Google and Blackstone's $25 Billion AI Cloud Bet: A TPU Empire Built on Private Equity Money

On May 18, 2026, Blackstone and Google announced a joint venture to create a new, as-yet-unnamed U.S.-based company that will sell Google's Tensor Processing Units — custom AI chips purpose-built for training and running machine learning models — as a compute-as-a-service offering [1]. Blackstone is committing $5 billion in equity capital, with total investment reaching approximately $25 billion when leverage is included [2]. The company expects to bring 500 megawatts of data center capacity online by 2027, with plans to scale from there.

The venture will be led by Benjamin Treynor Sloss, a Google executive with more than two decades of experience building and operating Google's global infrastructure [1]. Google will supply the hardware — its TPU chips — along with software and services, while Blackstone provides the capital and is expected to hold a majority ownership stake [3].

The deal lands at a moment when the four largest hyperscalers are collectively spending more than $330 billion on capital expenditure in 2025 alone, and when questions about whether AI infrastructure investment has outrun actual demand are growing louder.

The Deal Structure: Who Controls What

The ownership arrangement places Blackstone as the majority financial stakeholder, with Google contributing technology rather than equity [3]. This structure echoes but meaningfully differs from other recent tech-finance partnerships.

Source: Company Earnings Reports

Data as of Oct 1, 2025CSV

The closest precedent may be Microsoft's relationship with OpenAI, though the comparison has limits. Microsoft invested billions and held roughly 27% of OpenAI on an as-converted diluted basis following OpenAI's recapitalization, with governance rights "generally limited to typical financial investor protections" and no board appointment rights [4]. The Google-Blackstone arrangement flips this dynamic: the financial partner holds the majority stake, while the technology partner supplies chips and services without a reported equity position.

Another reference point is Blackstone's own 2021 acquisition of QTS Realty Trust for $10 billion, which gave the private equity firm control of the largest independent data center operator in the world [5]. QTS's development pipeline has grown more than 25 times since that acquisition as hyperscalers race to expand AI capacity [5]. The new Google venture appears designed to extend this playbook — but with Google's proprietary TPUs rather than commodity server hardware.

The appointment of Treynor Sloss as CEO suggests Google intends to maintain significant operational influence regardless of equity split. Treynor Sloss is widely credited inside Google with building the company's Site Reliability Engineering practice and has overseen global infrastructure operations for years [6].

Why Google Needs Blackstone's Money (Or Does It?)

Google Cloud posted $15.2 billion in revenue in Q3 2025, up 34% year-over-year, with a backlog that surged 46% quarter-over-quarter to $155 billion in reserved AI compute contracts [7]. Full-year 2024 revenue reached approximately $41 billion, and analysts project $58 billion by 2026 [7].

Source: Alphabet SEC Filings

Data as of Oct 1, 2025CSV

Alphabet has already committed $93 billion in capital expenditure for 2025 — up from an initial projection of $75 billion at the start of the year — with two-thirds flowing to GPUs, TPUs, and servers [8]. Google also announced a separate $25 billion investment to build out data center and AI infrastructure across the PJM Interconnection power grid spanning 13 states [9].

So why does a company spending $93 billion a year on infrastructure need a partner to put up $5 billion?

The answer likely involves balance sheet management. By routing new TPU capacity through a Blackstone-controlled entity, Google can monetize its chip technology — collecting hardware supply fees and service revenue — without booking the data center construction as its own capital expenditure. For a company whose investors are already uneasy about the sheer scale of AI spending, offloading even a fraction of the buildout to private equity capital has clear appeal.

The structure also lets Google reach customers it might not otherwise serve efficiently. The new company is positioned to compete with CoreWeave and similar "neocloud" providers that sell AI compute to enterprises, AI labs, and financial firms that don't want to commit to a full hyperscaler relationship. CoreWeave's revenue backlog grew to $66.8 billion by the end of 2025, more than quadrupling over the year, and the company landed a $21 billion contract with Meta in April 2026 [10]. That market is real and growing fast.

Blackstone's Infrastructure Empire

The Google partnership is one piece of a much larger Blackstone infrastructure strategy. In July 2025, Blackstone announced it would invest more than $25 billion in Pennsylvania's digital and energy infrastructure, projecting over 6,000 jobs created or supported annually over an estimated 10-year construction timeline [11]. The firm also formed a joint venture with PPL Corporation to invest in natural gas power generation to supply electricity for data centers [11].

In May 2026, just days before the Google announcement, Blackstone launched the IPO for Blackstone Digital Infrastructure Trust (BXDC), a data center REIT targeting $1.75 billion in proceeds. BXDC plans to acquire stabilized, newly-built data centers leased to investment-grade hyperscalers under 10-to-20-year triple-net leases with 2-3% annual rent escalations [12]. The projected property yields range from 5.75% to 7% per year [12].

Blackstone has invested in data center assets worth more than $130 billion since 2018 [12]. The firm's private funds target internal rates of return (IRRs) of 15% or higher on greenfield development, while stabilized assets are channeled into vehicles like BXDC at lower but steadier yields [13].

For a private equity firm accustomed to 5-to-7-year exit horizons, the Google TPU venture offers a plausible path: build out capacity with cheap leverage, sign long-term compute contracts with enterprise and government clients, then potentially contribute the stabilized assets to a public vehicle like BXDC or sell to infrastructure investors at a premium.

Who Bears the Risk

The $5 billion equity commitment comes from "funds managed by Blackstone" [1] — meaning the capital originates from Blackstone's limited partners: pension funds, sovereign wealth funds, endowments, and high-net-worth individuals. If the venture underperforms, those LPs absorb the losses, not Blackstone's corporate balance sheet.

This distinction matters. As the Private Equity Stakeholder Project has noted, public pension funds increasingly shoulder risk in private equity infrastructure bets that are tied to AI's projected demand growth [14]. If that demand doesn't materialize at the pace underwriting the investment, ratepayers and retirees whose pensions allocated to Blackstone infrastructure funds face the consequences.

Google's risk, by contrast, appears more contained. The company supplies hardware and services — a revenue stream regardless of whether the venture's end customers generate the hoped-for returns. Specific contractual protections, minimum revenue guarantees, or put options in the deal have not been publicly disclosed.

The Antitrust Question

Google is already under significant regulatory pressure. In April 2026, the DOJ finalized remedies from its landmark search monopoly case, prohibiting Google from entering exclusive distribution contracts for Search, Chrome, and its AI products [15]. The remedies explicitly extended to GenAI products — a signal that regulators view Google's AI ambitions through the lens of its established market power [15].

Separately, the DOJ cleared Google's $32 billion acquisition of cybersecurity firm Wiz in November 2025 [15], suggesting that antitrust scrutiny is applied case by case rather than as a blanket block on Google expansion.

The Blackstone joint venture creates an interesting structural dynamic. Because Blackstone holds the majority stake and an independent CEO runs the company, it is not a Google subsidiary. Regulators could view this as reducing Google's direct market concentration in AI compute — the new company competes with Google Cloud, at least in theory. Alternatively, the FTC or DOJ could argue that Google's exclusive supply of TPU hardware and software creates a de facto control arrangement that the equity structure merely obscures.

No regulatory challenge to the venture has been publicly signaled as of the announcement date. But given the current administration's stated commitment to competitive AI markets [15], the structure will likely receive scrutiny if the company grows to significant scale.

The Target Market: Who Will Buy This

The venture's compute-as-a-service model aims squarely at the growing market for AI inference — the process of running trained models to generate predictions, text, images, or other outputs. Google's TPUs reportedly deliver 4x better performance-per-dollar for inference compared to Nvidia GPUs [16], a claim that, if accurate, gives the venture a compelling cost advantage.

Potential anchor clients include AI laboratories that currently rent GPU capacity from CoreWeave or hyperscalers, capital markets firms running latency-sensitive trading models, federal agencies pursuing sovereign AI compute, and regulated industries — healthcare, finance, defense — that need dedicated infrastructure with specific compliance guarantees.

Google has noted that its TPUs already "power workloads for many of the world's top AI labs and capital market firms" [1]. The joint venture would extend that reach to customers who want TPU access without a full Google Cloud commitment.

On jobs, Blackstone's broader Pennsylvania infrastructure push projects over 6,000 positions annually over a decade of construction [11]. The Google-specific venture's job creation figures have not been separately disclosed, but 500 MW of data center capacity typically requires several thousand construction workers during buildout and hundreds of permanent operations staff.

The Bear Case: Overbuilding in an Era of Falling Costs

The most pointed criticism of deals like this comes from the supply side. AI inference costs have collapsed: the cost of running a model matching GPT-3.5 performance dropped from $20 per million tokens in November 2022 to $0.07 in October 2024 — a 280x decrease [17]. Nvidia H100 cloud pricing fell 64-75% between Q4 2024 and Q1 2026 [17]. Hardware costs have declined roughly 30% annually while energy efficiency improves about 40% per year [17].

Enterprise GPU utilization rates tell a stark story. VentureBeat reported that enterprise GPU utilization sits at approximately 5% — meaning companies locked in capacity during the AI scramble that they are not actually using [18]. Even industry benchmarks for efficient operations target only 65-75% average utilization, with inference workloads typically achieving 40-50% due to request variability [18].

Goldman Sachs projects data center demand to grow roughly 50% to 92 GW by 2027 [19]. But efficiency improvements — liquid cooling, model specialization, AI-driven optimization — could reduce infrastructure requirements and contribute to overbuilding risk in some locations [19]. By 2027, inference workloads are expected to become the dominant AI compute requirement, and the economics of inference favor efficiency gains over raw capacity additions [19].

The bear case, then, is straightforward: if inference costs continue falling at 10x per year, and if efficiency gains reduce the compute required per query, then $25 billion in new capacity could arrive into a market where supply exceeds demand. The counterargument — made by Google, Blackstone, and their peers — is that falling costs will expand the addressable market faster than efficiency shrinks it, a dynamic that held true for cloud computing, bandwidth, and storage over the past two decades.

What This Means

The Google-Blackstone joint venture is best understood as three things simultaneously: a product company (selling TPU compute-as-a-service), a financial vehicle (letting Google monetize chip IP while Blackstone deploys LP capital into long-duration infrastructure), and a competitive response (matching CoreWeave and similar players who have proven there is enterprise demand for dedicated AI compute outside the traditional hyperscaler model).

Whether the $25 billion bet pays off depends on variables that neither company fully controls: the trajectory of AI inference demand, the pace of hardware efficiency gains, the regulatory environment, and whether customers will pay a premium for TPU-specific performance advantages over increasingly cheap commodity GPU capacity.

For Google, the downside is limited — it collects hardware and services revenue regardless. For Blackstone's limited partners, the stakes are higher. Their capital is underwriting a bet that AI compute demand will grow faster than AI compute costs fall. That bet has been correct for the past three years. Whether it remains correct through 2030 is the $25 billion question.