Google has just made its boldest move yet to break Nvidia's stranglehold on AI computing. At Cloud Next 2026 on April 22, the company announced two new custom chips, the TPU 8t for training and TPU 8i for inference, that deliver roughly three times the computing power of the previous generation while promising significantly better value per dollar spent. The announcement signals a fundamental shift in how hyperscalers are building AI infrastructure, and it comes at a moment when Nvidia controls more than 80% of the $400 billion AI accelerator market.

What Makes Google's New Chips Different From Nvidia's Approach?

The TPU 8t superpod, codenamed Sunfish and designed by Broadcom, packs 9,600 liquid-cooled chips to deliver 121 FP4 exaflops of peak compute power. To put that in perspective, that is roughly 2.85 times the computing capacity of Google's previous generation TPU, called Ironwood, which delivered 42.5 exaflops. The real advantage, however, isn't just raw speed. Google emphasizes price-performance, meaning the TPU 8t offers up to 2.8 times better training cost-efficiency than Ironwood, a metric that matters far more to companies running production AI workloads than peak computing speed alone.

The TPU 8i, codenamed Zebrafish and designed by MediaTek, takes a different approach entirely. Rather than chase training speed, it is engineered specifically for inference, the process of running a trained model to generate responses. Each TPU 8i delivers 10.1 FP4 petaflops of compute, fed by 288 gigabytes of high-bandwidth memory and 8.6 terabytes per second of memory bandwidth. A pod of 1,152 chips delivers 11.6 FP8 exaflops of compute, a 9.6 times increase over Ironwood's inference performance. More importantly, Google engineered the chip around latency, the time it takes to generate a response. A dedicated Collectives Acceleration Engine reduces on-chip collective latency by up to 5 times, a feature specifically designed for AI agents that make thousands of small tool-use calls.

How Do These Chips Compare to Nvidia's Blackwell?

On paper, Nvidia's B200 chip still edges out Google's TPU 8t on raw per-chip computing throughput, delivering roughly 20 petaflops of FP4 compute compared to the TPU 8t's 12.6 petaflops per chip. But Google wins decisively at scale. Nvidia's NVLink interconnect, which connects multiple GPUs together, caps systems at 72 chips before performance drops significantly due to switching overhead. Google's proprietary interconnect fabric scales to 9,600 chips in a single logical cluster without that penalty. For trillion-parameter model training, that scale advantage compresses wall-clock training time by an estimated 35 to 45 percent relative to equivalently priced Nvidia systems.

The inference story is even more lopsided. Google claims the TPU 8i delivers 80 percent better inference performance per dollar than Ironwood, a metric that compounds with every Gemini query, YouTube recommendation, and Anthropic Claude request running through Google's data centers. That efficiency advantage directly translates to lower operating costs for hyperscalers running inference at massive scale.

Why Is Anthropic's Commitment to Google's Chips So Important?

Anthropic, the AI company behind Claude, is the anchor customer validating Google's silicon roadmap. In October 2025, Anthropic committed to use up to 1 million Ironwood TPUs, with access to more than 1 gigawatt of computing power in 2026 and 3.5 gigawatts by 2027. The deal, structured as a multi-year Google Cloud commitment reportedly valued in the tens of billions of dollars, makes Anthropic the single largest customer of Google's custom silicon and the only external company with guaranteed access to TPU 8t when it ships in late 2027.

This commitment is a lifeline for Google Cloud, which still trails Amazon Web Services at 31 percent market share and Microsoft Azure at 24 percent in the broader cloud infrastructure market. Google Cloud's share hovered around 12 percent in the first quarter of 2026, but its TPU-driven AI workload growth has outpaced rivals significantly. AI compute revenue at Google Cloud grew 71 percent year-over-year in that quarter, and TPU utilization has run above 90 percent at every major data center region since January.

What Does This Mean for Nvidia's Market Position?

Nvidia's dominance in AI accelerators is not in immediate danger. The company still controls more than 80 percent of the $400 billion AI accelerator market in 2026, and its CUDA software ecosystem, built over 20 years, remains the industry standard. However, Google's announcement signals that the era of general-purpose AI accelerators is ending. Broadcom, the design partner for TPU 8t, is projected to generate $21 billion in Google and Anthropic AI revenue in 2026 and $42 billion in 2027, according to estimates cited in Google's Cloud Next briefing. That represents a meaningful redistribution of silicon contracts away from Nvidia.

The structural shift also matters. By splitting its silicon strategy into training and inference chips, Google is signaling that one-size-fits-all accelerators no longer make sense at hyperscale. Nvidia's B200 is a general-purpose chip designed to handle both workloads reasonably well. Google's approach, by contrast, optimizes each chip for its specific job. That specialization allows Google to deliver better price-performance in both categories, even if Nvidia's raw per-chip throughput remains competitive.

How to Understand the Supply Chain Implications

• Design Partnership Shift: Broadcom continues its relationship with Google as the design partner for TPU 8t, while MediaTek, best known for smartphone chips, has emerged as the design partner for TPU 8i. This dual-partner model signals that Broadcom's near-monopoly on hyperscaler AI custom silicon has ended, opening opportunities for other semiconductor design firms.
• Manufacturing Concentration: Both TPU 8t and TPU 8i will be fabricated by Taiwan Semiconductor Manufacturing Company on its 2-nanometer process, with late 2027 availability targeted for customers outside Google. This concentrates manufacturing risk in a single foundry but ensures access to the most advanced process technology available.
• Broadcom's Revenue Opportunity: Broadcom's ASIC content per TPU 8t exceeds $4,000, meaning a single 9,600-chip superpod carries more than $38 million in Broadcom silicon alone. This represents a substantial revenue stream for the company and demonstrates how custom silicon creates new opportunities for semiconductor suppliers beyond Nvidia.

The announcement also reveals a broader competitive dynamic. Meta has a standing TPU rental arrangement with Google Cloud, an unusual concession from Mark Zuckerberg's company given Meta's own $35 billion MTIA custom silicon deal. That Meta is willing to rent Google's TPUs alongside developing its own chips suggests that even companies building custom silicon recognize the value of having multiple sources of AI accelerators.

When Will These Chips Actually Be Available?

Both TPU 8t and TPU 8i are targeting late 2027 availability for customers outside Google. That timeline is important because it means Google has roughly 18 months to validate the chips internally before external customers can deploy them at scale. Anthropic's 1 million TPU commitment provides a massive internal test bed, but external validation from other hyperscalers will be crucial to establishing these chips as a credible Nvidia alternative.

The delay also reflects the engineering complexity of scaling custom silicon to this level. Google's announcement stresses that the TPU 8t's networking fabric supports near-linear scaling to one million chips in a single logical cluster. Achieving that level of scalability requires solving problems that Nvidia solved decades ago with CUDA. Google is essentially building a new software ecosystem from scratch, which takes time.

What's clear is that Nvidia's unchallenged dominance in AI accelerators is ending. Google's TPU 8t and 8i represent the most credible challenge to Nvidia's market position in years, backed by the computing power of one of the world's largest hyperscalers and validated by Anthropic's massive commitment. Whether these chips will actually dent Nvidia's market share depends on execution, software maturity, and whether other hyperscalers follow Anthropic's lead. But the strategic message is unmistakable: custom silicon is no longer a niche play. It is the future of AI infrastructure.