The news is still just an agency headline, but its substance is a manifesto. Meta has announced its entry into the visual AI race, a segment dominated by generative models for images and videos that promise to transform content, advertising, and digital creativity. Almost simultaneously, a number leaks from ByteDance: Seedance, the Chinese company’s visual AI platform, has reportedly reached gross margins of 90 percent. A staggering figure that makes any SaaS envy, and which is not just a commercial triumph but a sign of a tectonic shift in how these services are built and monetized.

The root of such high margins lies not in aggressive pricing or a magical patent, but in vertical integration between hardware and software. ByteDance, forced by data sovereignty rules and geopolitical tensions to build self-sufficient datacenters and often custom chips, has effectively built an inference machine optimized down to the last watt. In a service like video generation – where every second of output can require the energy expenditure of an entire GPU for hours – slashing the cost per token (or per frame) is the decisive lever. Those who can do this on their own metal, cutting out public cloud intermediaries and designing accelerators tailored for visual workloads, can achieve margins that are still a mirage in the world of language models.

If confirmed, this news overturns many dominant narratives. For years, it has been debated whether generative AI is a zero-sum game due to uncontrollable inference costs. Seedance suggests the opposite: when infrastructure control is total, the marginal cost of a generation becomes negligible, turning a potentially expensive service into a near-pure profit machine. And the ByteDance case is not isolated: Apple with its Neural Engine, and Google with TPUs, have shown that computational efficiency translates into defensible margins.

Now it’s Meta’s turn. Mark Zuckerberg’s company is not starting from scratch: it already has MTIA chips for model training in production, but planetary-scale visual inference – think Instagram, Facebook, or Ray-Ban Meta glasses – requires a completely different class of accelerators and software stack. Entering this market will almost certainly mean accelerating the in-house hardware program to avoid leaving NVIDIA or others with the highest cost in the value chain. The ByteDance parallel is instructive: if Seedance achieves a 90% margin, it’s because it has vertically integrated everything from silicon to service. Meta will have to do the same, or risk being trapped by inference costs that erode any scale advantage.

This scenario paints a future where visual generation becomes an accessible commodity, but only for those who own the physical means of production. For the rest of the industry – startups, SMEs, public bodies – the lesson is equally clear: no generalist cloud provider can offer comparable long-term efficiency, because its business model relies on a margin on rented hardware. Consequently, for high-volume projects, the choice between cloud and on-premise becomes a strategic obligation: either internalize the hardware, with all its complexities, or accept thin margins and third-party dependence. For those evaluating on-premise deployment of LLMs and visual models, tools like those covered on AI-RADAR in the section dedicated to evaluation frameworks become essential to calibrate investments.

The game, in short, is no longer played on the most creative model or the largest dataset. It’s played on datacenters, inference pipelines, and chip design. And those who don’t understand this will be left watching.