Claude on Vertex, Sidekick on Shopify, and the quiet question of who owns the agent

Anthropic showed up at Google Cloud Next this week with a set of enterprise demos, the flagship one being Shopify's Sidekick, a commerce agent running Claude through Vertex AI, with the stagey detail that Shopify can swap model versions in under 24 hours. The sessions around it were about multi-agent orchestration on Google's infrastructure for long-running autonomous workloads. That is the surface. The structural story is the stack underneath: a frontier lab being sold through a hyperscaler into a SaaS vendor whose own product is being redefined by the thing it's embedding. Three layers, three different sets of economics, and a quietly interesting question about who ends up owning the customer relationship.¹

Start with what was actually on stage. Sidekick is Shopify's merchant-facing agent, it takes multi-step instructions ("set up a Mother's Day campaign, discount the top three SKUs by category, draft the email") and executes them across Shopify's surface. The architecture, as presented, is Claude via Vertex AI Model Garden, with Shopify handling the application layer and Google handling the serving. The "under 24 hours" model-update claim is the interesting one: it implies Shopify is not fine-tuning, not doing heavy eval gating, and is treating the underlying model as a fungible runtime. That is a specific choice, and it cuts against the narrative, popular in 2024, that serious enterprise AI deployments would be deeply customised, RLHF'd, constitutionally tuned things with long qualification cycles. Shopify is doing the opposite. It is treating Claude as a utility.

The architecture, as presented, is Claude via Vertex AI Model Garden, with Shopify handling the application layer and Google handling the serving.

Which brings us to the first frame: inference economics. If you're Shopify and you're running an agent that does multi-step workflows for millions of merchants, your cost of goods sold is now substantially an inference bill. The Vertex route means Google is taking a margin on the serving, Anthropic is taking a margin on the model, and Shopify is left with whatever's between the merchant's subscription and those two line items. Shopify does not disclose Sidekick unit economics and I would not expect them to, but the direction of pressure is clear: every time a merchant lets the agent run a ten-step workflow instead of clicking through ten screens, Shopify's cost per active merchant goes up and the merchant's perceived value of the subscription also goes up. Whether those two lines cross in the right direction depends on pricing, which Shopify has been conspicuously quiet about. Sidekick is bundled. Bundled things don't stay bundled when they become the product.²

This is where the second frame gets interesting: SaaS apocalypse, applied not to Shopify's customers but to Shopify itself. The per-seat logic of commerce SaaS is that you pay for access to a surface, dashboards, editors, analytics, that humans use. If the agent is the primary user of that surface, the surface becomes an API and the seat count becomes an abstraction. Shopify's current pricing is tiered by feature and transaction volume, not seats, which partly insulates it, but the question sharpens: if Sidekick does the work, what is the merchant paying Shopify for? The commerce rails, presumably. But the rails are the commoditised bit. The interface was the moat. I would watch for Shopify to either start charging for Sidekick usage directly (consumption pricing, the honest answer) or to quietly absorb the cost and bet that agent-mediated commerce lifts GMV enough to carry the subscription (the hopeful answer). The blueprint presentation didn't resolve this, which I think is the tell.

Now the third layer, and the one that Anthropic is actually doing here. FDE market structure: how does frontier-lab capability reach the enterprise? There are three models in the field. OpenAI's model is direct plus Azure, sell to the enterprise yourself, let Microsoft handle the ones that want a hyperscaler-mediated procurement. Google's model is vertical, its own models, its own cloud, its own apps. Anthropic's model, increasingly, is horizontal partnership: sit inside both AWS (Bedrock) and Google Cloud (Vertex), let the hyperscaler own the enterprise relationship, take the model margin and run. The Shopify demo is an advertisement for this posture. Anthropic isn't in the room with Shopify's procurement team; Google is. Anthropic is the ingredient. It is a specific strategic bet, that in enterprise AI, the lab that makes itself most available through the channels enterprises already buy from will win more seats than the lab that tries to own the channel itself.³

This is not obviously good for Anthropic's margins. Being the ingredient means Google takes a cut, and Google's Vertex pricing on Claude is meaningfully higher than Anthropic's direct API pricing for the same model, with the delta flowing to Google. Anthropic accepts this because the alternative, building its own enterprise sales motion at hyperscaler scale, would cost more than the margin it gives up. It is the classic ingredient-brand trade, and it works until the hyperscaler decides the ingredient is substitutable. Google has Gemini. Google has a strong commercial incentive to move Vertex customers onto Gemini over time, because the economics of Google selling Google are better than Google selling Anthropic. The Shopify deployment is therefore a slightly strange arrangement: Anthropic's model, running on Google's infrastructure, serving a customer whose ability to swap models in 24 hours is itself a headline feature. Shopify has, helpfully, disclosed the switching cost. It is low.

What this is a case of: the unbundling of the AI stack into three distinct profit pools, foundation models, serving infrastructure, application layer, each with its own margin structure and its own substitution risk. The Shopify/Vertex/Claude triangle is one of the cleanest examples yet of all three layers visible in a single deployment. It is also a structure in which the model layer is the most substitutable and the application layer is the most defensible, which is the opposite of where most of the 2023–2024 capital went.

What to watch. Shopify's next pricing disclosure, specifically whether Sidekick gets a usage meter. Anthropic's enterprise revenue split between direct, Bedrock, and Vertex, if Vertex grows faster than direct, the ingredient-brand bet is working on volume and the question becomes margin. Google's Gemini attach rate inside Vertex customers that started on Claude. And the 24-hour swap claim itself: if Shopify ever exercises it publicly, we'll know what the actual switching cost was.

Footnotes

¶