Cerebras SVP Angela Yeung dropping a bombshell: AI-assisted kernel development is eroding NVIDIA's moat faster than anyone expected.
The old argument? "$NVDA has 15-20 years of CUDA investment, you can't catch up." Programmability was the fortress.
Now? AI is writing kernels at ridiculous speed. Cerebras hired interns with almost zero kernel experience this summer. Within weeks, they were bringing up models independently. That's not supposed to happen.
The implication: CUDA's software lock-in might not be the insurmountable barrier it once was. If you can bootstrap junior devs to production-level kernel work in weeks instead of years, the entire competitive landscape shifts.
This is the meta-move: using AI to accelerate AI infrastructure development. The feedback loop is getting tighter.
The old argument? "$NVDA has 15-20 years of CUDA investment, you can't catch up." Programmability was the fortress.
Now? AI is writing kernels at ridiculous speed. Cerebras hired interns with almost zero kernel experience this summer. Within weeks, they were bringing up models independently. That's not supposed to happen.
The implication: CUDA's software lock-in might not be the insurmountable barrier it once was. If you can bootstrap junior devs to production-level kernel work in weeks instead of years, the entire competitive landscape shifts.
This is the meta-move: using AI to accelerate AI infrastructure development. The feedback loop is getting tighter.