Gemini Spark isn’t a chatbot upgrade. It’s a different architectural model entirely.
According to Google’s official announcement, Spark runs on dedicated virtual machines inside Google Cloud, meaning the agent’s tasks don’t terminate when a user closes their laptop or locks their phone. The session doesn’t end, Google keeps it running. That’s the core design claim: the AI isn’t resident on your device; it’s resident in Google’s infrastructure, working on your behalf around the clock.
The product is available in public beta to Google AI Ultra subscribers in the United States, per Google’s announcement. Those two qualifiers, AI Ultra tier, US geography, aren’t confirmed from independent sources in this package, but they’re consistent with Google’s established rollout pattern for premium AI features. The beta launched in late May 2026, approximately ten days after the initial announcement at Google I/O 2026.
How the VM architecture works in practice. Multiple technology publications reporting from Google’s announcement describe the same setup: Spark runs on a dedicated Google Cloud VM assigned to the user’s account. When you initiate a task, drafting a document, monitoring a service, processing a queue of emails, the VM continues executing even after your device goes dark. You return to a completed result rather than a paused state.
Disputed Claim
One source indicates the underlying model is Gemini 3.5. That detail hasn’t been confirmed from Google’s primary announcement text, so treat it as directional rather than definitive.
What Google claims Spark can do beyond persistence. Google states that Spark can navigate a remote browser to interact with signed-in websites, adding items to a cart, reading dashboard data, submitting forms, and execute code on the user’s behalf. These capabilities exist in commercial AI agents from other vendors, so the category is established. But independent verification of Gemini Spark’s specific implementation of remote browser execution isn’t available from sources in this package. “Google states” is the right framing, not “Spark can.”
The catch is that vendor-claimed browser interaction capabilities often look different at production scale than in announcement demos. Latency across the VM-to-remote-browser handoff, session timeout behavior on third-party sites, and the practical scope of “signed-in website” access are all details Google hasn’t addressed publicly yet.
The pattern this fits. Gemini Spark is the third named persistent-agent product to surface in as of publication. xAI’s Grok Build CLI and OpenAI’s Codex @Computer command both represent variations on the same architectural bet: move AI workload off the user’s device and into always-on cloud infrastructure. The competitive pressure is visible. Google isn’t reacting to a trend, it’s racing inside one.
What to Watch
That convergence matters for developers building on any of these platforms. Long-running workflows that previously required a persistent local process or a custom cron job now have a vendor-hosted alternative. The design question shifts from “how do I keep this running?” to “do I trust this vendor with always-on access to my accounts?”
What to watch. The AI Ultra subscriber tier is currently the access gate. Watch for whether Google expands Spark to lower subscription tiers or opens an enterprise API, either signal would indicate commercial traction beyond the early-adopter base. Independent evaluations of the remote browser execution claim will matter before enterprise teams consider building workflows on top of it. Google’s June roadmap, if published, should clarify the expansion timeline.
Don’t expect production-ready enterprise tooling from a public beta in late May. The architecture is real. The capability scope is still Google’s word.