Technology and innovation · Corporate strategy ·
Google's own AI chips reach a seventh generation, then an eighth
Google Cloud announced general availability for Ironwood, its seventh-generation TPU, in November 2025 and an eighth generation in April 2026, making the case for designing its own chips while still offering customers Nvidia's.
- Published
- Sources checked
- Sources 2, all primary
- Where North America · Cloud computing and semiconductors
A curated item, not a live report: it was checked on 3 November 2026 and does not update by itself. Later events may have changed the picture.
- Ironwood chips linked in one superpod Source [1]
- 9,216
- Peak performance of Ironwood against TPU v5p, as Google claims Source [1]
- 10x
- Better inference performance per dollar for TPU 8i than the previous generation, as Google claims Source [2]
- 80%
Section 01: What happened?
Sourced facts only, each with the date it refers to and the document that states it.
The development
-
Google Cloud announced that Ironwood, its seventh-generation Tensor Processing Unit (TPU), would be generally available in the coming weeks. Google said Ironwood offers ten times the peak performance of TPU v5p and more than four times the performance per chip of TPU v6e for both training and inference.
-
Ironwood chips can be linked in a superpod of 9,216 chips sharing 1.77 petabytes of high-bandwidth memory, and superpods can be joined into clusters of hundreds of thousands of TPUs.
-
Google said it built its first TPU ten years earlier, and that it designs its processors with model research, software and hardware development under one roof. It also announced virtual machines on its own Arm-based Axion processors, which it said offer up to twice the price-performance of comparable x86 machines.
-
The announcement described a shift in demand from training AI models to running them for users, which Google called the age of inference, and said this creates new opportunities for custom chips.
Since then
-
At Next '26, Google Cloud said Ironwood and Axion were generally available and introduced eighth-generation TPUs: TPU 8t for training, which it said has three times the processing power of Ironwood, and TPU 8i for inference, which it said delivers 80% better performance per dollar than the previous generation.
-
Google described its range of computing hardware as running from its own TPUs and Axion processors to the latest Nvidia GPUs, so that customers can choose the right architecture for each job.
From here to the discussion question, the page gives the platform's reading of the facts above: the context, the concepts, the choices, the implications and the lessons. It is interpretation, not reporting, and it adds no new facts about the organizations.
Section 02: What is the business context?
The organization, its industry and the situation it was in, read from the facts above.
Companies that rent computing to others, such as Google Cloud, need very large numbers of processors to train AI models and to run them for users. Most businesses that need computing power buy their chips from specialist suppliers; Google has also designed its own for a decade, and it says it develops them alongside its models and software under one roof. Ironwood is the seventh generation of that line.
The demand behind it is changing. Google describes a shift from training models to running them for users, which it calls the age of inference, and argues that this opens new opportunities for custom chips. Yet Google still offers customers Nvidia's processors alongside its own. The context is therefore not a choice between making and buying, but a decision about how much of a critical input to make, at a time when the cost of running AI may become a main basis of competition.
Section 03: Which management concepts are relevant?
Where this development meets the IB course, and where it reaches beyond it, with links to the subtopics, terms, frameworks and tools that explain it.
In the IB course
-
5.8 Research and development HL
Long-term research and development, and incremental innovation that may add up to something more disruptive.
-
Cost to make against cost to buy, for an input as critical as the chips that run AI.
-
Backward vertical integration into the supply of an input.
-
5.9 Management information systems HL
Artificial intelligence, data centres and cloud computing as critical infrastructure.
-
Differentiation; a cloud provider competing on hardware its rivals cannot buy.
Key terms
Toolkit and frameworks
-
Contribution IB toolkit HL
Make or buy analysis compares the cost of making an input with the price of buying it in.
Key concepts: Creativity, Change
Beyond the course
Source: EnrichmentFor undergraduate, MBA and professional readers. These ideas are not part of the IB course and are not assessed by it.
-
Designing its own chips moves Google further back along its value chain, from buying an input to making it.
Terms beyond the course
Section 04: What choices did the organization face?
The realistic options, with the case for and against each. Where the organization's choice is known from the facts above, it is marked.
-
Buy all its chips from specialist suppliers
No design costs and access to the market leader's products, but dependence on suppliers' prices, plans and capacity, and no hardware advantage over rivals who buy the same chips.
-
Design its own chips and offer nothing else
The most control and the largest potential cost advantage, but every gap in its own range becomes a gap in its offer, and customers who want other chips go elsewhere.
-
Design its own chips and keep offering others The option taken
Gains the benefit of its own designs where they are strongest while keeping customer choice and a fallback, at the cost of running two hardware strategies at once.
-
Partner with a chip designer instead of designing in-house
Shares the cost and the risk, but shares the advantage too, and the partner may sell the same design to rivals.
Section 05: What were the implications?
What followed, or may follow, for the organization, its stakeholders and its industry.
Designing its own chips is a long-term commitment to research and development, and a form of backward vertical integration: making an input Google could buy. The case Google makes is that designing chip, software and models together gives performance and cost that a part bought off the shelf cannot match. In the language of production planning, Google is judging that, at its scale and counting performance, the cost to make these chips is below the cost to buy them.
Each generation improves on the last, which is incremental innovation; less than six months after announcing Ironwood's availability, Google introduced an eighth generation split into chips for training and for inference. The cumulative effect could be more disruptive if the cost of running AI models becomes the main basis of competition, because rivals who buy the same chips from the same suppliers cannot match a cost advantage built into the hardware. For customers the implication is choice: Google still offers Nvidia's chips beside its own, which also gives Google a fallback if its own designs fall short.
Section 06: What can a student or manager learn?
- Research and development is a long game. Google's first TPU came a decade before the generation described here.
- Make or buy turns on cost and on control. At enough scale, making an input can be both cheaper and a source of differentiation.
- Treat a company's performance figures as claims. Independent tests are needed before products can be compared.
- For managers, vertical integration is not all or nothing; a business can make some of an input and buy the rest.
Section 07: Discussion question
Using the concepts of cost to make and cost to buy, discuss why a cloud computing provider might design its own chips while continuing to offer chips made by other suppliers.
Command term: Discuss (AO3). The platform's own question for discussion; it is not an IB examination question and has no official markscheme.
Section 08: The limits of the evidence
What the sources do not show, so that no one takes the analysis for more than it is.
- Every performance figure here is Google's own claim; no independent test was read for this item.
- The pages give no figures for what Google spends on designing chips, or for how much of its computing runs on its own chips.
- The November 2025 post said Ironwood would be generally available within weeks; the pages read do not give the exact date it became available.
Section 09: Sources
Every source was opened and read on the date shown. Each entry says which facts on this page it supports.
-
[1] Company blog · Primary source
Announcing Ironwood TPUs General Availability and new Axion VMs to power the age of inference (external site)
https://cloud.google.com/blog/products/compute/ironwood-tpus-and-new-axion-based-vms-for-your-ai-workloads
Supports
- Ironwood's announced availability, and Google's performance claims against TPU v5p and TPU v6e.
- The size of an Ironwood superpod, its shared memory, and clusters of many superpods.
- The first TPU ten years earlier, Google's in-house co-design, and the Axion virtual machines and their claimed price-performance.
- Google's description of the shift from training to inference and the opportunity it sees for custom chips.
-
[2] Company blog · Primary source
Welcome to Google Cloud Next '26 (external site)
https://cloud.google.com/blog/topics/google-cloud-next/welcome-to-google-cloud-next26
Supports
- The general availability of Ironwood and Axion, the eighth-generation TPU 8t and TPU 8i, and Google's claims for them.
- Google's range of hardware from its own TPUs and Axion processors to Nvidia GPUs.
Arjun Agrawal | The Business Brief, https://the-business-brief-dun.vercel.app/intelligence/google-ironwood-custom-chips/. Original teaching material, not an IB publication. The organizations in the teaching cases are fictional unless a case says otherwise.