Vendor Spends Quarter Proving Carbon Footprint Beats Competitor

Proxima AI released a forty-eight-page sustainability analysis this week claiming its flagship language model consumed sixty percent less energy per inference than a competing system from Anthropic. The report, titled 'Measured Progress', measured Proxima's own model running on a custom chip in a data center outside Portland. It measured Anthropic's model by calculating theoretical power consumption based on a paper published three years ago, then halving the estimate to account for what the Proxima analysis called 'architectural inefficiencies we are reasonably confident exist.'

We measured our model on a Tuesday, so the numbers are definitely comparable to anything else anyone has ever measured.

"We measured our model on a Tuesday, so the numbers are definitely comparable to anything else anyone has ever measured," said Marcus Chen, Proxima's Vice President of Sustainability Communications. Chen is forty-one, based in San Francisco, and has never worked at a utility company or anywhere that generates actual electricity. He joined Proxima five months ago from a role in corporate communications at a ride-sharing platform, where he spent two years explaining why the company's carbon accounting had not changed despite increased vehicle miles.

The report excluded cooling energy, transmission losses, and the power draw of Proxima's redundant systems, citing industry-standard measurement practices that Proxima appeared to have invented. It included a marginal note suggesting that the competitor's infrastructure 'likely' ran on cheaper, dirtier grid power, though Anthropic's servers are physically located in multiple countries. A footnote on page thirty-seven observed that Proxima's methodology 'may not be directly comparable to other sustainability reports due to scope, assumptions, and the fundamental difference between actual measurements and reasonable speculation.'

At press time, Anthropic had not released a competing analysis, citing the time and expense required to conduct actual measurements. Instead, the company published a statement noting that comparing models across different hardware, software, and measurement protocols was like comparing the fuel efficiency of two cars when one was measured on a highway and the other was estimated by someone who had only read the manual. The statement included a link to its own sustainability page, which had not been updated since last year and contained a single number with no methodology attached.