google io At Google I/O this week, the Chocolate Factory argued for its AI supremacy, making the case with benchmark-topping machine learning models, developer tools, and a few promising products.
How it might recoup the billions it is spending to build out AI infrastructure was left to the imagination.
There were a few glimmers of brilliance. Real-time language translation in Google Meet is great. It’s a vision of the Star Trek future many of us hope for, as opposed to the Torment Nexus nihilism fueled by social media rage-bait and surveillance capitalism.
Google Beam, a 3D conferencing system, at least has novelty appeal. Google pitches it as “a new platform designed to create more meaningful connections,” messaging that sounds a lot like Facebook before that brand took on so much baggage it metastasized into Meta. Let’s just say that personal connections are more meaningful when not mediated by technology.
But much of the show focused on promoting the use of Google’s Gemini model family, which just happens to come at a cost. It’s not a coincidence that some of the biggest boosters of AI developer services – AWS, Google, and Microsoft – happen to run cloud platforms.
That may be a tough sell for hobbyists or entrepreneurs aspiring to bootstrap a business into existence. Cost estimates for running an AI app vary, depending on the project specifications and requirements, but they’re not trivial, particularly when open source models running locally may be an option.
Even Google’s AI subscription fees can be consequential at up to $250/month for all the latest and greatest features, plus another $299/year for the recommended Premium Developer Program membership – introduced last November.
Businesses can afford to spend more and are doing so. According to CloudZero’s report The State of AI Costs 2025, shared with The Register, businesses are ramping up their AI spending. Based on a survey of some 500 software professionals, the average monthly AI spend among respondents’ companies went from $62,964 in 2024 to $85,521 in 2025, representing an increase of 36 percent.
At the same time, only about half of those organizations (51 percent) said they could evaluate the return on investment for their AI commitments. That’s about the same as when Gartner looked at the issue last year and found 49 percent of their survey participants had trouble demonstrating the value of AI projects.
Maybe agents will save the day
Google, like its peers, is betting on the “agentic era” to help AI usage take off. Though that’s a common industry refrain – we were told recently that within a few years, most internet traffic will come from software agents interacting with one another – there are reasons to be skeptical.
We have had the ability to automate online interactions for years through traditional imperative programming. The result is our current state of affairs – bot traffic in April 2024 for the first time surpassed human traffic on the internet, according to Imperva.
Bot traffic, however, may not be wanted. Why would a news publisher, for example, want to pay the bandwidth cost for an AI crawler that generates no advertising views or subscriber revenue to scrape the site? As a further disincentive, Google might use that data to create AI Overview search results that reduce click-through rates and actually deprive the publisher of visitors and attendant revenue.
So in an environment where AI model makers harvest training data in an adversarial manner, we’re expected to believe that bots called AI agents will interoperate without barriers or friction. That fantasy of harmonious vendor-agnostic interoperability requires new infrastructure like the Agent Name Service – to separate trusted agents from malicious ones – and the abandonment of decades of self-preferencing, the very thing that keeps generating antitrust lawsuits against AI industry leaders like Google and Microsoft.
For businesses in regulated industries, concerns about security severely limit how AI can be deployed. The Register was speaking recently with some executives at an AI security firm about how CIOs at banks and health companies just can’t move forward with their various AI pilot tests because they aren’t confident that their projects meet compliance requirements.
Those concerns only grow when the issue is granting AI agents the ability to communicate with other agents for decisions of any consequence.
There’s a place for automation, no doubt. But automation that isn’t deterministic is tricky – does anyone really want AI models taking unanticipated actions to solve problems when the results might be undesirable? Maybe when the agent is dealing with software unit testing or drafting pull requests that can be reviewed, but probably not when the agent is taking actions with real-world ramifications. And if the AI is only taking anticipated actions, why does it need to be AI as opposed to a simple programmatic decision tree?
Software agents may not be as useful as Google and its peers hope – or may take a lot longer than expected to get there. Concerns about competition, security, and cost can’t just be wished away.
Google is trying to do so anyway. The company that pioneered the hoarding of personal data to fuel its ad business wants Google Account holders to authorize its Gemini models to read the data stored in Google services in order to personalize their AI output.
Here’s how Google CEO Sundar Pichai described it in a blog post: “If your friend emails you for advice about a road trip that you’ve done in the past, Gemini can do the work of searching your past emails and files in Google Drive, such as itineraries you created in Google Docs, to suggest a response with specific details that are on point. It will match your typical greeting and capture your tone, style and even favorite word choices, all to generate a reply that’s more relevant and sounds authentically like you.”
Seriously, if your idea of friendship is sending people AI slop flavored with favored words culled from documents in Google Drive and Gmail, try asking Gemini how to be a better person. ®