Google's $185bn answer to a question nobody asked out loud

Google I/O 2026 ran for two hours this morning and produced four product announcements: Gemini Omni (frontier model, "world-model" framing), Gemini 3.5 Flash (efficient tier, available in APIs at launch), Gemini Spark (a persistent personal agent sold through a new AI Ultra subscription), and Antigravity 2.0 (agentic developer platform, globally available day-one). Sundar Pichai opened with "we are entering the Gemini era," which is the sort of sentence that a CEO has to say when the company has spent the previous eighteen months explaining that, no, really, it is entering the Gemini era this time.¹

The product news is not the news. The news is the number Pichai disclosed in passing, which is that Google intends to spend $180–190 billion on AI infrastructure in 2026.¹

What the capex number is a case of. Microsoft and OpenAI's Stargate commitment is $500bn over four years, which annualises to roughly $125bn. Google has just announced a single-year figure that exceeds Stargate's annualised pace by something like 50%, and has done so as a forward commitment rather than a trailing report.² Google Cloud's 2025 annualised revenue was approximately $43bn. The spend-to-cloud-revenue ratio implied by a $185bn capex year is structurally implausible as a margin story; it only resolves as either (a) an acceleration thesis about AI revenue arriving at a scale that retires the question, or (b) a willingness to carry the loss as positioning spend until the question is retired by someone else.

This is the AI performativity frame in its purest form. The spend is the message. Whether the capex is recoverable on the timescales that ordinarily govern capital allocation is, for the moment, a category error. Google is buying optionality and competitive deterrence; the recoverability question is a 2028 problem.

What was actually shown. Gemini Omni was framed by Demis Hassabis as a "world-model" with "persistent understanding of physical and digital environments."³ This is consistent language with Hassabis's prior public position that AGI is "years, not decades" away. It is not, at launch, supported by published third-party evaluations. Google disclosed no benchmark numbers, no API pricing, and no developer access timeline for Omni in the keynote materials I've been able to read. The video-generative world-model category, Genie 2, Sora, Movie Gen, has accumulated a reasonable amount of marketing language about world-modelling without anyone resolving the question of whether generating coherent video constitutes modelling the world in any sense that a robotics researcher would recognise. Omni may be the answer; I'd want to see the evals.

Gemini 3.5 Flash, by contrast, shipped. It is in the APIs today. This inverts the usual frontier-model commercial sequence, where the headline model is the launch artefact and the efficient tier follows months later. Google is shipping the margin tier first and keeping the capability tier as a showcase, which is the right move if you believe that the binding competitive constraint is inference economics rather than benchmark leadership. (It is also the right move if your benchmark leadership is contested, which Gemini's has been through most of 2025 and 2026.) No Flash pricing was disclosed at launch — a small but notable absence. Pricing a Flash-tier model into a market where GPT-4o-mini and Claude Haiku already have developer share is a decision Google appears to want to make in response to competitor pricing rather than ahead of it.

Gemini Spark is the interesting one. Spark is sold as a personal AI assistant available through a new Google AI Ultra subscription tier, with the specific framing that it operates "24/7" in the background, including when the user's laptop is closed.³ This is a workflow automation product wearing consumer-assistant packaging. A persistent agent with device-level access that executes tasks while the user is not present is structurally the same product that enterprise SaaS vendors have been quietly trying to build into their suites for two years, except sold direct to consumers via a subscription.

The SaaS apocalypse frame predicts that per-seat pricing compresses as agents replace human users in the loop. Spark is the consumer-side version: if a single Spark agent can manage email, calendar, files and search for a household, the consumer-productivity SKU stack (Microsoft 365, Google Workspace personal, the various calendar and email tools) starts looking like it has the same seat-count problem at the bottom of the market that the enterprise tools have at the top. Google is, in effect, cannibalising the consumer side of Workspace pre-emptively, on the theory that it would rather cannibalise itself than have OpenAI do it.

The "laptop closed" framing is also where the consent and data-access questions live, and the launch materials do not address them. An agent that operates on a device when the user is not present has, by definition, access to whatever the device has access to. EU AI Act and UK ICO ambient-AI guidance are both live constraints here and have a reasonable chance of shaping Spark's scope in major markets before it reaches scale.

Antigravity 2.0 is the platform lock-in move, and the day-one global launch is the tell. Toolchain availability decisions are market-structure decisions: you launch globally on day one when you are trying to capture mindshare before the developer audience has settled on a different toolchain, not when you have a comfortable lead. LangChain, LlamaIndex, and OpenAI's Assistants API have a collective head start of eighteen to thirty-six months in production deployments. Antigravity arrives into a contested field with the table-stakes assumption that Google's distribution can overcome incumbent switching costs. Maybe. The AWS-style "lock in workloads via managed services" playbook is a real pattern, but it worked for AWS because the infrastructure parity question was unresolved. In the agent toolchain layer, infrastructure parity is largely resolved; what's contested is developer ergonomics and model-tool fit, which are not problems that day-one global availability solves.

What to watch.

• Gemini 3.5 Flash pricing when it arrives, and whether it's set below GPT-4o-mini and Claude Haiku or at parity. Below means Google is buying ecosystem; at parity means Google is defending margin.
• AI Ultra tier pricing for Spark. A consumer-priced tier ($20–30/month) signals upsell; an enterprise-priced tier ($50+/month) signals displacement.
• Whether Omni produces third-party benchmark data within 60 days. If it doesn't, "world-model" was positioning.
• The 10-Q capex execution against the $185bn commitment. Headline capex announcements in growth years have historically run ahead of executed spend; the gap between disclosure and execution is itself a signal.
• Antigravity adoption metrics. Google rarely discloses these voluntarily; the absence of disclosure within a quarter is the disclosure.

Footnotes

¶