What Google announced

At Google Cloud Next 2026 in Las Vegas, Google made two distinct announcements. It confirmed general availability of Ironwood, its seventh-generation TPU, the first purpose-built for what Google calls the “agentic era” of inference at scale. It also previewed its eighth-generation architecture: two purpose-built chips, TPU 8t for large-scale training and TPU 8i for high-speed inference, both targeting TSMC 2nm manufacturing and expected to reach general availability later in 2026.

A TPU is Google’s custom alternative to NVIDIA’s graphics processing unit (GPU). Where a GPU is a general-purpose workhorse, a TPU is a specialist chip built from the ground up for AI calculations. Google has been building them since 2016. The eighth generation is its most ambitious split yet, and the first time the company has designed separate chips for each half of the AI lifecycle.

The TPU 8t training pod reportedly delivers nearly three times the compute of an equivalent Ironwood pod, with double the performance per watt. The TPU 8i inference chip is designed to serve millions of AI agents simultaneously for enterprise customers.

That last part carries a structural implication. On a recent earnings call, CEO Sundar Pichai indicated that as TPU demand grows from AI labs, capital markets firms and high-performance computing applications, Google would begin delivering TPUs to select customers in their own data centres. Google is no longer content to keep its silicon advantage internal.

Anthropic’s compute strategy

Anthropic, the AI company behind Claude, has confirmed a major infrastructure deal with Google covering access to up to one million Ironwood TPU chips. The commitment is worth tens of billions of dollars and was formally announced by both companies.

Understanding that deal requires understanding Anthropic’s compute strategy in full.

The multi-platform picture matters because coverage elsewhere has occasionally characterised this Google deal as Anthropic “switching” from NVIDIA. That framing understates the deliberate architecture of Anthropic’s compute strategy. The Google deal is an expansion, not a departure from either AWS or NVIDIA.

Why this matters beyond the benchmark

On a per-chip basis, the current generation comparison is closer than the headlines suggest. Ironwood, now in general availability, delivers approximately 4.6 petaflops of FP8 computing power. NVIDIA’s Blackwell B200 delivers roughly 4.5 petaflops at FP16, although cross-precision comparisons require care, as the two figures are not measured on an identical scale.

But benchmark comparisons miss the bigger story.

At pod scale, where these chips are actually deployed, the gap widens. An Ironwood superpod of 9,216 chips delivers 42.5 exaflops. The eighth-generation TPU 8t pod, at 9,600 chips, targets 121 exaflops at FP4 precision. Google also claims near-linear scaling to one million chips inside a single logical cluster. For hyperscalers running hundreds of thousands of chips simultaneously, pod-level economics matter far more than per-chip benchmarks.

The NVIDIA position

NVIDIA currently controls an estimated 81% of the AI data centre chip market, according to IDC. That is an extraordinary concentration of market power, and the near-term demand picture has remained resilient.

Recent analyst expectations have pointed to strong NVIDIA earnings growth, supported by elevated demand for AI infrastructure and broad adoption of the Blackwell platform. NVIDIA has itself guided for a combined US$1 trillion in Blackwell and upcoming Vera Rubin orders across 2026 and 2027.

AMD is developing rack-scale server systems and has gained meaningful ground. Estimates from analysts, including IDC, suggest AMD may now hold approximately 10% of the AI accelerator market, up from low single digits two years ago. Amazon and Google continue to expand custom chip businesses. The combined chip operations at Amazon alone, covering Trainium, Graviton and Nitro, have crossed a US$20 billion annual revenue run rate, growing at triple-digit percentages year over year, with nearly 40% sequential growth in Q1 2026.

The bull case for NVIDIA remains clear: demand has stayed strong, and NVIDIA’s ecosystem remains deeply embedded across the AI compute stack.

The longer-term question is less about near-term earnings and more about pricing power in the next upgrade cycle. Every reporting period in which Google, Amazon and Microsoft gain confidence in their own silicon is another data point in that debate. The incentive structure is powerful: these companies have every reason to reduce dependency on a single supplier, and the capital to act on it.

Stocks and sectors to watch

For NVIDIA, the near-term earnings story and the longer-term competitive story are pulling in different directions. Strong results may validate the current cycle. But the structural dynamic, where major customers build their own silicon, is unlikely to reverse.

For Alphabet, the Ironwood general availability and eighth-generation preview represent a potential monetisation opportunity well beyond advertising. Google Cloud grew 63% year over year in Q1 2026, among the fastest growth rates of any major hyperscaler. TPU-as-a-service, with confirmed anchor customers including Anthropic and Meta, could extend that runway materially if enterprise inference workloads continue migrating to Google infrastructure.

The less obvious plays are in the supply chain. TPU 8t and 8i are both targeting TSMC 2nm manufacturing, with Broadcom designing the training chip and MediaTek the inference chip. TSMC may remain a critical enabler regardless of which chip architecture gains share in each cycle, as may advanced packaging suppliers, liquid cooling companies and data centre real estate investment trusts (REITs).

Power infrastructure, liquid cooling suppliers and data centre REITs may also be exposed to sustained capital expenditure growth. Combined hyperscaler capital expenditure from the four major cloud providers is tracking toward US$700 billion or more in 2026, nearly double the US$388 billion spent in 2025. That scale of investment, sustained over multiple years, represents a different kind of macro signal.

Supply chain plays: If neither NVIDIA nor Google “wins” the chip war outright, infrastructure may still benefit. TSMC already manufactures both Ironwood and the upcoming eighth-generation chips. Advanced packaging suppliers, liquid cooling companies and data centre REITs may benefit regardless of which chip architecture gains share in each cycle.

Where the risks sit

Higher AI infrastructure spending does not automatically translate into stock gains. Several factors complicate a straight line from “chip war” to “buy everything”.

What investors might take away

The AI chip war is not a story of one winner and one loser. It is a story of a market that is too large and too strategically important for any single company to own indefinitely.

NVIDIA built its lead through genuine technical excellence and a decade of software investment. That lead is real, and near-term earnings are likely to continue reflecting it.

But the challengers are no longer startups with benchmark slides. They are trillion-dollar companies with their own silicon, their own cloud infrastructure and every incentive to reduce their dependency on a single supplier, along with the capital expenditure commitments to show they are following through.

One way to frame the longer-term question is this: demand for AI compute may not be the primary variable for investors to focus on. Who captures the margin from that demand, and at what valuation multiple, may matter just as much. Those are questions each investor may need to weigh against their own risk profile and objectives.