Google's TPU Strategy Shift: Selling Custom AI Chips to Neoclouds

For nearly a decade, Google kept its Tensor Processing Units (TPUs) as a proprietary advantage — available only through Google Cloud as a managed service since 2018. That changed in 2025–2026. At Google Cloud Next '26 (April 22, 2026), Google for the first time explicitly stated it would sell TPUs externally to selected third-party providers . On April 29, 2026, Alphabet's parent company formally announced its intention to sell TPUs to select clients who would incorporate the chips into their own data centers . This marks a strategic pivot toward competing with NVIDIA by expanding TPU distribution beyond Google's own cloud walls .

The Fluidstack / TeraWulf Deal: $3.2B Backstop and 14% Stake

This complex deal is often conflated. Here are the facts from authoritative sources:

• August 2025: TeraWulf signed 200+ MW, 10-year hosting agreements with Fluidstack, an AI cloud provider. Google backstopped $1.8 billion of Fluidstack's lease obligations to support project debt financing and received warrants to acquire approximately an 8% equity stake in TeraWulf .
• August 18, 2025: Google added an additional $1.4 billion in backstop support, bringing its total financial commitment to approximately $3.2 billion. In return, Google obtained additional warrants, elevating its total stake in TeraWulf to 14%, making it the largest shareholder .
• Crucial clarification: The $3.2 billion backstop and 14% equity stake are in TeraWulf (a bitcoin miner converting to AI hosting), not in Fluidstack directly. Google struck a deal with Fluidstack to host TPUs at TeraWulf's Lake Mariner, NY facility, and the backstop secures Fluidstack's lease obligations .
• Fluidstack also has two 10-year hosting agreements with TeraWulf totaling $6.7 billion in contracted revenue starting in 2026, supported by Google's financial guarantee .

Crusoe and CoreWeave: Approaches Without Firm Deals

Google approached Crusoe and CoreWeave about hosting TPUs in their data centers, as reported by The Information in September 2025 . However, no publicly announced finalized deal terms were found for Crusoe. More significantly, CoreWeave publicly stated it does not expect to deploy Google's TPUs any time soon, along with Nebius and Lambda .

The Blackstone Joint Venture: $5B for 500 MW of TPU Capacity

On May 18, 2026, Blackstone and Google announced a joint venture to create a new U.S.-based company offering Google Cloud TPUs as compute-as-a-service . Key terms:

• Blackstone's commitment: $5 billion in initial equity
• Capacity target: 500 MW of TPU capacity online by 2027
• Google's role: Supplies TPUs, software, and services
• The deal was described as "the largest single-day AI-infrastructure capital event" and the first major neocloud anchored by Google's TPUs
• Reports suggest the total venture value including leverage could reach ~$25 billion

Ironwood (TPUv7) Specs

Google's latest TPU generation, Ironwood (also documented as TPU7x), was announced in April 2025 at Google Cloud Next:

• 192 GB HBM capacity per chip — confirmed in official Google Cloud documentation
• 6x increase in HBM capacity over the prior Trillium generation (which had 32 GiB per chip)
• 7.37 TB/s HBM bandwidth per chip
• 9,216 chips per superpod with shared memory of 1.77 PB
• Dual-die architecture with 4,614 TFLOPS FP8 performance per chip
• 42.5 exaflops of FP8 compute at full superpod scale

Known TPU Customers

• Anthropic — Committed to deploying over one million TPU chips to support its Claude model
• Apple — Named in industry chatter as testing external TPU demand
• Meta — Mentioned in context of a multibillion-dollar rental deal with potential future purchases
• Safe Superintelligence (SSI) — Listed among customers, though specific deal terms were not detailed in the available sources

How This Compares to Microsoft and Amazon

• Microsoft — Its custom AI silicon (Maia) is used exclusively within Azure. Microsoft has not made Maia available to third-party cloud providers.
• Amazon — Its custom Trainium and Inferentia chips are available through AWS but are not sold or licensed to third-party cloud providers for independent deployment.
• Neither company has announced any equivalent strategy of selling custom AI chips to neoclouds, hosting them in third-party data centers, or creating a joint venture to offer their custom silicon as third-party compute-as-a-service.

Notable Uncertainty and Pushback

As of July 2026, major neoclouds Nebius, Lambda, and CoreWeave have all said they are unlikely to buy Google TPUs any time soon, creating a potential distribution bottleneck for Google's strategy .

Note on Morgan Stanley Projections: The search budget was exhausted before reaching the original Morgan Stanley report. The sources gathered reference supply chain analysts seeing "massive upward revisions for Broadcom/Google TPU orders" for 2026 and projections of TPU v6 physical sales beginning, but the specific Morgan Stanley projection of 7 million TPU units by 2028 generating $13 billion in new revenue was not confirmed in the sources captured. This specific figure requires separate verification from the original Morgan Stanley research note.