TPU Shipment Projections: A Heated Debate

Mizuho's June 2026 channel survey forecast Google TPU shipments at approximately 4.3 million units in 2026, scaling to over 35 million units by 2028 — roughly an eightfold increase . This figure has become a lightning rod for debate. Other estimates are more conservative: Morgan Stanley projects ~5 million in 2027 and ~7 million in 2028, while Wolfe Research sees ~7 million annually by end of 2028 . The Mizuho 35M figure has been openly questioned by industry watchers as 2.3x the broader industry's total forecast and "seriously detached from supply chain reality" . TrendForce itself issued a more tempered estimate in January 2026, projecting 5 million units in 2027 and 7 million in 2028 . The actual trajectory will be a critical indicator of Google's ability to execute.

The Custom Chip Market Inversion

Industry-wide data from TrendForce shows custom AI ASIC shipments growing 44.6% year-over-year in 2026, nearly triple the 16.1% growth rate projected for merchant GPUs — the first inversion of the AI era . The broader custom AI accelerator market is projected to surge roughly 45% in 2026 and is on a path to $118 billion by 2033 . This structural shift directly benefits Google's TPU strategy. Broadcom, Google's long-time ASIC design partner, is projected to capture roughly 60% of the custom AI accelerator market by 2027 .

Diversifying the Supply Chain: Intel and Marvell

Google is reducing its dependence on a single foundry and design partner. The company has reportedly tapped Intel to manufacture more than 3 million TPUs in 2028, according to reporting from The Information cited by Reuters and TrendForce . This marks a strategic diversification away from sole reliance on TSMC and Broadcom for manufacturing. Additionally, Google has held talks with Marvell Technology for co-designed AI chips, intensifying competition in the inference chip market . This would add Marvell as another design partner alongside Broadcom and MediaTek, potentially building a four-partner supply chain .

TorchTPU: The Attack on CUDA's Moat

At Cloud Next 2026 (April 24), Google officially announced TorchTPU, a software layer that lets PyTorch workloads run natively on TPUs without the legacy PyTorch/XLA bridge . This is a direct challenge to Nvidia's CUDA ecosystem — the primary software moat that locks developers into Nvidia hardware. Google is collaborating with Meta (PyTorch's steward) on the project . The approach aims to lower switching costs for developers and break Nvidia's programming model lock-in. With PyTorch commanding an estimated 85% of deep learning research, making it a first-class citizen on TPUs could fundamentally alter the competitive landscape .

The Unverified $35 Billion Deal

Not verified. The specific claim of a "$35 billion deal involving Google, Broadcom, and Anthropic financed by Apollo and Blackstone" did not surface in the available search results. Evidence exists for Google's $3.2 billion TeraWulf backstop and Broadcom's role as Google's TPU ASIC design partner, but the $35B deal figure was not found in the sources retrieved. This detail should be treated as unconfirmed until a reliable primary source is located.

TPU Shipment Projections: A Heated Debate

Mizuho's June 2026 channel survey forecast Google TPU shipments at approximately 4.3 million units in 2026, scaling to over 35 million units by 2028 — roughly an eightfold increase . This figure has become a lightning rod for debate. Other estimates are more conservative: Morgan Stanley projects ~5 million in 2027 and ~7 million in 2028, while Wolfe Research sees ~7 million annually by end of 2028 . The Mizuho 35M figure has been openly questioned by industry watchers as 2.3x the broader industry's total forecast and "seriously detached from supply chain reality" . TrendForce itself issued a more tempered estimate in January 2026, projecting 5 million units in 2027 and 7 million in 2028 . The actual trajectory will be a critical indicator of Google's ability to execute.

The Custom Chip Market Inversion

Industry-wide data from TrendForce shows custom AI ASIC shipments growing 44.6% year-over-year in 2026, nearly triple the 16.1% growth rate projected for merchant GPUs — the first inversion of the AI era . The broader custom AI accelerator market is projected to surge roughly 45% in 2026 and is on a path to $118 billion by 2033 . This structural shift directly benefits Google's TPU strategy. Broadcom, Google's long-time ASIC design partner, is projected to capture roughly 60% of the custom AI accelerator market by 2027 .

Diversifying the Supply Chain: Intel and Marvell

Google is reducing its dependence on a single foundry and design partner. The company has reportedly tapped Intel to manufacture more than 3 million TPUs in 2028, according to reporting from The Information cited by Reuters and TrendForce . This marks a strategic diversification away from sole reliance on TSMC and Broadcom for manufacturing. Additionally, Google has held talks with Marvell Technology for co-designed AI chips, intensifying competition in the inference chip market . This would add Marvell as another design partner alongside Broadcom and MediaTek, potentially building a four-partner supply chain .

TorchTPU: The Attack on CUDA's Moat

At Cloud Next 2026 (April 24), Google officially announced TorchTPU, a software layer that lets PyTorch workloads run natively on TPUs without the legacy PyTorch/XLA bridge . This is a direct challenge to Nvidia's CUDA ecosystem — the primary software moat that locks developers into Nvidia hardware. Google is collaborating with Meta (PyTorch's steward) on the project . The approach aims to lower switching costs for developers and break Nvidia's programming model lock-in. With PyTorch commanding an estimated 85% of deep learning research, making it a first-class citizen on TPUs could fundamentally alter the competitive landscape .

The Unverified $35 Billion Deal

Not verified. The specific claim of a "$35 billion deal involving Google, Broadcom, and Anthropic financed by Apollo and Blackstone" did not surface in the available search results. Evidence exists for Google's $3.2 billion TeraWulf backstop and Broadcom's role as Google's TPU ASIC design partner, but the $35B deal figure was not found in the sources retrieved. This detail should be treated as unconfirmed until a reliable primary source is located.