# Enjoy the Bubble While It Lasts: The Secret Behind Your AI Bill > Source: > Published: 2026-07-28 11:48:11+00:00 Somebody else is paying for your AI subscription. Not most of it. Almost all of it. That's the open secret of this whole era, and it's not on the pricing page. The arithmetic is inconvenient enough that nobody selling you a plan is in a hurry to walk you through it. SemiAnalysis [did the obvious experiment](https://x.com/SemiAnalysis_/status/2064815044085318040): buy every plan OpenAI and Anthropic sell, then spend a month actually hammering them - long agentic runs, code generation that doesn't stop. A fully maxed-out ChatGPT Pro subscription burns through **$14,000 a month** of tokens at API prices. Claude Max lands near **$8,000**. You're paying $200. Serving costs are lower than list, so the real hole is [nearer $3,500 a month](https://arize.com/blog/ai-model-subsidies-ending-llm-inference-costs/) - which is still seventeen times what you handed over. *Plan price vs. maximum possible spend at API pricing. Source: SemiAnalysis.* The utilization math is worse than the headline. OpenAI's margin on Plus and Pro [goes underwater past ~11.4% of the available limit](https://www.techspot.com/news/112759-openai-anthropic-cant-afford-have-everyone-use-ai.html). On the top plan, it hits zero at around 5.7%. Everything above that is loss. The business model, right now, is that most subscribers barely use what they bought - and the few of us who do are the cost center. If you're a software engineer using AI seriously, you are exactly the user these plans lose money on. Agentic coding tasks consume [roughly 1,000x the tokens of a standard chat query](https://arize.com/blog/ai-model-subsidies-ending-llm-inference-costs/), with frontier models burning over 100,000 output tokens on a single task. Every time you point an agent at a repo and let it run, you're spending someone's capital. That someone is a pension fund, a sovereign wealth fund, a hyperscaler's balance sheet, a chip vendor financing its own customers. Which is to say: ultimately a fairly small number of very rich people and the institutions that move their money. Enormous piles of capital betting that habits formed today convert into pricing power tomorrow. The bet may be correct. It may be not. Either way, the capital has already been spent, and the receipts are landing in your terminal. This is not a scam and it's not new. It's the same playbook that gave us $5 cross-town rides and two-day free shipping and cloud compute below cost. The difference is the multiple. Uber subsidized maybe 50% of a ride. This is 40-70x. Except we're not getting away completely clean, and this half doesn't make it into the launch keynote. Go price a RAM kit. A 32GB DDR5 kit that ran [$60-80 in early 2025 now asks $180-250](https://shattered.io/ram-prices-ai-memory-shortage-2026/). That's not a supply hiccup - Samsung, SK Hynix and Micron make over 95% of the world's DRAM, and they're pointing their wafers at high-margin HBM for AI accelerators. Datacenters are on track to absorb [something like 70% of high-end memory output this year](https://wccftech.com/roundup/memory-crisis/), up from 20-30% a few years ago, and the squeeze runs [through at least 2027](https://www.networkworld.com/article/4093752/server-memory-prices-could-double-by-2026-as-ai-demand-strains-supply.html). So there *is* a tax, just not on the invoice with our name on it. The giants underwrite our inference, and we underwrite their memory supply chain by paying triple for a component we've bought without thinking for twenty years. A few hundred dollars every few years against thousands a month of frontier compute is still an excellent deal - for us. Worth noting who else is in that "us", though. The markup is on the memory, not on the AI plan, so it lands on anyone buying a laptop, a phone, a console, a prebuilt. Most of them will never open an agent. The subsidy flows to people who use AI, but the memory tax is spread across everyone who buys a computer. But it's a tell. When memory - the one input you'd need to run models yourself - is what's being priced out of reach, the cheap path and the independent path are quietly diverging. "We're in a bubble" is now the safest opinion in tech, which usually means it's about to stop being useful. The interesting question isn't *whether* - it's what happens next, and there are roughly three shapes: **It bursts.** Capital dries up, the loss-leader plans die, and inference gets priced at what it costs. Uncomfortable, but not the end - the models don't unlearn anything. **It deflates.** Prices creep up, limits tighten quietly, the generous tiers get renamed and repriced. Most likely, honestly. Nobody announces the end of a subsidy; they just adjust the quota. **The margins catch up.** Per-token cost is falling [somewhere between 9x and 900x per year depending on the benchmark](https://epoch.ai/data-insights/llm-inference-price-trends), with a median around 50x, and [API gross margins are already moving in the right direction](https://arize.com/blog/ai-model-subsidies-ending-llm-inference-costs/). It's genuinely possible efficiency outruns the burn before the money runs out. I don't know which one. Neither does anyone tweeting confidently about it. But all three end the same way: **the gap between what you pay and what it costs closes.** Only the third one closes it in your favour. Burning tokens for sport is the obvious way to waste this. Probably the better move is to treat it as a window with a closing date you don't control. Learn the workflows now, while experimenting is nearly free. The muscle memory of working alongside an agent - knowing what to delegate, what to verify, where it reliably lies to you - is the part that survives any pricing change. Skills are not repriced. Quotas are. Build the things that only work at today's prices. Some ideas need so many tokens that they would never pay for themselves at the real cost of inference. Right now you can still run them - so run them, and find out what they are worth while the finding out is cheap. And keep an exit. This is [the argument I made about small models](https://dev.to/) from a different angle: the same task that costs $14,000/month at frontier scale often costs approximately nothing on a 4B model running on your own hardware. Not for everything - for the long tail of focused, repetitive work that makes up most of what we actually automate. Knowing which of your workflows would survive a 10x price increase is cheap insurance, and you can find out in an afternoon. Buy the RAM before it gets worse, if you were going to anyway. We're living through a strange, brief moment where the most expensive computation ever built is available to individual developers at a rounding error of its cost, because a lot of very rich people are making a very large bet. The secret was never that any single one of these numbers was hidden. It's that they're never put next to each other: what your plan costs to serve, what fraction of it you'd have to use before you turned unprofitable, and what happened to the price of memory. Each one is public, and each one got covered on its own - the compute economics as an AI story, the memory squeeze as a hardware story. Put them in the same paragraph and they describe a deal nobody is going to walk you through, because the labs have no reason to hand you the ruler. Enjoy it. Use it hard. Learn everything you can from it. Just don't build your entire practice on the assumption that a subsidy is a permanent feature of the universe - and don't mistake the current price for the actual value of what you're using. The bill is real. It's just not yours yet - most of it, anyway.