Eric Seufert: Meta's image model is its most compelling AI investment thesis — but the API business may be a mistake

Meta's AI investment thesis, the API mistake, and the Prosperous Society

Eric Seufert argues that Meta's proprietary image model is the clearest way the company has ever had to make its AI investment tangible to skeptical investors — and that the API business built around Muse 1.1 is a distraction from where the real value lies.

The image model as investor narrative

The core problem Meta has faced is that its most important AI investments — GEM for ranking, Andromeda for retrieval, Lattice for transfer learning — are invisible. They show up in spreadsheets but not in anything an analyst can point to. Meta posted 33% advertising revenue growth last quarter on roughly $55–60 billion in revenue, outpacing Google Search (19%) and Amazon advertising (24%), yet Seufert says investors at his research dinners remain unconvinced. When he asked one what it would take to change their mind, the answer was 40% growth — a number, he says, pulled from thin air.

The image model changes the narrative mechanics. It is trained, not fine-tuned, on conversion data that flows back to Meta through CAPI and the Pixel — ROAS signals, post-click behavior, app events — at a volume no other frontier lab can replicate. Seufert's argument is that Meta can eventually point to an ad in a user's Facebook or Instagram feed and say that creative only exists because of data only Meta holds. That's a concrete artifact investors can tie to the CapEx. No other lab can fine-tune for this use case, because no other lab has the data.

The advertiser bottleneck reinforces the thesis. The most common answer advertisers give when asked what limits their Meta spend is creative volume and quality. A foundation image model trained on actual performance outcomes — not third-party tools working without ROAS feedback — directly removes that constraint. As Seufert frames it, the bid in a second-price auction should reflect true value, and better creative gets you closer to that.

“If they can point to the output and say, look, this ad that you saw in your Facebook, Instagram feed was created only as a result of our ability to train on this data that only we have, then I think you can kinda make the case... My sense is it's a mistake [selling the API]. I think it's a capitulation. I think you're gonna get much more value out of that compute if you apply it to your own core business, which is advertising.”

The API business is a mistake

Seufert is direct that selling Muse 1.1 over API is the wrong move. His position is that Meta would extract more value from that compute by pointing it entirely at advertising. The API business is a capitulation to a narrative problem, not a solution to a productivity one.

The numbers illustrate the mismatch. A $3 billion AI API business built from scratch in a year would be, by any historical standard, an extraordinary outcome. Against Meta's overall revenue base, it barely moves the needle on the growth rates investors are demanding.

Seufert also sees a structural ceiling on API switching. Swapping a model variable in Vertex AI takes thirty seconds, but sampling from a new distribution changes output qualitatively, which triggers AB testing, QA cycles, and long-term retention analysis. That friction means many enterprise customers will stay on older models that work rather than upgrade, which pushes Muse 1.1 toward commodity pricing from day one.

On internal workloads, there's a more legitimate case. Meta reportedly spends something in the range of $10 billion annually on external models from Google, Anthropic, and OpenAI. If it can migrate those workloads to its own infrastructure, the efficiency story is real — though Seufert notes that coding and agent workflows, where Meta's internal tooling is genuinely sophisticated (he cites an internal self-learning agent system called Confucius), still benefit from frontier-quality models that Meta's own may not match.

ByteDance and the video ad timeline

On AI video advertising, ByteDance is the benchmark worth watching. Seufert describes a paper ByteDance published roughly a month ago detailing photorealistic 3D avatars selling products on TikTok Shop, trained on approximately 12,000 hours of live human product interactions. The key finding: when users can tell creative is AI-generated, ad performance drops. When they can't tell, AI creative outperforms human creative. ByteDance identified hand-object collision as the primary uncanny-valley trigger and built a dedicated visual interpretability model to fix it.

General-purpose photorealistic video for all ad formats is probably still years away. Use-case-specific video — infomercial style, controlled environment — may be achievable now, or close. YouTube deploying VO3 broadly across video ad inventory is a different, harder problem.

The Prosperous Society

The broader intellectual framework Seufert has been developing — a three-hour, four-part podcast series — is an economic bull case for AI framed as a response to John Kenneth Galbraith's The Affluent Society (1958).

Galbraith's "dependence effect" argued that advertising manufactures demand to justify mass production, which made sense in a postwar economy selling washing machines and refrigerators to newly suburbanized households. Seufert argues that framing is obsolete. The binding constraint in modern commerce is no longer production — it's distribution. AI investments flowing into recommendation systems and digital advertising push that constraint further out, enabling commerce across a much longer tail of specific, previously unserved demand.

The implication is heterogeneous product development at scale: producers who couldn't reach a viable audience before can now reach exactly the people who want what they make, at prices those people are willing to pay. Consumer surplus compresses, but welfare expands. The Pareto principle — where 20% of products drive 80% of commerce — weakens as the economics of niche distribution improve.

Seufert also addresses the labor displacement narrative directly. The data, including a recent Financial Times analysis he cites, does not support the claim that AI is reducing employment. Hiring is up. Entry-level displacement may warrant policy attention, but the broad white-collar displacement story, he argues, has no empirical foundation and should be dropped from serious debate.

The political framing matters to him too. He argues that much anti-AI sentiment is really anti-capitalism sentiment using AI as a proxy, and that anchoring the bull case in Enlightenment liberal economics forces critics to either engage on those terms or reveal that their objection isn't really about AI at all.

---

Takeaway: Seufert's sharpest investment-relevant claim is the narrowest one — Meta's image model is the first AI asset the company has built that investors can see, touch, and tie to revenue. The API business solves a narrative problem Meta doesn't actually have, while pulling compute away from the advertising platform that's already printing 33% growth on a $55 billion base.