Nvidia Keynote 2025: Missing: Consumer Use Cases

Nvidia’s Jensen Huang delivered another riveting keynote reminiscent of Steve Jobs’ iconic events 15 to 20 years ago. Key differences would be length, focus (industrial or enterprise vs. consumer), and consumability - for me. I learned a lot. Here are a few of my takeaways from a consumer perspective: 

• Technology is incredibly powerful especially if it is not bound by financial- or power constraints.  Nvidia proposed a sweet spot of just over 500K TPS/MW (i.e., throughput) to generate just over 800 tokens per second (TPS) for one user. Example: Jensen Huang popped this prompt into two models - a traditional LLM and a reasoning model (DeepSeek): “I need to seat 7 people around a table at my wedding reception, but my parents and in-laws should not sit next to each other. Also, my wife insists we look better in pictures when she’s on my left, but  Ineed to sit next to my best man. How do I seat us on a round table? But then, what happens if we invite our pastor to sit with us?” The traditional LLM answered in a couple of seconds using 439 tokens with a slightly wrong answer. The reasoning model used 8,559 tokens to offer a correct answer in about 40 seconds. Assume about 200 TPS. The reasoning model used 20x the number of tokens and 150x the compute power. Demo was meant to show the speed versus quality trade-offs. There you go.
  

• Most of the use cases today are enterprise or industrial - for agentic AI as well as physical AI (i.e., robots). Even the example above doesn’t pass the common sense test. A human could solve this problem faster (and more inexpensively) than the reasoning model. Consider also that someone had to type in an accurate prompt. FWIW … I am keeping an open mind about the robots in the home emptying a dishwasher, pouring a cup of coffee, or stirring a pot of soup on a stove. For these robots to make sense in a home, they’d have to cost closer to what a vacuum cleaner does today. Could be  Dyson.
  

• Technologists are still demo’ing a very limited number of agentic use  cases in lab type environments. Most of these use cases involve software/coding, math problems, and now machines trained with computer vision (expensive - think autonomous vehicles). Most of these agentic examples have right or wrong answers - not answers that require judgement or extensive context (i.e., memory) 

Things history has taught us about technology (and humans): 

• Just because the technology makes it possible doesn’t mean anyone wants to buy it. 
  

• Capabilities of technology often lead humans’ ability to consume it i.e., change their behavior to do old things in new ways let alone new things. (quoting J. McQuivey)
  

• Employees won’t rush to adopt productivity and efficiency tools - that is what enterprises do. Humans need time and motivation to form new habits. Besides, humans are still responsible or accountable for the errors, outcomes, and more. To oversimplify dramatically, the conversations need to expand to include the human factor. 
  

• Nothing really goes away. There are more than 300 digital touch points yet we still receive paper mail. We receive hundreds of Emails each day that we don't read. We still use paper money. Humans still ride horses. Agents may start to replace enterprise apps, but we are far far off from a wholesale replacement - even if it does happen one day. Keep in mind that automation and other "old" technologies work really well. So do humans.

Otherwise:

• The technology, vision, and ecosystem of partners is impressive - I really have no words. Blew me away. Nvidia is using all the techniques (e.g., powerful hardware, pruning, distillation, parallel processing, distributed inference, pruning, etc.) to accelerate inference. 
  

• Building AI infrastructure including AI Factories (new term for me) is a very, very big business. Understanding the numbers seems to require knowledge of physics, math, and a new vocabulary of Greek prefixes recently introduced. 
  

• Technologists still tend to focus on what the technology might do rather than what consumers or humans might need (or want). Their marketing teams are also introducing so much hyperbole about what the technology might do … without any timelines, safety-, risk-, climate- or financial impact … that it comes across as irresponsible science fiction. It is almost as if folks feel free to just make stuff up.