▶ 43:00 · AI Infrastructure · Grid & Power Infrastructure · US data center flops utilization at 15% creates national security crisis per AMP founder

episode briefing

AMD CEO Lisa Su Live on TBPN

2026-07-23 · 17 company · 17 thematic

sentiment

15 bull0 bear2 neu

speakers

lisa su

CEO of Advanced Micro Devices since 2014; led AMD's transformation into AI and high-performance computing leader with MI300/MI355/MI450 accelerator roadmap.

anne

Founder of AMP venture firm and infrastructure arm building 2GW US AI compute capacity; advocates 'output maxing' efficiency; investing in nuclear (Maver) to solve energy bottleneck for 2030 AI scaling.

seth cohen

Co-founder of Maver (AI infrastructure/nuclear); former DOE Chief Counsel for Nuclear Policy and NRC conservator; led nuclear waste narrative shift resulting in 27 governors competing for fuel cycle campuses.

oliver cameron

Founder of Odyssey building foundational world models; previously at Cruise working on self-driving cars; raised $310M for world model development across robotics, gaming, and autonomy.

muhammad

Founder of Ideogram building AI foundation for enterprise design and content; previously started Google's Imagen; released open-weight frontier image model for on-prem enterprise deployment.

Founder of AMP, a venture firm (AMP Foundry) and infrastructure platform (AMP Grid) building ~2 GW of US AI compute capacity; investor in Maver and Periodic Labs; focused on energy, compute utilization, and frontier AI applications.

now playing · AI Infrastructure · Grid & Power Infrastructure

World Models & Simulationtailwindscore 8/10oliver cameron

World models at GPT-2 phase will replace hand-tuned autonomy in robotics and driving

Broad-distribution world models lightly tuned for tasks outperform narrow vision-language-action models with orders of magnitude less data; applications span gaming, robotics, driverless ca…

Robotics & Physical AItailwindscore 8/10oliver cameron

World models pretrained on broad internet data generalize to robotics with orders-of-magnitude fewer demonstrations

Training world models on diverse video/game/simulation data then fine-tuning for specific embodiments (cars, humanoids, drones) dramatically reduces real-world data requirements versus visi…

Autonomous Vehiclestailwindscore 7/10oliver cameron

World models enable Tesla-level autonomy diffusion to legacy auto by 2030 via supplier model

Specialized autonomy companies will sell turnkey driverless stacks to OEMs (Honda, Ford) by 2030; legacy auto must choose between adopting external tech or burning capital on internal failu…

Open Source AItailwindscore 8/10muhammad

Frontier open-weights models enable enterprise on-prem deployment and data sovereignty

Compact, high-quality open-weights models (Ideogram, DeepSeek) allow regulated enterprises to self-host generative AI, removing IP leakage risk and reducing inference costs, accelerating B2…

Enterprise AI Adoptiontailwindscore 7/10muhammad

Enterprise image generation requires consistency, brand DNA, and on-prem data sovereignty

Brands need 100% accurate logo/typography translation and product photography; open-weight compact models enable on-prem licensing solving IP concerns for design, manufacturing, and defense…

AI Infrastructuretailwindscore 9/10aj

US data center utilization at 15% flops efficiency creates national security crisis

Only 15% of provisioned flops are productively used due to 30-40% scheduling losses and 15% model flop utilization (MFU); fixing this software-defined waste is as critical as building new c…

AI Infrastructuretailwindscore 9/10anne

US data center flops utilization at 15% creates national security crisis per AMP founder

Only 15% of paid flops are utilized due to 30-40% scheduling losses and 15% chip MFU; solving this efficiency gap is as critical as building new capacity for US AI competitiveness vs China.

Grid & Power Infrastructuretailwindscore 9/10aj

AI infrastructure players procuring 2030 energy capacity now, nuclear essential for US competitiveness

Multi-gigawatt AI campuses require 5-7 year energy lead times; AMP's 2030 procurement horizon and nuclear investment (Maver) highlight that power permitting and generation, not chip supply,…

AI Agentstailwindscore 8/10anne

RL with formal verification drives next consumption leg beyond coding agents

Industrial-grade RL with verifiable rewards (unit tests, physics simulations) enables rapid capability gains in domains like materials science (Periodic Labs superconductors); verification…

AI Agentstailwindscore 9/10aj

Reinforcement learning with verifiable rewards drives next leg of AI capability expansion beyond coding

Industrial-grade RL with formal verification (unit tests, physics simulations, X-ray diffraction) creates tight feedback loops that rapidly expand model capabilities in any domain with grou…

Nuclear Energytailwindscore 9/10seth cohen

Spent nuclear fuel reframed as strategic energy asset worth 4x Saudi oil reserves

100,000 tons of commercial spent fuel retains 97% of its energy potential; policy shifts have turned 27 governors into advocates for hosting full fuel cycle facilities, unlocking a massive…

Nuclear Energytailwindscore 9/10seth cohen

Spent nuclear fuel reframed as strategic energy asset worth 4x Saudi oil reserves

100,000 tons of commercial spent fuel retains 97% energy content; policy shift turned waste liability into asset with 27 governors competing for fuel cycle campuses, unlocking domestic ener…

AI Bubble / Capex Debatetailwindscore 8/10john coogan

OpenAI's $750B 2030 compute budget and $20B single data center signal sustained capex conviction

The 25% increase in OpenAI's projected infrastructure spend to $750B through 2030, anchored by a $20B Georgia campus, reflects leadership's confidence that scaling laws continue to deliver…

Semiconductorstailwindscore 8/10lisa su

AI-assisted chip design cuts 3-6 months from silicon bring-up, creating compounding hardware-software flywheel

Anthropic using Claude to optimize kernels for AMD MI355 demonstrates LLMs accelerating the hardware development loop; faster silicon iteration enables faster model training, which further…

AI Hardware & Chip Architecturetailwindscore 8/10lisa su

AMD bets on heterogeneous compute (CPU/GPU/FPGA) over single killer chip for AI workloads

No one-size-fits-all silicon exists; AMD's portfolio approach with chiplets, networking optimization, and acquisitions (Xilinx, Pensando, ZT) enables right-sized compute per workload, valid…

AI Hardware & Chip Architecturetailwindscore 8/10lisa su

Heterogeneous compute (CPU/GPU/FPGA/NPU) wins over GPU-only for diverse AI workloads

No single accelerator dominates all AI workloads; AMD's portfolio approach — combining x86 CPUs, CDNA GPUs, Xilinx FPGAs, Pensando DPUs, and ZT Systems rack-scale integration — matches the…

Semiconductorstailwindscore 7/10lisa su

AMD roadmap extends to MI500/MI600 with 3-5 year design cycles accelerated by AI

AI is compressing chip design cycles (Anthropic's Claude optimizing MI355 bring-up); AMD working with top customers 3-5 years out to embed workload flexibility into MI500/MI600 architecture…

←

▶ 43:00 · AI Infrastructure · Grid & Power Infrastructure · US data center flops utilization at 15% creates national security crisis per AMP founder

episode briefing

TBPN

AMD CEO Lisa Su Live on TBPN

2026-07-23 · 17 company · 17 thematic

sentiment

15 bull0 bear2 neu

speakers

lisa su

CEO of Advanced Micro Devices since 2014; led AMD's transformation into AI and high-performance computing leader with MI300/MI355/MI450 accelerator roadmap.

anne

seth cohen

oliver cameron

Founder of Odyssey building foundational world models; previously at Cruise working on self-driving cars; raised $310M for world model development across robotics, gaming, and autonomy.

muhammad

Founder of Ideogram building AI foundation for enterprise design and content; previously started Google's Imagen; released open-weight frontier image model for on-prem enterprise deployment.

now playing · AI Infrastructure · Grid & Power Infrastructure

World Models & Simulationtailwindscore 8/10oliver cameron

World models at GPT-2 phase will replace hand-tuned autonomy in robotics and driving

Broad-distribution world models lightly tuned for tasks outperform narrow vision-language-action models with orders of magnitude less data; applications span gaming, robotics, driverless ca…

Robotics & Physical AItailwindscore 8/10oliver cameron

World models pretrained on broad internet data generalize to robotics with orders-of-magnitude fewer demonstrations

Training world models on diverse video/game/simulation data then fine-tuning for specific embodiments (cars, humanoids, drones) dramatically reduces real-world data requirements versus visi…

Autonomous Vehiclestailwindscore 7/10oliver cameron

World models enable Tesla-level autonomy diffusion to legacy auto by 2030 via supplier model

Specialized autonomy companies will sell turnkey driverless stacks to OEMs (Honda, Ford) by 2030; legacy auto must choose between adopting external tech or burning capital on internal failu…

Open Source AItailwindscore 8/10muhammad

Frontier open-weights models enable enterprise on-prem deployment and data sovereignty

Compact, high-quality open-weights models (Ideogram, DeepSeek) allow regulated enterprises to self-host generative AI, removing IP leakage risk and reducing inference costs, accelerating B2…

Enterprise AI Adoptiontailwindscore 7/10muhammad

Enterprise image generation requires consistency, brand DNA, and on-prem data sovereignty

Brands need 100% accurate logo/typography translation and product photography; open-weight compact models enable on-prem licensing solving IP concerns for design, manufacturing, and defense…

AI Infrastructuretailwindscore 9/10aj

US data center utilization at 15% flops efficiency creates national security crisis

Only 15% of provisioned flops are productively used due to 30-40% scheduling losses and 15% model flop utilization (MFU); fixing this software-defined waste is as critical as building new c…

AI Infrastructuretailwindscore 9/10anne

US data center flops utilization at 15% creates national security crisis per AMP founder

Only 15% of paid flops are utilized due to 30-40% scheduling losses and 15% chip MFU; solving this efficiency gap is as critical as building new capacity for US AI competitiveness vs China.

Grid & Power Infrastructuretailwindscore 9/10aj

AI infrastructure players procuring 2030 energy capacity now, nuclear essential for US competitiveness

Multi-gigawatt AI campuses require 5-7 year energy lead times; AMP's 2030 procurement horizon and nuclear investment (Maver) highlight that power permitting and generation, not chip supply,…

AI Agentstailwindscore 8/10anne

RL with formal verification drives next consumption leg beyond coding agents

Industrial-grade RL with verifiable rewards (unit tests, physics simulations) enables rapid capability gains in domains like materials science (Periodic Labs superconductors); verification…

AI Agentstailwindscore 9/10aj

Reinforcement learning with verifiable rewards drives next leg of AI capability expansion beyond coding

Industrial-grade RL with formal verification (unit tests, physics simulations, X-ray diffraction) creates tight feedback loops that rapidly expand model capabilities in any domain with grou…

Nuclear Energytailwindscore 9/10seth cohen

Spent nuclear fuel reframed as strategic energy asset worth 4x Saudi oil reserves

100,000 tons of commercial spent fuel retains 97% of its energy potential; policy shifts have turned 27 governors into advocates for hosting full fuel cycle facilities, unlocking a massive…

Nuclear Energytailwindscore 9/10seth cohen

Spent nuclear fuel reframed as strategic energy asset worth 4x Saudi oil reserves

100,000 tons of commercial spent fuel retains 97% energy content; policy shift turned waste liability into asset with 27 governors competing for fuel cycle campuses, unlocking domestic ener…

AI Bubble / Capex Debatetailwindscore 8/10john coogan

OpenAI's $750B 2030 compute budget and $20B single data center signal sustained capex conviction

The 25% increase in OpenAI's projected infrastructure spend to $750B through 2030, anchored by a $20B Georgia campus, reflects leadership's confidence that scaling laws continue to deliver…

Semiconductorstailwindscore 8/10lisa su

AI-assisted chip design cuts 3-6 months from silicon bring-up, creating compounding hardware-software flywheel

Anthropic using Claude to optimize kernels for AMD MI355 demonstrates LLMs accelerating the hardware development loop; faster silicon iteration enables faster model training, which further…

AI Hardware & Chip Architecturetailwindscore 8/10lisa su

AMD bets on heterogeneous compute (CPU/GPU/FPGA) over single killer chip for AI workloads

No one-size-fits-all silicon exists; AMD's portfolio approach with chiplets, networking optimization, and acquisitions (Xilinx, Pensando, ZT) enables right-sized compute per workload, valid…

AI Hardware & Chip Architecturetailwindscore 8/10lisa su

Heterogeneous compute (CPU/GPU/FPGA/NPU) wins over GPU-only for diverse AI workloads

No single accelerator dominates all AI workloads; AMD's portfolio approach — combining x86 CPUs, CDNA GPUs, Xilinx FPGAs, Pensando DPUs, and ZT Systems rack-scale integration — matches the…

Semiconductorstailwindscore 7/10lisa su

AMD roadmap extends to MI500/MI600 with 3-5 year design cycles accelerated by AI

AI is compressing chip design cycles (Anthropic's Claude optimizing MI355 bring-up); AMD working with top customers 3-5 years out to embed workload flexibility into MI500/MI600 architecture…