Build1 publisher2 min readPublished
Crux AI's reported $22bn TPU facility pledges the chips and the customer contracts as security
The reported collateral is Google's proprietary processors plus contracts with customers Crux has not named. A lender's recovery case therefore rests on someone else wanting TPU capacity outside Google Cloud.
The Engineer · Build desk
What happened
- Bloomberg reported on September 16 that 10 banks are providing financing for Crux AI, secured by the chips themselves and by Crux AI's customer contracts.
- 9fin had earlier described roughly $23 billion of debt for the venture and said the facility would most likely be structured as a bridge loan.
- Google and Blackstone disclosed an initial $5 billion equity commitment and a target of bringing 500 MW of TPU capacity online in 2027.
- Bloomberg Law reported on September 9 that the venture had hit delays at major planned data-center locations while keeping its 2027 target for renting Google processors to customers.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- exposure Customers who sign with Crux become part of the collateral package, because their contracts are pledged to the banks alongside the accelerators.
- constraint Growth past the 500 MW already disclosed depends on lenders accepting used accelerators and contract quality as security, so the pace of TPU capacity sold outside Google Cloud gets set in credit committees.
- decision A buyer comparing Crux to a Google Cloud commitment has to price the counterparty's leverage as well as the silicon, since the roadmap sits on debt at 4.4 times the disclosed equity.
- contradiction The reported debt figures differ and the structure is described as a bridge. Whether this is permanent capital or a placeholder that has to be refinanced before the capacity ships is unresolved.
The security described in Bloomberg's report is two assets with different failure modes. One is hardware: Google's proprietary processors, which Google agreed to supply to the venture under the May deal [6]. The other is Crux's book of customer contracts [1]. Crux exists to sell those processors as a compute service outside the standard Google Cloud platform [7]. A lender that enforces against the hardware ends up holding TPUs whose re-lease market is the same narrow set of buyers who wanted TPU capacity outside Google Cloud in the first place [21]. The runtimewire account frames the test the same way: chip value, customer contracts and re-leasing risk become the questions for lenders funding an eventual multiple-gigawatt expansion [15].
The contract half cannot be priced from outside the deal. Crux launched under its current name in September with materials describing an integrated service across power, data centers, TPUs, networking, orchestration software and operations. Those materials did not disclose customers, pricing or facility locations, and the accessible reports leave the banks unnamed and the facility's status open: closed, committed or still contemplated [10][4]. The two reported debt figures sit about a billion dollars apart, a rounding difference at this size [20]. A bridge gets repaid by a refinancing, so the same contracts get read a second time by whoever takes it out.
Run the reported facility against the capacity that has actually been disclosed and it comes to $44 million per megawatt for the 500 MW targeted for 2027 [18]. Against the 2 GW that Data Center Dynamics reported Crux plans to reach, the same money is $11 million per megawatt [8][19]. Neither number is a build cost. Bloomberg described the facility as funding TPU purchases, not power, land or construction [1].
The reported facility is 4.4 times Blackstone's initial equity commitment [9]. On those terms Crux would be an unusually debt-intensive infrastructure build from an early stage. The organisation is still being assembled around it. Data Center Dynamics counted 17 open positions across construction, energy, operations, supply negotiation and capital markets [11], and Crux hired Alan Duong as chief development officer after more than 12 years at Meta leading data-center engineering and construction [16].
Sloss joined Google in 2003 to lead its emerging Site Reliability Engineering organization, and Google's own SRE history credits him as the creator of the discipline [13]. He said SRE is "what happens when you ask a software engineer to design an operations function" [14]. The Crux assignment extends that list to securing power, delivering construction, obtaining processor supply and, if the reported financing closes, matching debt maturities to customer contracts [17].
What to watch
- Identification of the lender group, or a signed facility agreement that moves this from reported to committed.
- A named anchor customer with a contract term long enough to survive the refinancing of a bridge.
- Whether the 500 MW target for 2027 slips again after the delays reported at planned data-center sites.