Ideogram CEO on building AI image generation for enterprise design and on-prem data sovereignty

Ideogram: Enterprise image generation and the on-prem opportunity

Ideogram is building AI image and design generation tools for enterprise adoption, with a particular focus on brand, marketing, and product design workflows. Mohammad, the company's CEO, came to the role from Google, where he launched the image generation product Imagen. He founded Ideogram believing he could have greater impact outside Google.

The company's pitch sits at a specific inflection point: consumer AI image generation has matured enough that most people have seen AI-generated images without knowing it, but enterprise-grade design generation remains unsolved. The gap is consistency. A brand running a global marketing campaign needs typography that translates accurately across languages, logos reproduced at 100% fidelity, and product photography that matches what customers actually receive. None of that is fully reliable yet. Mohammad says Ideogram expects to close the consistency gap within the next few months.

“With the newest model that we released, it's actually a relatively compact model that's at the frontier and it's open weight. And that creates a new set of opportunities for us in terms of licensing and partnership where companies can host these models on prem. That brings down their cost, but also data sovereignty is really important. And then on the brand side, consistency is an issue, but we think we're gonna solve that in the next few months.”

On-prem and data sovereignty

The most concrete near-term demand Mohammad describes from enterprise customers is on-premises deployment. Design IP is commercially sensitive in ways that make cloud-based generation a hard sell — an automaker won't send next-generation vehicle designs to a third-party API. Ideogram's response is a new frontier model released as an open-weight, relatively compact architecture. That unlocks a licensing and partnership model where customers host the model themselves, reducing both cost and data exposure.

A second, less obvious customer segment is emerging: companies generating synthetic training data for manufacturing and defense applications, where real-world examples of rare scenarios simply don't exist. On-prem is equally critical there, since the generated data itself is sensitive.

Hardware and model design

Ideogram has been running on AMD infrastructure since the MI 300 generation and went live with its 4.0 model on AMD hardware a few months ago. Mohammad says the team designs model architecture with specific hardware characteristics in mind — price, performance, and latency targets — then trains to fit. Certain architectural choices have no impact on model quality but meaningfully affect downstream inference performance, and those are the levers the team adjusts against the roadmap. AMD's upcoming MI 450 and MI 455 chips are already part of Ideogram's planning horizon.

Team structure

Ideogram runs a roughly 50/50 split between researchers and product engineers on what Mohammad describes as a small overall headcount. Half the team is focused on the model; half on the product.

The competitive framing Mohammad offers on IP and open weights is pragmatic: American image generation companies are performing well at the frontier, and enterprise customers primarily want protection guarantees rather than the most permissive possible model. That makes the alleged IP permissiveness of some international open-weight alternatives less of a direct threat to Ideogram's customer base than it might be for consumer-facing products.