Anthropic+OpenAI compute (IT power, not FLOP/s) will 2.5x-2.3x every year through 2028 (growth only slightly slows), say the projections given by SemiAnalysis in a talk delivered 1 Oct 2026. In the slide at 4:15, incremental compute (new compute that becomes available that year) for Anthropic+OpenAI is given for 2025-2028, in the context of global incremental compute (AI+non-AI): 2.9 GW in 2025, 5.7 GW in 2026, 13.2 GW in 2027, 29.9 GW in 2028. This is 16%, 21%, 26%, and 36% of global incremental compute (17.8 GW in 2025, 27.5 GW in 2026, 50.0 GW in 2027, 83.8 GW in 2028) respectively. The next slide at 4:40 gives the aggregate compute for Anthropic and OpenAI individually. In particular, Anthropic's implied incremental compute for 2028 is 18.2 GW. And since the aggregate compute for end of 2028 for Anthropic+OpenAI is 52.6 GW, while the incremental compute for 2025-2028 is 51.7 GW, it follows that their aggregate compute for end of 2024 was about 0.9 GW. Thus aggregate compute (Anthropic+OpenAI) is 3.8 GW end of 2025 (4.2x Y/Y), 9.5 GW end of 2026 (2.5x Y/Y), 22.7 GW end of 2027 (2.4x Y/Y), and 52.6 GW end of 2028 (2.3x Y/Y).

The slide at 18:20 gives global incremental AI compute for 2026-2030: 19.3 GW in 2026, 35.8 GW in 2027, 40.7 GW in 2028, 63.3 GW in 2029, 79.2 GW in 2030. Comparing to the first slide for 2026, 2027, and 2028, this is respectively 70%, 72%, and 49% of global incremental compute (AI+non-AI). And in turn the Anthropic+OpenAI incremental compute for these years is then 30%, 37%, and 73% of the global incremental AI compute respectively. There's something funny going on with 2028 here, as if the satisfied demand for AI compute that's not for Anthropic+OpenAI goes down relative to both Anthropic+OpenAI demand, and to non-AI global compute demand (like a supply shortage that gets resolved in the direction of giving almost everything to Anthropic+OpenAI). But this happens in a way that maintains the year-over-year trend for Anthropic+OpenAI compute (down to only 2.3x end of 2028, from 2.4x end of 2027), so this is a reason to doubt that the trend really does hold through end of 2028 (that is, the 2.3x Y/Y end of 2028 might be an overestimate).

If Anthropic+OpenAI compute continues to 2.5x in 2029 and 2030, they get 78.9 GW of incremental compute in 2029 and 197.3 GW in 2030, which is much more than the global incremental AI compute for those years. Thus they won't be able to 2.5x their compute in those years (it's already 2.3x in 2028, not 2.5x). If they get 80% of global incremental AI compute in 2029 and 2030, their aggregate compute becomes 103 GW end of 2029 (2.0x Y/Y) and 167 GW end of 2030 (1.6x Y/Y). So the year-over-year compute growth in terms of IT power gets significantly slower after 2028 (2.0x in 2029, 1.6x in 2030), but over 2026-2028 it only changes slightly, decreasing from 2.5x in 2026 to 2.3x in 2028.