# Business Insights

By [DYLIT Chronicles](https://dylit.info/user/dylitmediabuzz)

[AI - Beyond the Hype](https://dylit.info/pr/ai-beyond-the-hype/6a9efac02e92664f4d50cf9d) > [Business Insights](https://dylit.info/ch/business-insights/6a9efac02e92664f4d50cfac)

The Frontier AI Profitability Problem AI is very likely to reshape how the world works. That part of the story gets most of the attention. A far less settled question is whether the companies building the frontier models will ever make good money doing it. Those sound like the same bet. The gap between them may be the most interesting open question in AI right now. Is the pie as big as it looks? Around 6 billion people use the internet today, according to the ITU. Let's be wildly generous. Say 5 billion of them paid $20 a month for AI. That's $1.2 trillion a year. Most won't pay anything, of course. Some will pay less, a few far more. But it's a useful ceiling. Add maybe $0.2 trillion from small businesses and $0.8 trillion from large enterprises, and the result is an illustrative market of about $2.2 trillion a year. These are rough assumptions for the sake of the argument. The real number could be higher or lower. Either way, enormous. And yet the size of the market says surprisingly little about who keeps the profit. Software that behaves like infrastructure Traditional software has one beautiful trait. Once it's built, serving the next customer costs almost nothing. That's why the best SaaS businesses have such fat margins. Frontier AI doesn't get that luxury. Every question asked burns compute. Every word it writes burns compute. Harder reasoning burns more. And every new generation of model needs a bigger, more expensive training run than the last. The result is something with the reach of software and the cost structure of a utility. It also creates an odd loop. Models get better, so people use them more and for harder tasks, so compute demand grows even faster... while the price of each unit of intelligence keeps sliding down. Great news for users. Less obviously great for whoever is paying the compute bill. Success itself becomes expensive: the more people rely on a model, the bigger the bill gets. Everyone is buying shovels from the same store Here's what makes it even trickier. OpenAI, Anthropic, Google, Meta, xAI and the rest compete fiercely for customers and talent. But they all queue up at the same few suppliers for chips. Analyst estimates put NVIDIA's share of the AI accelerator market somewhere between roughly 70% and 86%, depending on how it is measured. AMD, Google's TPUs and other custom chips are fighting for the rest. NVIDIA doesn't need to pick a winner. Whichever lab comes out on top, open source included, still needs compute. The supplier gets paid in every scenario. That's a very comfortable seat at the table. Intelligence is getting cheaper, fast  The cost of a given level of AI capability is falling at a pace that's hard to grasp. Stanford's 2025 AI Index found that running a model at roughly GPT-3.5 level dropped from $20 per million tokens in November 2022 to $0.07 by October 2024. That's about 280 times cheaper in under two years. Epoch AI's analysis found the price of equivalent performance falling anywhere from 9x to 900x a year depending on the task. Wonderful for society. Awkward for producers. Imagine a lab finds a way to cut its inference cost by 10x. If competitors can get close, customers will expect much of that saving to show up in lower prices. Usage might explode and revenue might climb. Profit is unlikely to grow anywhere near 10x. That's how commoditization usually plays out. The efficiency is real. It just doesn't all stay with the inventor. And the chasing pack is closing in. Stanford's same report found the gap between the best open-weight and closed models on the Chatbot Arena leaderboard narrowed from about 8% to 1.7% in a single year. When a near-equivalent model is available cheaply, or even free, holding premium prices gets very hard. The number that actually matters So a narrower question deserves more attention: How much profit does each extra dollar of compute actually produce? The compute bill never really ends. Nobody builds a giant GPU cluster once and calls it done. New chip generations keep arriving. Rivals build bigger clusters. Reasoning models chew through more compute per answer. Customers want everything faster. Much of that spending is simply the price of staying in the race. Fall a generation behind and customers can switch to a rival's model with a few lines of code. So the honest question for any lab becomes: if another $100 billion goes into compute, how much gross profit will it generate over its life, after running it and eventually replacing it? If the answer is $400 billion, the economics are spectacular. If it's $120 billion, they're mediocre at best. That difference is the whole debate. The case for optimism To be fair, the bull case is strong and deserves to be taken seriously. The biggest argument is that AI stops being a $20 subscription and starts being labor. If an agent can reliably do work that costs a company $100,000 a year, charging $10,000 or $20,000 for it is entirely reasonable. Then the market is measured against global wages, which run to tens of trillions of dollars a year. That changes the math completely. A coding agent, a customer support agent or a research assistant priced as a fraction of a salary is a very different business from a chatbot priced like a streaming subscription. Customers also tend to stick with whatever is wired into their workflows, which could finally give labs some pricing power. The second argument is efficiency. If algorithms and hardware improve faster than demand grows, the cost of producing intelligence could fall faster than its price. MIT researchers estimate algorithmic efficiency alone is improving roughly 3x a year. And then there are the giants. Google, Microsoft, Amazon and Meta don't need AI to be a great standalone business. Google can earn it back through Search, YouTube and Cloud. Microsoft through Azure and Office. Amazon through AWS. Meta through better ads. They can live with economics that would sink an independent lab. A standalone model company has no search engine or cloud business to fall back on. The model is the product, and every dollar of compute lands directly on its own income statement. Which is why "AI will be profitable" and "frontier model companies will be highly profitable" are two very different claims. A lot of the confusion in this debate comes from blurring them. The uncomfortable possibility Now picture an AI industry earning $2 trillion a year. Sounds incredible. Then $1.5 trillion goes to compute and infrastructure, and another $400 billion to everything else. What's left is $100 billion of operating profit. A 5% margin. AI would be one of the most important technologies in human history... and still earn mediocre returns on the capital it consumes. It wouldn't be the first time. Airlines transformed how the world moves, and the industry has struggled for most of its history to earn back its cost of capital. Fast progress plus heavy capital plus fierce competition is a tough combination for profits. So where does the value go? The lazy version of the bear case, that AI simply won't make money, doesn't hold up. AI will make money. Lots of it. The sharper question is this: Will the value frontier labs capture grow faster than the capital it takes to stay at the frontier? If yes, the world may be watching some of the greatest businesses ever built. If no, the value still gets created. It just leaks out along the chain. Users get it through falling prices. App builders get a share. Cloud providers and data centers take theirs. Chipmakers take a big slice. And the model companies end up in the middle of it all, creating extraordinary value while fighting to keep enough of it. None of this is a reason to bet against AI. It's a reason to be precise about who actually gets paid. Which way it goes is still an open question. Where the margin finally settles may tell the real story of the AI era. Sources ITU, Facts and Figures 2025: https://www.itu.int/en/mediacentre/Pages/PR-2025-11-17-facts-and-figures.aspx Stanford HAI, 2025 AI Index Report:  https://hai.stanford.edu/ai-index/2025-ai-index-report Epoch AI, LLM inference price trends (2025):  https://epoch.ai/blog/llm-inference-price-trends Gundlach et al., MIT FutureTech, "The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference" (2025):  https://arxiv.org/abs/2511.23455 NVIDIA AI accelerator share, compiled analyst estimates incl. IDC and Silicon Analysts (2026):  https://commandlinux.com/statistics/ai-gpu-market-share-nvidia-vs-amd-vs-intel/ Market sizing and margin scenarios are illustrative assumptions.