Foundation labs structurally undercut the startups building on them

“Anthropic kills startups” shouldn't surprise anyone by now. But “they have distribution” is the lazy version of the reason, and it misses the part that actually makes them hard to beat. The real edge is the subscription: one price, an ever-growing pile of tools, and a single AI budget shared across all of them. You get to sell one tool, and you can't share anything with anyone.

Single-feature wrappers were always going to die

Early on, nobody knew how the market would shake out, so we got a wave of “ChatGPT for X” wrappers. As the base models got better, a single feature stopped being a business. Perplexity was a real step forward and still exists - but the moment OpenAI shipped search, the moat got thin fast. None of this is new. Thin wrappers die for the same reason they always have.

AI-native products are the hard case

Agents for coding, video, sales, design, support. The software around the model API is genuinely complex now, so there's real room to compete. The problem is the economics are backwards compared to normal SaaS.

Normal software had good margins: pay for servers, maybe an API, price above cost, done. AI-native products lose money. Free trials either cost you real money, ship a worse model, or don't exist. Everyone builds their own metering. And the main input - tokens - often costs more wholesale than what users will pay you.

The killer is that you can't predict it. You have no idea how much AI any given user will burn, so you price defensively - higher than you'd like - to avoid losing money on the heavy ones. High prices wreck conversion. That's how you end up with low conversion, $200/mo tiers, and users subscribing to five separate tools whose usage is really split across all five.

That last part - usage split across many tools - is exactly what decides the fight.

One subscription, one shared budget

Anthropic doesn't just sell a model anymore. It sells a subscription that keeps absorbing products - Claude Design, Claude Cowork, Claude in Excel and PowerPoint. One price, and every new tool is included. Crucially, the user's spend is one pool, shared across everything in the bundle.

That pooling is the moat. They buy inference at cost instead of retail, and then they spread that cost across every tool a subscriber touches. Add a tenth product and the price to the user doesn't move - the marginal tool is basically free, because it draws from a budget that's already paid for and shared with everything else. The user gets all of it for the price they were already paying.

You don't get into that pool. You sell one product and you cover its inference by yourself. It doesn't matter if yours is an order of magnitude better - you're pricing a single tool against a bundle where the next tool costs the user nothing extra. And the bundle prices calmly, because across millions of users and dozens of tools the usage averages out. You, pricing one product for one unpredictable user, can't. That's the same high-price-kills-conversion trap from before, except now your competitor doesn't have it.

This is why they work so hard to block third-party apps from using the subscription: keeping that shared pool closed to outsiders is the advantage. So this isn't “big company has distribution,” which would be nothing new. The unit of competition quietly became the bundle - and you can only field one product.

What this looks like in practice

Claude Code is bloated. A big chunk of its context window goes to static safety system prompts that mostly burn tokens and leave you less to work with. That's the gap lean tools like Forge CLI exist to fill, and I use one daily. But to make it affordable I'd have to wire my own Claude subscription into it - against the ToS, and risking a ban - because there's no legitimate way for a third-party tool to buy inference at subscription rates. The fact that people do this anyway shows how the lock-in works. It also hurts those tools with enterprises, which is probably the point.

Pricing floors. Take an AI video editor like usecardboard.com: no free trial, cheapest plan $32/mo. Fair enough on its own - they have a bill to pay and no way to predict your usage, so they price for the worst case. But every tool like that prices for its own worst case, so stack ten of them and a solo founder trying to ship a first demo just walks. Anthropic charges once and folds all ten jobs into the same shared budget. So when they ship a barely-usable video editor billed to your existing subscription, you don't switch because it's better. You switch because it's already in the bundle you pay for.

The counterargument

Specialization is real. Cursor didn't die because Claude Code exists, plenty of AI-native companies are doing well on top of these models, and a single lab can't build every app well. The API is also a real business - they want you building on it.

But none of that closes the gap. The lab can pick any vertical once it's big enough and fold a “good enough” version into the shared budget millions already pay for. “They make money from the API” doesn't help here - it's exactly why they keep that pooled, subscription-rate inference closed to you. You're allowed to compete. You're just not allowed to share costs the way they do.

What to actually do

1. Build things that don't depend on expensive inference. Analyzing a single stock with a tight prompt on a cheap model scales fine. A frontier model running on every keystroke does not.
2. Build a moat that isn't “my software is complex.” Data, distribution, workflow lock-in, a network - something a lab can't copy in a sprint.
3. Push the inference cost onto the user. Raw BYOK does this, but it can be clunky. Just don't forget to explain clearly how it works if your users aren't technical. Alternatively, you could try my thing (note: my opinion is definitely biased here) - Wattfare, which is essentially a simpler BYOK in an OAuth-like format. Either way, if the user funds their own usage, you stop guessing at how much they'll burn - and the bundle's shared budget has nothing to undercut.

I don't have a clean answer. The short version: if you build AI-native, you're building on top of the company that sells you your inputs - and bundles them under one shared budget you can't get into. Being better isn't enough when your competitor's next product is free to the user. Pick your battles accordingly.

- Bartek