Consider the connection between computation and commerce. Alphabet Inc.’s custom Tensor Processing Units (TPUs) have graduated from being a purely internal accelerator for AI workloads to a cornerstone of its cloud growth, competitive differentiation, and financial resilience. What began as a quiet program in 2015 and was publicly revealed in 2016 24,28 has now reached an inflection point: the eighth-generation chips, unveiled at Google Cloud Next in April 2026, introduce a decisive split into training-optimized (TPU 8t) and inference-optimized (TPU 8i) variants 3,4,8,25,33. This design pivot is expressly tailored for the “agentic era,” delivering superior price-performance and lower energy consumption for specific AI tasks 5,12,16,20. Production ramps are underway 53, and the company projects a demand for 35 million TPUs annually by 2028 35, with orders for over 3 million units from Intel for that year alone 18,19,21,34,51.
A watershed moment arrives with Anthropic’s commitment to use up to one million Google Cloud TPUs, with over one gigawatt of capacity online in 2026, in a deal valued at tens of billions of dollars 10,23,31. Expansive as that is, an additional 3.5 gigawatts of TPU capacity will come online starting in 2027 through Broadcom 2,6,23. This engagement—projected at $200 billion over five years—validates TPUs as a credible alternative to Nvidia GPUs for frontier model training 10,31. Meanwhile, Alphabet is now selling TPUs directly to enterprise customers for use in their own data centers, a shift from its prior cloud-only access model 7,8,11,14,25. Initial revenue is expected in late 2026, with the majority in 2027 36. A new TPU Developer Hub provides architectural guidance, software stacks, and optimization tooling, supporting both bare-metal and cloud consumption 24,27, while Ubuntu TPU support moves upstream 13.
Competitively, the landscape is intense. Alphabet’s TPUs compete directly with Nvidia GPUs, and internal strategy aims to replace Nvidia hardware where possible 32,52. TPUs are seen as more power-efficient and profitable 9, with some claims asserting ~4x price-performance advantage over H100 GPUs for certain inference tasks 23. Yet Google’s dual approach—developing TPUs while still using Nvidia GPUs—offers flexibility 17. Meanwhile, Meta Platforms’ nascent cloud ambition introduces both competition and dependency: Meta plans to launch “Meta Compute” to sell excess GPU capacity 30,47,49,54,55,56, directly challenging Google Cloud 42,43,44, but it also uses Google’s Gemini models and TPU clusters 29,37, and is reportedly considering purchasing TPUs directly 1,50. This intricate web suggests that while Meta could become a formidable rival, Alphabet may benefit as a supplier, and the Anthropic deal provides a massive, sticky revenue stream.
Financially, Morgan Stanley projects 4 GW of first-party TPU sales by 2028 40,41, with gross margins on Anthropic TPU sales potentially exceeding 40% 45. TPU sales have already driven a 63% year-over-year growth in Google Cloud 14, and expanded capacity is expected to improve unit economics across Search, Cloud, and other services 24, while acting as a hedge against supply chain volatility 24. Challenges remain: the concurrent ramps of TPU V8, Amazon Trainium4, and Meta MTIA 400 may strain a limited supplier base 38, and ecosystem dependency risks customer lock-in 24.
The Decade-Long Ascent from Proprietary Engine to Cloud Foundation
The TPU program’s evolution reflects a deliberate, multi-year arc. From its 2015 inception and 2016 public debut 24,28, each generation has refined performance and efficiency. The eighth-generation leap is particularly instructive: the introduction of separate training and inference variants responds directly to the divergent demands of model building and serving in the agentic era 3,4,5,8,12,16,20,25,33. With volume production already ramping 53, the infrastructure required is staggering: Google is aiming for 35 million TPUs per year by 2028 35, and has already placed orders for over 3 million units from Intel for 2028 18,19,21,34,51. This scaling is underpinned by strategic manufacturing partnerships with TSMC, Intel, Broadcom, and MediaTek 15,26,32,39,53, and gigawatt-scale liquid cooling across more than 2,000 TPU Pods deployed over seven years 36.
The Anthropic Milestone: Validating an Alternative Architecture
The Anthropic deal represents a great leap in commercial validation. The commitment to use up to one million TPUs, with over one gigawatt of capacity brought online in 2026 10,23,31, and an additional 3.5 GW coming in 2027 via Broadcom 2,6,23, demonstrates the capacity and trust the market places in TPU infrastructure. Valued at $200 billion over five years 10,31, this arrangement not only secures a massive revenue stream but also establishes TPUs as a credible alternative to the Nvidia GPU ecosystem for training frontier models. Alphabet’s ability to deliver at this scale reflects its robust supply chain architecture and large-scale cooling deployments 36.
Enterprise Direct Sales and the Ecosystem Play
The decision to sell TPUs directly to enterprises for on-premises use marks a fundamental shift from Google’s traditional cloud-only model 7,8,11,14,25. By offering TPU capacity and maintenance without transferring physical hardware in most deals 22, Alphabet preserves its intellectual property while embedding itself into hybrid cloud architectures. The launch of the TPU Developer Hub further lowers barriers, providing architectural guidance, software stack coverage (JAX, TensorFlow, XLA, Kubernetes, Borg), and optimization tooling 24,27. Support for bare-metal and cloud consumption modes 27 and upstream Ubuntu support 13 signal an intent to broaden adoption across diverse environments. This externalization is projected to generate initial revenue in late 2026 and scale significantly into 2027 36.
Competitive Dynamics: Nvidia Rivalry and Meta’s Ambiguous Co-opetition
TPUs are positioned as a direct competitor to Nvidia GPUs, with Google’s internal strategy explicitly targeting replacement where possible 32,52. Claims of ~4x price-performance advantage over H100 GPUs for certain inference tasks 23 underscore the architectural efficiencies possible. Yet Alphabet’s simultaneous use of Nvidia hardware 17 provides a dual-sourcing hedge and flexibility. The rise of Meta as a potential fourth hyperscaler 48 adds complexity. Meta’s “Meta Compute” cloud business is a direct competitive threat 30,42,43,44,47,49,54,55,56, but Meta’s reliance on Google’s Gemini models and TPU clusters 29,37 and reported interest in purchasing TPUs directly 1,50 create a co-opetitive dynamic. This could ultimately benefit Alphabet as both a cloud provider and component supplier, while the immense Anthropic commitment helps insulate Google’s cloud growth.
Financial Projections and Scaling Risks
The financial implications are substantial. Morgan Stanley projects 4 GW of first-party TPU sales by 2028 40,41, and gross margins on Anthropic TPU sales could exceed 40% 45. The 63% YoY growth in Google Cloud already attributed to TPU sales 14 indicates the revenue power of this strategy. Moreover, expanded TPU capacity is expected to improve unit economics across Alphabet’s broader service portfolio 24 and serve as a hedge against supply chain volatility 24. However, the industry-wide simultaneous ramp of TPU V8, Amazon Trainium4, and Meta MTIA 400 could strain a limited supplier base 38, and deep ecosystem lock-in 24 remains a watchpoint. Alphabet’s ability to navigate manufacturing scalability and integration complexity will determine whether these ambitious targets are met.
Implications for the Connected Future
Alphabet’s TPU program has crossed from internal differentiator to offensive cloud weapon, poised to capture a projected 27% market share by 2027 46. The Anthropic mega-deal and direct enterprise sales represent a multi-year revenue backlog that rivals traditional hyperscale capital commitments. The split into training and inference variants and the Developer Hub reflect a maturing product mindset aimed at diverse workloads. While competition from Nvidia and Meta looms, the co-dependence with Meta—as a potential TPU customer and Gemini user—creates a buffer. The ultimate test will be execution: scaling to 35 million units annually, managing supply constraints, and convincing enterprises to embrace a Google-dominated compute fabric. Much like the telephone network once became the nervous system of human conversation, these silicon architectures are becoming the nervous system of global AI commerce.
