Token Prices Fell 280x. AI Bills Rose 320%. How Founders Should Read the Cost Split

DeepSeek cut V4-Pro to a quarter of its old price while Meta raised 2026 AI capex to $145B. One side pushes inference toward free; the other pours money into infrastructure. Token prices dropped 280x in two years, yet enterprise AI bills climbed 320%. That contradiction is rewriting every founder’s cost structure and moat.

What Happened

Two numbers landed in the same quarter. DeepSeek slashed V4-Pro token prices by 75%, and the cut hardened into a permanent quarter of the old price after May 31, 2026. Output tokens went from $3.48 to $0.87 per million; cache-hit input fell to $0.003625. This was not a promotion, it was an engineering shift. The company says V4-Pro was built to run long-context inference at roughly a quarter of the per-token compute and a tenth of the memory footprint. Meta walked the opposite road in the same window. It raised 2026 capex guidance to $125B-$145B, nearly double the $72.2B it spent in 2025. Asked for ROI signals, Mark Zuckerberg deflected, “that’s a very technical question”, and Meta’s stock dropped more than 6% after hours. One camp is making inference almost free; the other is spending astronomically on the infrastructure that runs it. The market is splitting in two.

What This Means for Founders

There is a trap hidden here. Per-token prices fell 280x in two years. Yet enterprise AI bills rose 320% over the same period. Cheaper tokens triggered an explosion in volume. An agentic workflow calls the model dozens of times per run, so falling unit prices do not lower the bill. For startups the squeeze is worse. With inference eating roughly 23% of revenue, AI-native gross margins sit around 52%, 20 to 30 points below the 75-85% mature SaaS enjoys. The “token tax” structurally erodes the margin. That is why raw model access is no moat. The moment DeepSeek cuts prices to a quarter, a thin wrapper over someone else’s API loses its differentiation overnight. The real moat lives elsewhere: proprietary data, integration buried deep in the workflow, and a cost design that hits the same result with fewer tokens. Cursor shipping its own code-tuned inference model to escape the token tax is the signal. Founders now compete not on “which model do you use” but on “how do you save every token.”

What You Can Do Now

Start by exposing your product’s per-token economics in hard numbers. If you don’t know how many tokens one user or one task burns and what it costs, you cannot set a price. Second, make caching a first-class part of your cost design. The gap between cache hits and misses runs tens of times. Routing repeated prompts and context through cache alone changes the bill. Third, split models by task. Stop running the priciest model on every call, push simple jobs like classification and extraction down to a low-cost model. The existence of cheap models like DeepSeek is exactly what makes that choice possible. Fourth, tie pricing to outcomes, not tokens. Usage-based billing absorbs the full risk of falling prices, while outcome-based pricing restores margin. Fifth, never stake your moat on model access. Build the company on what survives a competitor cutting prices to a quarter tomorrow, data, integration, cost efficiency.