Google is splitting its TPU chip designs between AI training and inference tasks.
The company is now eyeing a custom CPU as its next focus.
This could make AI workloads more efficient and cost-effective.
What happened
Google is optimizing its TPU AI accelerators for different workloads, separating training and inference tasks. Its chief technologist for AI infrastructure, Amin Vahdat, said at SEMICON Taiwan 2026 that the CPU is the next focus for potential custom chip development.
Why it matters
As AI model training, inference, and AI agent workloads grow, specialized chips can improve efficiency and performance. Google's willingness to develop additional dedicated chips for specific needs signals a broader shift toward task-specific hardware in AI infrastructure.
What to watch
Whether Google builds a custom CPU hinges on a workload reaching sufficient scale and the economics of custom silicon proving viable. The test is whether AI infrastructure demand justifies the investment, with no product or timeline set.
Ask the AI about this article →
Summaries like this, in your inbox every morning.
The announcement underscores a broader industry trend toward purpose-built AI hardware. Google's TPUs are already specialized for AI, but splitting them further into training and inference variants reflects the differing computational demands of each stage. Training requires massive parallel processing, while inference prioritizes low latency and throughput. By potentially adding a custom CPU, Google could offload other data-center tasks, improving overall system efficiency.
However, the move is not a definitive commitment. Vahdat qualified that additional chips would only be developed if a workload reaches sufficient scale and the investment is economically viable. This suggests Google is exploring the option but will make decisions based on market demand and cost-benefit analysis. For businesses relying on Google Cloud, more specialized hardware could mean faster and cheaper AI services in the long run, though concrete products remain unannounced.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
OpenAI's product lead Tibo Sotiou posted on X on September 6 that GPT-6 Astra's 'low' setting outperforms GPT-…
OpenAI published two blog posts on September 6: a research acceleration report and an essay by Chief Scientist…
A swarm of OpenAI agents hacked a German website this spring, according to Reuters
Stanford University reports that the performance gap between top US and Chinese AI models has narrowed sharply…
Apple is set to release iOS 27 around mid-September, likely on Sept
Anthropic is launching AI commerce agents in partnership with Visa and Mastercard, according to a Finextra Res…
