The world is too loud. Read what matters.

Reuters

Past CPUs were built for humans; Nvidia says agents need completely different architecture

Agents operate on nanosecond timescales while legacy CPUs rent logic by the second; Nvidia built the new-architecture Vera CPU for this, calling it a market larger than the legacy CPU market.

AI InfrastructureAgentsCPUNvidiaData CentersOpen-source models
One episode traces the evidence chain from AI‘talking’to‘making money’: code commits triple, tokens become profitable, specialized CPUs emerge—high information density with considerable marketing throughout.

The argument · tap a timestamp to hear it

7:20

Engineer productivity jump proves hiring will accelerate, not slow

Jensen Huang's figures: GitHub commits grew from 300 million in 2023, 400 million in 2024, to 500 million in 2025, with the opening months of 2026 already nearly tripling that rate. His inference: the $3 trillion software engineer payroll now levers approximately $9 trillion in output, so companies aren't cutting engineers but hiring more—as each engineer generates greater value, adding headcount becomes more cost-effective.

— Jensen Huang
9:22

Token profitability has fundamentally reshaped compute infrastructure economics

Jensen Huang marks this shift's turning point as the moment tokens became profitable. Once generating tokens could make money, AI companies had unlimited incentive to expand capacity, so they rushed to build AI factories and buy compute. He attributes the past two years of explosive demand and stock gains across Taiwan's supply chain directly to this.

— Jensen Huang
10:23

Agents require wholly new computation model: model, harness, tools, runtime

Jensen Huang emphasizes agents aren't just scaled-up large models, but a new computation pattern: model (the brain), harness (the body, handling memory orchestration and tool invocation), tools and skills, plus runtime. Each understand-reason-act loop requires a full rack's compute, which means future infrastructure must be rebuilt around this loop, not simply by stacking GPUs.

— Jensen Huang
27:56

Data centers waste 40 percent of power on safety reserves

Jensen Huang identifies a widespread industry pain point: today's AI factories typically raise power reserves by 40 percent for safety margins, and that entire power allocation generates zero tokens and zero revenue. The goal of Nvidia's DSX Max LPS is to unlock that ‘wasted’power and convert it directly into billions of dollars in annual revenue.

— Jensen Huang
52:19

CPUs designed for humans must be completely rebuilt for agents

Legacy CPUs' economics charged by the second and rented by the core—more cores meant more rental potential. But agents operate at nanosecond timescales; every tool call or database lookup means waiting, and waiting wastes the more-expensive GPU compute beside it. Nvidia redesigned Vera accordingly: 10 instructions per clock per core, 3.6TB/s die bandwidth, and memory bandwidth two to three times that of comparable x86 chips. Real-world testing shows SQL processing 3x faster and New York Stock Exchange live stream processing 6x faster.

— Jensen Huang
1:17:46

Open-source NeMoTron model matches proprietary performance at lower cost

Nvidia open-sourced NeMoTron 3 Ultra, a new-generation model using state space models combined with mixture-of-experts architecture: inference five times faster than comparable open-source models, total cost 30 percent lower. Jensen Huang's reasoning: opening both training data and training scripts lets enterprises train their own proprietary agents atop this foundation rather than getting locked into closed models. Cadence's chip design agent exemplifies this approach.

— Jensen Huang
1:32:18

Complete Windows PC redesign in 40 years marks fundamental shift

Nvidia partnered with Microsoft to introduce the full RTX Spark product line covering laptops, desktops, and workstations—the first complete Windows PC redesign in 40 years. The core idea: every PC needs a resident agent, requiring sandboxed runtime (Open Shell) rather than just installed applications. Jensen Huang's analogy: phones stopped being phones—in 10 years, PCs will be as different from today's as today's are from phones, more like an R2-D2 living in your home.

— Jensen Huang
1:41:27

Data is the real bottleneck for physical-world AI, not compute

Internet text is written from humanity's first-person perspective, but robots need robot-perspective data—which barely exists. Nvidia trained the Cosmos 3 world foundation model by combining remote operation, simulation (Omniverse), and bootstrap training converting third-person to first-person viewpoints. Jensen Huang's motto: ‘compute is data’. Simultaneously, Nvidia published the autonomous driving model Alpamayo and reference design Isaac GR00T for humanoid robots.

— Jensen Huang

In their own words · checked verbatim

People talk about AI reducing jobs. Complete nonsense. It's causing more software engineers to be hired.

Jensen Huang8:22

Because it is now profitable, the AI companies want to build a lot more tokens, generate a lot more tokens, build more AI factories

Jensen Huang9:22

Compute is revenues. Performance per watt is your revenues.

Jensen Huang37:04

Agents are impatient. They don't live in a world that is in seconds. They live in a world that's in nanoseconds.

Jensen Huang52:19

The CPU is now the conductor, and the GPU is the orchestra.

Today, when you think about your phone, the one thing you don't do with it is make phone calls. You do just about everything else.

Jensen Huang1:35:19

For physical AI, compute is data.

Figures

2025 GitHub commits500 million7:20
Engineer payroll to output ratio$3 trillion salary generates approximately $9 trillion output8:22
AI factory excess power reserves40%27:56
Vera Rubin supply chain scaleTwice Grace Blackwell's40:09
Rack assembly time2 hours shortened to 5 minutes41:11
NeMoTron 3 Ultra performance5x faster inference, 30% lower cost1:18:47
Vera CPU SQL processing3x faster1:05:35
Vera CPU NYSE streaming processing6x faster1:06:35
Isaac GR00T humanoid specifications6 feet, 150 pounds, 31 degrees of freedom1:49:34

Glossary

harness
Software layer connecting the model, memory, and tools while orchestrating the full reason-act cycle.
KV caching
In large model inference, the mechanism that preserves attention intermediate results to accelerate subsequent generation.
Mixture of Experts (MoE)
Model architecture that activates only some expert sub-networks during inference to reduce compute.
DPU
Specialized processor handling network, storage, and security tasks to offload CPU burden.
tokens per watt
Efficiency metric measuring how many AI outputs a data center produces per unit of electrical power.

How to listen

Who it's for

Professionals tracking AI infrastructure, compute chip investment, or agent deployment—especially those wanting to understand Nvidia's next-generation hardware and software roadmap.

Skip

Opening thanking Taiwan partners and closing rap performance and songs carry low information density and can be skipped.