Stronger privacy protection—yet Google moved training onto the servers
When people hear “privacy AI,” their intuition is usually: data never leaves the phone, so the cloud loses a share of business. But newer technical routes don’t necessarily go in that direction.
On October 2, Google Research introduced a new-generation federated learning system: devices encrypt training data first, then upload it to a server-protected computing environment that runs according to pre-authorized rules. Gboard has already used it to launch next-word prediction models for English and Japanese.
The most noteworthy line in this official article for traders is that the training bottleneck shifts from device availability and phone compute power to protected computing resources on the server. It’s not saying that all phone data can be freely handed to cloud vendors, nor is it promising absolute security.
My take is that protecting data and keeping computation local are two design questions that can be separated. If the phone does less work, the server may do more; you can’t see the four words “on-device AI” and automatically conclude that cloud needs must therefore shrink.
That said, this progress can’t be directly translated into new chip orders. You still need to look at the scope of adoption, the actual amount of training, and how many resources each task consumes under the new approach.
It’s the same with AI narratives like RENDER, FET, and NEAR: first check where the task is truly executed and who can be charged. Google’s research hasn’t announced participation in these projects, so it can’t replace token-allocation orders.
Compute demand hasn’t disappeared—the bottleneck may have just moved.
$RENDER $FET $NEAR
Click my avatar to view live trading with orders
When people hear “privacy AI,” their intuition is usually: data never leaves the phone, so the cloud loses a share of business. But newer technical routes don’t necessarily go in that direction.
On October 2, Google Research introduced a new-generation federated learning system: devices encrypt training data first, then upload it to a server-protected computing environment that runs according to pre-authorized rules. Gboard has already used it to launch next-word prediction models for English and Japanese.
The most noteworthy line in this official article for traders is that the training bottleneck shifts from device availability and phone compute power to protected computing resources on the server. It’s not saying that all phone data can be freely handed to cloud vendors, nor is it promising absolute security.
My take is that protecting data and keeping computation local are two design questions that can be separated. If the phone does less work, the server may do more; you can’t see the four words “on-device AI” and automatically conclude that cloud needs must therefore shrink.
That said, this progress can’t be directly translated into new chip orders. You still need to look at the scope of adoption, the actual amount of training, and how many resources each task consumes under the new approach.
It’s the same with AI narratives like RENDER, FET, and NEAR: first check where the task is truly executed and who can be charged. Google’s research hasn’t announced participation in these projects, so it can’t replace token-allocation orders.
Compute demand hasn’t disappeared—the bottleneck may have just moved.
$RENDER $FET $NEAR
Click my avatar to view live trading with orders
