leased at attractive rates, lifetime economics improve. If they cannot, the first contract carries more of the burden. Finally, pricing power is the largest open question. If compute pricing declines smoothly as supply enters the market, returns compress. If demand continues to outrun supply, operators with chips, power, customers, and execution speed can earn scarcity returns for longer. ## Conclusion The AI infrastructure debate becomes clearer when analyzed as a set of gigawatt-level projects rather than as one giant capex number. CoreWeave gives us a grounded model: roughly $12 billion of annual revenue per gigawatt, $27 billion of GPU capex per gigawatt, long-term contracts, matched debt, and an estimated 20% unlevered project IRR. The SpaceX/xAI discussion describes a more aggressive case: faster deployment, higher monetization per gigawatt, and a possible long-term reduction in non-silicon costs through orbital compute. The question is not simply whether AI capex is too high. The better question is how many gigawatts can earn enough revenue, quickly enough, under contracts strong enough to support the capital being deployed. That is where the debate should sit: not in the headline capex number alone, but in the spread between what each gigawatt costs and what each gigawatt can reliably earn. ## Source notes Built from [[26-07-03 substack.com CoreWeave Unit Economics, Margin Profile, and Leverage Analysis]] and [[26-07-04 youtube.com The SpaceX IPO, Fable 5, AI Capex Update & Market Check w Gavin Baker, Andrew Fox & Clark Tang]].