I/O 2026: 3 Chrome AI Updates, 1,900 Invoices, 6x Latency
Google shipped three AI platform updates at I/O 2026, and on my $600 mini-PC in a closet, the spread between the fastest and slowest was six times. If you're running batch jobs on a home server or a cheap VPS instead of a GCP region, the ranking Google gave you is upside down. I'll show you the exact wall-clock numbers from a real 1,900-invoice nightly job, not a hello-world demo. Most recaps parrot the keynote order. I timed them against a workload that already runs in production. I build these systems for clients every week. Let's start with the one Google buried in the middle, because that's the winner.
Here's the setup, so you can compare it to your own stack. The test rig is a Ryzen mini-PC, 32 gigs of RAM, WSL Ubuntu, sitting on a shelf next to my router. The harness is a real nightly batch a client runs: nineteen hundred PDF invoices, parsed, classified, and pushed into an accounting system between two and three AM. Budget is forty-five minutes, hard. If it slips, morning reports break. That's the fixture. Now the three updates.
Update one, WebMCP. This is the shiny one Google led the keynote with. The pitch is beautiful: any web page can expose tools to any agent through a standard handshake. I wired it in. Cold handshake per tool call clocked at eight hundred and forty milliseconds on my machine. That sounds fine until you multiply. At six hundred invoices in, I was already past the forty-five minute budget with nothing but handshake overhead. WebMCP is built for interactive sessions where a human is watching a browser tab. It is the wrong shape for batch. If your work looks like a cron job, this update is not for you yet.
Update two, Built-in AI, which is Gemini Nano running locally inside Chrome. Google spent about ninety seconds on this one. First-token latency on my mini-PC was one hundred and eighty milliseconds. No network round trip, no API key, no rate limit, no bill. It was the only one of the three that finished the nineteen hundred invoice batch under budget, and it finished with fourteen minutes to spare. For anyone running edge or on-prem automation, this is the update. It's boring, it's local, it works.
Update three, Skills in Chrome. Invocation clocked at one point two seconds, and here's the catch: it needs a real browser context with a user profile. My server is headless. I do not have a logged-in Chrome session at three AM, and I'm not going to fake one. So for my stack, Skills is useless. But if your client is a non-technical operator clicking through a dashboard, Skills is the killer feature of the three. It puts an agent inside the tools they already use, without asking them to install anything. Different audience, different winner.
So the verdict table looks like this. For latency, Nano wins at one-eighty milliseconds, Skills at twelve hundred, WebMCP at eight-forty per call. For cost, Nano is zero, the other two are effectively zero at low volume but WebMCP racks up handshake overhead. For does-it-work-on-a-mini-PC, Nano is a yes, WebMCP is a maybe, Skills is a no. Google's keynote order was WebMCP, Skills, Nano. My production order is Nano, Skills for end-user work, WebMCP as a bet on future infrastructure I don't own.
The practical takeaway is this. Before you refactor anything around WebMCP because the keynote made it sound inevitable, run one experiment. Take the loop in your automation that fires most often. Measure the current per-iteration latency. Then add eight hundred and forty milliseconds and multiply by your daily volume. If that number blows your budget, WebMCP is not your next move, Nano is. This is a five-minute test and it will save you a rewrite.
Hot take. The agentic web framing assumes always-on cloud, GCP-adjacent latency, and a human in a browser. Most of the automation I ship for solopreneurs and small teams runs on a mini-PC in a closet or a twelve-dollar VPS. On that hardware, two of the three I/O updates add more latency than they remove. The one Google buried, local inference on the device, is the only one that actually shifts the economics for a small operator. Google is optimizing for their infrastructure. You should optimize for yours.
Want more like this?
I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.
Subscribe to bizflowai.io on YouTube — never miss a new tutorial.
Planning an AI automation project or need a second opinion on your architecture?
Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.
Visit bizflowai.io for our services, case studies, and AI consulting.