Labs have subsidized your access for years. The IPO queue is about to hand the bill back — and the physics of the hardware underneath it all makes the math worse than the pitch decks let on.
Start with a number: OpenAI is projected to lose $14 billion in 2026 — nearly triple what it lost the year before. The company collects around $13 billion in revenue and spends roughly $22 billion to do it, burning approximately $1.69 for every dollar it earns. That is not a temporary funding crunch. That is a structural condition, and it has been deliberately designed that way.
The deal AI companies struck with users was always a quiet one: subsidize access, flood the market, figure out profit later. That arrangement now has a deadline. The IPO queue — OpenAI, Anthropic, xAI — is forcing a reckoning. Public markets don’t absorb “we’ll figure it out” as a line item. They require disclosed burn rates, explained unit economics, and a credible path to profit with a date attached.
Token Deflation: The Paradox Nobody Mentions Out Loud
The “tokens are getting cheaper” narrative is real — but the numbers tell a more interesting story about who it actually helps. GPT-4 launched at $30 per million input tokens. Today’s mid-tier models run below $2. Budget-tier options approach $0.08. On a per-unit basis, the cost of intelligence has collapsed.
| Model Era | Input / 1M tokens | Output / 1M tokens | vs. GPT-4 Launch |
|---|---|---|---|
| GPT-4 (2023) | $30.00 | $60.00 | — |
| GPT-4o (2024) | $2.50 | $10.00 | ↓ 92% |
| Claude Sonnet 4 (2025) | $3.00 | $15.00 | ↓ 90% |
| Budget tier (2026) | $0.08–0.15 | $0.30–0.60 | ↓ 99% |
The catch is simple: cheaper tokens expand usage faster than they reduce the bill. Every enterprise that ran one AI pilot in 2024 runs ten in 2026. Every developer who tested one integration ships five. The aggregate compute demand compounds upward even as the unit price falls. Efficiency gains get absorbed by demand, not banked as margin.
There’s also a pricing sleight-of-hand baked into how most companies budget AI: output tokens cost 3–10x more than input tokens and most API calls are output-heavy. Teams anchoring budgets to the advertised input price routinely receive invoices at a substantial multiple of their estimate.
The Physics Problem Wall Street Isn’t Pricing
The financial conversation about AI costs keeps orbiting the same abstractions — burn rate, margin expansion, revenue multiples. What it consistently underweights is that AI infrastructure now carries a hard physics constraint, and physics doesn’t negotiate valuation multiples.
According to the IEA’s Energy and AI report, global data center electricity consumption already hit 415 terawatt-hours in 2024 — equivalent to 1.5% of all electricity generated on Earth. By 2030, that figure doubles to 945 TWh. In the United States alone, data centers are on pace to consume more electricity than all energy-intensive manufacturing combined — aluminum smelting, steel production, cement, chemicals, everything. AI inference, not training, now drives roughly 80–90% of that load, and it grows with every user who opens a chat window.
Wholesale electricity prices near major data center clusters have risen as much as 267% versus five years ago. Residential bills in parts of the U.S. have climbed specifically because of data center grid draw. The efficiency gains happening at the chip level are real — but they get partially absorbed by the new thermal management infrastructure required to house the chips. The net effect on operating cost is considerably less dramatic than the chip benchmarks imply.
The Two-Track Race: OpenAI vs. Anthropic
According to WSJ financial documents, the two leading AI labs are running fundamentally different bets. OpenAI keeps its burn rate at 57% of revenue through 2027, betting that dominance justifies the cost. Anthropic, by contrast, projects its burn falling to 9% of revenue by 2027 — and expects to break even by 2028, roughly two years ahead of OpenAI’s timeline.
Anthropic’s path has a different texture entirely. TechCrunch reported that the company projects gross margins reaching 77% by 2028, up from roughly 40% in 2025 — though The Information noted that inference costs on Google and Amazon cloud infrastructure already came in 23% above internal estimates. The margin story is real, but it runs on assumptions that haven’t fully cleared yet.
Before Q3: A Practical Checklist
The price hikes aren’t theoretical — they’re structurally embedded in the cost stack of every lab heading toward public markets. Here’s how to get ahead of them:
AI Stack Audit — Do This Now
- Route by task, not by habit. Most production workloads don’t require frontier models. Mid-tier models perform identically to premium ones on 70–80% of tasks at a fraction of the cost. Map every API call to the minimum capable model.
- Model output token spend separately. Output tokens cost 3–10x more than input. Any budget built on input-token pricing alone will come in significantly under the real number.
- Enable prompt caching. Repeated system prompts can be cached for up to 90% cost reduction on those tokens — a material saving on any high-volume workflow.
- Watch your cloud layer. Inference costs on third-party cloud infrastructure (AWS Bedrock, Google Vertex) carry their own markup on top of model pricing. Know both numbers.
The Amazon Parallel — and Where It Ends
The standard comfort narrative is Amazon: build the network, subsidize access, capture the market, monetize later. The argument has structural logic. Amazon built its empire on low prices, free shipping, and avoided sales tax — and eventually charged enough to cover costs.
But Amazon sold physical goods and logistics, where scale economies are durable and compounding. The open question for AI labs is whether inference economics follow the same curve. The early evidence is mixed: chips get more efficient, models get larger. Context windows grow, compute requirements scale with them. Every prior efficiency gain has been absorbed by capability expansion rather than banked as margin. There is no clear mechanism by which the next threshold breaks this pattern.
The subsidized era of AI had a good run. The bill is due — and for companies that haven’t mapped where they’re spending, it’s going to arrive without much warning.
Related: Nvidia’s $26B AI Bet: Why the Chip Giant Is Entering the Open-Weight Model Wars
