When Building Is Cheap, Taste Is the Moat

The scarce asset in an AI-native company is no longer the ability to build. It is the judgment to decide what is worth building, what to leave out, and why a customer should care. AI has collapsed the cost of production. It has not produced taste, architecture, or customer empathy. Those stay scarce. Scarcity is where margin lives.

That is the through-line of Tony Fadell's career. He shipped the iPod, helped build the iPhone, and founded Nest, which Google acquired for $3.2 billion in 2014. In a recent conversation about building in the AI era, his argument was blunt. When everyone can produce, the premium moves to the people who know what to produce. We build AI-native companies in Brazil and Latin America for a living, and his frame matches what we see in the field every week.

This is not a nostalgia piece about a hardware legend. It is a working thesis about where value accrues once the model itself is cheap for everyone.

The cost of building collapsed

Start with the number that changes everything. For a model of equivalent performance, the cost of inference is falling by roughly 10x a year. Andreessen Horowitz named it LLMflation and measured it. The cost of LLM inference dropped by a factor of 1,000 in three years. Hitting an MMLU score of 42 cost about $60 per million tokens with GPT-3 in November 2021. By late 2024 an open model reached the same score for about $0.06 per million tokens, per [a16z](https://a16z.com/llmflation-llm-inference-cost/). Independent measurement from Epoch AI puts the median decline across benchmarks at about 50x per year, and rising.

Read the second-order effect, not just the first. When the cost of building falls this fast, building stops being the bottleneck. The bolt-on AI feature, the chat box next to the old product, is now available to everyone at a price that approaches zero. Cheap to build is the same sentence as not defensible. If you can ship it in a weekend, so can the next ten teams. The collapse that makes AI-native companies possible is the same collapse that makes most of them disposable.

So the interesting question is no longer what the model can do. It is what you decide to do with it, and what you refuse to do.

Taste is a test, not a vibe

Fadell's word for the scarce asset is taste. It sounds soft. It is not. Taste is the discipline of knowing what to leave out.

His clearest example is the iPhone keyboard. The data did not settle the debate between a physical keyboard and a virtual one. Someone had to decide against the evidence of the moment and ship a glass screen with no keys. Breakthrough 1.0 products cannot be fully validated by user research, because users cannot judge an experience they have never had. In a new category, consensus kills differentiation. Conviction makes it.

AI sharpens this, because AI makes addition free. You can bolt on every feature, every integration, every clever capability the model exposes. The hard job becomes subtraction. A product that does everything is a product no one can describe, and a product no one can describe does not get bought. The founder's real work is deciding what the product is not.

One honest caveat. Taste is not the same as ego. Informed taste is judgment accumulated through doing the work and staying close to the customer. Founder delusion wears the same clothes and skips the work. The test is whether the conviction survives contact with real users, not whether it feels bold in the room.

Sell the painkiller, not the model

The antidote to feature-chasing is to start with pain, then add the new technology. The question is never what the model can do. It is what expensive, frequent, budgeted pain can now be solved differently because the model exists.

This shows up most clearly in how a company talks. We use agents is not a story. It describes your supply chain, not the customer's life. We cut claims processing from ten days to ten minutes is a story, because it names a pain the buyer already pays for. The strongest AI products are marketed around the human job to be done, not the model capability that powers it.

The screening question we apply to every venture is simple. Is this a painkiller, a vitamin, or a toy? Painkillers attach to a budget line and a measurable cost. Vitamins are nice and get cut first in a downturn. Toys get demoed and never bought. AI makes toys cheaper to build than ever, which is exactly why the discipline of finding real, budgeted pain matters more than ever.

The product is the whole system

Here is the trap that catches AI startups. They believe the product is the model plus an interface. Fadell's whole career argues the opposite. The winning product is the entire customer journey.

The iPod was not a music player. It needed iTunes and the iTunes Music Store to become the iPod. The iPhone needed the App Store. Nest needed a new retail motion, a new install experience, and design that made a thermostat worth showing a friend. The object was never the product. The system around it was.

The lesson for AI-native companies is direct. A thin interface over someone else's model owns nothing. Anyone can rent the same model at the same falling price. The company that owns the workflow, the data, the onboarding, and the outcome owns a position. The company that owns only the prompt owns a feature that the model provider can absorb in its next release.

Where the moat actually lives

Models commoditize. The cost curve guarantees it. When every competitor can call the same model at the same falling price, the model cannot be the moat. Defensibility moves to what the model touches.

Three places hold up, and none of them come from the model. They come from judgment about how the model meets a specific customer's reality.

• Proprietary data and network effects. Data generated inside a real workflow, that compounds with every use, and that no competitor can buy. The product gets better as customers use it, and the gap widens on its own.
• Vertical workflow ownership. A company that redesigns one painful, regulated, messy process end to end is harder to displace than a horizontal copilot that floats above everyone's work and owns none of it.
• Trust. AI products are becoming intimate. Copilots, assistants, agents that act. Transparency, permissions, auditability, and human override stop being compliance overhead and become product features that win the deal. The most trusted product often beats the most aggressive one.

How Avante builds for taste

Avante Ventures is a venture studio building AI-native companies in Brazil and Latin America. We do not bet on a model. We build the loop around it. Every venture is AI-native from day one, with a model in the core product loop and a copilot positioned to capture proprietary data inside a real workflow. That is the recurring pattern across the portfolio: [copilot to data to fund](/library/copilot-to-data-to-fund-flywheel). Build the copilot, generate the data, use the data to raise and deploy capital. More on the thesis at [why Avante](https://avanteventures.com/why-avante).

The studio model is itself an exercise in subtraction. We solve company plumbing once, centrally, which routes roughly $300K to $500K of effective capital per venture into product and traction instead of overhead. The same logic that drops inference cost 10x a year, applied to the company itself. Do the expensive thing once and let every venture launch lean. We launch three to four ventures a year through six stages, Research, Partner, Build, Traction, Revenue, Compound, with $500K to $1.5M deployed per venture and co-founder economics retained.

The structural payoff is the studio model's track record, covered in full in [Why Venture Studios Outperform VC in LATAM](/library/why-venture-studios-win-latam). The model rewards exactly what the AI era rewards: judgment, iteration, and capital efficiency over raw build speed.

One last Fadell idea, because it sets the right expectation. Category-defining products take three generations. First you make the product, then you fix the product, then you fix the business. The iPod, the iPhone, and Nest were not iconic at launch. They earned it through iteration. The job of a studio is to underwrite that iteration capacity, not to bet on a perfect version one.

In an AI-saturated market, anyone can build. Few decide well. Taste is not a soft virtue. It is the leverage that turns cheap production into a durable company. Browse the rest of [the Library](https://avanteventures.com/library).