Ten banks lend $22B to fund Alphabet and Blackstone's TPU cloud

Share
Ten banks lend $22B to fund Alphabet and Blackstone's TPU cloud

The compute buildout just got its largest chip-backed loan yet — and the collateral is silicon only one company's software can use.

A consortium of 10 banks is providing a $22 billion loan to Crux AI, the cloud venture Blackstone and Alphabet formed in May, to buy Google's custom Tensor Processing Units, Reuters reported Wednesday, confirming Bloomberg's earlier account. Goldman Sachs, Sumitomo Mitsui Banking Corp, Barclays, BNP Paribas and Bank of Nova Scotia are among the top lenders, according to a person familiar with the deal who was not authorized to speak publicly. None of the banks, Blackstone, Alphabet or Google would comment.

The structure is the interesting part. The debt is secured by the value of the TPUs themselves plus Crux AI's customer contracts, Bloomberg reported, and the banking group is still syndicating the facility to spread the risk. The report also says the loan could later be refinanced into long-term institutional money via the investment-grade bond market — a sign the lenders see this as a bridge, not a permanent exposure.

Crux AI formally launched last week with $5 billion of equity from Blackstone; Google contributes the TPUs, software and services. The venture said it will bring its first 500 megawatts of capacity online in 2027, and it is pitching dedicated compute at AI labs, enterprises and governments. The chips it is now borrowing against are the whole product.

Why it matters: chip-backed lending is becoming the default way to finance AI capacity, and it carries a risk a data-center mortgage does not. Nvidia-powered clouds already ran this playbook — Lambda raised roughly $1 billion of short-dated private debt in August against GPUs that Microsoft would lease. Crux AI is the same instrument with a different vendor. We flagged the shape of this financing wave earlier — The Take — The $3T isn't Enron. It's a no-exit clause — and the bull case here is that the debt is matched to contracted revenue rather than to a hope.

The bear case is the collateral. A $22 billion facility secured against TPUs assumes those parts hold resale value, and the secondary market for Google's silicon outside Google's own stack is thin to nonexistent — Nvidia's GPUs at least trade across clouds. The repayment side leans on customer contracts signed with labs that are themselves under pressure to explain their spending, in a month when the industry's own executives spent more time talking about pacing than about buying more compute. Cheap credit does not make demand certain; it just moves the moment of truth to a syndication book and a bond prospectus.


Nvidia has shipped CUDA Rust, two official tracks for writing GPU kernels natively in Rust instead of wrapping code written somewhere else. The tile track compiles Rust to PTX and lets the compiler decide how work maps onto each architecture, needing only CUDA 13.3, stable Rust 1.89, compute capability 8.0 or later, and no custom LLVM; the SIMT track, which is the one for readers who want to manage memory and threads by hand, still wants a pinned nightly toolchain.

The part worth caring about is what the compiler now catches. Thousands of threads touch the same buffers in no guaranteed order, so when one of them writes while another reads, the result depends on timing — the class of bug that passes every test and then fails in production. Both tracks refuse to compile a kernel whose output buffer is passed as one of its own inputs, whether or not that kernel would actually race. Nvidia is leaning the whole pitch on that bargain: the systems layer around AI is increasingly written in Rust, and the kernel was the last holdout. For readers who want the vocabulary underneath this, we explained what a kernel actually is back in August — AI 101 — What is a learned kernel?. The caveat is adoption: this is early, it is NVlabs work published as open repos, and neither track's source is the language most CUDA code in production is written in today.

What to watch: whether Crux's syndication fills at $22 billion and what spread the remaining lenders demand; and whether Nvidia's tile track gets adopted by a major inference stack, which would be the first real evidence that GPU code can leave C++ behind.

Would you lend $22 billion against chips only one company's software can run? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Utilities sign on to let Google and Nvidia dim data centers on demand

Utilities sign on to let Google and Nvidia dim data centers on demand

The power bill is now the AI buildout's real bottleneck — and the industry just answered it with the same move it makes on safety: a voluntary alliance. Emerald AI, Google and Nvidia launched the AI Energy Management Alliance on Wednesday with 18 launch partners, including Anthropic and utilities AES, National Grid, Constellation, NRG and Germany's RWE, to make AI data centers cut their electricity draw when the grid strains.

Huawei's chairman says its computing goal is to become Nvidia

Huawei's chairman says its computing goal is to become Nvidia

Two infrastructure stories today: the ambition inside Huawei's compute business finally goes on the record, and a security vendor puts a hard number on whether detection rules actually fire. Huawei's supervisory board chairman Guo Ping told new employees that the company's target in ICT and computing is to become the Nvidia of the field — and that its real advantage is not owning a frontier model, but running everyone else's.

Cornelis raised $205 million. Its pitch: the network, not the chip

Cornelis raised $205 million. Its pitch: the network, not the chip

Monday's news ran from the inside of the AI rack to the inside of the White House — a networking challenger that says the GPUs are not the bottleneck, a private Altman-Trump meeting that surfaced the same day the two split publicly on slowing down, and a mathematician's attempt to say what the profession should be for. Cornelis Networks announced roughly $205 million in funding alongside Active Compute Fabric, an open networking architecture that puts programmable compute inside the fabric…

Alibaba plans a 10T-parameter model and a 500,000-chip cluster

Alibaba plans a 10T-parameter model and a 500,000-chip cluster

Alibaba's cloud conference in Hangzhou opened this morning with a three-part commitment that is harder to wave off than the usual keynote arithmetic: a model bigger than anything trained so far, the chip to run it, and the power to feed both. Huawei spent the same week putting a coding model inside its own operating system, and a small project on the local-AI forums quietly showed that a model's refusals can be switched off without touching a single weight. Alibaba says its next Qwen models…

Firebird's Armenia AI factory lands 70,000 Nvidia chips by 2027

Firebird's Armenia AI factory lands 70,000 Nvidia chips by 2027

A data center that doubles as a diplomatic instrument, and a quiet upgrade to the tool most people already use to run models on a laptop. Firebird's data center in Armenia is set to reach 300 megawatts and more than 70,000 Nvidia chips by the end of 2027 — hardware that Bloomberg reports the Trump administration agreed to let into the country as part of a wider diplomatic push. The San Francisco startup began operating the facility in August.

Chanos asks why Nvidia sells the GPUs if renting them is this good

Chanos asks why Nvidia sells the GPUs if renting them is this good

Jensen Huang says Nvidia's compute is "fungible, durable and highly rentable." One of Wall Street's best-known short sellers heard that and asked the obvious question: then why are you selling it? Two days later the argument has become the cleanest framing yet for the AI capex question. Jim Chanos, who called Enron, replied to Huang's X post with four words that cut straight to the accounting: "Then why not rent them out yourself?