OpenAI Just Obliterated a Whole Generation of Image Gen Startups
Foundation models aren't stopping at infrastructure — they're coming for the applications too. The only path forward? Expert agents, data flywheels, and narrow but defensible value propositions.
This week, OpenAI rolled out the upgraded image generation feature inside GPT-4o. The results? Not only are the visuals more photorealistic, but the model can now automatically place text in images — precisely where it should go — based on scene composition and character relationships.
For anyone who's been following the image generation space, this is huge. The kind of polished detail people have struggled with for years? Solved overnight. The feature instantly went viral, spawning waves of AI-generated images that look like they came straight out of a Ghibli film.
But behind the excitement, there's an uncomfortable truth:
Billions of dollars and years of engineering effort from image-gen startups were just wiped off the map.
It’s hard not to draw a parallel to when DeepSeek open-sourced their base model and instantly tanked the valuations of dozens of closed-source foundation models.
The Agent Gold Rush Faces the Same Question
While doing market research for a VC recently, I kept hearing the same questions:
“Is there a benchmark to predict which Agent products will succeed?”
“As long as we find the right user needs and scenarios, are we good?”
Honestly? I'm cautious.
My view is simple: Until we see a real leap in foundation model capabilities, most of today’s Agents are really just placeholders — prototypes to test if certain needs even exist. In some cases, they’re little more than cannon fodder.
The Unspoken Consensus Among Builders
Over the past year, the ecosystem has quietly converged on a few hard truths:
Foundation model companies can't rely on model performance alone to maintain their lead — technical and market moats erode fast. (See Stanford’s Foundation Model Transparency Index 2023.)
We're running out of high-quality, publicly available data for training. The small fraction of truly valuable, high-impact data might be less than 1%. (See Ethan Perez et al.'s Scaling Data-Constrained Language Models.)
That said, private, expert, and synthetic datasets still have room to grow. Meta and DeepMind are already making moves.
Pretraining + fine-tuning is hitting a wall when it comes to driving toward AGI. The academic world is now exploring RLHF 2.0, self-reflection, tool-augmented learning, and other next-gen approaches.
Open-source models are crushing inference costs (think Qwen, Llama 2, DeepSeek-V2). The old "API = money printer" business model is breaking down. Token-based monetization is becoming unsustainable, stretching out payback periods indefinitely. (See OpenAI Dev Day 2023.)
What Happens Next?
It’s clear: Foundation model companies will increasingly ship applications themselves, going directly after both consumer and enterprise markets.
Why?
The public data gold rush is over. The most valuable data now comes from real user interactions — the kind Agents naturally generate.
Applications drive frequent user interaction, generating consistent token consumption and offering more flexible monetization options (think usage-based, outcome-based, or task-completion-based pricing).
The Shift is Already Happening
Take a look around:
Deep Research quietly stopped offering APIs and embedded their models directly into products.
Cursor struggled as Claude's coding capabilities became unreliable (some speculate this was intentional throttling).
Cursor and Perplexity are both training their own small models to reduce dependence on foundation models.
Perplexity, for example, fine-tuned on DeepSeek to build its own search functionality.
In short:
Without your own small model, custom tools, or a closed data loop, Agent projects are going to have a hard time surviving.
Case Studies Worth Noting
Cursor treats its internal code completion model as a core asset.
The Windsurf team built a native Codeium-powered code model.
Perplexity trained its own search model based on DeepSeek fine-tunes.
My Take
The foundation model landscape is already set.
What comes next is pure competition — plus the added complexity of US-China tech rivalry.
Agent teams that are still just "wrapping LLMs into products" may soon find themselves swept away once foundation model companies fully commit to shipping applications.
But There Are Playable Strategies
If I were betting on this market, I’d focus on:
1️⃣ Becoming a vertical partner or capability outsourcer
Help foundation model companies handle pretraining, fine-tuning, or RAG for niche industries or specialized scenarios. This could make you an attractive acquisition target or a key player in the ecosystem.
2️⃣ Building specialist Agents
Most Agents today are orchestration tools at best, far from handling open-ended tasks.
There’s real opportunity in crafting focused, expert Agents that solve well-defined, valuable problems.
3️⃣ Owning the data flywheel
Products that can continuously accumulate dense, real-world interaction data will be the few capable of building lasting moats in this space.
One Last Thought for Founders and Investors:
The major players will eventually dominate the Agent interface.
Instead of trying to beat them head-on, focus on creating value atop their platforms.
The future may be less about a single "super Agent" and more about networks of specialized, expert Agents collaborating.
Don't just optimize for productivity.
Become the Agent that actually solves real problems — and build the data flywheel that could become your most valuable asset.