Google Turns to Third-Party Compute as Power Constraints Slow its Own Data Center Build
Renting external capacity signals AI demand is running well ahead of what hyperscalers can build themselves, with power access now setting the pace.
Google's most recent earnings call was dominated almost entirely by discussion of Cloud infrastructure and TPU capacity — analysts submitted no questions on the company's core advertising business, Bloomberg Intelligence noted in a recent podcast segment.7
The episode illustrates a concrete supply problem. Demand for AI compute is outrunning Google's ability to deploy infrastructure at the required pace, pushing the company to rent capacity from third-party operators. Power availability, grid interconnection timelines, and equipment procurement delays have emerged as the binding constraints on AI data center deployment across the US and globally, according to a July 16 (2026-07-16) report from DataM Intelligence. Chips can be ordered. Substation lead times and utility interconnection queue positions cannot be compressed on demand.6,7
Google has been aggressive on its own build. The company has contracted more than 22 GW of clean energy since 2010, struck a hydropower framework with Brookfield covering up to 3 GW nationally, and commissioned a 1-GW-plus co-located data center and generation complex in the Texas Panhandle, Power Magazine reported on June 4 (2026-06-04). It also holds 1 GW of demand response capacity under long-term contracts with the Tennessee Valley Authority, Entergy Arkansas, DTE Energy, and others, enabling it to curtail or shift workloads during peak periods.3
That build program has not kept pace with demand. PJM's first reformed interconnection queue cycle drew 811 projects totaling 220 GW of proposed capacity, Power Magazine noted, signalling how congested the race for grid access has become across the broader market.3
Third-party operators are absorbing the overflow. CoreWeave reported a record revenue backlog of $99.4 billion in the first quarter of 2026, with quarterly revenue of $2.08 billion — up 112% year-over-year, oilprice.com reported in July (2026-07-02). That backlog represents forward commitments from companies that need compute capacity they cannot yet build themselves, and shows how much of the hyperscaler infrastructure ramp is being routed through external operators.5
The energy numbers attached to this buildout are large and accelerating. Data centers already account for more than 1% of global electricity use, according to the IEA. Between 10% and 20% of US data center power is currently consumed by AI workloads, with that share expected to rise significantly, TIME reported in May (2026-05-19). McKinsey projects AI-linked data center electricity demand reaching nearly 12% of total US power consumption by the end of the decade, oilprice.com noted.1,5
For gas-fired generation, the scale is legible. A data center complex in the 4-6 GW range would consume roughly 1 billion cubic feet per day of natural gas depending on turbine efficiency, Atlantic Council analysis published June 1 (2026-06-01) estimated. NYMEX Henry Hub front-month closed Friday (2026-07-24) at $2.87/MMBtu. At that price, gas-fired power remains competitive as a data center generation source, but fuel cost is not the binding variable here — grid access is.2
Some independent operators have already reorganised around that constraint. Bitzero CEO Mohammed Bakhashwain told oilprice.com in June (2026-06-18) that the company secures power access, grid positioning, and pricing frameworks before building infrastructure, the reverse of the industry's long-standing norm of constructing first and fighting for interconnection later. The company's forward deal pipeline reached up to $2.6 billion, with as much as 85% expected to convert to net income.4
Google's move into rented external compute is an operational bridge: the company routes workloads through operators that already hold live grid connections while its own builds catch up. The risk for power markets is that AI compute demand keeps accelerating while the interconnection queue — 220 GW in PJM alone — clears too slowly to absorb it, leaving hyperscalers and independent operators competing for a fixed and slowly-growing stock of energised capacity.3,6