Ironwood Is Live—and Google’s TPU Strategy Hits Nvidia Harder Than AMD

Google’s Ironwood TPU is no longer merely a promised accelerator: TPU7x reached general availability in March 2026, with broader Compute Engine integration following in June. That makes the competitive question more concrete. Google can now offer customers a production alternative to Nvidia-based infrastructure for large AI training and inference workloads.

The pressure is greater for Nvidia than for AMD because custom accelerators challenge the market position Nvidia already monetizes at enormous scale. AMD can also lose potential Instinct deployments to TPUs, but it does not have an Nvidia-sized incumbent revenue base for Google to displace.

Ironwood’s status changed after the 2025 announcement

Google’s November 6, 2025 announcement described Ironwood as its seventh-generation TPU and said general availability would follow “in the coming weeks.” It also claimed more than four times the per-chip training and inference performance of Trillium and ten times the peak performance of TPU v5p.

The subsequent release history is more precise than the launch headline. The official Cloud TPU release notes record TPU7x entering preview on November 24, 2025, reaching general availability on March 31, 2026 and gaining generally available Compute Engine API and managed-instance-group support on June 1.

This changes the practical analysis. In late 2025, Ironwood was an announced product entering preview; in 2026, it became infrastructure that eligible Google Cloud customers could provision for production workloads. Google’s documentation describes a cloud resource, however, not a merchant chip offered for unrestricted installation in customer-owned data centers.

TPUs attack the part of Nvidia that matters most

The TPU threat is not simply that another processor can execute matrix operations. Google controls the accelerator, compiler path, cloud orchestration, networking environment and models that generate a large share of its internal demand. It can therefore optimize a complete service around selected workloads and decide how much external GPU capacity those workloads require.

Nvidia’s fiscal 2026 Form 10-K reports $215.9 billion in total revenue and $193.7 billion in Data Center revenue, says two direct customers represented 22% and 14% of total revenue, and identifies NVLink Fusion as a way for hyperscalers and custom-ASIC designers to integrate their own processors with Nvidia’s platform.

Those figures reveal both the exposure and Nvidia’s response. A cloud provider that shifts a suitable training or inference workload to custom silicon reduces demand in Nvidia’s largest end market. At the same time, NVLink Fusion gives Nvidia a route to retain networking, interconnect and platform value when the central accelerator is not an Nvidia GPU.

Why the same development is less damaging to AMD

AMD competes for many of the same deployments through Instinct accelerators, so TPUs are not good news for the company. If a Google Cloud customer selects an Ironwood slice instead of an Instinct-based service, AMD loses an opportunity just as Nvidia would. The difference is the position from which each supplier enters the contest.

Nvidia is the incumbent with a much larger pool of data-center revenue that custom silicon can bypass. A workload moving away from Nvidia threatens an established sale and may weaken the advantage created by its integrated GPU, networking and software stack. For AMD, the same workload is more likely to represent a prospective share gain against the incumbent rather than revenue already secured.

There is also a competitive asymmetry. Wider acceptance of non-Nvidia accelerators can erode the assumption that every serious AI deployment must begin with a CUDA-compatible GPU. That fragmentation may create openings for AMD even while Google captures some workloads for itself. TPUs can therefore substitute for AMD hardware in specific cases while making a multi-accelerator market easier for AMD to contest.

Ironwood does not make general-purpose GPUs obsolete

TPUs are strongest when a workload fits Google’s supported frameworks, numerical formats and distributed system. GPUs remain more portable across clouds, vendors and a wider range of accelerated applications. Organizations with extensive CUDA software, specialized kernels or deployment requirements outside Google Cloud face migration and engineering costs.

General availability also does not guarantee that a TPU will be the economical choice for every model. Utilization, model architecture, batch size, latency targets, data movement, regional capacity and engineering effort all affect total cost. Vendor peak-performance comparisons should not be treated as universal application benchmarks because they describe particular generations and test conditions.

Google also continues to provide Nvidia GPU infrastructure alongside TPUs. The two architectures can coexist within one cloud portfolio: a customer may train one model on GPUs, serve another on TPUs and reserve GPU capacity for software that depends on Nvidia’s ecosystem.

The deeper risk is hyperscaler bargaining power

Ironwood’s largest strategic effect is that it gives Google another credible option when deciding whether the next unit of AI capacity should use an external accelerator. Even if TPUs never replace most GPUs, that choice can influence purchasing volumes, deployment timing and the price buyers will accept for comparable computing capacity.

For Nvidia, the contest is therefore broader than chip-versus-chip performance. It must preserve enough value in CUDA, networking, systems and developer tooling that customers keep buying its platform when cloud providers can direct optimized workloads to their own silicon. Its support for connecting custom accelerators reflects preparation for that mixed-hardware future.

AMD faces the same architectural shift but has less incumbent business for Ironwood to erode. Google’s TPU program consequently poses the larger threat to Nvidia: it targets the economic center of the market leader while changing the market structure in which AMD is still trying to gain ground.

Also read:

• Gemini Failed Twice in Google’s Live Demo—Its 2026 Role Is Much Bigger
• Marketing Is Harder Now: More Channels Create a Bigger Proof Problem
• Google Product Listings Are Still Free—but Visibility Is Not Guaranteed