Google Releases Eighth Generation TPU Chips
At the Cloud Next conference, Google unveiled its eighth generation TPU chips (TPU 8t and TPU 8i) and the Gemini Enterprise Agent platform. The TPU 8t is aimed at large model training, with nearly 3 times the computing power per pod; the TPU 8i focuses on inference efficiency, featuring 1152 chips per pod, with performance cost improvements of up to 80%. Additionally, the enterprise-level Agent platform supports the construction and large-scale deployment of AI agents, with partners including Oracle and Salesforce.
Google disclosed that approximately 75% of its new internal code has been generated by AI, with model processing capabilities exceeding 16 billion tokens per minute. This news has driven Alphabet's stock price upward. English media generally view this as a key milestone in Google's dual advancement in "computing power + application layer."
Source: Public Information
ABAB AI Insight
This release's core is not a single product but rather the formation of an integrated closed loop of "computing power - model - application." TPU 8 compresses both training and inference costs, while the Agent platform directly transforms model capabilities into enterprise production tools, indicating that AI is no longer competing at the model level but entering "production system-level replacement." When the costs of computing power decrease and application layer standardization occurs simultaneously, the diffusion speed will significantly accelerate.
The fact that 75% of code is generated by AI is a direct signal of changes in the production function. Software development is shifting from "humans writing code" to "humans defining tasks + models generating execution," elevating labor's position in the value chain while automating the underlying execution. This not only enhances efficiency but also compresses the marginal value of mid- to low-end engineering labor, strengthening the control of top engineers and platforms.
Google's self-developed path for TPU is reshaping the structure of the computing power market. Compared to relying on general-purpose GPUs, Google reduces unit computing power costs through vertical integration, essentially competing for "computing power pricing rights." As model scales and inference demands continue to grow, whoever can provide lower-cost computing power will dominate the profit distribution of the AI ecosystem.
On a deeper level, the Agent platform signifies a change in the form of enterprise software. In the past, SaaS was a tool; in the future, Agents will be "executors." Once enterprise processes are taken over by Agents, software may no longer charge based on functionality but rather based on "the results of completed tasks." This will reconstruct the structure of enterprise IT spending and alter the business models and valuation logic of software companies.