Exaflops and Empty Pockets: The AI Grid Without a Load
OpenAI’s $500 billion Stargate isn’t just a compute cluster. It’s a leveraged infrastructure play, and the cash flows haven’t arrived
We’re witnessing the most capital-intensive tech buildout since the electrical grid: hyperscale GPU clusters, sovereign-aligned data centers, multibillion-dollar power commitments. At the center is Stargate, a $500 billion facility backed by OpenAI, Oracle, and SoftBank. But the fundamental question is not technical. It’s financial. Who pays for all this? And what if the demand never comes?
I. Two Capital Stacks, Two Sets of Risks
Stargate isn’t one financial entity. It’s two interlocking capital profiles, and OpenAI sits in both.
OpenAI-as-Operator
OpenAI will run the facility. That means it must fund everything that makes Stargate operable:
GPU racks, interconnects, liquid cooling
Power distribution, load management
Software orchestration layers, security infrastructure
Think of this like a triple-net commercial lease: the landlord provides the shell, and the tenant modifies it at great expense. These are leasehold improvements, but at sovereign scale: likely tens of billions in tenant-side capex, front-loaded and non-recoverable.
OpenAI-as-Owner
OpenAI is not just a tenant. It is also a co-owner of the Stargate infrastructure vehicle, alongside Oracle, SoftBank, and others. That means it has equity exposure to the total $500 billion investment:
Returns depend on Stargate’s utilization and monetization
Losses are shared if demand falters or capacity is stranded
This dual role creates a circular cash flow loop:
OpenAI-the-operator pays for infrastructure use
Those payments help generate yield for OpenAI-the-owner
Which means OpenAI is exposed at both layers of the stack:
If usage is high and monetized, it benefits as both user and equity partner.
If usage is low or pricing collapses, it takes a hit on both its income statement and its balance sheet.
One possibility to consider is that OpenAI can control demand via model access, gating usage to ensure Stargate utilization. But this assumes unchallenged pricing power and regulatory latitude. Should OpenAI face antitrust constraints or developer churn, that control becomes brittle.
II. Historical Analogues: When Infrastructure Came First
We've seen this pattern before: infrastructure built ahead of proven demand:
🔦 Dark Fiber (1999–2001)
Billions were spent on optical fiber anticipating internet explosion. Demand eventually came, but it came too late for many investors. Assets were stranded for years, repurposed later at steep discounts. The glut enabled future broadband, but only after massive capital destruction.
🔥 Liquid Natural Gas (LNG) Export Terminals (2005–2020)
Massive capex spent on liquefying and shipping natural gas. These facilities only worked because of sovereign offtake agreements: 20-year, take-or-pay contracts that made demand bankable.
LNG became a hedgeable commodity because flows were predictable, contract-backed, and priced into futures.
AI lacks those offtake structures. No one is guaranteeing demand at sovereign scale. The infrastructure is being built on speculation, not contractual commitment.
☁️ AWS Data Center Expansion (2006–2016)
Amazon built ahead of demand, but embedded compute into a seamless developer workflow. Monetization was immediate, granular, sticky. Crucially, AWS built both the infrastructure and the developer funnel.
OpenAI has the former. The latter is still murky.
Critics might say this is simply how Big Tech builds. But Stargate differs in capital structure. AWS had internal cash flows and granular developer monetization. Stargate is co-financed, externally exposed, and developer-indirect. It has no real equivalent to AWS's sticky usage loop.
III. The Real Risk: Not a Lack of AI, but a Lack of Cash Flow
OpenAI’s operator costs are high. Its obligations to the owners will be long-term and inflexible.
So what’s the plan to fund it all?
It requires:
Microsoft to keep writing checks for Copilot, Azure integration, and downstream resale
Enterprises to deploy copilots and agents at scale, not just in pilot
Governments to underwrite sovereign-grade demand — as customers or regulators
Consumers and developers to keep generating token volume at scale
That revenue must then:
Cover OpenAI’s leasehold capex and operational overhead
Satisfy the partnership’s IRR expectations
De-risk future expansion phases
If any link in that chain fails, the system becomes overbuilt, under-monetized, and capital-negative.
AI enthusiasts will counter that demand is already exploding: usage is real, not speculative. But the distinction here is critical: real usage is not yet structured into forecastable, contractual, or hedgeable flows. Present activity does not equal anchor-grade demand.
IV. A Capital Stack Without Anchors Becomes Leverage
The viability of Stargate doesn’t just depend on growing demand. It depends on structuring that demand into contractual, forecastable, and hedgeable flows. Otherwise, the capital stack collapses under its own weight.
Some early signals:
GPU leasing platforms (CoreWeave, Lambda) are testing spot and term markets
Futures exchanges (e.g., DRW’s compute.exchange)1 are prototyping compute hedging
Long-term contracts between hyperscalers and compute providers are starting to form
But these are embryonic. There is no robust forward curve for compute. No standardized unit of pricing. No take-or-pay contracts.
Some might argue that model progress itself creates demand. AI, in other words, is an endogenous flywheel. But growth does not eliminate financial structure. Infrastructure is not funded by potential; it's funded by cash flow.
Until compute becomes priceable like power or LNG, in which it is modeled, collateralized, and hedged, every new megawatt is a speculative long position.
Stargate is not a SaaS company. It’s a leveraged utility.
V. Sovereign Leverage: Infrastructure Subsidies Come With Strings
The Trump administration has made clear it wants to accelerate AI infrastructure through:
Fast-tracked permitting
Grid interconnection priority
Energy allocation guarantees
Public-private infrastructure compacts
That sounds like a subsidy. But it’s a bargain:
In return for speed, the Trump administration will likely demand:
Onshore supply chains (e.g., U.S.-based fabs)
Strategic carveouts for defense and intelligence
Export geofencing, especially toward China
Data governance compliance and auditability
This is not free-market infrastructure. It’s sovereign-industrial accommodation.
Stargate becomes:
A national asset
A defense-adjacent utility
A compliance-heavy facility vulnerable to regime change and regulatory shocks
Structured like a utility. Financed like a startup. Governed like a defense contractor.
It could be argued that sovereign backing reduces risk, but this cuts both ways. When infrastructure is embedded in national strategy, it becomes subject to non-market volatility: elections, geopolitical reprioritizations, export control shocks. Anchors become liabilities.
VI. Final Word: The Infrastructure Is Real. But Is the Load?
Stargate is not just a compute cluster. It’s a financial instrument.
A multi-tiered structure with OpenAI exposed on every axis:
As operator and tenant, it must fund and maintain the functional core by generating sufficient demand for its inference output
As owner, it depends on its own and others’ payments to hit IRR thresholds
If demand falls, OpenAI eats the loss from both ends: revenue shortfall and equity impairment.
Add in:
Monetization opacity
Customer concentration (chiefly Microsoft)
Regulatory fragility
Political entanglements from federal subsidies
And Stargate becomes:
Structured like a utility
Financed like a startup
Governed like a defense contractor
Exposed like a speculative REIT
An easy response here is that Microsoft will backstop all this. Microsoft is, after all, a strategic investor with deep pockets. But dependency is not a hedge; it is a concentration risk. If Microsoft reprioritizes or cuts support, the recursive structure implodes. And the WSJ recently reported that Satya Nadella and Sam Altman are “drifting apart”.
It might work. It might change the world.
But if Microsoft defects, if enterprises stall, if sovereigns balk, then Stargate won’t just be a stranded asset.
It will be a $500 billion monument to recursive exposure.
Because in the end, this isn’t just a question of whether AI scales.
It’s a question of whether anyone can afford to build and sustain the grid.
VII. Addendum: What Would Make the Grid Bankable?
If this infrastructure is to be viable, demand must harden. That means:
AI Capacity Purchase Agreements (AICPAs): Long-term, fixed-price compute contracts
Developer-native monetization channels, like AWS circa 2010
Sovereign guarantees modeled on CHIPS Act incentives
Pricing standardization and commodity-style benchmarking for compute
Only then can this cease to be a speculative capital burn and become a genuine infrastructure investment class.
Until then, Stargate is the world’s most expensive bet that software will catch up to hardware in time to pay for the grid that both needs.
To be more clear, DRW’s venture capital arm invested capital in compute.exchange.
Great laterals. In LNG they want decades of visibility before the first shovel hits the ground.