Google and Blackstone have put a name on the TPU-focused cloud venture they first disclosed in May 2026: Crux AI. The bigger news landed alongside it. On September 10, 2026, the joint venture confirmed it had hired Alan Duong, Meta’s longtime vice president and head of data center engineering and construction, as its new chief development officer, according to Datacenter Digest and IDCNOVA. Data Center Dynamics broke the story first, and the announcement quickly moved through trade press and data-center trackers including DC Atlas.
The timing matters. Two days earlier, Bloomberg reported that the same venture, known internally as “Project Braid,” was already facing delays at planned data center sites. So Crux AI is entering public view carrying both a marquee hire and a credibility question, at the exact moment the AI industry’s appetite for compute capacity shows no sign of slowing. Here is what is actually confirmed, what it means for the neocloud market, and what to watch next.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Is Crux AI? Inside the Google-Blackstone TPU Neocloud
Crux AI is a joint venture between Alphabet’s Google and investment giant Blackstone, built to offer Google’s Tensor Processing Units as a rentable cloud service outside of the standard Google Cloud platform. Data Center Dynamics first detailed the structure in May 2026: Blackstone commits equity capital from its funds, Google supplies the TPU hardware plus supporting software and services, and the resulting platform lets enterprise and AI customers access TPU compute without going through the Google Cloud console.
That “platform other than Google Cloud” framing is the whole point. Google has offered Cloud TPUs for years as one line item inside its broader cloud business. Crux AI is a separate, purpose-built company designed to scale TPU capacity aggressively and sell it the way independent GPU clouds already sell Nvidia capacity. Bloomberg’s reporting adds a detail that had not previously surfaced: the venture’s internal codename is “Project Braid,” and it is aiming to start renting Google’s AI processors to customers in 2027.
Datacenter Digest frames Crux AI plainly as “a neocloud provider offering TPU-based compute capacity to enterprise and AI customers,” and argues that its early staffing moves signal serious intent to compete with GPU-focused neoclouds rather than simply extend Google Cloud’s existing TPU offering under a new brand.
Who Is Alan Duong? Meta’s Data Center Architect Changes Sides
Alan Duong spent more than 12 years at Meta, most recently as vice president and head of data center engineering and construction, according to IDCNOVA. In that role he led the team that redesigned Meta’s data center program for the AI era, a process that at times meant scrapping and reworking facilities that had been planned around a pre-AI blueprint, as covered previously in our look at Meta’s push to automate data center construction.
At Crux AI, Duong now holds the title of chief development officer. Servnet UK reports he is responsible for delivering “multiple GW of data center capacity over the coming years,” while IDCNOVA describes his mandate as the end-to-end delivery of multiple gigawatts of capacity. Both descriptions point to the same thing: Crux AI hired Duong specifically for his experience building at gigawatt scale under AI-era power and timeline pressure, not for a conventional facilities-management background.
It is a notable departure. Meta has spent the past two years publicly reorienting its own infrastructure roadmap around AI, including the ongoing ramp of its in-house Iris AI chip program. Losing the executive who oversaw that data center redesign to a Google-backed rival is the kind of signal that data-center recruiters and rival hyperscalers will be reading closely for months.
The $5 Billion Blackstone Bet
When Data Center Dynamics first reported the joint venture in May 2026, the headline number was Blackstone’s commitment: $5 billion in equity capital from its funds, aimed at the new TPU-based cloud platform. Bloomberg’s September reporting reiterates that figure, describing Crux AI as having “launched with $5 billion from Blackstone.” That equity is no longer the whole financing picture: Reuters reported in September 2026 that a syndicate of 10 banks, including Goldman Sachs, arranged a separate $22 billion loan for Crux AI specifically to fund purchases of Google TPUs, layering debt financing on top of Blackstone’s original equity commitment.
Neither the original DCD report nor the more recent coverage discloses an overall company valuation for Crux AI, nor a separate contract value between Crux AI and Google Cloud for the TPU supply arrangement. What is confirmed is the structure: Blackstone brings capital, Google brings chips, software, and services, the joint entity builds and operates the data centers in between, and, as of September 2026, a 10-bank lending group has layered on $22 billion in debt that Reuters reports is backed by the underlying chip value and Crux AI’s customer contracts rather than by real estate alone. That is a materially different arrangement from Blackstone’s typical real-estate-heavy data-center financing plays, because it ties both the equity and the new debt directly to a specific chip architecture rather than a general-purpose colocation footprint.
From 500MW to 2GW: Crux AI’s Buildout Roadmap
Data Center Dynamics reports that under the joint venture, Google and Blackstone aim to bring 500MW of capacity online in 2027, with plans to expand from there. Servnet UK adds that Crux AI’s own public materials describe a goal of scaling to 2GW of capacity overall, and Crux AI’s own September 2026 launch announcement, covered by Business Wire, explicitly frames that initial 500MW deployment as just the first phase of a broader multi-gigawatt roadmap rather than a standalone project. Below is a timeline of how the venture has moved from unnamed joint venture to a fully staffed, publicly branded, debt-and-equity-financed neocloud in under four months.
| Date | Milestone | Source |
|---|---|---|
| May 2026 | Google and Blackstone confirm a joint venture for a TPU-based cloud platform; Blackstone commits $5 billion in equity capital | Data Center Dynamics |
| September 8, 2026 | Bloomberg reports delays at planned data center sites for the venture, internally codenamed “Project Braid” | Bloomberg |
| September 10, 2026 | Venture officially named Crux AI; Alan Duong announced as chief development officer | Data Center Dynamics, Datacenter Digest |
| September 10-11, 2026 | Coverage of the Duong hire spreads across data-center trade press and industry trackers | IDCNOVA, Servnet UK, DC Atlas |
| 2027 (targeted) | First 500MW of capacity planned to come online; customers expected to begin renting Google TPUs through Crux AI | Data Center Dynamics |
| Multi-year (stated goal) | Company states an ambition to scale to roughly 2GW of total capacity | Servnet UK |
Two gigawatts is a large number in isolation, but it is worth sizing against the rest of the industry’s build-out. Our earlier coverage of Meta’s Iris AI chip program put that company’s own ambitions at roughly 14GW of eventual AI infrastructure capacity, and Nvidia-aligned buildouts covered on this site have run into similar multi-gigawatt territory. Crux AI’s 2GW target is meaningful, but it is entering a field where the biggest players are already planning several times that scale.
Why TPUs, Not GPUs? Google’s Compute-as-a-Service Strategy
The distinguishing feature of Crux AI is the chip underneath it. Where most of the recent wave of neoclouds, including the GPU-focused operators referenced in Datacenter Digest’s coverage, have built their businesses around renting Nvidia GPUs, Crux AI is built entirely around Google’s own Tensor Processing Units. That is a strategic choice with two implications.
First, it gives Google a way to monetize TPU capacity through a vehicle that is not Google Cloud itself, reaching customers who may want TPU-class performance without adopting the full Google Cloud stack. Our prior deep dive into Google’s TPU 8t and 8i chips covered how aggressively Google has scaled its custom silicon roadmap in partnership with Broadmcom and MediaTek; Crux AI is the commercial vehicle for putting more of that silicon into outside hands. Second, it positions Google to compete more directly with Nvidia’s dominance of the AI compute market, a rivalry we detailed in our comparison of Nvidia’s Rubin platform against Google’s TPU roadmap.
Data Center Dynamics’ original JV reporting is explicit that the arrangement lets customers “access the TPUs as a compute-as-a-service offering via a platform other than Google Cloud.” In practice, that means Crux AI functions less like a Google Cloud product line and more like an independent chip landlord that happens to have Google as its chip supplier and equity partner.
Bloomberg’s Delay Report Complicates the Launch
The Crux AI naming and Duong hire did not happen in a vacuum. Two days before the announcement, Bloomberg reported that the Google-Blackstone venture was already facing delays at major data-center locations meant to run Google’s chips, describing the setbacks as underscoring broader tension in the pace of AI infrastructure build-outs. Bloomberg’s report referred to the venture by its internal codename, Project Braid, before the Crux AI name became public.
That sequencing is worth sitting with. A high-profile executive hire announced within 48 hours of a report about site delays reads, at minimum, as a company trying to control its own narrative. It does not mean the venture is in trouble. Large data-center projects routinely slip on permitting, power interconnection, and construction timelines, and Duong’s entire professional track record at Meta was built on solving exactly that class of problem. But it does mean Crux AI’s public launch is happening under real execution pressure, not a clean slate.
In His Own Words: Alan Duong on Building “End-to-End”
Duong’s own comments on joining Crux AI, as reported by IDCNOVA, lay out the operating philosophy he says he is bringing from Meta:
“I spent the last 12 years at Meta building data centers, and the clearest lesson from that time is that success in this business isn’t one thing done well. It’s a chain that has to hold, end-to-end: originating sites with real power and real runway, showing up in host communities as a long-term neighbor rather than a transaction, designing for efficient use of power, water and materials, building safely with the trades who put the steel in the ground, and handing over facilities that operators can run for decades. Break any link and the rest doesn’t save you. That end-to-end view is what I’m bringing to Crux AI, and what we’ll build the organization around as we scale.”
Alan Duong, Chief Development Officer, Crux AI – IDCNOVA
The framing is telling in the context of the Bloomberg delay report. Duong is not describing a single bottleneck to fix; he is describing site selection, community relations, power procurement, construction execution, and long-term operability as one continuous chain. That is a direct answer, whether intentional or not, to the kind of multi-cause delay Bloomberg’s sources described at the venture’s early sites.
Inside Meta’s AI-Era Data Center Redesign
To understand why Google and Blackstone wanted Duong specifically, it helps to look at what he was describing during his time at Meta. In earlier remarks about Meta’s shift toward AI-era facility design, reported by Data Center Dynamics, Duong explained the scale of the pivot in blunt terms:
“We saw the writing on the wall for the potential scale of this technology about two years ago.”
Alan Duong, then Global Director of Data Center Engineering, Meta – Data Center Dynamics
He described Meta’s response as a deliberate hedge rather than an all-in pivot to AI-only facilities: “During that two-year journey, we did consider doing AI-dedicated data centers, and we decided to move towards more of a blend, because we do know there’s going to be this transition,” he told Data Center Dynamics. In a separate analysis piece, he framed the underlying pressure on Meta’s design teams even more directly: “We’re experiencing growth in AI tech, [so] we need to make sure that our data center can adapt to something that’s still evolving,” adding that the required scale-up was roughly “4x.”
That combination, hedging design choices while planning for roughly a fourfold capacity increase, is precisely the skill set Crux AI needs if it intends to move from 500MW to 2GW without repeating the kind of scrapped-and-rebuilt facility cycle Duong oversaw at Meta.
The Neocloud Boom: Crux AI Enters a Crowded Field
Bloomberg’s report places Crux AI inside a broader category it calls “neoclouds,” entities that lease chips and computing capacity rather than operating full hyperscale platforms. Most of the neoclouds that have grown quickly over the past two years, and that regularly appear in this outlet’s AI infrastructure coverage, are built around renting Nvidia GPUs to AI labs and enterprises that cannot or do not want to build their own data centers.
Crux AI’s TPU-first approach sets it apart structurally, even if the sales pitch to customers, rentable AI compute without hyperscaler lock-in, sounds similar. The table below lays out how Crux AI’s model differs from both a traditional Google Cloud TPU subscription and the GPU-centric neoclouds that currently dominate headlines.
| Model | Chip Focus | Ownership Structure | Access Point |
|---|---|---|---|
| Crux AI | Google TPUs | Joint venture: Google supplies chips/software, Blackstone supplies $5B in equity capital | Independent platform, separate from the Google Cloud console |
| Google Cloud TPU | Google TPUs | Wholly owned by Google, part of Google Cloud | Google Cloud console and APIs |
| GPU-focused neoclouds (category, per Datacenter Digest) | Nvidia GPUs | Independently financed, various backers | Direct rental platforms and APIs |
| Core hyperscaler clouds (AWS, Azure, Google Cloud) | Mixed: GPUs plus custom silicon | Wholly owned by the respective hyperscaler | Full-stack cloud consoles and enterprise contracts |
Datacenter Digest’s assessment is that Crux AI’s early moves, especially the Duong hire, “signal serious intent to compete with GPU-focused neoclouds,” rather than function as a quiet side project for Google’s excess TPU capacity. If that read is right, Crux AI is not just a financing vehicle, it is a market entry.
Talent War: Why Big Tech’s Infrastructure Chiefs Are Job-Hopping
Duong’s move is part of a wider pattern that has been building across the AI infrastructure sector for roughly two years: senior data-center engineering and construction executives, once a relatively quiet corner of big tech, have become some of the most fought-over hires in the industry. The reason is straightforward. Building AI-scale data centers fast, safely, and on budget requires a narrow set of people who have actually done it before, and there are far more gigawatt-scale projects underway right now than there are executives with that track record.
Meta itself has not been shy about the scale of its own ambitions in this area, a theme this outlet has tracked through Meta’s Iris AI chip production ramp and its broader data-center automation push. Losing the executive behind its own AI-era redesign to a rival-adjacent, Google-backed venture is a reminder that even companies spending tens of billions on infrastructure are not immune to poaching when a well-capitalized joint venture comes calling with a chief-officer title and a multi-gigawatt mandate attached.
Market Impact: What This Means for Google, Blackstone, and Meta
For Google, Crux AI represents a way to turn TPU manufacturing capacity into a revenue stream that does not depend entirely on Google Cloud’s own sales motion. It also gives Google a more direct answer to the neocloud boom that has, until now, been almost entirely a showcase for Nvidia’s GPU ecosystem. Every dollar of TPU capacity that Crux AI rents out is, implicitly, a dollar of AI compute demand that does not flow to an Nvidia-based provider.
For Blackstone, the $5 billion equity commitment extends a broader strategy of treating AI data-center capacity as a distinct, investable asset class, similar in spirit to its other large-scale digital infrastructure bets, but notable here for tying capital directly to a single chip architecture rather than a general-purpose facility. The September 2026 arrival of a $22 billion, 10-bank loan alongside that equity, per Reuters, also signals that mainstream lenders are now willing to underwrite TPU-specific infrastructure debt against chip value and customer contracts, not just against traditional real estate collateral.
For Meta, the immediate impact is a personnel gap at a moment when the company is simultaneously racing to scale its own in-house AI chip infrastructure. None of the available reporting suggests Meta is behind schedule as a result, but replacing 12 years of institutional knowledge about how Meta specifically builds AI-era facilities is not a same-quarter fix.
Comparing Crux AI to Existing Sovereign and Distributed Cloud Plays
Crux AI’s joint-venture structure also echoes, without directly copying, the sovereign and distributed cloud arrangements this outlet has covered among the major hyperscalers, including our comparison of AWS, Microsoft, and Google’s sovereign cloud commitments and our look at AWS Outposts versus Azure Local and Google Distributed Cloud. In each of those cases, hyperscalers have found it useful to spin up semi-independent infrastructure vehicles, whether for regulatory, capital, or go-to-market reasons, rather than running every new capacity push directly through their core cloud brand. Crux AI fits that same pattern, just with an outside financial partner and a single-chip focus rather than a geography-specific mandate.
Historical Context: From Internal TPUs to a $5 Billion External Bet
Google’s TPUs began as an internal project, custom silicon built to run the company’s own machine-learning workloads more efficiently than general-purpose GPUs. Over the following decade, Google gradually opened that silicon to outside customers through Cloud TPU offerings inside Google Cloud, expanding the program alongside newer TPU generations detailed in our earlier coverage of the TPU 8t and 8i chip family.
Crux AI represents the next step in that arc: rather than simply offering TPUs as one more Google Cloud SKU, Google is now willing to bring in an outside financial partner, commit to an independent brand, and hire executive talent away from a direct AI infrastructure competitor to scale TPU access as its own standalone business. It is a meaningful escalation from “TPUs are available on Google Cloud” to “TPUs are the product of a dedicated, multi-gigawatt company.”
Predictions: What’s Next for Crux AI and the TPU Neocloud Race
- Expect more named executive hires before the 2027 capacity target. Duong’s chief-development role covers site origination and construction; Crux AI will likely need to announce leadership for sales, cloud operations, and customer engineering as it approaches its first commercial capacity.
- The Bloomberg-reported site delays will likely surface again in follow-up reporting. Large-scale data-center build-outs rarely resolve permitting and power interconnection issues quietly, and journalists already tracking Project Braid’s early hiccups are likely to keep checking in on specific sites.
- Google will likely continue distinguishing Crux AI from standard Cloud TPU offerings in its messaging. The “platform other than Google Cloud” framing from Data Center Dynamics’ original report suggests Google wants Crux AI perceived as an independent compute option, not a rebrand.
- Other hyperscalers with custom silicon programs may explore similar external joint ventures. If Crux AI’s model proves it can attract capital and customers around a non-Nvidia chip, it becomes a template other companies with proprietary AI silicon could study.
- Talent movement between Meta, Google, and infrastructure-focused investors is likely to continue. Duong’s move follows a broader pattern of senior data-center executives becoming a scarce, highly recruited resource across the AI infrastructure buildout.
These are analytical projections based on the confirmed facts above, not guaranteed outcomes, and each depends heavily on how quickly Crux AI resolves the site issues Bloomberg flagged just before the Duong announcement.
Frequently Asked Questions
What is Crux AI?
Crux AI is a joint venture between Google and Blackstone that offers Google’s Tensor Processing Units as a cloud compute service, accessible through a platform separate from Google Cloud, according to Data Center Dynamics.
Who is Alan Duong?
Alan Duong is a data-center executive who spent more than 12 years at Meta, most recently as vice president and head of data center engineering and construction, according to IDCNOVA. He now serves as chief development officer at Crux AI.
How much has Blackstone invested in Crux AI?
Blackstone has committed $5 billion in equity capital from its funds to the joint venture, as reported by Data Center Dynamics and reiterated by Bloomberg. In September 2026, Reuters reported that a 10-bank group led alongside Goldman Sachs added a further $22 billion loan to finance Google TPU purchases, on top of Blackstone’s equity stake.
When will Crux AI’s data centers come online?
Data Center Dynamics reports a target of bringing 500MW of capacity online in 2027, with Servnet UK reporting a longer-term company goal of scaling to roughly 2GW overall.
Is Crux AI the same as Google Cloud TPU?
No. Crux AI is a separate joint venture designed to offer TPU compute through its own platform, distinct from the existing Cloud TPU service inside Google Cloud, according to Data Center Dynamics’ original reporting on the venture’s structure.
What is “Project Braid”?
Project Braid is the internal codename Bloomberg used to refer to the Google-Blackstone venture before it was publicly named Crux AI.
Did Crux AI’s data center sites face delays?
Yes. Bloomberg reported on September 8, 2026, that the venture was facing delays at planned data-center locations meant to run Google’s chips, two days before the Crux AI name and Duong hire were announced.
How does Crux AI differ from GPU-focused neoclouds?
Crux AI is built specifically around Google’s TPUs rather than Nvidia GPUs, which Datacenter Digest describes as a structural difference from the broader category of GPU-focused neoclouds that has driven much of the recent AI infrastructure funding boom.
Related Coverage
- How to Use Google Imagen 4 API: 12 Steps, 90 Min [2026]
- Siri Runs 5 Gemini-Based Models, Not Google's App [2026]
- Meta Launches Muse AI Agent: $20 and $100 Tiers [2026]
- Google Pics Rivals Canva Weeks After $10B Valuation Cut [2026]
- Mistral AI Valuation Tops €21B After Samsung €3B Deal [2026]
