AI is often described as software: models, applications, assistants, images, video, music or code. But the more AI enters creative work, the more visible its physical base becomes.
01 / Core idea
AI is not just cloud and interface
On the surface, we see a prompt, an image and a fast answer. Underneath, there are graphics processors, memory, advanced packaging, lithography, electricity and large data centers.
Every prompt consumes compute. Every large model needs specialized chips. Every data center needs cooling, energy and network infrastructure.
The future of creative work does not depend only on the best application. It also depends on cheap, fast and available compute.
02 / Compute layer
GPUs are central, but not the whole story
Nvidia remains the symbol of AI infrastructure. Its GPUs are a foundation for training and running large models. But that dominance opens another question: custom ASICs, chips designed for specific use cases.
Why large AI companies want their own chips
Cloud platforms and AI companies do not want to depend forever on one chip type and one supplier ecosystem. AI infrastructure is becoming layered: general GPUs for flexibility, custom chips for efficiency, specialized memory for bandwidth and advanced packaging to connect the system.
03 / Hidden bottleneck
Memory is the quiet hero of the AI boom
A large part of the story is not only about GPUs, but about memory. HBM, or high-bandwidth memory, is critical because models do not only compute. They constantly move enormous amounts of data between memory and processing cores.
HBM
High-bandwidth memory without which large AI workloads quickly hit a wall.
CoWoS
Advanced packaging that connects accelerators and memory into a working system.
Energy
Data centers are not abstract cloud. They need power, cooling and grid access.
Strong demand does not mean zero risk
Memory is cyclical. Historically, it can look cheap precisely when profits are near a peak. A low multiple does not automatically mean a cheap stock.
04 / Physical world
Data centers are not a metaphor
The AI supply chain is often reduced to the question of who makes the best chip. In reality, the weak point may be packaging, materials, manufacturing capacity, the electric grid or cooling.
Data centers are physical buildings. They need stable power, cooling and network access. That is why large technology companies are signing long-term energy agreements, including nuclear capacity.
The fundamentals did not change; positioning changed. That distinction matters when technology and investing meet.
05 / Creative impact
What this means for creative work
At first glance, market corrections may seem far from design, video, music or writing. In reality, this infrastructure is their base. Creative tools will depend on whether cheap and fast compute is available.
Availability will shape who can use AI fully
If AI chips and memory remain bottlenecks, the most advanced features will stay expensive, centralized and mostly available to large platforms. If infrastructure expands and becomes more efficient, AI can become an ordinary layer of creative software.
Conclusion
Performance has no taste by itself
I do not see the future of creative work as a simple replacement of people by machines. I see it as a new distribution of attention: some work becomes faster, some becomes curatorial, and some requires even more discipline.
Semiconductors sit at the beginning of that chain. At the end, the question is still what the image should say.