The second most persistent myth about AI is that leading companies are selling at a loss, burning cash to fuel hype, and relying on endless investor funding to stay afloat. This narrative paints the frontier AI labs as overvalued bubbles, doomed to collapse once the money dries up.
But the reality, backed by recent independent analysis, is quite different. Frontier AI companies are already generating substantial gross margins on inference — the process of running models to deliver responses — often in the 40-50% range, with clear paths to higher profitability as costs continue to plummet.
A January 2026 report from Epoch AI ("Can AI companies become profitable?"), one of the most respected independent research groups tracking AI progress (alongside METR, a nonprofit focused on evaluating frontier model risks), dives deep into this question using public data and estimates, particularly around OpenAI's GPT-5 era.
Key findings from the analysis:
- Gross margins (revenue minus direct inference compute costs) for running models like GPT-5 appear to hover around 48-50%. This is lower than the 60-80% typical for mature software companies but still robust — and far from "selling at a loss" on the core product.
- For OpenAI specifically, inference costs were estimated at around $3.2 billion against $6.1 billion in relevant revenue during GPT-5's active period, yielding roughly $2.9 billion in gross profit (≈48% margin).
- This aligns with other reports: OpenAI's compute margins (a proxy for inference profitability) reportedly climbed from ~35% in early 2024 to as high as 70% for paid users by late 2025. Anthropic, meanwhile, saw gross margins improve dramatically from negative figures in 2024 to around 40% in 2025 projections, with expectations of reaching 50-77% in coming years as inference efficiency improves.
These margins come from "turning sand into intelligence" — leveraging cheap compute, optimized hardware (e.g., custom clusters), quantization, distillation, and massive scale. The inference business is profitable on a gross basis, even if overall operations (including massive R&D, salaries, marketing, and data-center builds) remain loss-making for now.
Why the "Loss-Leader Hype" Myth Persists
Several factors keep this misconception alive:
1. Psychological denial and fear of change
Rapid disruption is uncomfortable. The idea that AI labs are just hyping vaporware to raise funds offers a comforting explanation: "This won't really change everything; it'll fizzle out." Cognitive bias favors narratives that preserve the status quo.
2. Mind-blowing cost declines
The cost of inference for GPT-4-level intelligence has fallen dramatically — by factors of 100x to 1,000x in just 2-3 years, depending on the benchmark and model. Epoch AI and others track trends showing 10x drops per year in many cases (e.g., price per token for high-performance models). This pace is faster than historical tech revolutions (PC compute, internet bandwidth), creating sticker shock and skepticism: "If it's getting so cheap, how can anyone make money?" Yet falling unit costs are exactly what enables profitable scaling — lower prices drive exponentially higher usage, which in turn funds further optimization.
3. Misunderstanding Silicon Valley culture
Outside tech hubs, it's hard to grasp the "growth über alles" philosophy. Companies like Uber, early Google, Amazon, and Twitter (now X) lost money for years while capturing market share. Losing temporarily to dominate is rational in winner-take-most markets. If an AI lab prioritized short-term profits over aggressive investment in better models and infrastructure, a faster-moving competitor would capture the future.
Frontier labs are raising and spending tens (soon hundreds) of billions on training next-gen models and building gigawatt-scale data centers — not because their core economics are broken, but because winning the race requires massive upfront bets.
Also read:
- The 73% Collapse: How AI Is Erasing Entry-Level Tech Jobs and Rewriting the Career Ladder
- All Roads Lead to Fanvue: Why AI Influencers Are Flocking to the Platform in 2026
- Why Is Fertility in South Korea So Low?
- The Egg That Hatched "Living" Robots: Tamagotchi at 30
The Bigger Picture in Early 2026
Frontier AI companies are not running unsustainable Ponzi schemes. Inference is a high-margin activity that covers direct costs and leaves room for growth. Losses stem from enormous R&D and capex — training runs now cost billions, and building proprietary clusters is capital-intensive — but these are investments in future dominance, not signs of failure.
As inference costs keep falling (driven by hardware advances, algorithmic efficiencies, and economies of scale), gross margins should expand toward traditional software levels (70%+). Revenue is exploding too: OpenAI and peers have seen annualized figures surge into the tens of billions, far outpacing forecaster expectations.
The myth that AI leaders are just "hype machines losing money" ignores the evidence. They're building real, profitable businesses — albeit ones that require unprecedented investment to stay ahead. The economics aren't broken; they're just different from legacy software. In a world where intelligence is becoming commoditized at lightning speed, the winners will be those who scale fastest and smartest.
