Carlos Sanchez: Adobe Builds Websites That Assemble Themselves in 1.1 Seconds — BigGo Finance

The Latency Wall Has Fallen

For years, the promise of AI-generated websites has collided with a simple physical constraint: users will not wait. Marketing teams know that every additional second of page load time bleeds conversion rates, which meant that real-time, per-user website generation — however theoretically appealing — was commercially non-viable. The math was unforgiving. If a large language model takes five seconds to compose a page, the user has already left.

That calculus has now changed. According to Carlos Sanchez, principal scientist at Adobe working on Adobe Experience Manager, speaking on the AI Engineer podcast, his team has brought full page generation down to just over one second. The figure is not a prototype aspiration but a measured benchmark: using Cerebras inference hardware running Google’s Gemma 4 model, average page generation latency came in at 1.1 seconds. The next-best provider took 4.6 seconds.

Sanchez frames the dual criteria in plain operational terms. “We look for two things. Accuracy. That’s typically what people look for, but also we want the speed because we don’t want the site generation to take more than 1 or 2 seconds, right? Because people this is already proven that people want the faster the site, the more conversions it generates or the better the experience it is for the user.”

The speed gap that Sanchez demonstrated is not incremental. It is the difference between a feature that feels responsive and one that feels broken in a modern web context. His live benchmark told the story directly: a query for “a coffee machine to prepare coffee while camping” generated a fully personalized page in 1.64 seconds total round-trip, with the LLM call itself consuming roughly one second at 2,300 tokens per second. A second run confirmed the consistency with 2,200 tokens per second.

“We get 2,300 tokens per second. Which is not bad. I would say. This is something that we only dreamed about before,” he noted.

The strategic insight underlying this choice is counterintuitive in an era of ever-larger models. Sanchez argues that site generation does not require a frontier-scale LLM. “You don’t need a huge LLM to do this sort of work because you are generating text, you are deciding where to put blocks and how to organize the website, you don’t need a lots of information for that.”

Architecture: Grounded Generation, Not Blank-Page Fantasy

The most important design decision Sanchez made was to reject the idea of generating entire websites from scratch. Marketing organizations have strict brand guidelines, and hallucination risk in customer-facing copy is unacceptable. Instead, the system personalizes discrete blocks — hero cards, product feeds, navigation menus, calls-to-action — while grounding all generated text in a retrieval-augmented corpus built from the existing site content.

This approach confines the LLM’s creative latitude to the areas where it adds value: selecting which blocks to show, composing copy for a specific user intent, and organizing page structure. The brand voice, product facts, and visual identity all come from the grounded corpus, not from the model’s parametric memory.

The system architecture separates concerns cleanly. The front end runs on Adobe Experience Manager’s edge delivery services, where content is served from the edge and blocks are composed dynamically. The backend handles LLM evaluation, prompt execution, and the retrieval-augmented generation layer, running across Google and Cloudflare infrastructure with a vector database supporting the RAG pipeline.