Crux AI secures ~$22B chip-collateralized bank loan for Google TPU data centres

On Wednesday 16 September 2026, Bloomberg reported — citing people with knowledge of the matter — that a group of 10 banks is providing a $22 billion chip loan to support Crux AI, the cloud venture created by Blackstone and Alphabet/Google, and that the debt would be used to purchase Google Tensor Processing Units (TPUs) and would be backed by the value of those chips and Crux AI's customer contracts. The same day, Reuters independently confirmed via its own unnamed source familiar with the matter that a group of 10 banks is providing a $22B chip loan, naming the lead institutions: Goldman Sachs Group, Sumitomo Mitsui Banking Corp (SMBC), Barclays, BNP Paribas SA and Bank of Nova Scotia (BNS) — five of the ten lenders; the source declined to be named because discussions were still private.

Corroborating detail across the two wires and their relays:

• The loan is the latest mega-debt deal in the race to finance AI processors, joining a wave of data-centre, chip, and electricity capex financing.
• The lending group was still syndicating the credit — "in the process of bringing in more lenders to share the risk" — and Bloomberg sources said the short-term facility could later be refinanced through longer-term issuance in the investment-grade corporate bond market.
• A separate $1 billion revolving credit facility accompanies the $22B term piece (reported by multiple relays of the Bloomberg story; not in the Reuters wire text).
• 9fin had earlier described approximately $23 billion of debt for the venture and characterized the facility as likely a bridge loan (reported by RuntimeWire).
• The venture itself: announced 18 May 2026 by Blackstone and Google — Blackstone committing an initial $5B equity investment and targeting 500 MW of TPU capacity online in 2027, with Google supplying TPUs, software and services; Benjamin Treynor Sloss, a two-decade Google infrastructure executive (creator of Site Reliability Engineering), leads the company as CEO. Crux AI formally launched in early September 2026 (Business Wire release dated 9–10 September 2026; Reuters: "formally launched last week" on 16 September).
• No party confirmed or denied on the record: BNP Paribas, Barclays and SMBC declined to comment; Blackstone, Alphabet, Goldman Sachs and Bank of Nova Scotia did not respond to Reuters.

• Before: Chip-collateralized lending existed but at smaller, GPU-only, arm's-length scale. The market's reference deals were CoreWeave's GPU-backed facilities — including the March 2026 $8.5B delayed-draw term loan (DDTL 4.0), the first investment-grade-rated (Moody's A3 / DBRS A(low)) HPC-infrastructure financing, explicitly non-recourse, secured by GPU clusters plus a customer contract, with Blackstone on the lender side (anchor investor via Blackstone Credit & Insurance). Google's TPUs were sold primarily through Google Cloud; no bank consortium had financed a TPU-focused neocloud at this scale.
• Change (event): A 10-bank consortium (Goldman Sachs, SMBC, Barclays, BNP Paribas, BNS, plus five unnamed) is reported to be providing ~$22B (plus a $1B revolver; 9fin had flagged ~$23B) for Crux AI — a vendor-linked borrower: Alphabet owns/backs the borrower, Google makes the chips being purchased, and the collateral is the TPUs themselves plus Crux's customer contracts. Blackstone has moved from being the lender in chip deals (CoreWeave) to being the equity sponsor of the borrower. The reported facility is ~4.4× Blackstone's $5B equity commitment.
• After: AI accelerators are now an established, rated, large-ticket collateral class with the vendor, the sponsor and the lender all in a vertically integrated deal; the same structure (asset-value + contracted-revenue collateral, syndicated across banks, refinancing path toward the IG bond market) is positioned to scale to other neoclouds and potentially to securitization (KBRA has flagged cross-collateral GPU/TPU structures as a future step). The systemic question shifts from "can chips be financed?" to "who bears the residual-value and re-leasing risk when the equipment's economic life outpaces its useful life?"

• Borrower and sponsors. Crux AI is a U.S.-based company created by Blackstone (equity sponsor, initial $5B commitment) and Google (supplier of TPUs, software, services; Alphabet is the parent). It sells integrated, dedicated AI compute — power, data centres, networking, TPUs, orchestration software and operations — as a compute-as-a-service offering for AI labs, technology companies, enterprises and governments, outside the conventional Google Cloud route.
• The loan. A ~$22B term facility (9fin had described ~$23B, likely a bridge structure) from a 10-bank group — Goldman Sachs, SMBC, Barclays, BNP Paribas, Bank of Nova Scotia named, five others unnamed — with a separate ~$1B revolving credit facility. Proceeds fund the purchase of Google TPUs.
• The collateral. Two layers, per Bloomberg: (1) the value of the TPUs being purchased — i.e., the hardware itself as security; (2) Crux AI's customer contracts — committed revenue from tenants of the capacity. This mirrors the CoreWeave DDTL model (HPC infrastructure + associated customer contract) transposed to TPUs with a vendor-linked borrower. Note the valuation subtlety: TPUs have no deep independent resale market (unlike Nvidia GPUs), so the "market value" pledged is effectively Google's transfer price — a point raised by analyst commentary, not by the parties.
• Syndication and refinancing path. The 10-bank group was still syndicating risk to additional lenders as of 16 Sep; Bloomberg sources said the short-term facility could later be refinanced in the investment-grade corporate bond market — the natural end-state given CoreWeave demonstrated the asset class can be rated IG (A3/A(low)).
• What the money buys. TPU inventory and the infrastructure to deploy it: 500 MW first capacity targeted online 2027, with a stated multi-gigawatt roadmap (Crux AI's own site says a 2 GW/yr capacity ramp beginning 2H 2027).

• AI chips now finance themselves as collateral at mega-scale. The reported facility is roughly 2.6× the size of CoreWeave's landmark IG-rated DDTL and nearly 4.4× Crux's own equity base. It marks the point where banks underwrite AI hardware primarily on the resale/re-lease value of the silicon itself plus contracted tenant revenue — a structural change in how AI capex is funded, with the credit risk moving onto bank balance sheets rather than hyperscaler or sponsor balance sheets.
• Vertical integration concentrates the chain. Alphabet designs the chips (via Broadcom), co-owns the borrower, and supplies the collateral being financed; Blackstone both sponsors the borrower and (in the CoreWeave deals) defined the lender playbook. The marquee prospective customers are AI labs in Alphabet's own orbit (e.g., Anthropic, in which Alphabet is an investor). Fewer arm's-length parties means less independent price discovery for both the hardware and the risk.
• It validates neocloud economics beyond Nvidia GPUs. TPU-based dedicated cloud capacity (off-Google-Cloud) now has bank financing at the same scale as GPU neoclouds — a direct competitive answer to CoreWeave's model and to Nvidia-centric financing, and a strategic win for Google's "sell TPUs everywhere" push.
• Systemic-risk implications for lenders. KBRA (June 2026) warned that the central credit question for GPU/TPU financings is not physical function but economic relevance — whether capacity can be re-leased at attractive prices after the original contract, given chip fungibility, switching costs and refresh cycles. A $22B exposure concentrated across ten banks (and later the IG bond market) turns that analytical warning into a live, large-ticket test.
• Refinancing fate = market verdict. Whether this bridge facility converts to investment-grade bonds will effectively be the market's rating of "TPU-backed debt" as an asset class — the same checkpoint that made CoreWeave's DDTL the reference point for GPU debt.

• ~$27B-scale (debt + equity) TPU-focused neocloud build-out — 500 MW first capacity in 2027 on a path to multi-gigawatt scale (2 GW/yr ramp from 2H 2027 per crux.ai), financed without loading Alphabet's or Blackstone's core balance sheets.
• Google TPU distribution outside Google Cloud at hyperscale — banks, not Google, funding the inventory that lets AI labs rent dedicated TPU pods directly.
• A bankable template for vendor-linked chip financing — chipmaker + sponsor-owned borrower + chip-and-contract collateral + syndication + IG refinancing path. Expect replication by other chip vendors (e.g., custom-silicon plays) and other neoclouds.
• An IG-rated path for TPU-backed paper — following CoreWeave's A3/A(low) GPU DDTL, TPU-backed debt can plausibly reach the public bond market, opening a new asset class to institutional investors.
• Cross-collateral structures (GPU + TPU in one securitization) — which KBRA already anticipates as diversification across chip generations.

• Has the facility formally closed and funded, and on what terms (rate, maturity, advance rate, LTV, covenants, amortization)?
• Who are the other five banks, and what is the final syndicate composition?
• Is the facility non-recourse to Blackstone/Alphabet, and what limited-recourse exceptions (if any) apply — as in CoreWeave's "bad acts" limited guarantee?
• Who are Crux AI's contracted customers, at what tenors and minimum commitments — and how much of the ~$22B collateral is contract-backed vs chip-value-backed?
• How is TPU "market value" appraised, by whom, and how often is it re-marked?
• Which specific sites host the first 500 MW, and what is the power/transformer/permitting status of each?
• Will the bridge convert to investment-grade bond issuance, and at what rating (CoreWeave's DDTL achieved A3/A(low))?
• What does Google's TPU supply agreement commit — volumes, generations, exclusivity — and who bears generation-refresh risk?
• Does the deal trigger any regulatory review (systemically important financial institutions' AI-exposure caps, transfer-pricing scrutiny, bank concentration rules)?

• Syndication completion: the 10-bank group (five named) finalizes participation and pricing; watch for further lender names and possible repricing given the Fed's September rate move.
• Closing confirmation: expect either a company statement or syndication-wire detail confirming whether the facility is committed/closed — currently unconfirmed.
• Contract disclosures: Crux's first named customers (and contract tenors) will be the most market-moving disclosures; the collateral's second layer depends on them.
• Refinancing decision: whether the bridge converts to IG bond issuance (and at what rating — CoreWeave set the A3/A(low) precedent for GPU debt) will define TPU-backed debt as an asset class.
• Delivery milestones: 500 MW in 2027 requires construction starts, transformer orders and power agreements through 2026–2027; Bloomberg Law had already reported location delays — milestones will be public pressure points.
• Copycat deals: expect other chip vendors (custom ASIC plays, Tier-2 accelerator makers) and neoclouds to attempt similar structures; watch KBRA/ratings-agency guidance updates for the asset class.
• Regulatory attention: bank concentration, transfer-price collateral valuation and related-party structure (Alphabet=chipmaker+co-owner; Blackstone=sponsor+former lender) may draw investor or supervisory questions.

The headline is correct in spirit: AI chips have become a collateral class, and a 10-bank group led by Goldman Sachs, SMBC, Barclays, BNP Paribas and Bank of Nova Scotia is reported to be financing $22B of Google TPUs for the Blackstone–Alphabet venture Crux AI, secured by the chips and the venture's customer contracts. But discipline is required on three fronts. First, the deal is reported, not confirmed: two wires corroborate the existence and headline terms, yet no party has commented, closing/syndication remains in progress, the "$1B revolver" and "$23B" (9fin) figures come from secondary relays, and "non-recourse" — the discovery record's framing — appears nowhere in the sourcing, however plausible the analogy to CoreWeave's explicitly non-recourse DDTL. Second, get the parties right: it is Bank of Nova Scotia, not Bank of America. Third, the real story is the risk transfer: Alphabet gets a TPU distribution channel funded off its balance sheet; Broadcom books sales up front; Blackstone sponsors the borrower after defining the lender playbook at CoreWeave; and the banks — and eventually IG bond investors — carry the KBRA-flagged exposure to chips whose economic life may end well before their useful life, collaterized at a "market value" set by the vendor's own transfer price. If the 2027 capacity arrives contracted and the refinancing clears, chip-backed debt becomes a permanent pillar of AI infrastructure finance; if delivery slips and tenancy softens, the first truly systemic test of the accelerator-collateral era will be this deal's lenders. Either way, the economic center of gravity of the AI build-out has moved from the model labs to the credit market — and that is the week's most important infrastructure signal.