About 40 GW of first-party compute might be available to an AI company by the end of 2030, enabling pretraining for about 2e29 FP8 FLOPs (using 10 GW). This follows from the data on incremental IT power in 2027-2030 given in a new article by SemiAnalysis, and some guesses about AI company compute share and Feynman performance that I detail below.
The figures the article gives for (what I'm going to assume is) incremental IT power using 800V DC racks and the fraction of their use are 16.5 GW (38%) in 2027, 24.3 GW (60%) in 2028, 34.6 GW (74%) in 2029, and 38.5 GW (78%) in 2030. The complement to these figures (to 100%) is the rest of (I'm assuming) incremental AI datacenter capacity. Together with the complement, it's 43.4 GW in 2027, 40.5 GW in 2028, 46.8 GW in 2029, and 49.4 GW in 2030. This is steady incremental compute of about 45 GW per year over 2027-2030, it's not getting much higher year-over-year.
OpenAI and Anthropic will each have 2-4 GW of compute at the end of 2026, and about 10 GW by the end of 2027, meaning 6-8 GW of incremental compute in 2027. This is about 16% of the above estimate for the total incremental compute [1] . Maybe this fraction steadily increases to 25% by 2030 for a more successful AI company. Then the incremental compute that company adds in 2028 is 7.7 GW, then 10.3 GW in 2029, and 12.4 GW in 2030, for a total of 17.7 GW by the end of 2028, and 40.3 GW by the end of 2030.
If 2025's 1-2 GW suggest pretraining runs of 0.3 GW and 1e27 FP8 FLOPs [2] (using Hopper or Trainium 2; 200K H100 chips with 1 compute die, 2e15 FLOP/s per chip), then 2028's 17 GW suggest pretraining with 5 GW and 8e28 FP8 FLOPs (Rubin, since there's probably not enough Blackwell; 2.9M compute dies [3] either in 2-die or 4-die chips, 8.7e15 FLOP/s per compute die), and 2030's 40 GW suggest 10 GW and 1.6e29-2.0e29 FP8 FLOPs (either 5.8M Rubin compute dies, or 4.5M Feynman dies, assuming 30% more power per Feynman die compared to Rubin, which maybe produce 14e15 FP8 FLOP/s per compute die).
The 14e15 FP8 FLOP/s figure for Feynman dies is estimated as follows. First, grounding for a model of how Blackwell-to-Rubin change works (this is not directly part of the Rubin-to-Feynman estimate). Rubin improved BF16 by 60%, extracted IO chiplets that used maybe 30% of the area of the compute dies in Blackwell, and in terms of transistors went from 208B (with the 4NP node) to 336B (with 3NP; assuming this includes IO chiplets), so 25% more transistors for the compute dies (excluding the IO chiplets). The power draw increased by 64% (1,400 W to 2,300 W), or by 26% for the compute dies (excluding the IO chiplets). Performance typically increases by half of the power increase (so by 13% for the 26% power increase), and there are 62% more transistors, which predicts an 83% increase in performance from Blackwell to Rubin. The increase in FP8 and FP4 is disproportionately larger, but that's a change in emphasis (the bet is that BF16 is no longer very important, so some of its potential in improved performance went to FP8).
Now, if Feynman increases power by 30% over Rubin, giving a 15% increase in performance, and uses A16 (jumps two steps, N3P to N2, then N2 to A16), giving maybe another 35% increase in performance from logic density, and if there are no more IO or BF16/FP8 rearrangements, then Feynman might improve FP8 performance by 55% per compute die. With 8.75e15 FP8 FLOP/s per die in Rubin, that's 13.6e15 FP8 FLOP/s per die in Feynman.
Note that the figures of incremental compute likely refer in part to compute that doesn't yet go online within that year, because 16.5 GW of 800V DC in 2027 seems too early, since Kyber will mostly go online over 2028. But the fractions are probably approximately right, since the total amount of incremental compute (of all types) is steady. ↩︎
Assuming 40% FLOP/s utilization for 3 months, here and below. ↩︎
With 225 kW per rack, and 10% IT overhead (transceivers and switches), that's 4,040 racks (144 compute dies each) per 1 GW of IT power. ↩︎