Crux AI lines up reported $22B TPU loan, status still unclear

Why it matters

[Google and Blackstone disclosed](https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/blackstone-tpu-cloud/?ref=runtimewire) an initial $5 billion equity commitment and a target of bringing 500 MW of TPU capacity online in 2027. If Bloomberg's reported financing closes on the described terms, chip value, customer contracts and re-leasing risk would become central tests for lenders funding an eventual multiple-gigawatt expansion.

Crux AI, led by longtime Google infrastructure executive Benjamin Treynor Sloss, is pursuing a reported $22 billion facility to buy Google's tensor processing units. Bloomberg reported on September 16th that 10 banks are providing the financing, backed by the chips and Crux AI's customer contracts. 9fin previously described approximately $23 billion of debt and said the facility would likely be a bridge loan. The accessible reports do not identify the banks or establish whether the facility has closed, is committed or remains contemplated.

The scale of Bloomberg's reported financing would move Crux AI far beyond the joint venture Blackstone and Google described four months ago. In its May 18th announcement, Blackstone committed an initial $5 billion in equity, while Google agreed to supply TPUs, software and services. The partners created Crux as a U.S.-based company offering data-center capacity, operations, networking and Google processors as a compute service outside the standard Google Cloud platform.

Blackstone said the first 500 megawatts of capacity would come online in 2027, followed by further expansion. Crux's longer-term roadmap reaches multiple gigawatts: Data Center Dynamics reported that the company plans to scale to 2 GW and that its development chief is responsible for delivering multiple gigawatts over the coming years. Those ambitions remain distinct from the disclosed 500 MW expected online next year.

The reported facility is 4.4 times Blackstone's initial equity commitment. If Bloomberg's reported facility closes on those terms, it would make Crux AI an unusually debt-intensive infrastructure build from an early stage.

Crux publicly launched under its current name in September. Its launch materials describe an integrated service spanning power, data centers, TPUs, networking, orchestration software and operations, but do not identify customers, pricing or facility locations. Data Center Dynamics reported 17 open positions across construction, energy, operations, supply negotiation and capital markets. A September 9th Bloomberg Law report said the venture had encountered delays at major planned data-center locations while retaining its 2027 target for renting Google processors to customers.

Sloss takes reliability doctrine outside Google

Sloss joined Google in 2003 to lead its emerging Site Reliability Engineering organization. Google's SRE history credits him as the creator of Google SRE, the discipline that applied software engineering to the work of keeping large systems running.

In a Google interview about the model, Sloss described SRE as "what happens when you ask a software engineer to design an operations function." He later oversaw networking, data centers and reliability engineering across Google's infrastructure. Crux AI gives him a related assignment with its own capital stack: combine power, data centers, networking, processors and operations into capacity that outside customers can rent.

That background makes Sloss a logical operator for the venture. Crux AI needs to coordinate facilities and hardware on schedules measured in years while selling computing capacity into a market where chip performance and customer requirements change much faster. Reliability now includes securing power, delivering construction, obtaining processor supply and, if the reported financing closes, matching debt maturities to customer contracts.

Sloss has also recruited executives who have already managed infrastructure at hyperscale. Data Center Dynamics reported that Crux AI hired Alan Duong as chief development officer after more than 12 years at Meta, where he led data-center engineering and construction. Duong is responsible for delivering multiple gigawatts of capacity, according to the report.

Google opens a second door to its TPUs

Crux AI gives customers a route to Google's proprietary processors outside the standard Google Cloud platform. That makes the venture distinct from AI clouds built primarily around Nvidia GPUs and gives Google another distribution channel for hardware it has designed and operated internally for more than a decade.

The division of labor is direct. Google supplies the processors and technical layer. Blackstone supplies equity and data-center development experience. Sloss and Duong are responsible for making those pieces function as an operating cloud rather than a collection of projects and supply agreements.

The model also gives Blackstone a new place to deploy large pools of capital into assets tied to AI demand. For Google, Crux AI can broaden TPU adoption without requiring every customer relationship to follow the conventional Google Cloud route. Crux AI sits between those incentives, carrying the execution work and, if the reported facility closes, the resulting debt obligations.

Its peer set includes CoreWeave, Crusoe, Lambda and Nebius, specialized computing providers that have largely built around Nvidia GPUs. Crusoe combines energy sourcing, data-center development and cloud services. Lambda raised more than $1.5 billion in November 2025 to expand its AI infrastructure. Crux's access to Google TPUs and Blackstone's capital separates its supply strategy from those GPU-focused providers, although it has yet to disclose customers or contracted utilization.

Chips and contracts could carry the loan

Bloomberg's description of the proposed collateral package puts two variables at the center of the financing: the future value of Google's TPUs and the durability of Crux AI's customer agreements. Physical data centers may remain useful for decades, while the structure described by Bloomberg would expose lenders to a risk KBRA has identified across AI compute financings: TPUs may remain functional while losing economic value faster than the facilities that house them.

Credit-rating firm KBRA outlined that tension in June research on AI compute financing. KBRA said lenders need to assess workload suitability, chip fungibility, switching costs and whether older processors can be leased again at attractive prices after an initial contract ends. The report discussed the asset class broadly and did not evaluate Crux AI specifically.

Microsoft Research's March 2026 paper on the AI data-center lifecycle found that hardware refresh decisions dominate long-term fleet costs. The researchers estimated that coordinated management of construction, IT provisioning and operations could reduce total cost of ownership by as much as 40% compared with conventional siloed approaches. For Crux, the useful life of the buildings and electrical systems could therefore extend well beyond the most competitive period for a particular TPU generation.

Customer contracts can reduce financing risk by linking debt to committed revenue. Their value depends on details including contract length, customer credit quality, pricing, renewal provisions and obligations if a deployment is delayed. Those terms would determine whether the proposed loan behaves like infrastructure debt backed by predictable cash flow or an equipment bet exposed to a rapid processor cycle.

Crux AI is entering a market where CoreWeave entered an $8.5 billion delayed-draw facility to fund GPU servers and related infrastructure required to perform a customer contract. CoreWeave disclosed the facility in a March 30th SEC filing. Crux AI's reported package would be more than twice that size and centers on TPUs instead of GPUs.

The reported loan would give Sloss substantial purchasing power, but Crux still has to bring its planned 500 MW of initial capacity online in 2027 and prove that customers will fill it. Reaching the longer-term multiple-gigawatt target adds another layer of site, power, construction and processor-supply risk. The company must keep the clusters occupied and preserve enough value in the hardware and contracts to satisfy lenders if the financing closes.

Sloss spent more than two decades designing systems around the assumption that failures will happen. At Crux AI, that doctrine now covers a 500 MW rollout planned for 2027, a longer-term multiple-gigawatt roadmap and a reported financing package whose status and structure remain unresolved.