Crux AI Explained: Google and Blackstone’s $5 Billion TPU Cloud Aimed at Nvidia and CoreWeave
TL;DR
- Google and Blackstone announced a joint venture on May 18, 2026 to rent out Google’s custom TPU chips as a cloud service that sits outside Google Cloud. It is now called Crux AI. Internally it was known as Project Braid.
- Blackstone committed $5 billion in equity and is reported to hold the majority. Bloomberg put the total at about $25 billion once debt is counted, and in September it reported that ten banks had lined up a $22 billion chip loan.
- The target is 500 megawatts by 2027, with talk of scaling to 2 gigawatts after that.
- The build is already running into trouble: site delays, a project Google took away from its developer, a Texas pause on new grid connections, and executives who now put the odds of hitting delivery dates at about 50%.
- CoreWeave’s shares did fall on the news, by roughly 4% before the open and about 7.5% on the day, according to market coverage. The “more than 10%” figure that circulates online isn’t supported by what we found.
- Crux AI has not publicly named a customer. Everything here is reported or announced, not audited, and none of it is investment advice.
Reviewed by Imran Khan Pathan, Editor at AI Tech Safar. Our editorial team built this from Blackstone’s own announcement, CNBC, Data Center Dynamics, Reuters (via KSL), The Next Web, and Anthropic’s and Google Cloud’s statements, plus Bloomberg reporting as relayed by other outlets. Bloomberg and the Wall Street Journal are paywalled, so we cross-checked their figures against at least one other source. We keep what the companies announced separate from what outlets report, and we label anything we couldn’t confirm. About our editorial approach and disclaimer.
Published and last updated: October 19, 2026. We’ll note any correction here. This is not financial or investment advice.
What Crux AI Is
Crux AI is a US company that will rent out Google’s tensor processing units, the AI chips Google has been building into its own data centers since the mid-2010s. Customers get chips, data center space, networking and day-to-day operations as one package, delivered through a platform that is separate from Google Cloud. Blackstone’s announcement framed it as another way to reach TPUs alongside Google Cloud, not a replacement for it.
The split of roles is simple. Google supplies the TPUs, plus software and services. Blackstone supplies the money and its experience building and running data centers. The Wall Street Journal reported that Blackstone holds the majority stake. Blackstone’s own statement didn’t spell out the ownership structure, and Google’s exact share hasn’t been published.
Blackstone says it is the world’s largest private owner of data centers, and it manages more than $1.3 trillion in assets. Crux AI is only the second deal to come out of BXN1, the unit the firm set up to pull its AI investing under one roof.
Timeline
| Date (2026) | What happened |
|---|---|
| February 26-27 | The Information reports Meta has signed a multi-billion-dollar, multi-year deal to rent Google TPUs, and that Google is lining up a joint venture to lease TPUs to others (unconfirmed by the companies) |
| April 24 | Google confirms it will invest up to $40 billion in Anthropic, with 5 GW of TPU capacity over five years |
| May 18 | Blackstone and Google announce the joint venture: $5 billion in equity, 500 MW targeted for 2027, Benjamin Treynor Sloss named CEO |
| May 19 | CoreWeave and Nebius shares fall on the announcement |
| June-July | Crusoe pauses its 1.8 GW Cheyenne, Wyoming campus; Google takes the project over the following month |
| August 3 | Texas Governor Greg Abbott orders a pause on new data center approvals in the grid interconnection process pending an audit |
| September 8 | Bloomberg reports delays at several planned sites for the venture, then known internally as Project Braid |
| Early September | The Crux AI name becomes public; Meta data center executive Alan Duong joins as chief development officer |
| September 14-16 | Reports that Blackstone may commit tens of billions more; Bloomberg reports a $22 billion bank loan to buy TPUs |
| October 9 | Bikash Koley, Google’s VP of global infrastructure and capacity, leaves to become Crux’s chief technology and product officer |
| October 19 | DigiTimes reports a separate, much larger Blackstone financing for Anthropic’s use of Google TPUs (see below) |
Why Google Wants a TPU Cloud Outside Google Cloud
For most of their history, TPUs stayed inside Google. They trained and served Google’s own models and were rented to cloud customers, but you couldn’t really buy them as a product. That has been changing all year. On Alphabet’s April earnings call, CEO Sundar Pichai said Google would start delivering TPUs to a select group of customers for use in their own data centers. Crux AI is the other half of the same idea: instead of shipping chips to a customer’s building, Google and Blackstone build the building and rent out what’s inside.
The demand is real, at least on the biggest accounts. In October 2025, Anthropic announced access to up to one million TPUs and well over a gigawatt of capacity in 2026, in a deal both companies described as worth tens of billions of dollars. In April 2026 the arrangement grew to about 3.5 GW of TPU capacity from 2027 through a partnership with Google and Broadcom, and Google then said it would invest up to $40 billion in Anthropic alongside 5 GW of TPU capacity over five years. The Information also reported that Anthropic committed to spending around $200 billion on Google Cloud over that period, a figure neither company has confirmed.
The hardware is moving too. Google made its seventh-generation chip, Ironwood, generally available in April and previewed an eighth generation that splits into separate training and inference chips, targeted at TSMC’s 2nm process in late 2027, according to The Next Web.
The capital problem is the part that explains Blackstone. Buying and powering chips at this scale costs far more than any one division wants to carry, and Google Cloud has itself been described as running at capacity. A partner with deep pockets and a data center playbook lets Google put more TPUs into the world without building every site itself. That is our reading of the logic, not a statement from either company.
The Money: What’s Committed and What’s Reported
| Item | Figure | Source and status |
|---|---|---|
| Blackstone equity | $5 billion initial commitment | Blackstone announcement (confirmed) |
| Total with leverage | About $25 billion | Bloomberg (reported) |
| Chip loan | $22 billion from ten banks, plus a separate $1 billion revolving facility | Bloomberg, September 16 (reported); another outlet put the debt at $23 billion |
| Lenders named | Goldman Sachs, Sumitomo Mitsui, Barclays, BNP Paribas, Bank of Nova Scotia, among others | Bloomberg via secondary coverage |
| Collateral | The value of the TPUs and Crux AI’s customer contracts | Bloomberg (reported) |
| First capacity | 500 MW in 2027 | Blackstone announcement (target) |
| Later ambition | About 2 GW | Data Center Dynamics (reported target) |
| Candidate sites | Around 29 locations | Bloomberg (reported) |
For scale, a common utility rule of thumb is that one megawatt serves about 750 US homes, so 500 MW works out to roughly 375,000 homes. That is a comparison, not a measure of how the data center will actually draw power.
The collateral detail matters more than the headline. A loan secured by chips and customer contracts is only as good as the contracts. Crux AI hasn’t disclosed who those customers are, and in mid-September Blackstone was also reported to be weighing a far bigger commitment in the tens of billions of dollars. We couldn’t find a primary source for that second report, so treat it as unconfirmed.
Who Is Running It
Benjamin Treynor Sloss is the CEO. He joined Google in 2003 and built its Site Reliability Engineering function, and he later ran the company’s global network backbone and data center operations. Alan Duong, chief development officer, spent more than a decade at Meta leading data center engineering and construction. Bikash Koley joined in October as chief technology and product officer after more than 16 years at Google across two stints, and earlier served as CTO of Juniper Networks.
The hiring pattern says something. This is a team assembled to build and operate physical sites quickly, not to write a new machine learning stack. In the launch statements, Blackstone’s Jon Gray called it a generational chance to invest at scale in AI infrastructure, and Google Cloud CEO Thomas Kurian said the venture would help meet rising TPU demand and widen access to accelerated computing.
How CoreWeave Reacted, and What It Means
CoreWeave fell about 4.3% in early trading on May 19, alongside Nebius, according to Invezz. A Yahoo Finance piece reported a 7.51% drop on May 19 and about 9.5% for the week. Both outlets read it as investors pricing in a TPU-based rival to the GPU-rental model.
Keep the size of that reaction in perspective. Crux AI had no live capacity then and still has none. Neocloud shares also swing on things that have nothing to do with Google: on October 8, CoreWeave fell nearly 8% in a selloff set off by a report on OpenAI’s revenue, which we cover in our breakdown of OpenAI’s $20 billion revenue gap. Our view is that the May move was a sentiment hit, and that the real test comes in 2027 when a TPU-first competitor either has racks running or doesn’t. As Propmodo noted, most neoclouds still run Nvidia hardware.
The Delays
This is where the story gets harder. On September 8, Bloomberg reported that the venture had hit delays at several planned data center sites. The details, as relayed by Data Center Dynamics and others:
- Cheyenne, Wyoming. Developer Crusoe had been building a 1.8 GW campus there, designed to scale as high as 10 GW. In June, Crusoe said it had paused work “at the request of our customer.” Google took the project over the next month. Bloomberg’s sources said Google had lost confidence in Crusoe’s ability to deliver, and Data Center Dynamics later reported the project had been scaled down.
- Texas. On August 3, Governor Abbott ordered state regulators to audit data centers moving through the ERCOT interconnection process, and paused approvals until that review is done. ERCOT was weighing roughly 474 GW of requested new demand at the time, about 90% of it from data centers. Bloomberg reported that the pause disrupted some of the venture’s plans. BloombergNEF estimated it could delay close to 50 GW of new data center load across Texas.
- Odds. Executives at the venture now put the chance of data center projects meeting their delivery dates at about 50%, down from about 90% three years ago, per Bloomberg.
- Backdrop. A JPMorgan report from May found that more than 60% of data center capacity planned for completion in 2027 hadn’t broken ground, with shortages of electrical equipment and skilled labor as the bottlenecks.
One counterpoint is worth keeping. The venture reportedly has about 29 candidate locations, which means a slipped site can be swapped for another. That cushions the schedule but it also means headline pipeline numbers shouldn’t be read as new capacity on top of what’s already planned. We couldn’t confirm the current status of the Texas pause as of this update, so check it before relying on it.
Blackstone’s Other Google-Chip Bet: Anthropic
Crux AI isn’t Blackstone’s only exposure to Google’s chips. In May, Bloomberg reported that Apollo and Blackstone were arranging roughly $36 billion in debt to buy TPUs that Anthropic would then lease, with Broadcom backstopping payments on the largest portions. A later report said Blackstone had pitched at least another $36 billion, and that Google had taken a stake of about 20% in a separate Anthropic data center project in Hubbard, Texas, agreeing to cover costs if Anthropic defaulted. DigiTimes reported on October 19 that Anthropic was offering Blackstone a return above 10% on a proposed $60 billion package covering about 1.5 GW of TPUs, reportedly code-named Project North Star. We found only that one report, so it is unconfirmed.
The point for Crux AI is that Blackstone is lending against Google’s chips in two different ways at once, one through the neocloud it co-owns and one through leases to a single large customer. That concentrates a lot of private credit around the same hardware. The Financial Times has estimated that Google’s financing programme for Anthropic alone covers more than $150 billion of TPUs.
Our Read
Crux AI is more a financing and construction story than a chip story. Whether TPUs are good enough isn’t the open question, since Anthropic has trained and served models on them for years. The open questions are whether the sites get power on time, whether the debt can be serviced, and whether customers beyond the anchor accounts show up.
For Nvidia, the near-term effect looks small. A 500 MW first phase is a fraction of the 5 GW Google has committed to Anthropic alone, and Meta is buying heavily from Nvidia and AMD while it explores TPUs. The pressure shows up more in pricing and negotiating leverage than in lost sales this year. For CoreWeave and its peers, the threat is slower and bigger: a well-funded competitor with a different chip and a hyperscaler’s supply chain behind it.
The bear case is concrete. Delivery odds of about 50%, power-grid politics, and a loan secured partly by contracts nobody has seen add up to real execution risk. The bull case is just as concrete: a team of operators who have built infrastructure at Google and Meta scale, and a chip with at least one demonstrated mega-customer.
What to Watch
- The first named customer outside Google’s existing TPU relationships.
- First power on a site. Energization, not announcements, is the real milestone toward the 2027 target.
- Whether the $22 billion loan closes on the reported terms, and whether Blackstone’s larger commitments are confirmed.
- The Texas audit and any change to interconnection approvals.
- Cheyenne’s scaled-down plan and whether the 2 GW ambition survives.
- Pricing. Any public rate card that can be compared with GPU rental from CoreWeave or Nebius.
What We Still Don’t Know
- Customers. No signed customer has been named publicly.
- Ownership. Blackstone’s majority is reported, but Google’s exact stake isn’t public.
- Pricing and terms. Nothing has been published on what Crux AI will charge or how it will contract capacity.
- The sites. Aside from Cheyenne, the locations haven’t been named.
- The real size of Blackstone’s commitment. The $5 billion is confirmed; the larger figures are reports.
FAQ
What is a neocloud, and how is it different from AWS, Azure or Google Cloud?
A neocloud is a newer cloud company built around renting AI computing power, usually GPU clusters, instead of offering the huge catalog of databases, storage and software services the big three provide. CoreWeave and Nebius are the best-known examples. They typically sell capacity in large blocks on multi-year contracts and finance their hardware with debt secured against those contracts, which is the same pattern as the reported Crux AI loan. Customers trade a narrower set of services for faster access to scarce chips.
Can a startup rent TPUs today, or does it have to wait for Crux AI?
The current route is Google Cloud, where Ironwood, the seventh-generation TPU, became generally available in April 2026, according to The Next Web. Crux AI hasn’t announced a public product, pricing or waitlist. Earlier statements described its customers as hyperscalers, frontier AI labs, sovereign customers and large enterprises, so it isn’t obviously aimed at small teams. Its first capacity is also targeted for 2027, so a startup that needs TPUs now should look at Google Cloud first.
Can code written for Nvidia GPUs run on TPUs?
Not as it stands. Nvidia GPUs are programmed through CUDA, while TPUs work through Google’s XLA compiler and are usually reached with JAX or PyTorch/XLA. Teams on high-level frameworks can often port with moderate effort, but custom CUDA kernels have to be rewritten and performance tuning starts again. That switching cost is part of why Nvidia’s position is sticky. It also explains why a company like Anthropic, which already trains on TPUs, is an easier early fit for Crux than a team built around CUDA.
Are TPUs actually cheaper than Nvidia GPUs?
Google and its biggest customer say they offer better price-performance, but no neutral public benchmark settles it. Anthropic said in October 2025 that it chose TPUs for their price-performance and efficiency, and Google Cloud’s CEO made the same case. Real cost depends on the model, the software stack, how fully the chips are used and the contract terms. Be skeptical of any percentage savings quoted online unless it comes with the test setup.
Can ordinary investors buy into Crux AI?
Not directly. Crux AI is a private company and no listing has been announced. Public exposure is indirect, through its parents Blackstone (NYSE: BX) and Alphabet, or through companies in Google’s TPU supply chain such as Broadcom. Whether that is a sensible way to back one venture that hasn’t named customers is a separate question, and this is general information, not advice.
Related Reading on AI Tech Safar
- OpenAI’s $20 Billion Revenue Shocker: How One Number Wiped Billions Off AI Stocks
- Anthropic IPO 2026: Inside the $2 Trillion Nasdaq Listing That Could Beat SpaceX’s Record
Useful Sources
- Blackstone – Announces Joint Venture with Google to Create New TPU Cloud (May 18, 2026)
- CNBC – Blackstone to invest $5 billion in AI infrastructure venture with Google
- Data Center Dynamics – Blackstone and Google team up for $5bn TPU-based cloud platform
- Data Center Dynamics – Google and Blackstone JV experiences delays in data center projects
- Data Center Dynamics – Bikash Koley heads to TPU neocloud Crux AI (October 9, 2026)
- Bloomberg – Banks line up $22 billion chip loan tied to Blackstone, Alphabet
- Propmodo – Data center supply chain bottlenecks slow AI infrastructure rollout
- Reuters via KSL – Texas governor orders pause on new data center approvals pending audit
- Invezz – CoreWeave stock falls 4% after Google-Blackstone deal
- Yahoo Finance – Google and Blackstone bet $5 billion on new neocloud to challenge CoreWeave
- The Stack – Google is building an AI cloud from scratch with Blackstone
- The Next Web – Google launches Ironwood TPU and previews eighth-gen chips
- Anthropic and Google Cloud – Anthropic to expand use of Google Cloud TPUs (October 23, 2025)
- Cybernews – Google to invest up to $40B in Anthropic (citing Bloomberg)
- Storyboard18 – Meta signs multi-billion dollar AI chip rental deal with Google (citing The Information)
- Silicon UK – Blackstone eyes major debt deal for Anthropic chip access
- DigiTimes – Anthropic offers Blackstone more than 10% return in US$60 billion AI chip financing deal (October 19, 2026)
For the broader context and related reading, explore our AI Business Guide: Stocks, Infrastructure and Layoffs.


