At a glance

  • Google’s TPU chips can be rented outside Google Cloud for the first time, through a US$5 billion Blackstone joint venture targeting 500MW by 2027.

  • Blackstone already owns AirTrunk and, with Coatue, funds Firmus. One balance sheet now sits behind both the GPU and the TPU side of Australia’s AI build.

  • Anthropic, the largest external TPU customer at up to one million chips, has signed an Australian government MoU and is shopping for local data centre capacity.

  • Meta’s reported plan to resell spare GPU capacity knocked about 15% off neoclouds CoreWeave and Nebius on 1 July. The TPU venture adds supply to the opposite end of the compute market.

  • Whether Blackstone routes TPU capacity through AirTrunk’s Australian sites is the question that decides how fast a US deal becomes an Australian build.

A decade-old captive chip goes merchant

Google has run TPUs in production for more than a decade, but the only way to rent one was through Google Cloud. That changed on 18 May 2026, when Blackstone (NYSE: BX) and Google announced a joint venture creating a new US company that offers data centre capacity, networking and Google TPUs as compute-as-a-service. Google’s statement put it plainly: the venture “will give customers another option to access cloud TPUs in addition to using them through Google Cloud.”

Blackstone is committing US$5 billion of initial equity and expects to bring the first 500MW online in 2027, with the company saying it plans to scale significantly from there. Trade coverage puts the full build, including debt, near US$25 billion. Benjamin Treynor Sloss, the Google engineer behind its site reliability practice, runs the business as CEO. Google supplies hardware, software and services; Blackstone supplies capital and real estate, and holds a majority stake, on the Wall Street Journal’s reporting.

It was Blackstone’s second AI joint venture inside a month, after the firm joined Anthropic, Hellman & Friedman and Goldman Sachs on 4 May in an enterprise AI services company with commitments reported near US$1.5 billion. That puts one firm beside both the TPU’s designer and its largest disclosed customer.

The name writing the cheques already runs deep through Australia’s AI infrastructure.

One balance sheet now sits on both sides of the compute market

Blackstone manages over US$1.3 trillion and describes itself as the largest global provider of data centres. The Blackstone-led consortium, alongside Canada’s CPP Investments, acquired AirTrunk in a transaction valued above A$24 billion, completed in December 2024. The deal, the largest recorded in the sector, handed Blackstone the biggest hyperscale platform in Asia Pacific, with sites in Sydney, Melbourne and Canberra. We ranked AirTrunk first in our top APAC data centres analysis.

Blackstone’s Australian exposure runs past the landlord layer into the compute layer. The firm, with Coatue, leads the US$10 billion debt facility behind Firmus, the operator behind Project Southgate, which we have tracked across our neocloud market report. Firmus is a Nvidia GB300 build. The Google joint venture is a TPU build. The same financier now sits behind both the GPU lane and the TPU lane of the same market.

Blackstone position

Role in the stack

Compute type

Australian exposure

AirTrunk (with CPP Investments)

Hyperscale data centre platform

Chip-agnostic capacity

Sydney, Melbourne, Canberra and APAC sites; largest platform in the region

Firmus (US$10bn debt facility)

GPU neocloud build

Nvidia GB300

Project Southgate across Tasmania, Victoria and other states

Google TPU joint venture

Merchant compute-as-a-service

Google TPU

US-sited at launch; Australian relevance via Blackstone’s local assets

Source: Blackstone and Google joint venture announcement, 18 May 2026; AirTrunk transaction announcements, 2024; Firmus financing release, February 2026. Australian exposure column is Certified Strategic analysis.

A merchant TPU cloud needs land, power and operations, and Blackstone owns one of the largest pools of those assets in this region. The US announcement is an Australian one in waiting.

Anthropic anchors TPU demand, with Meta and Apple reported behind it

In October 2025 Anthropic committed to up to one million Google TPUs, an expansion valued in the tens of billions of dollars and expected to bring well over a gigawatt of capacity online in 2026, the largest external TPU commitment disclosed to date. In April 2026, on CIO Dive’s reporting, Google and Broadcom signed a further agreement with Anthropic adding multiple gigawatts of TPU capacity from 2027. Anthropic is not a party to the Blackstone venture, but for the financiers backing a standalone TPU cloud, commitments on that scale are the demand evidence. Meta is reported to be in talks to lease TPUs from 2026 and to buy systems from 2027, and Apple and Safe Superintelligence have been reported among TPU users. Broadcom co-designs the silicon behind all of it.

The chip roadmap is built for the faster-growing half of the market. McKinsey puts 2026 demand at about 31GW each for training and inference, with inference roughly tripling to 93GW by 2030 while training doubles to 62GW. Google has split its eighth generation to match: TPU 8t for large-scale training, TPU 8i for high-speed inference, both on TSMC’s 2nm process with general availability targeted for late 2027. The current generation, Ironwood, is Google’s first TPU positioned for inference at scale. Inference runs close to users, and latency pulls it onshore.

Several Nvidia-aligned neoclouds, including Nebius, Lambda and CoreWeave, say they have no immediate plans to adopt TPUs, and reporting suggests most accelerator demand still flows to Nvidia GPUs. A merchant ASIC channel existing is not the same as the market switching to it.

Meta’s resale plan lands on the other side of the split

On 1 July CNBC reported that Meta is building a business, internally called Meta Compute, to resell the GPU capacity it does not use. CoreWeave and Nebius fell about 15% in the session, and Australian-founded IREN fell about 5 to 6.5%, partly on a separate broker note, a repricing we covered in Meta’s move to become its own neocloud. Meta now sits on both sides of the compute trade, reported to be in talks to lease TPUs while preparing to rent out spare GPU cycles.

The two announcements add supply to opposite ends of a compute market that is splitting in two. Rental rates for the newest, contracted capacity are still rising while older, interchangeable GPU capacity gets cheaper, a divide we mapped in our analysis of the two compute markets. Meta’s plan, if it proceeds, would release existing GPU fleets into the rental market; the neoclouds repriced in a day. The Blackstone venture is financing new capacity that has yet to name a customer, a US$5 billion commitment priced on the view that frontier demand stays ahead of supply.

What it means for a neocloud tier built on GPUs

Australia’s neocloud tier is, by design, a Nvidia GPU tier. Firmus, Sharon AI, IREN, ResetData and Polaris built around GB300, Hopper and Blackwell silicon, aligned with the Nvidia Cloud Partner program and the software stack enterprise customers already target.

A merchant TPU channel introduces a second supply curve for the inference half of demand. For operators that lease capacity rather than pick chips, the building case is intact: TPU pods are dense and liquid-cooled like GPU pods, so power, land, cooling and connectivity remain the binding constraints whatever sits in the rack. For operators whose economics depend on a single accelerator’s pricing, a credible alternative introduces price competition that did not exist a year ago. The Australian independents mostly sell contracted, current-generation capacity, so the 1 July selloff moved share prices while the signed contracts underneath did not change.

One Australian operator already sits on the ASIC side of that line. SouthernCrossAI runs SambaNova inference silicon, with a sovereign Sydney node live from January 2026, and the Blackstone and Google venture validates the category it chose early.

How the options market is reading the shift

Situational Awareness LP, the fund run by former OpenAI researcher Leopold Aschenbrenner, disclosed in its Q1 2026 13F filing roughly US$8.5 billion in notional put options against the semiconductor layer, including about US$1.6 billion against Nvidia and further positions against Broadcom, AMD, TSMC and ASML, paired with long exposure to physical infrastructure and Nvidia-based neocloud equity. The thesis is a hedge against the whole accelerator layer rather than a wager on TPUs over GPUs; the Broadcom position even cuts against the TPU’s co-designer. The narrower signal is simpler: a credible second chip supplier weakens Nvidia’s pricing power, and the fund’s puts are a bet against that pricing power holding. Certified Strategic does not offer investment advice; the filing is noted as a market signal.

What to watch

Whether Blackstone routes any TPU capacity through AirTrunk’s Australian footprint is the first question, and the joint venture’s US-only framing at launch leaves it unanswered. Whether Google itself places Ironwood capacity onshore in Sydney or Melbourne also remains unconfirmed in current Google documentation.

A Blackstone-financed merchant model is one route by which TPU capacity could reach this market without Google building it directly. Google committed A$1 billion through its Digital Future Initiative in 2021, and its larger reported data centre ambition has stalled, covered in Google’s paused A$20 billion plan.

Demand is the variable that would settle it. Anthropic has signed an Australian government MoU and is running a process to secure Australian capacity, a contest we track in the contenders for Anthropic’s Australian compute contract. Where that demand physically lands, and on which silicon, is the signal that would turn this US joint venture into an Australian build.