Mathew's back-of-envelope ties the labs' revenue to the power capacity they run on and sets it against a Goldman estimate of the AI revenue that 2026-27 hyperscaler capex needs to earn its return; Zephyr argues new hardware lifts revenue per gigawatt well above today's. Both posts went up hours before the FT report and assume the higher OpenAI figure, which makes this the clearest way to see what the lower number does, and doesn't do, to the compute-demand case.
Blended revenue per GW will pass $60B+ (due to new hardware and efficiency gains), and dedicated inference capacity should be around 10GW-12GW by EOY27 As long as demand is there, $600B+ in combined lab ARR by the end of 2027 is possible Automating more parts of knowledge work
Shanu Mathew · (@ShanuMathew93) · 2026-10-08
Gigabrains help me out:
GS puts required annual hyperscale AI revenue at $526bn, based on average annual capex across 2026 and 2027 and a 20% hyperscaler ROIC.
I assume OpenAI and Anthropic have similar ARR, GW capacity and inference share, with $20bn of revenue per blended GW and $50bn per monetized GW (40% of capacity monetized).
Today, combined ARR is ~$75bn on 3.5-4GW, or ~$20bn per GW for Ant (OAI is close to $70bn now so lets just assume the same for both). Each lab goes from ~2GW to 5GW by YE26 and 10GW by YE27.
YE26: 5GW x $20bn = $100bn per lab. The monetized route agrees: 40% x 5GW = 2GW x $50bn = $100bn.
YE27: 10GW x $20bn = $200bn per lab (4GW monetized x $50bn).Two labs at YE27 give $400bn of implied lab ARR. That leaves ~$126bn of the $526bn for everyone else: non-lab hyperscaler revenue, agentic software, Muse/Instinct, other labs, routers and harness companies. Attainable or not? That all assumes pricing and revenue per GW don't come down as well for the labs.
Am I thinking about this right? Feel free to critique/debate anything I laid out @JaredKubin @azeem @GavinSBaker @WarrenPies @FredaDuan @altcap @_clarktang @TBU12345678 @TMTLongShort @zephyr_z9
Mathew's back-of-envelope ties the labs' revenue to the power capacity they run on and sets it against a Goldman estimate of the AI revenue that 2026-27 hyperscaler capex needs to earn its return; Zephyr argues new hardware lifts revenue per gigawatt well above today's. Both posts went up hours before the FT report and assume the higher OpenAI figure, which makes this the clearest way to see what the lower number does, and doesn't do, to the compute-demand case.
Tweets, videos, filings, articles, and source pages tied to this read.