For the past few years, the price of AI has moved in one direction: down. OpenAI, Anthropic, Google — they've been in a race to the bottom on API costs, subsidised by venture capital, burning cash to win market share. That era appears to be ending.

TechCrunch's weekend deep-dive coins a term that's already spreading: the "Tokenpocalypse." The thesis is straightforward. The major AI labs — OpenAI, Anthropic, and others — are preparing for IPOs. Public markets demand a path to profitability. The easiest lever to pull is pricing, and several labs have already begun pulling it.

What's actually changing? Token prices — the per-unit cost of sending text to and receiving it from an AI model — have crept up in recent months after years of cuts. The increases are modest so far, but the direction has reversed. For individual developers building hobby projects, this is mildly annoying. For enterprises running millions of AI calls a day, it's a line item that can reshape a business case entirely.

The timing is not coincidental. OpenAI has been telegraphing an IPO pathway for much of 2026, and Anthropic has had similar conversations with investors. When you're telling public market investors a story, "we're the cheapest option" is not the story you want to tell. "We have pricing power" is.

The deeper problem is dependency. Many companies — from scrappy startups to large enterprises — have built products and workflows on the assumption that AI inference costs would continue to fall. Some have even priced their own products on that assumption. A sustained reversal puts those businesses in a difficult position: absorb the margin hit, raise prices on their own customers, or rebuild on cheaper alternatives.

Those cheaper alternatives do exist. Open-source models like Meta's Llama family and Mistral's releases have made self-hosting increasingly viable. But self-hosting carries its own costs — compute, engineering time, reliability overhead — that aren't zero. For most mid-sized companies, switching is a genuine project, not an afternoon's work.

Watch the Notion/Anthropic situation as a canary here. This weekend, Notion experienced a service disruption tied to its Anthropic integration. The outage was brief, and Notion's head of product seemed genuinely surprised by the social media response. But the episode illustrates a structural risk hiding inside every AI-native product: when your core feature runs on someone else's infrastructure and someone else's pricing, your reliability and your margins are both partially out of your hands.

For Australian businesses, the stakes are real. Australian AI adoption has accelerated sharply across professional services, government, and the technology sector. Many of those deployments sit on US-priced, US-hosted API services. A sustained price increase in USD terms, compounded by any AUD/USD movement, hits harder on this side of the Pacific. Procurement teams that haven't stress-tested their AI cost assumptions against a 2–3× price increase should probably do that this quarter.

The AI price war was never going to last forever. The question now is how fast the pendulum swings — and who gets caught without a chair when the music stops.