The days of needing photographers, studios, and production staff for a single product image are ending. E-commerce teams now generate hundreds of production-ready images in minutes. fal.ai compiled a report based on over 600 million inference calls, and a16z added their investment lens. The most striking fact: enterprise production deployments use a median of 14 different models simultaneously.
What is this?
This is an analysis published in February 2026 by a16z infrastructure partner Jennifer Li and AI investing partner Justine Moore. They layer investment-grade interpretation onto fal.ai's "State of Generative Media Report Volume 1."
fal.ai has a uniquely privileged vantage point — their inference engine serves 600+ models to millions of users who collectively generate billions of assets. In 2025 alone, 985 new models were integrated into the fal platform (video 450, image 406, audio 59, 3D 35, speech 35).
What changes?
1. There is no "god tier" model — multi-model is the default
In the LLM market, OpenAI, Gemini, and Anthropic command 89% of enterprise spend. But the image/video market is entirely different — it's intentionally fragmented, and that makes sense. A model that excels at photorealistic images won't necessarily handle background removal, sound generation, or multi-shot narratives.
2. The unit of work isn't one model — it's a workflow
Producing a single polished asset is rarely a single inference call. In practice: image generation → background removal → upscale → recolor → style LoRA. Even a short branded film requires chaining scene generation, camera motion, character persistence, dialogue synthesis, sound design, and post-production.
| Traditional pipeline | AI orchestration pipeline | |
|---|---|---|
| Product photography | Photographer + studio + weeks of editing | AI image gen → background swap → upscale (minutes) |
| Ad creative | Agency, 2-4 weeks | Hundreds of A/B variations (hours) |
| Game assets | 3D artist, weeks of work | Text-to-3D + auto-texture (minutes) |
| Video pre-viz | VFX team, weeks | Text-to-video + native audio (hours) |
3. Not all pixels are worth the same
For high-volume thumbnails and feed assets, use fast and cheap models (Flux). For hero assets like ad campaigns and brand imagery, pay for premium models (Nano Banana Pro). Cost optimization (58%) outranks model availability (49%) and generation speed (41%) as the top infrastructure selection criterion.
4. Industry adoption — advertising, gaming, and e-commerce lead
| Industry | Adoption | Primary use cases |
|---|---|---|
| Advertising | 56% | Campaign visuals, banners, social graphics at scale |
| Entertainment/Media | 43% | Storyboarding, pre-viz, VFX, promo clips |
| Creative Software | 31% | AI features in design platforms and editing tools |
| Education | 30% | Interactive learning videos, animated explainers |
| Retail/E-commerce | 19% | Automated product photography, virtual try-on |
75% of marketing organizations have adopted generative AI, but 80% still use it on less than half their work. The biggest barrier? 94% cite intellectual property ownership and liability.
Getting started: the essentials
- Don't lock into a single model
The enterprise median is 14 models. Separate by use case (high-volume assets vs hero assets) and leverage multi-model infrastructure like fal.ai or Replicate. - Design workflows as pipelines
Think generation → editing → upscaling → style application. A unified API interface across models is key. - Build a cost-quality matrix
Not every image needs a premium model. Define speed-cost-quality balance by use case and automate high-volume assets first. - Seriously consider open-source models
Flux and Qwen Image Edit closed the quality gap faster than expected. For brand consistency requiring fine-tuning on your own data, open-source can be the better choice. - Focus ROI on specific use cases
Broad experimentation yields poor returns. Companies that focused on high-value use cases (automated product photography, A/B creative testing) saw 65% achieve ROI within 12 months.
Watch out: IP risk
94% of marketing organizations cite IP ownership and liability as adoption barriers. Verify training data licenses, build audit trails for generation processes, and add human creative contributions to critical assets.
