Hyperscaler backlog hits $1.7T as Citi conference flags shift to agentic inference
sanjaykalra · x · 2026-09-08
Takeaways from Citi's Global TMT Conference (250+ management teams):
- Compute: Hyperscaler backlog near $1.7T at end of Q2, up 151% YoY; AI capex tracking toward $3.8T by 2030; new AI power capacity at 19 GW this year, 45 GW by 2028.
- Inference shift: AMD's Matt Ramsay says inference is now the majority driver of AI computing, with a "radical change from chatbot inference to agentic inference" — agent runs call tools and retry dozens of times per task, breaking seat-based cost models.
- Enterprise: Cognizant describes an "A0 to A4" model with 40% of engineering AI-assisted; Kyndryl's CEO likens the gap to "a 200 mph bullet train on tracks built for 30 mph" and runs 1,600 agents internally.
- Other: Edge compute decisions hinge on latency and data residency; agentic commerce turns product data and APIs into the storefront as agents replace search.
Advice for CIOs: own the routing/orchestration control plane, protect context and evals, budget per completed task, and instrument agent fleets before scaling.
Related event: Citi: Hyperscaler Backlog Nears $1.7 Trillion as AI Capex Eyes $3.8T(3 posts)→
More from AGI Musings
- Curious How AI Solves Infinity? — karmicoder · 2026-09-09
- Dallas Fed: AI-heavy sectors use 16% of US work hours but drive 40% of productivity gains — bittingthembits · 2026-09-09
- Alignment reduces to preference elicitation — and preference elicitation is collaboration — clarejtbirch · 2026-09-09
- When NVIDIA's CEO says AGI has arrived, researchers still can't even define intelligence — annetgriffin · 2026-09-09
- ARR reviewers caught using AI for peer reviews amid two-week compulsory deadlines — tallinzen · 2026-09-09
- Pro-AI doesn't mean pro-billionaire: rebutting the anti-AI caricature — AIandDesign · 2026-09-09