CrbonFree Wants to Meter Your AI Tokens in Grams of CO2

CrbonFree launched on Product Hunt this week with a single sentence of positioning: audit-grade carbon numbers for every AI token you use. That is the entire pitch as listed. No disclosed methodology, no named customers, no published emission factors. What is new here is not carbon accounting for data centers. That market has existed for years and is dominated by consultancies and a handful of reporting platforms. What is new is the unit of account. CrbonFree proposes to bill emissions at the token, which is the same granularity at which AI providers bill revenue.

That choice matters more than the product does right now. Tokens are the only unit the AI industry has agreed on. OpenAI charges per token. Anthropic charges per token. Every inference provider, from Together to Fireworks to the hyperscaler endpoints, quotes a per-token rate. If carbon gets denominated in the same unit, it stops being a facilities problem owned by the data center team and becomes a line item owned by whoever writes the prompt. That is a real shift in where responsibility sits inside a company.

The measurement problem is the product

The obvious objection is that per-token carbon is not directly measurable. It is estimated, and the estimate depends on a chain of assumptions: which model, which hardware, which region, what time of day, what the grid’s marginal generation mix was at that moment, and what power usage effectiveness the facility ran at. Google has published its own per-prompt estimates for Gemini. Microsoft has published figures for its own models. Those numbers come with wide error bars and methodological caveats, and they come from the companies that own the servers.

CrbonFree has not, in the material available at launch, disclosed how it derives its factors. That is the whole ballgame. A carbon number that cannot be traced to a grid mix, a hardware SKU, and a utilization assumption is not audit-grade. It is a vibe with a decimal point. The company may have this work done and simply not have published it yet. Until it does, the phrase “audit-grade” is doing unpaid labor.

There is a defensible version of this business. The EU’s Corporate Sustainability Reporting Directive phases in over 2024–2028, and it pushes large companies toward Scope 3 disclosure, which includes purchased cloud and compute services. A vendor that can hand a sustainability team a defensible per-token factor, with a documented methodology and a named assurance path, sells into a compliance budget rather than a marketing budget. That is a much better business than a dashboard.

Why the AI industry is the hard case

Carbon accounting for traditional cloud workloads is tractable because workloads are relatively stable and utilization is high. Inference is not that. A frontier model serving production traffic may run at low utilization on expensive accelerators, or it may batch aggressively and run hot. The same token, generated on the same model, can carry wildly different emissions depending on whether it was served from a region with hydro or a region with coal at 3 a.m. A per-token number that does not vary by region and time is close to meaningless.

Then there is the training side, which CrbonFree’s positioning does not appear to address. Training runs are where the headline emissions are, and they are lumpy, one-off, and concentrated in a small number of labs. Allocating training emissions across the tokens a model later serves is an accounting choice, not a measurement. Every provider will make that choice differently, which means cross-provider comparisons will be apples to oranges unless someone standardizes the method.

That is the real opportunity, and it is not a Product Hunt product. It is a standards body. The Greenhouse Gas Protocol is the incumbent for corporate accounting, but it has no token-level guidance. The ISO has no AI-specific emission factor registry. If CrbonFree wants to be the default, it needs its methodology to be the one auditors accept, not the one its own dashboard displays.

What to watch

Three things would separate CrbonFree from the long tail of sustainability dashboards. First, a published methodology with named emission factors and a stated uncertainty range. Second, at least one named customer willing to put its logo next to a number, which is a much higher bar than a pilot. Third, an integration with the actual billing systems that AI teams use, because a carbon number that lives in a separate tool will be ignored. If the number shows up next to the invoice, it gets managed.

The competitive risk is that the model providers do this themselves. Google, Microsoft, and Amazon all have sustainability reporting teams and all have an incentive to control the narrative on their own emissions. If they publish per-token factors, third-party metering gets squeezed toward the assurance and audit layer, which is a smaller and slower business.

The strategic risk is different. If CrbonFree’s numbers are wrong in either direction, the whole category gets discredited. A carbon metric that overstates emissions gets ignored by engineers as noise. One that understates them gets ignored by regulators as greenwashing. The product only works if the number is boring and defensible.

For AI builders, the practical read is this: carbon is about to become a procurement question, not an engineering one. Enterprise buyers with CSRD obligations will start asking inference vendors for emissions data, the same way they now ask for SOC 2 reports. Vendors that cannot produce a number will lose deals to vendors that can, regardless of whether the number is good. CrbonFree is early to a market that will be defined by whoever writes the methodology first. The token is a clever unit of account. The audit trail is the product, and it is not on the page yet.

{/* TODO: verify CrbonFree’s disclosed methodology, emission factors, and any named customers — launch listing on Product Hunt did not include them. /} {/ TODO: comment sought from CrbonFree; none received at time of writing. /} {/ TODO: verify per-prompt carbon figures attributed to Google and Microsoft — searched, found company sustainability reporting exists but did not confirm specific per-prompt numbers. */}