Tactical insights for first-time founders to outsmart the burn, the churn & the breakdown.

Hey Founder,

Every big technology shift creates new winners, but their edge rarely looks exciting at first.

In AI, it might be something as unsexy as cost structure.

Right now, most companies can access the same intelligence. Two products can feel equally smart, often running on the same models underneath.

But one answers a query for $2.
The other does it for $0.20.

At 10 customers, nobody notices.
At 10,000 queries, that becomes the business.

One compounds profit. The other compounds cost.

The tricky part is that AI costs hide inside model calls, retrieval, embeddings, retries, and humans cleaning up the output until you realise you’re running economics you never designed.

The best model matters, but the cheapest way to deliver the same outcome may matter more.

This issue is about why cost structure is becoming one of AI’s real moats, and how to build it before your margins start leaking.

Let’s dive in. 

For most of the software era, the moat was the product.

Build a better feature and competitors needed months or years to catch up.

AI is changing that.

Today, companies can rent similar intelligence from the same model providers. As capabilities converge, the battle shifts.

Not to who has the coolest demo.
But to who can deliver the same outcome at the lowest cost.

In other words: who has the better COGS. 

Imagine two AI travel agents.

They book the same hotels, offer the same experience, and charge the same price.

Company A

Company B

The customer sees the same product.

The P&L doesn't.

At the same subscription price, the lower-cost company can reinvest more, lower prices when needed, survive mistakes, and compound profits faster. 

That's the cost flywheel: Lower cost → better margins or lower prices → more customers → more scale → lower unit costs.

NVIDIA is already playing this game. By using production data from its internal support agent, it distilled a 70B-parameter model into 1–8B models with comparable accuracy, cutting inference costs by roughly 98% and latency by around 70x.

As intelligence becomes abundant, the advantage shifts.
Not to the company with the smartest AI.
To the one that delivers it most efficiently. 

AI costs are messy, and they’re very easy to lie to yourself about.

A model can be cheaper per token and still more expensive per outcome if it needs more context, more retries, or more calls to finish the job.

Your provider prices in tokens. Your P&L lives in outcomes. If you don’t translate that into true cost-to-serve, your margins are already distorted. 

One customer resolves a task in two queries. Another burns thousands.
Under each request sit model calls, retrieval, tools, storage, observability, and human fixes.

Usage unpredictability isn’t a bug. It’s built into AI. Left unchecked, it compounds fast.

Inference, vector databases, retrieval, and AI-specific tooling all sit inside COGS and grow with adoption.

If you bury them inside “cloud costs,” your gross margin is lying to you. And as AI prices fall, your margins only improve if your unit costs fall faster.

In SaaS, heavy usage was usually good news.
In AI, it can be the thing that burns your runway.

Your cost to serve doesn’t stay fixed per user. It grows with usage, complexity, and edge cases.

That’s why expensive-to-run AI products may look useful now, but structurally weak later.  

Tokens, compute, and model calls aren’t just costs to minimise, they’re productive assets.

The winners won’t use the least intelligence, they’ll create the most value from every unit they spend.

Founders already understand this with capital: the goal isn’t to spend less, it’s to turn every $1 into $10 or $100 of enterprise value.

AI needs the same lens.

Not: “How much intelligence are we using?”
But: “How much value are we creating per unit of intelligence consumed?” 

Most founders track CAC, LTV, burn, and gross margin. They should. But AI adds three numbers you can’t ignore:

1. Consumption

How much intelligence you burn: tokens, API calls, GPU seconds, retries, retrieval.

2. Value per unit of consumption

What you create from that burn: tickets resolved, productive sessions, active seats, revenue.

3. Gross profit per unit of consumption

How much profit you keep for every unit of intelligence used.
That’s the real scoreboard.

The less intelligence you need to deliver the same customer outcome, the stronger your business becomes. 

Pick one outcome your product delivers: a resolved ticket, a summarized document, a finished draft, a booked meeting, an active seat.

Then trace what it takes to create it: model calls, retrieval, embeddings, external APIs, retries, and human review.

That is your real unit of consumption. Until you define it, you can’t manage it. 

Your AI costs are probably buried inside one cloud bill.

Split production from dev/test first. Then separate inference, retrieval, vector DBs, orchestration, monitoring, and human review.

Tag those costs to COGS, not generic OpEx.
Otherwise your gross margin is telling you a prettier story than the business is living. 

Once you see the costs, calculate gross profit per outcome:

Then segment it by customer type, plan, and usage intensity.

You’ll probably find at least one segment that’s underwater. That’s not failure. That’s finally seeing the business clearly.

Now attack the denominator: same outcome, fewer units burned.

Route simple work to cheaper models. Cache repeated or similar queries. Cut redundant retries and unnecessary checks. Stop sending irrelevant context on every request.

Same outcome, fewer tokens.

The product doesn’t get worse. The economics get better. 

Cost cutting has a floor. Value creation doesn’t.

Price around the outcome, not the token bill. Reuse one piece of intelligence across multiple workflows. Know which margin band your AI business type can realistically support.

The goal isn’t just cheaper intelligence.

It’s more value from every unit you consume. 

A founder DM’d me:

The person who chooses the highest salary will usually leave for the next highest salary.

That's not who you're looking for.

The best people are often optimizing for things money can't fully buy: ownership, growth, meaningful work, and the people they'll build with.

That's one game a bootstrapped founder can actually win.

Three things I'd do:

1. Be painfully specific.

Skip "we're changing the industry." Instead explain the exact problem you're solving, why you think you'll win, and why this role matters. Great people don't buy vague missions. They buy problems worth solving.

2. Be honest about the money.

Don't dance around it.

"I can't win the cash fight. I can give you meaningful ownership."

Then do the math together. Show the upside, acknowledge the risks, and let honesty become part of your offer.

3. Sell the company before you sell the job.

Don't just send an offer.

Pull them into a real problem. Introduce them to the team. Let them experience how decisions get made and how fast the company moves.

Momentum is hard to fake.
And it's often what they're really buying.  

Send it my way, just hit reply.

Some reads worth your time:

Before you go…

The AI winners won’t necessarily have the smartest models, biggest budgets, or most tokens. They’ll have the strongest economics.

That’s the moat.

See you next Thursday,

— Mariya

Hit reply and let me know. I read every single one (for real).

Hey, I’m Mariya, a startup CFO and founder of FounderFirst. After 10 years working alongside founders at early and growth-stage startups, I know how tough it is to make the right calls when resources are tight and the stakes are high. I started this newsletter to share the practical playbook I wish every founder had from day one, packed with lessons I’ve learned (and mistakes I’ve made) helping teams scale.

Mariya Valeva

Find me on LinkedIn