I was listening to a podcast recently where Anthropic CEO Dario Amodei spent about 30 minutes trying to explain how his company thinks about profitability. Honestly, he made it more complicated than it needed to be. But buried in that conversation was something genuinely useful for accountants — a real-world example of a concept you already know from manufacturing, showing up in one of the most talked-about industries in business right now.
So I made a lesson about it.
The Question Everyone Is Asking
AI companies like Anthropic, OpenAI, and Google DeepMind are spending billions of dollars a year and losing money. At the same time, their executives keep saying their underlying unit economics are healthy. How can both of those things be true?
It sounds contradictory. But once you understand the structure of these businesses, it actually makes sense — and it has real implications for accountants advising clients, doing valuations, or evaluating investments in this space.
Two Costs, Not One
The first thing to understand is that AI companies have a dual cost structure that doesn’t exist in traditional software businesses.
Cost number one is inference. Every time you ask ChatGPT or Claude a question and get a response, that costs the company real money in GPU compute. It’s not like traditional software where you build something once and the marginal cost to serve each new user is basically zero. With AI, every single query burns compute.
Cost number two is training. Building the next model — the one that keeps you competitive — costs billions. OpenAI’s GPT-4 reportedly cost over $100 million to train. Next-generation models are projected to cost far more. And you have to keep doing it. Stop training and your model falls behind. Fall behind and your revenue disappears.
So you have ongoing operational costs (inference) sitting right next to massive R&D costs (training). Both are real. Both are large. And they hit the income statement very differently.
The Gross Margin vs. Operating Profit Gap
Here’s where it gets interesting for accountants. Amodei described inference gross margins of around 75%. That’s strong — comparable to a healthy SaaS business. So on the inference side of the business, the unit economics genuinely are good. Revenue is coming in well above the cost to deliver it.
But then you layer in the training spend — which generates no direct revenue — and suddenly you’re looking at a multi-billion dollar operating loss.
Both numbers are real. The company isn’t lying when it says gross margins are healthy. It’s also not lying when it reports massive losses. They’re describing two different layers of the income statement.
For any accountant doing valuation work or advising a client on an AI-adjacent investment, this distinction matters a lot. A client who tells you their gross margins are 75% is telling you something very different from what their income statement shows. Your job is to bridge that gap.
Underabsorbed Overhead — Sound Familiar?
The second concept Amodei touched on — without using this term — is underabsorbed overhead. And this one every CPA already knows.
AI labs have to purchase or contract for GPU compute capacity many months — sometimes years — in advance. They commit to a fixed cost before they know how much demand they’ll actually see. Sound familiar? It’s the same problem a manufacturer faces when it builds out a factory and then has to fill it.
If demand comes in lower than expected, that unused capacity hits the P&L as an unabsorbed cost. The company reports a loss — not because the business model is broken, but because they bought more factory than they needed this year.
Amodei made this point explicitly. Whether an AI lab is profitable in a given year depends largely on how accurately they predicted demand when they committed to compute. Over-predict demand and you’re profitable. Under-predict and you’re not. The business model itself doesn’t change — just the utilization rate.
This is a timing issue. Not a structural one. And knowing that difference is exactly what clients need from their accountants right now.
What the Path to Profitability Actually Looks Like
Anthropic is targeting profitability by 2028. OpenAI isn’t expecting to break even until 2030. Both companies are currently burning billions annually while reporting strong revenue growth — Anthropic recently hit $19 billion in annualized revenue.
The path to profitability, in Amodei’s model, isn’t about slashing R&D. It’s about revenue growing fast enough to consistently fill the compute capacity being purchased. When that happens — when demand prediction gets more accurate and inference revenue reliably covers both inference costs and a portion of training — the underlying economics show through on the income statement.
Whether that happens on schedule is a forecasting question. And forecasting questions are something accountants are very good at evaluating.
Key Takeaways
- AI labs have a dual cost structure: inference (ongoing, per-query) and training (large, periodic R&D). These hit the income statement differently.
- Strong gross margins on inference and large operating losses can coexist. They’re not contradictory — they describe different layers of the P&L.
- Compute capacity must be purchased before demand is known. When demand misses, the result is underabsorbed overhead — a timing issue, not a structural failure.
- Profitability follows when revenue growth consistently fills purchased compute capacity. That’s the mechanic behind Anthropic’s 2028 target.
- Accountants advising on AI-adjacent valuations, lending, or investment should separate gross margin analysis from operating margin analysis — and understand what’s driving the gap.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply