Radek Bartyzal's blog

[Video notes] Jensen Huang interview 2025-01

A look into the future with Jensen Huang (2025-01)

Video here

AI 5-layer stack

  • Applications (Cursor, chatbots etc,)
  • Models
  • Infrastructure (both HW = datacenters and SW = libraries)
  • Chips
  • Energy

Chips and Moore Law

Moore’s Law is dead, transistor gains are 10s% in 1-2 years, not 200%.

  • gains must be made in SW layer
  • ties it nicely to why NVIDIA’s general GPUs are going to be still relevant because the SW will keep changing to get the gains, who knows :)

Example: MoE (Mixture of Expert models) are difficult for inference compared to classic one big Transformer, that is architectural innovation that removes some specialized chips advantages.

There are specialized chips that optimize for inference of specific architectures like Transformers. NVIDIA’s main product is general computing HW that allows them to be relevant as long as the architectures evolve:

  • Diffusion vs token-by-token inference of Transformers
  • State space models vs Transformers
  • CNNs still run everywhere = the GPU is general = backward compatible with technologies

Fits with the narrative why NVIDIA acquired Groq:

Industries to look at in 2026

Biology

ChatGPT moment for digital biology = protein sequences

  • Multi-modality + long context + Synthetic generation
  • Protein understanding but also protein generation
  • Chemical understanding and generation
  • Synthetic data, new infrastructure for testing, generating data, doing experiments on designed proteins
  • Ultimately foundational model for proteins, cells

Cars

Reasoning = thinking unlocks being able to handle unknown edge cases, out of distribution situations which is the current limitation of end-to-end trained models.

4 generations:

  • smart sensors with hardcoded rules
  • ML models for each task = planning, image recognition
  • end to end models
  • end to end models with reasoning = NEW

Robotics = Verticalization

AI is general, it will embody any robot.

In consumer apps, you are fine with 90% success rate.

But in industrial application you need 99.999 for kitchen appliances etc. so you will get verticals for each domain to get these 9s right.

Verticals like Harvey, Cursor = wrapper of general AI

  • Same with Robotics = embody all those different types like hands, claws, delivery robots, kitchen appliances etc.

Energy

  • Energy growth is crucial for any growth.
  • For the next decade, we need to take all we can get = gas turbines, any fossils etc. combined with renewables.

US vs China 2026 - (from CEO of NVIDIA that sells GPUs to China)

  • Relationship US vs China will improve
  • Decoupling not feasible.
  • Happy that US keeps selling GPUs to China.

AI Bubble

  • Accelerated computing is used everywhere not just for AI.
  • NVIDIA diversifies to quants, AV (self-driving), Biology, Robotics - all using AI
  • Biology companies: R&D were spending on wetlabs, now they are building supercomputers.
  • 2% of Economy is R&D => if it will be in supercomputers then it’s 2T USD per year

Purpose vs Task

Each job has a purpose and consists of numerous tasks. AI can accelerate the task but shouldn’t replace the purpose.

Example: Radiologists

  • task = analyze scans, purpose = diagnose patients
  • there are more radiologists because the quality nad quantity of service goes up = there is more work to be done.

My take:

  • this a nice framework to think about your job, the takeaway is true for areas where the purpose cannot be automated of course :D