Google and Blackstone Are Building a New AI Cloud Company. Here's What $25 Billion Buys.

How much does it cost to become the backbone of the AI economy? For Blackstone, a number is now on the table: $5 billion in equity, and potentially $25 billion counting the debt financing stacked on top.

On May 19, 2026, Blackstone and Google announced a joint venture to create a new U.S.-based AI cloud company. Blackstone takes a majority stake. The new company will give organizations access to Google's proprietary Tensor Processing Units, or TPUs, as a compute-as-a-service offering. It's the first time that level of TPU access will be available outside of Google's own cloud platform.

Big tech companies are expected to collectively spend around $700 billion on AI infrastructure in 2026 alone. This joint venture is one piece of how that spending materializes into actual capacity.

What the New Cloud Company Will Actually Do

TPUs are chips Google designed from the ground up for AI. They're not adapted from graphics processors the way most of the industry's AI chips started. Google has been building and deploying TPUs internally for more than a decade. They power Gemini, Google's consumer AI products, and workloads for many of the world's top AI labs. The chips are optimized for the specific matrix math that underlies modern deep learning, which makes them faster and more power-efficient than general-purpose GPUs for the tasks they're built for.

The key difference from NVIDIA's H100 and newer chips, which dominate the current AI cloud market, is specialization. NVIDIA GPUs are flexible. They run any workload reasonably well, which is why most AI developers default to them. TPUs are narrower. They're purpose-built for transformer-based architectures, the kind that power most modern large language models and multimodal AI systems. For those specific workloads, TPUs run faster per watt and cost less to operate at scale. For workloads that sit outside the transformer paradigm, they're not the right tool.

Google has demonstrated the efficiency case at enormous scale. Most of the company's own AI inference, including the Gemini models used by billions of people, runs on TPUs rather than GPUs. The joint venture is essentially a bet that what Google does internally can be productized for outside customers at a comparable level of efficiency.

The new company will offer TPU capacity through a compute-as-a-service model, meaning customers pay for access to compute rather than buying hardware. The company handles data center space, power, cooling, networking, and operations. Customers get access to TPU pods sized to their workloads without needing to source or manage the physical infrastructure.

The first 500 megawatts of capacity is targeted for 2027. That number carries weight. A single large hyperscale data center typically consumes 100 to 300 megawatts of power. Reaching 500 MW in roughly 18 months from announcement requires moving fast on land acquisition, power agreements, construction, and hardware procurement simultaneously. It's an aggressive timeline that reflects how intense demand for AI compute has become.

Blackstone will hold a majority stake. Google contributes TPU infrastructure, technology, and expertise. The structure gives Blackstone durable exposure to AI compute demand without requiring it to build chip research or fabrication capabilities of its own. It gives Google a revenue channel for TPU capacity that doesn't depend on Google Cloud being the only path to market. According to Blackstone's press release, the company will be U.S.-based and its name hasn't been publicly disclosed yet.

Benjamin Treynor Sloss will lead the company as CEO. He spent more than two decades at Google, most recently as chief programs officer. Before that role, he built and ran Google's Site Reliability Engineering function. SRE is the engineering discipline that defines how large-scale distributed systems stay available and reliable under real-world load. Treynor Sloss is credited with inventing and formalizing it, and it's now standard practice across the tech industry.

That background matters here. Running a compute-as-a-service company at 500 MW scale isn't primarily a sales or product challenge. It's an operations challenge. Customers paying for AI compute need consistent uptime and reliable performance when they're running a 10,000-chip training job with a hard deadline. Treynor Sloss has spent his career solving exactly that kind of problem, at exactly the scale this company needs.

Google Cloud CEO Thomas Kurian said the partnership would "help meet growing demand for TPUs by offering organizations additional ways to access computing capacity." That framing presents the joint venture as additive to Google Cloud rather than a competitor to it, which makes strategic sense. Most potential customers of the new company already have some relationship with Google Cloud or Google's broader ecosystem.

Blackstone's $5 Billion Infrastructure Bet

Private equity firms don't announce $5 billion equity commitments casually. Blackstone's president and COO, Jon Gray, described the opportunity as "generational," citing "unprecedented demand for compute" as the foundation of the thesis. Jas Khaira, who leads Blackstone's infrastructure investment platform known as BXN1, called Google's TPUs "foundational to the AI economy" and exactly the kind of proven platform his team was built to back at scale.

Blackstone is not a tech company. It's an asset manager with deep expertise in physical infrastructure: data centers, logistics, energy, and real estate. That's precisely why this structure works. Building AI data centers at the scale this venture requires is fundamentally an infrastructure problem, not a software problem. You need land, power contracts, cooling systems, fiber, and the operational discipline to keep all of it running reliably. Blackstone has executed infrastructure builds at this scale many times before.

The compute-as-a-service model also fits how Blackstone evaluates long-term infrastructure assets. Rather than owning hardware that depreciates on a fixed schedule, the model generates recurring revenue from customers who need sustained access to AI compute. Enterprises training large models or running inference at scale don't make a one-time purchase. They need ongoing access to capacity. That recurring revenue profile is exactly what infrastructure investors have found attractive about the cloud sector for the past decade, and AI compute is increasingly where that dynamic is playing out most intensely.

This deal differs from Blackstone's prior data center investments. It's not acquiring existing capacity. It's building something new around a chip technology with a decade of proven production history. That reduces technology risk considerably. TPUs aren't an experiment. Google has used them to train and run some of the most capable AI systems in production today. The question isn't whether TPUs work at scale. The question is whether the market will pay for dedicated TPU access outside of Google Cloud's standard offerings. Blackstone and Google are betting it will.

The broader market context makes that bet look reasonable. AI compute demand is outpacing what hyperscalers can provision. CoreWeave built a multi-billion-dollar GPU cloud almost entirely on that gap, serving AI labs and enterprises that couldn't get capacity from AWS, Azure, or Google Cloud fast enough. CoreWeave went public earlier this year. The Google-Blackstone venture enters the same demand dynamic, but with a differentiated technology play built on chips that have been in production for over a decade.

The AI Cloud Market Now Has a Serious TPU Option

The current AI cloud market is dominated by NVIDIA hardware. Most GPU cloud providers, CoreWeave and Lambda included, run NVIDIA's chips because they're flexible, well-supported, and broadly compatible with the software tools most AI teams already use. TPUs require software written for Google's ecosystem, historically centered on JAX and Google's frameworks, rather than the PyTorch-native tooling that most teams have built around.

That gap is narrowing as large language model architectures standardize. When most production inference involves running a transformer of a known shape at consistent batch sizes, a chip purpose-built for that task accumulates efficiency advantages over time. Google has demonstrated this internally. The joint venture is a bet that the market will follow the same pattern. As more AI teams mature beyond prototype and into sustained production inference, the case for a specialized chip optimized for that exact workload gets stronger.

For enterprises evaluating their compute strategy, the new company creates an option that didn't previously exist. Organizations whose workloads align with what TPUs do well can now access dedicated capacity outside of Google Cloud's standard queues, with contract structures and service levels tailored to their needs. That's particularly relevant for AI labs and enterprises that need reserved capacity guarantees rather than spot-market access.

Switching costs in AI compute infrastructure are real. A team that builds its training pipeline on TPUs and runs inference on TPU pods isn't going to migrate quickly or cheaply. That's one reason the 2027 capacity target matters tactically. The customers who get in early tend to stay, and the infrastructure choices organizations make now will shape their AI stacks for years. Both Google and Blackstone have strong incentives to move fast and lock in early adopters.

AWS, Azure, and Google Cloud will continue to dominate the hyperscale market. Specialized providers like CoreWeave and this new venture will compete for customers who need capacity the hyperscalers can't always deliver on time, or who want specific chip advantages for their workloads. The AI infrastructure market is large enough for multiple winners, but the distribution of workloads across hardware types will depend on how quickly the TPU software ecosystem matures and whether the efficiency case holds as model architectures keep evolving.

For context on how the AI infrastructure market is developing right now, the AIntelligenceHub AI Infrastructure guide covers compute access, cloud selection, and the capacity constraints shaping real deployments in 2026.

The physical scale of what's being built is already creating strain on power grids. AI data centers drove a 76% power surge on America's largest grid, underlining how much physical infrastructure the AI economy actually requires. The Google-Blackstone venture is a large-scale, well-capitalized attempt to build more of that infrastructure faster, and to do it with a chip technology Google believes will define the next decade of AI compute. The $25 billion total commitment signals that Blackstone believes it too.