Social Graph VC said data center capital expenditures are expected to exceed $1 trillion in 2026, about twice the 2025 spending of the four hyperscale cloud providers. According to ChainCatcher, the article also said AI investment accounts for about 0.9% of global GDP and is projected to rise to about 1.4% by 2028.
The article said Fable 5 and GPT-5.6 each required less than 2 GW of training power and cost about $120 billion to train. It added that each generation of models can pay for itself, while lab losses come from continuously scaling training for the next generation by 10 times.
It also said computing cannot be purchased at a single price, with H100 rental rates around $1.95 per hour in the market and $8 to $9 per hour at hyperscale cloud providers. The article said hourly rates can fluctuate by 137% in a year, and that of roughly $3 trillion in global data center spending through 2028, there is a $1.5 trillion financing gap.
Separately, CME plans to list cash-settled monthly H100 and B200 rental futures on NYMEX on October 5, 2026, pending regulatory review, with settlement based on the Silicon Data index. ICE also announced GPU futures based on the Ornn index.