Er, How Are These AI Companies Going to Make Money?

I was in Berkeley a year ago, talking to someone who worked for one of the major AI companies, and when I questioned the financial sense of LLMs, they said this in response:

Alright, so imagine that we stall out, and the LLMs replace just 5% of the US workforce. That’s 5% of a huge amount of money.

Doing the napkin math after the fact, if there are around 160 million people in the US civilian labor force, earning an average of ~$70,000, that’s in the range of $500 billion to $600 billion a year.

This was something that really stuck with me, mostly because … when you replace a job, the company doesn’t get all that money, right? That is to say, if an LLM could make a position that a company is paying $70,000 a year for, the company that owns that LLM does not suddenly make $70,000 a year. Value created does not equal value captured. Economics has a name for this: consumer surplus.

I think containerized shipping is a good comparison. When Malcolm McLean introduced it, unloading cargo went from $5.86 a ton to 16 cents a ton, a 36-fold decrease. This made him and his company rich, for a time, as infrastructure was built out, but as soon as he’d proved the effectiveness of it, competitors rose up, and eventually this was no longer his advantage, because everyone was doing it. Prices fell, and the global economy benefitted … but this didn’t make shipping much more profitable than it had been, not in the long term.

The AI company charges some kind of monthly or yearly plan, or perhaps charges some amount of money per token. Some of this goes to pay for compute costs, some goes to administration, sales, and marketing, and some of it pays down the enormous cost of R&D.

So if AI is good enough to replace a $70,000/yr worker with $10,000/yr compute, this is great for the company employing that worker. They save $60,000/yr! It’s probably less great for the worker. Or, more realistically, the company makes each of their $70,000/yr workers more productive for that same $10,000/yr compute, augmenting instead of replacing. But how much does the AI company actually make?

In theory, with two AI companies competing against each other, prices are pushed down to a margin above compute, which then goes to paying down R&D costs (mostly training runs), staff, etc. It’s extremely difficult to guess what the margins are here, given that we’re still in early days and the companies are in the “burn cash” phase while also being extremely tight-lipped about their financials. I’ve seen all kinds of estimates, but a gross margin of 50% seems fairly reasonable for the frontier models, and net margin of maybe 25%.

So this means that a $70,000 job replaced by $10,000 in compute with a 100% markup becomes $20,000 in cost-to-consumer, with $5,000 actual profit. Or in other words, if this holds for the economy as a whole, then $500 billion to $600 billion a year becomes $35 to $45 billion. In a market with healthy competition, this gets split in many directions.

Competition is what produces margins like this, and competition in the AI space is steep. There are five or six credible frontier labs right now, along with state-funded models. There’s the threat of model distillation and weight theft, which are their own cans of worms. There are competing open weight models, which might someday be competitive and will at least cut down the market. It remains to be seen whether any of these companies can build a proper moat that makes them into a monopoly, something that most software companies have depended on for their high margins.

The core argument here is that the benefits of AI seem like they’ll be reaped by the consumers, assuming that the AI is doing actual work instead of revenue-neutral hype-driven integrations. And the AI companies will be left with something that looks more like a commodity market than a SAAS market.

For context, the estimated total capex spend for the eight Big Tech companies is going to be more than $600 billion this year, the majority of it AI-related. A lot of that is going toward data centers and chips. The market is growing, but total revenue was $37 billion in 2025. That ratio is staggering, and while capex spending necessarily precedes revenue, I keep asking myself “what revenue though”. Even in a favorable scenario, the economics for these companies seem grim. Unless one of these companies can build a durable moat, which seems unlikely given low switching costs and little difference in capabilities, the benefits flow to consumers rather than suppliers. The current breed of AI-skeptic seems to be arguing that the AI don’t work. I’m left wondering whether the LLMs might prove to be a revolution that doesn’t make particularly much money for the companies that supply them.