When Microsoft launched Maia 300 on October 12, the headline number was "40% more tokens per megawatt" — not teraflops. That wasn't a marketing accident. Power is now the binding constraint on how much AI the world can run, so the chip vendors have switched to advertising efficiency per watt. But if you're a founder renting inference, the metric is easy to misread — so here's what it actually tells you, and what it doesn't.

The energy per token is pennies — do the math#

Start with the arithmetic, because it kills the obvious wrong conclusion fast. Energy per token is just power divided by throughput:

energy per token (J) = rack power (W) ÷ throughput (tokens/sec)

Take a dense inference rack drawing 167kW — the figure Microsoft quotes for a 96-chip Maia 300 rack. Suppose it produces, illustratively, 300,000 tokens/second across all its concurrent requests. That's 167,000 ÷ 300,000 ≈ 0.56 joules per token. Convert to electricity: divide by 3,600,000 (joules per kWh) and multiply by a cents-per-kWh rate. At a representative industrial $0.10/kWh:

0.56 J/token ÷ 3,600,000 × $0.10 × 1,000,000 ≈ **$0.0155 per million tokens**

About a cent and a half per million tokens, in electricity. Compare that to what you pay for hosted inference — dollars per million tokens, sometimes tens of dollars. Raw energy is well under 1% of your bill. (The throughput number is illustrative; the point is the order of magnitude, which holds across any plausible figure.) So anyone telling you to pick a provider because its chip sips less power has the causality backwards at the line-item level.

So why does tokens-per-watt decide your bill anyway?#

Because it works indirectly, through two channels that both dominate the tiny energy cost.

Channel one — supply. Total power is capped. A data center has a grid connection measured in megawatts, and the scarce resource in 2026 is the power itself, not the chips — Morgan Stanley pegs most of this year's AI-infrastructure money as flowing toward power, not compute. A chip that delivers more tokens per watt lets a provider run more total inference under the same connection. More supply under a hard cap is downward pressure on the rental price. That's the real reason a bootstrapper should care that Maia, Trainium, TPUs, and photonic inference parts are all chasing efficiency: each one pushes the floor down.

Channel two — amortization. Your inference price is dominated by the fixed cost of the hardware, spread across every token it produces over its life, plus margin. In a fixed power envelope, more tokens per watt means the same expensive rack produces more tokens — so the big fixed cost divides across a larger denominator, and cost-per-token falls. Efficiency per watt is a lever on the dominant cost component, even though the electricity it literally saves is pocket change.

What to actually ask a vendor#

The metric that collapses all of this into one honest figure is tokens per dollar on your workload. It already contains the watts, the amortized silicon, and the provider's margin. Tokens-per-watt is upstream of it — a signal about where the floor is heading over quarters — but tokens-per-dollar is the number that predicts your October invoice.

So treat a "40% more tokens per watt" claim the way you'd treat any vendor benchmark: as a supply-side weather report, not a spec you can port into your cost model. Ask which model, which context length, and which batch size produced it — because the number swings with all three — and then ask the only question that bills you: what does a million tokens of my traffic cost? The rest is the chip vendors arguing about who gets to lower that number for you.