AI Diminishing Returns Are Here. Stop Chasing the Frontier.

Meta Just Blinked. You Should Pay Attention.

Also, meta reportedly delayed their next flagship AI model, codenamed “Avocado”, after sinking billions into training it. Indeed, the delay is framed as a technical setback. But I think it’s a signal. A very loud one. We are watching AI diminishing returns play out in real time, at the highest level.

Furthermore, this isn’t a story about Meta. Of course, Meta will be fine. After all, they have the money to try again. This is a story about the arms race eating itself. And what it means for everyone building with AI.

The Math Behind AI Diminishing Returns Is Brutal

So here’s what the training cost curves actually look like. GPT-2 cost around $40,000 to train. GPT-3 cost around $4 million. GPT-4 cost somewhere between $50M and $100M. The next generation? Estimates run north of $1 billion for a single training run.

Meanwhile, what did end users actually gain from GPT-3 to GPT-4? Meaningful improvement. From GPT-4 to whatever comes next? Most users cannot tell the difference on real-world tasks. The benchmarks improve. The user experience often does not.

That is the definition of AI diminishing returns. Each dollar buys less capability. Each capability gain delivers less real-world value. Yet the narrative keeps saying: you must chase the frontier or you’ll fall behind.

Moreover, who benefits from that narrative? The GPU manufacturers. Not you.

Why “Good Enough” Is Wildly Underrated

In fact, there’s a model available right now that writes production code. It summarizes documents, analyzes data, and generates marketing copy. Moreover, it costs fractions of a cent per call. It runs reliably. It’s probably not the newest model.

However, for the vast majority of business use cases, it’s completely sufficient.

In addition, think about the actual jobs companies are automating with AI. Drafting emails. Extracting data from PDFs. Summarizing support tickets. Categorizing feedback. Answering FAQ questions. For example, none of these tasks require frontier-level reasoning. They require a decent model plus a well-designed workflow.

Moreover, the gap between “good enough” and “frontier” is shrinking. Models that were considered cutting-edge eighteen months ago are now open-source and free to run locally. The frontier moves, but the useful middle keeps getting better too. AI diminishing returns means the frontier costs ten times more for improvements that matter ten percent more.

Similarly, most builders are optimizing for the wrong variable.

The Companies Actually Winning Aren’t Model Chasers

Meanwhile, look at the AI companies generating real revenue. They are not building their own frontier models. They are building workflows, interfaces, and integrations on top of existing models.

Specifically, they’re winning because they figured out the hard parts. The hard parts are not the model. The hard parts are reliable data pipelines. Consistent prompt engineering. Context management at scale. Human-in-the-loop checkpoints. Error handling. User trust. Fast iteration cycles.

In other words, the moat is operational. It’s about execution. It’s about ten thousand small decisions. Specifically, those decisions make AI useful inside a real product, for real users.

Indeed, a startup that spends six months chasing the best model is burning runway. A startup that spends six months building a rock-solid workflow on a good-enough model is building a business. The difference in outcomes is dramatic.

Furthermore, when a better model drops, workflow-first companies upgrade in a day. They swap the model out like swapping a supplier. But model-first companies are structurally dependent on the frontier. They have to chase it forever or die.

What This Means for Startups Specifically

In fact, if you’re a startup building with AI right now, here’s my honest take: stop reading model benchmark leaderboards. They are not your business.

Instead, ask these questions. What workflow are you automating? Notably, what data do you have that others don’t? In fact, what does “good enough” look like for your specific user, on your specific task? How do you make the output reliable enough that users actually trust it?

Because of this, the best AI startups I’ve watched succeed all share one trait. They got obsessive about the workflow long before they got obsessive about the model. They treated the model as infrastructure, important, but not the product.

Also, there’s a financial argument here. Running GPT-4-class models via API at scale is expensive. Running slightly older but still very capable models is a fraction of the cost. That margin difference compounds fast. A company with 40% AI infrastructure margins can reinvest in product. A company burning on frontier API costs cannot.

Still, I know what you’re thinking. But what about the use cases that actually need the frontier?

The Counter-Argument, And Why It’s Narrower Than You Think

Specifically, yes, frontier models matter. I’m not arguing they don’t. But let’s be honest about which use cases genuinely require them.

Consequently, complex multi-step reasoning chains. Novel scientific research assistance. High-stakes legal or medical document analysis where nuance is critical. Agentic systems making dozens of interdependent decisions. These are real. These exist.

However, these represent maybe 10-15% of actual AI use cases in production today. The other 85-90%? They’re doing things like: classify this support ticket, generate this product description, summarize this call transcript, answer this FAQ. For those tasks, frontier models are overkill. Sometimes they’re even worse, slower, more expensive, and prone to over-thinking simple tasks.

Yet most companies build as if they’re in that 10-15%. They justify frontier costs by imagining future complexity. In other words, they pay for headroom they never use.

Consequently, they’re essentially buying a Formula 1 car to commute to work. It’s impressive. It’s expensive. It does not make the commute better.

The Real Competitive Advantage in the Next Two Years

Likewise, here’s where I think this goes. The winners in the next wave of AI won’t be the companies with the biggest models. They’ll be the companies with the best workflows, the best data flywheels, and the fastest iteration loops.

Moreover, as open-source models continue to improve, the competitive advantage of proprietary frontier models keeps shrinking. Llama-class models are already good enough for a huge range of tasks. That range will expand. The gap between “free, open, local” and “expensive, proprietary, frontier” will narrow.

Therefore, any strategy that depends on permanent frontier access is fragile. The ground keeps shifting. The only stable ground is operational excellence.

Build great workflows. Own your data. Move fast. The model is a commodity. Your process is the product.

What Meta’s Delay Actually Tells Us

Meta’s “Avocado” delay isn’t embarrassing. It’s inevitable. When you’re spending billions to eke out marginal gains on benchmarks that don’t map cleanly to user value, delays happen. Because the returns are diminishing. Because the problems are getting harder faster than the compute is getting cheaper.

In addition, the companies watching that story and thinking we need to keep up are falling for a trap. The race is real. But it’s not your race. You are not Meta. You do not have their cash, their infrastructure, or their strategic reasons for owning frontier AI.

But you do have something they don’t. Instead, you can move fast, stay lean, and build for your users. No billion-dollar training runs required.

Use that.

The era of AI diminishing returns on frontier chasing is here. The builders who recognize it early will win. In contrast, the treadmill runners will spend a fortune to stay in the same place.

Stop chasing the frontier. Start building the workflow.

For additional context, see OpenAI’s research on AI capabilities.