Kai ChenGoogle TPU coverage is moving backward: velocity -0.724, acceleration -3.351. Yet Google split its eighth-generation chips into separate training and inference lines in 2026, just as Gemini spread into Android, video production and owned-book workflows. The hardware is losing attention while more of Google’s products depend on what it can make cheap and abundant.
Key takeaways
- Google is moving Gemini from a destination users choose into Android, media-production tools and owned-content workflows where it can become the default route for completing tasks.
- Embedding Gemini across existing products lowers adoption friction but turns every use into a recurring inference, capacity, latency and power obligation for Google.
- Google’s separation of training and inference TPUs reflects the underlying economics: models are trained periodically, while widely distributed products must be served continuously.
- Cheaper models and specialized inference hardware reinforce each other—efficiency supports broader deployment, and greater request volume increases the value of custom silicon.
- TPU does not need to be visible to users to confer an advantage; its strategic value is determining how broadly and cheaply Google can make Gemini feel like an ordinary product feature.
A model becomes more valuable when users stop choosing it
The standalone chatbot makes model competition unusually visible. Users pick a destination, enter a prompt and receive an answer. Every step is legible, including the moment they leave for a rival. Google’s more consequential Gemini releases reduce the number of times that explicit choice has to occur.
Gemini Intelligence bundles cross-app task automation and Android-widget creation into the phone environment. That turns Gemini from an application into part of an assistant operating layer that can act across applications rather than wait inside one of them. The competitive unit is no longer a model response but the preinstalled route through which a task gets done.
Google delayed its planned upgrade from Assistant to Gemini on most Android devices beyond its previous end-of-2025 target and into 2026. Distribution assets reduce the cost of reaching users, but they do not eliminate migration work, compatibility constraints or the burden of replacing behavior that already functions. Defaults have to work at installed-base scale, not launch-demo scale.
Google is taking the same route through narrower workflows. Gemini Omni 1.1 Flash is framed around production tasks such as extending a video scene, controlling first and last frames and upscaling to 4K. Gemini Notebook’s Expert Intelligence lets users import eligible titles they own in Google Play Books, ask grounded questions and generate podcasts. Both keep the model beside work or content the user already values, making the surrounding workflow harder to replace than response quality alone.
Every default creates a recurring service obligation
By putting Gemini into phones, media tools and content libraries, Google lowers adoption friction and takes on a standing service obligation. Each surface can generate more inference, and every successful workflow gives the user another reason to invoke it again.
A standalone AI product can ration expensive capabilities, meter access or charge directly for usage. An assistant embedded across a large platform estate has less room to behave like a boutique compute reseller. Users experience it as part of the product, while Google experiences every invocation as cost, capacity and power demand.
An embedded assistant needs low latency because it sits inside an interaction. It needs higher rate limits because distribution can create bursts of demand. It needs fewer tokens because small savings repeat across every request. As use becomes ambient, efficiency gives Google room to distribute more aggressively.
TPU’s fading visibility reflects a change in its job
Google introduced the Tensor Processing Unit in 2016 as a custom machine-learning chip tailored for TensorFlow. In 2018, Cloud TPUs entered beta at a starting price of $6.50 per TPU hour. Developers could buy a discrete resource with a name and a price.
Google assigned custom silicon a different role in its newer roadmap. In 2025, it described Ironwood as its first TPU for inference, offered in 256-chip and 9,216-chip configurations. In 2026, Google split its eighth-generation line into TPU 8t for training and TPU 8i for inference, with general availability planned later that year.
By splitting the line, Google mapped its hardware directly onto Gemini’s economics. Google trains a model periodically but serves the product continuously. TPU 8i targets the recurring obligation created by Gemini’s widening distribution.
Google is advancing that roadmap while public attention cools. Lower coverage does not mean weaker hardware: coverage reflects narrative heat, while Google can earn the return through repeated operation rather than repeated launches.
Meta reportedly signed a multiyear agreement to rent Google TPUs and discussed buying them for its own data centers from 2027. External demand could make the hardware more visible. For Google, however, TPU can remain both a merchant product and a source of internal operating leverage.
The chip need not disappear as a product; it only has to disappear from the user’s decision. A person extending a video, questioning an owned book or asking a phone to complete a task chooses an outcome, not an accelerator. The hardware can shape the service without becoming part of the pitch.
Inference cost is coordinating the industry
Investor materials reportedly showed OpenAI and Anthropic projecting inference costs above half of revenue. OpenAI and Broadcom unveiled an LLM-optimized inference chip, while Amazon Web Services continued advancing its Trainium accelerator line. Inference bills pushed all three toward custom hardware without a shared doctrine.
Google has also redesigned models around the cost of serving them. It positioned Gemini 1.5 Flash-8B at 50% below the price of 1.5 Flash, with twice the rate limits and lower latency on small prompts. Google later said Gemini 3.6 Flash used up to 17% fewer tokens and cost less per token than Gemini 3.5 Flash. Those changes determine how often Google can rationally invoke a model inside routine products.
As Google serves more requests, small efficiency gains become more valuable. Those gains support broader deployment, which creates more volume for specialized hardware. A provider at that scale can either control serving infrastructure or let a major share of its business depend on someone else’s costs and capacity.
Google also has to secure physical capacity. When the company introduced new Android app memory-use thresholds, it cited significant hardware supply constraints caused by the AI boom. Until Google can supply a feature across the installed base, it has a branded queue rather than a platform default.
The least visible layer sets the visible default
The model race stays visible because models have names, scores and release dates. The counterweight appears as smaller product decisions: an assistant migration, a video control, a book import, an inference chip split from a training chip. Each looks incremental; together they move competition from choosing a model to sustaining a default.
That is how TPU coverage can register a velocity of -0.724 and acceleration of -3.351 while Google gives inference a dedicated chip line. Users choose the outcome on the screen, not the accelerator behind it. Gemini wins attention at the surface; TPU determines whether Google can afford the default underneath.
Google coverage shifted from consumer themes toward research, 2024–2026
| Coverage theme | Earlier share | Later share | Change |
|---|---|---|---|
| Consumer | 36.3% | 31.0% | -5.3 points |
| Research | 15.5% | 22.5% | +7.0 points |
Frequently asked questions
Why is TPU important to Gemini becoming a default?
Default distribution can create enormous, recurring inference demand. TPU gives Google more control over the cost, capacity and latency required to serve Gemini across phones, media tools and content workflows.
How is Google distributing Gemini beyond a standalone chatbot?
Google is integrating Gemini into Android through assistant functions and cross-app automation, into video production through Gemini Omni 1.1 Flash, and into owned-book workflows through Gemini Notebook’s Expert Intelligence.
Why did Google separate training and inference chips?
Training and serving have different economic profiles: training happens periodically, while inference continues with every user request. A dedicated inference line targets the ongoing service burden created by Gemini’s widening distribution.
What could prevent Gemini from becoming the Android default?
Google still faces migration work, compatibility requirements, hardware constraints and the challenge of replacing established Assistant behavior. Its move from Assistant to Gemini on most Android devices slipped beyond the previous end-of-2025 target into 2026.
Does falling TPU coverage mean Google’s hardware strategy is weakening?
Not necessarily. The piece argues that TPU’s return increasingly comes from quietly lowering the cost of repeated operation, so the hardware can become more strategically important even as it receives less public attention.