GPU Rents Up 106%, Token Prices Down 58%: Who's Working for Whom?

a16z's chart this week presents a somewhat uncomfortable comparison: from mid-July to October 26, B300 GPU rents rose 106.0%, B200 rose 65.7%, and H100 rose 16.9%; over the same period, blended token costs fell 58.4%. The chart calls this The Great 'Rack Spread' Of 2026.

Let's fill in the context first. A token is the unit of pricing for model-generated text, code, and other outputs. GPU rent is the price of renting compute racks. B300, B200, and H100 are several generations of NVIDIA GPUs. Rack Spread refers to the widening gap between rack rental prices and token selling prices. The data source is Orn Data, and the chart is linked in a16z's Charts of the Week.

Translated into business terms: racks are more expensive, output is cheaper. For those who rent GPUs and sell tokens, upstream costs are rising while downstream selling prices are falling.

In the a16z replies, a commenter did a more intuitive calculation: B300 rent is about 2.06x, token price is about 0.42x, and the gap between the two is nearly 5x. Unless GPU throughput also increases in tandem, some margin will always be eaten. Another commenter's addition is closer to reality: new GPUs produce more tokens, so overall it may still be net positive. The two statements are not contradictory; efficiency gains and price wars can happen at the same time.

Some commenters said that even though token prices fell 58%, their API bills still went up. A drop in unit price does not equal a drop in total spend; usage, routing, retries, and plan structures can all pull the bill back up. This experience is not new, but placed next to this chart, it is jarring.

There are two typical situations in the replies: token buyers are catching a good time; sellers of fixed-price contracts are being squeezed from both sides. Someone described the bottleneck position very directly: cheaper tokens, more expensive racks, the bottleneck has moved upward. Another commenter asked how much longer old servers can appreciate, how much of a server's value is attributable to power availability, and when power efficiency will overtake server value and force replacement. This question is harder than the chart. If electricity prices and grid connection capacity are constrained, the value of old cards is not determined by compute alone.

A commenter responded to Rack Spread with a frog emoji, the emotion is direct.

Rack Spread is multiple price signals stacking together: tight GPU supply, improving inference efficiency, and model competition driving prices down. Another commenter mentioned that demand for open-source weights is pushing up GPU demand. In the short term, token buyers benefit; in the medium term, resellers and fixed-price sellers without a throughput advantage suffer the most. One commenter predicted that the spread will widen further in 2027, but the chart only goes to October 26, so predictions from the replies cannot be taken as conclusions.

What is more worth watching is unit economics: how many tokens can be produced per dollar of rack rent, and how much those tokens can be sold for. A 106% rent increase is not fatal; what is fatal is if throughput and selling prices do not keep up together.