AI inference costs in free fall: Google Ironwood goes commercial
On August 2, 2026, just as the US stock market experienced a fresh tech rebound, Google parent Alphabet dropped a bombshell. Google Cloud officially announced that its 7th-generation self-developed Tensor Processing Unit (TPU), codenamed "Ironwood," is now fully available to external cloud customers. This marks not only a milestone for Google in AI infrastructure but also signals that the inflection point has arrived where AI computing shifts from "training is king" to "inference efficiency gains."
According to data disclosed by Google, Ironwood's peak single-chip performance reaches an astonishing 4614 TFLOPS (FP8 precision), nearly four times that of NVIDIA's H100 chip. Even more staggering for competitors is the cost metric: for equivalent AI inference tasks, Ironwood's per-token inference cost is only one-tenth that of the previous-generation Trillium TPU—a 90% plunge.
Breaking the CUDA moat: Software ecosystem alliances
For a long time, NVIDIA GPUs have built a deep competitive moat with their CUDA software ecosystem. However, as large models advance from hundreds of billions to trillions of parameters and inference workloads explode, cost is becoming the primary consideration over raw performance.
The launch of Ironwood marks the full rise of ASICs (Application-Specific Integrated Circuits) on the inference side. To address software ecosystem gaps, Google is not fighting alone. Ironwood has been deeply adapted to PyTorch's XLA (Accelerated Linear Algebra) compiler and seamlessly integrates with mainstream open-source model hubs like Hugging Face. Developers can migrate open-source large models such as Llama and Gemma, originally running on NVIDIA GPUs, to the Ironwood platform with virtually no code changes.
This combination of "low cost + low migration barriers" directly targets the AI industry's biggest pain point—expensive inference computing. For many AI application companies seeking cost reduction and efficiency gains, Ironwood offers a highly tempting "alternative" solution.
Supply chain ripple effect: Hidden dividends for Broadcom and Wiwynn
Ironwood's commercialization benefits not only Google but also triggers a ripple effect across the US-listed AI computing supply chain. As Google's long-term core co-design partner for TPU chips, Broadcom is undoubtedly the biggest behind-the-scenes winner. Broadcom provides key ASIC IP cores and advanced packaging interconnect technologies for Ironwood. With the exponential growth in Google TPU shipments, Broadcom's custom AI chip revenue is poised for explosive growth in the second half of 2026.
Meanwhile, at the server level, because Ironwood adopts a liquid-cooling architecture, ODM partners such as Wiwynn have also secured substantial orders. This further confirms a shift in US AI computing investment logic: capital is spreading from pure general-purpose GPU speculation to niche segments like custom ASIC chips and liquid cooling.
NVIDIA's response and the dual-track computing landscape
Facing Google's strong challenge, NVIDIA is not sitting idle. Just before the Ironwood commercial news broke, NVIDIA had just released its Blackwell Ultra GB300 NVL72 rack-scale solution, attempting to solidify its irreplaceability in high-performance training clusters through higher memory bandwidth and scale-up capabilities.
Industry analysts point out that the future AI computing market will exhibit a clear "dual-track" structure. On the training side, NVIDIA GPUs will still dominate due to the need for extreme versatility and ecosystem compatibility; but on the inference side, especially for hyperscale and relatively standardized generative AI inference tasks, Google's TPU and Amazon's Trainium ASIC chips will steadily eat into market share with their extreme cost advantages.
For investors, Ironwood's commercialization is a clear signal: the second half of AI computing competition has shifted from a pure peak-performance race to a brutal elimination game centered on "performance-per-watt" and "total cost of ownership (TCO)." At this stage, focusing on ASIC supply chain players like Broadcom and Marvell Technology, as well as companies providing cooling and power solutions for these giants such as Vertiv, may offer greater excess return potential than simply betting on GPU leaders.
As of press time, buoyed by the Ironwood commercial news, Broadcom's stock edged up in after-hours trading, while some GPU-related stocks faced profit-taking pressure. The market is reassessing the market share landscape for AI computing hardware over the next three years.
