Phone Models, Agent Token Savings, and Grok’s Power Supply

Two approaches to running models locally

Google and Bonsai are targeting phone execution with different claims. Google announced Gemma 4 optimized for the Pixel 10’s TPU, tying its mobile model work to a specific hardware target. Bonsai introduced Bonsai 27B, which it calls the first 27B-class model capable of running on a phone. The former emphasizes hardware optimization; the latter emphasizes the size of the model that can run locally.

Both announcements point toward more computation happening on the device. For mobile application development, that makes local execution a concrete option to investigate, though the announcements establish different things: optimization for a named phone processor and a claim about model capacity. Neither claim alone establishes how a particular application will perform.

Token efficiency and electricity supply

Agent execution offers a narrower, potentially more accessible efficiency opportunity: an article reports that compiling an AI agent skill cut token use by 94%. That result suggests a way to reduce token consumption and potentially cost, with the percentage belonging to the reported approach rather than serving as a general expectation for agents.

At the infrastructure level, Elon Musk reportedly acquired gas turbine company APR Energy for $1 billion to power Grok. The acquisition addresses electricity supply, a different resource constraint from agent token consumption. Its stated purpose is to bolster Grok’s power capacity; the purchase itself does not establish an improvement in model performance or operating efficiency.

Service delivery changes beyond the model

AI adoption can also change how an outside provider delivers its work. A CIO analysis describes service vendors being rebuilt around AI, with more automation becoming part of their delivery models. That puts vendor operating methods alongside model and infrastructure choices as a consideration for engineering organizations. More automated delivery may improve resource management or customer experience, but those remain potential benefits, rather than outcomes established for every vendor.