AI data centers are buying the power stack before the grid can catch up
Anthropic, SoftBank and Vertiv point to the same infrastructure shift from different layers of the stack: AI capacity now depends on reserving generation equipment, financing it through volatile project plans, and securing the semiconductors that turn power into usable compute.
By Seth Stint · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not hold a real degree or possess firsthand experience.
Key points
- Anthropic’s generator agreement shows a frontier AI company moving beyond cloud contracts toward direct control of power-generation equipment, with the equipment expected to become available within the stated delivery window.
Sources: S1
- SoftBank’s reported losses on changes to U.S. equipment procurement plans show that reserving scarce power hardware can create a financial exposure well before a data center is operating.
Sources: S2
- Vertiv and Infineon’s supply agreement shifts attention downstream: power-generation capacity alone is insufficient without the semiconductor components needed for higher-voltage data-center distribution systems.
Sources: S3
The infrastructure constraint is becoming a procurement problem
The latest AI infrastructure race is often described in terms of chips, model capability and cloud leases. The more immediate practical contest may be for the equipment between a utility connection and a server rack. Anthropic has agreed to take over Caterpillar generator orders through OnePWR Solutions in a deal reported at $3.4 billion and tied to 1.7GW of power. The equipment had been ordered earlier by OnePWR, which needed financing to complete the transaction. OnePWR is expected to assist with engineering, construction and installation at Anthropic data centers, while the generators are expected to be available within the next two years.
Sources: S1
This is not simply another cloud-capacity commitment. The supplied report says Anthropic has cloud and data-center commitments through 2045 and has become more involved in facility construction and power deals. Its generator purchase therefore sits alongside, rather than replaces, leased-compute arrangements with providers including Google, Amazon, Microsoft, Nscale, Lambda and others. Several of those arrangements carry execution dependencies: Nscale must seek financing for Nvidia GPUs, while Lambda must secure its facility lease on time. Compute contracts can reserve demand, but they do not independently deliver an energized facility.
Sources: S1
SoftBank’s disclosures supply the clearest warning about the cost of treating equipment reservation as a solution in itself. It said it made advance investments in U.S. power-generation equipment because it expected shortages and long procurement lead times, then reported a loss tied to changes in procurement plans. Cumulative payments for generation equipment for its U.S. AI data-center projects had reached $9.48 billion at the end of September. The company also highlighted a long-term OpenAI lease for IT capacity at its Ohio project. That combination illustrates how commitments at the application, data-center and equipment layers can overlap while remaining exposed to changes in project design, timing or demand.
Sources: S2
Owning generators does not complete the electrical architecture
The comparison becomes more useful at the rack and distribution layer. Vertiv and Infineon have signed a three-year agreement to develop and scale supplies of silicon and silicon-carbide power semiconductors. Vertiv says the components are important to its 800VDC power systems and that the agreement is meant to support supply continuity as it expands production for digital infrastructure. This is a different kind of reservation from Anthropic’s generator order: it concerns components that condition, convert and distribute power after generation or grid supply is available.
Sources: S3
The practical dependency is concrete. A data-center operator may secure gas generators, utility service or both, but still need switchgear, conversion equipment, power distribution and thermal systems that can support the intended IT architecture. Vertiv’s announced component arrangement does not establish that a particular AI facility will receive 800VDC power, nor does it establish efficiency or performance gains in an operating deployment. It does show that a major infrastructure vendor is trying to secure access to a component category it considers necessary for the next generation of power architecture. Infineon’s existing work with other power-infrastructure suppliers further indicates that direct-current systems are being pursued across the supply chain.
Sources: S3
The supplied evidence therefore separates a launch claim from what has been measured. Anthropic’s agreement reports equipment scope and expected availability, not delivered data-center power or operating utilization. SoftBank reports payments and an accounting impact from procurement-plan changes, not a completed facility’s reliability or cost. Vertiv and Infineon describe a supply and development partnership, not field data from an 800VDC deployment. These are consequential commercial signals, but none of the reports provides a like-for-like operational benchmark for generator-backed AI capacity, grid-connected capacity, or a new distribution architecture.
What builders can use—and what they should not assume
For builders, the usable lesson is sequencing rather than imitation. Power procurement now has to be designed as a portfolio of linked dependencies: generation equipment, financing, construction delivery, power electronics, facility leases, cooling, grid arrangements and IT demand. Anthropic’s transaction suggests equipment ownership or control can reduce reliance on an intermediary that cannot close financing. SoftBank shows that early reservation can also turn into a material balance-sheet risk when procurement plans change. Vertiv’s agreement points to a third exposure: even a funded project may compete for specialized electrical components needed to deploy its chosen power system.
Inference: these developments indicate that AI infrastructure is becoming a power-system integration challenge rather than a straightforward expansion of server procurement. The evidence supports this inference because the reported actions span generation equipment, a data-center project’s advance payments, and semiconductors for distribution equipment. It does not prove that gas generation is the lowest-cost, lowest-emissions, or most reliable approach for every site. Nor does it show whether the cited hardware is intended for continuous prime power, backup use, or a mixed operating model. Those distinctions matter for fuel supply, permitting, emissions obligations, maintenance and the economics of an operating campus.
What could change the assessment is specific operational evidence. For Anthropic, useful disclosures would include commissioning progress, the role of the generators in the facility design, and whether the stated equipment becomes available as expected. For SoftBank, changes in equipment payments, revised procurement plans and progress at the Ohio project would indicate whether advance buying is turning into deployable capacity or further financial friction. For Vertiv and Infineon, shipment volumes, qualified 800VDC products, deployment conditions and measured results would reveal whether component supply is translating into scalable systems. Until then, the strongest conclusion is modest: securing power equipment is increasingly a strategic capability, but it is not the same as proving usable AI compute at scale.
The decision point is optionality, not just capacity
The cross-source picture is not that every AI builder should purchase generators or redesign around direct current. The reports concern different companies and different points in the value chain. Their common implication is narrower: project credibility increasingly depends on whether physical-power dependencies have been contracted, financed and integrated—not solely whether a customer has announced demand for AI compute. Builders should test each proposed capacity expansion against those dependencies before treating a lease, hardware order or supply agreement as available capacity.
That framing also clarifies the system effect. As major buyers reserve equipment and suppliers reserve critical semiconductors, availability may become a competitive variable alongside land, chips and financing. The disclosed actions show companies attempting to reduce that exposure through earlier commitments. But earlier commitments transfer risk; they do not remove it. The relevant watchpoint is whether those commitments convert into commissioned, supportable infrastructure under the operating conditions that customers actually need.
Why it matters
The strategic unit of AI infrastructure is shifting from the GPU order to the integrated power path. The evidence shows active attempts to lock down equipment at generation and power-electronics layers, while also showing that early procurement can create material execution and financial risk. For builders, announced capacity should be treated as conditional until the linked power, component, financing and construction dependencies are demonstrably in place.
Sources
- Anthropic signs $3.4bn gas generator deal to support 1.7GW of data center power — Data Center Dynamics ·
- SoftBank swings to $7 billion loss after U.S. data center hit and Intel share decline — CNBC Technology ·
- Vertiv partners with Infineon to scale power semiconductors for AI data centers — Data Center Dynamics ·