Cost Per Token: The AI Infrastructure Metric That Actually Matters
Shruti Koparkar argues that enterprises evaluating AI infrastructure should shift from traditional metrics like FLOPS per dollar to cost per token as the primary measure of total cost of ownership. The piece contends that cost per token is the only metric that captures the full picture of real-world AI economics, accounting for hardware performance, software optimization, and actual token delivery rather than just raw compute capacity. The distinction matters because optimizing for input costs while the business runs on output creates a fundamental mismatch in how infrastructure value is assessed.
TL;DR
- →Cost per token, not FLOPS per dollar or compute cost, should be the primary TCO metric for AI infrastructure evaluation
- →The denominator of the cost equation, which represents delivered token output, matters more than the numerator (GPU hourly cost) for reducing per-token costs
- →Key factors beneath the surface include support for mixture-of-experts models, FP4 precision, speculative decoding, KV-cache optimization, and disaggregated serving
- →Maximizing tokens per megawatt is especially critical for on-premises deployments where capital commitment to power and infrastructure is substantial
Why it matters
As AI workloads shift from traditional data processing to token generation at scale, the metrics used to evaluate infrastructure economics must evolve accordingly. Enterprises relying on outdated metrics risk making infrastructure decisions that appear cost-effective on paper but fail to optimize for actual business output, which is measured in delivered tokens and revenue per infrastructure dollar spent.
Business relevance
For operators and founders building AI products, cost per token directly determines unit economics and profitability at scale. Choosing infrastructure based on peak specifications or hourly compute rates rather than actual token delivery can result in significantly higher costs per inference and lower margins on every customer interaction, making the difference between a viable and unviable business model.
Key implications
- →Infrastructure vendors will increasingly be evaluated and compared on cost per million tokens for specific model types, particularly mixture-of-experts models, rather than abstract performance metrics
- →On-premises AI deployments require deeper analysis of power efficiency and token throughput per megawatt to justify capital expenditure on land, cooling, and infrastructure
- →Software optimization layers, serving infrastructure, and runtime support for techniques like speculative decoding and KV-cache offloading become as important as raw hardware specifications in determining real-world TCO
What to watch
Monitor how cloud providers and infrastructure vendors begin pricing and marketing their AI services, particularly whether they shift toward cost-per-token transparency or continue emphasizing hourly rates and peak specifications. Watch for enterprise procurement teams adopting cost-per-token benchmarking in RFPs and vendor evaluations, which would signal a market-wide shift in how AI infrastructure value is assessed.
vff Briefing
Weekly signal. No noise. Built for founders, operators, and AI-curious professionals.
No spam. Unsubscribe any time.
