Amazon CEO: AI Compute Sold Out Until 2028, Power & Delivery Are the Real Bottlenecks - AI Daily Brief
Audio in Mandarin Chinese · English transcript below
⚡ Amazon CEO Andy Jassy told a16z AWS compute remains sold out despite ~$220B capex this year, with 2026-2028 capacity mostly locked in long-term contracts. The real bottleneck isn't money—it's power, cooling, and physical chip delivery timelines.
The race for AI compute is increasingly constrained by physical infrastructure rather than capital. Andy Jassy revealed that AWS's roughly $220 billion investment this year remains far from sufficient to meet total demand, with capacity nearly sold out for the next three years. Notably, AWS has reached an annualized revenue of $169 billion with growth accelerating to 36.7%, yet the paradox between surging revenue and exhausted compute capacity underscores that power supply, cooling systems, and chip delivery timelines represent the true hard constraints facing the industry.
Today's Top 3 Headlines
- Opinion
Amazon CEO Andy Jassy: AI compute sold out through 2028, bottlenecks are power & delivery
Jassy says even ~$220B capex this year leaves unmet demand: 2026 sold out, 2027 nearly full, 2028 heavily booked, rejecting "not enough money in the world" takes. AWS grew 36.7% YoY to $169B annualized run rate, ~3x 2021 levels. For practitioners, power, cooling and chip delivery—not orders—are the bottleneck.
Source ↗ - Opinion
SemiAnalysis Kimbo Chen: Trainium3 Challenges Training Chip Throne with Native PyTorch Experience
Amazon's Trainium3 offers native PyTorch training via TorchNeuron—just swap .to("cuda") for .to("neuron") per SemiAnalysis. The team found Neuron Explorer superior to Nvidia/AMD tools, with AWS dedicating 80K physical cores to QA, though ecosystem maturity lags CUDA and MXFP8 missed day-one availability despite solid BF16 support.
Source ↗ - Opinion
Matthew Berman: Open-weight Kimi K3.1 nails 3D demos, still misses exact replication
Berman tested 3T-param open-weight Kimi K3.1 via paid desktop sub: 3D Rubik's cube sim (scramble + algorithm solve, keyboard control) passed among best he's seen, most open weights fail; but WebGL iPhone Duo render took 28m 45s, web search still inaccurate. Noted "eyebrow text" resembling Claude family. Boundary: single-user manual demo, not benchmark.
Source ↗
+8 more headlines
- 20VC: Code Review Trumps Code Gen, Time-to-Prod Still 10x Higher
- Lenny's Podcast: Asking LLMs "what they're thinking" covers only a tiny fraction of their internal processing
- 💡 Thought Lab: Hotel Robot's Taiwan Answer Hinges on LLM Provider & Operator
- Greg Brockman: Brotli Has Unauth RCE Flaw, Patch Coordination Complex
- Ethan Mollick: LLM makers earn more in finance, no need to train on client data
- Vinod Khosla: AI Agents Should Proactively Hire Humans to Complete Tasks, Fo Agent Already in Production
- Dan Shipper: The Best AI-Era Team Has Just Two People—Pirate and Architect
- Margaret Mitchell: LLM welfare is now an urgent issue we must face