Kai ChenRoughly half of Meta’s $62 billion in debt since 2022 arrived in 2025. Yet even that financing cannot buy a shortcut to the first token.
Silicon can be ordered; capacity has to be built
The chip race is not over. Nvidia has reportedly asked TSMC to ramp production of H200 accelerators and may hold allocations or inventory for 700,000 H200 chips in 2026. Silicon remains valuable and contested.
But securing accelerators no longer clears the entire deployment stack. Meta is enlarging its Hyperion compute campus in Richland Parish, Louisiana, to multi-gigawatt scale. Google has contracted for the entire initial output of the 1.6GW Steel River Energy Center solar project. xAI, meanwhile, faces an unconfirmed allegation that it installed more gas turbines for Colossus 2 than it publicly disclosed and did so without permits.
These are different companies pursuing different models in different places. Yet all three are making decisions about generation, construction and local authority before compute becomes usable. No one coordinated this. The load did.
The balance sheet is becoming construction equipment
Since 2022, Meta has raised $62 billion of debt, with roughly half raised in 2025. It also moved $30 billion of debt for AI data-center construction off its balance sheet through special-purpose vehicles. Hyperion is reportedly set to receive an almost $30 billion financing package, with Meta retaining a 20% stake.
xAI has made the same structural move at a different scale, shifting $20 billion of debt for AI data centers off its balance sheet through SPVs. The common mechanism matters more than the accounting presentation. Large compute campuses require enough capital, over enough time, that financing design becomes part of the infrastructure itself.
SPVs do not make the concrete free. They make the obligation easier to place.
This is the transition from procurement cycle to construction cycle. A chip purchase is an equipment decision. A multi-gigawatt campus is a portfolio of long-lived commitments involving power supply, buildings, financing and permission to operate. Once operators converge on project-finance structures, the competitive unit is no longer merely the accelerator. It is the completed system around it.
A 2029 power contract is a present-tense capacity move
Google’s agreement for Steel River Energy Center does not provide immediate electricity. The 1.6GW solar project is not due to operate until 2029. That limitation is not incidental; it identifies the constraint.
If power available years from now must be contracted in full today, energy procurement has moved ahead of model deployment in the decision chain. Google is not buying current compute capacity with the agreement. It is securing one input required for future capacity before that input exists.
Meta’s target carries a similar caveat. A compute target does not establish that Hyperion will be completed on schedule or operate above 5GW. But the target still reveals what must be assembled for the campus to become real. At that scale, a model roadmap depends on a construction roadmap.
A chip is inventory. A powered, financed and permitted campus is a position.
Permitting now allocates compute
The Colossus 2 turbine claim must be treated carefully. It is reporting analysis, not a confirmed regulatory finding. It is evidence of pressure around the permitting boundary, not proof that regulators have established a violation.
Even with that distinction, the allegation belongs in the same frame as Meta’s campus expansion and Google’s long-dated solar contract. If desired compute capacity requires on-site generation, then turbine installation and public authorization become part of the deployment path. Permitting is no longer an administrative epilogue after the technical work. It determines whether installed infrastructure can become operating capacity.
This changes what counts as a leading indicator. Accelerator orders describe potential compute. Financing structures describe whether the site can be funded. Energy contracts describe whether it can be supplied. Permits describe whether it can legally run. Companies announce models at the end of that chain; builders commit capital and land much earlier.
The scarce unit is a permitted megawatt
Meta, Google and xAI are absorbing the same constraint through three different systems: Meta is engineering around capital intensity, Google around future power supply, and xAI around the boundary between deployment speed and local approval.
Once usable electricity becomes harder to assemble than purchasable silicon, competition moves upstream. The unit of AI capacity is now a megawatt that is financed, supplied and allowed to operate.
Meta’s $62 billion matters less as a debt total than as a marker of how far the race now reaches before the first token.