The AI labs are done waiting for Nvidia to decide their fate.
The Summary
- Anthropic asked SK Hynix for supplies to make its own chips, according to SK Group Chairman Chey Tae Won — a signal that frontier AI labs are going vertical on silicon
- This mirrors moves by Google, Meta, and Amazon who all build custom training chips. OpenAI, backed by Microsoft, has been the holdout. Not anymore.
- The implication: compute independence is now existential for AI labs, not just cost optimization
The Signal
Anthropic approaching SK Hynix directly marks a turning point in how AI labs think about their infrastructure stack. SK Hynix makes High Bandwidth Memory (HBM), the specialized memory that sits next to GPUs in AI training clusters. It's the bottleneck in modern AI systems. More critical than compute cores in many workloads.
When an AI lab asks a memory supplier for help making chips, they're not just optimizing margins. They're securing supply. They're designing around their specific architecture needs. Most importantly, they're declaring independence from the Nvidia-dominated roadmap that has constrained everyone's training schedules for two years.
"Compute independence is now existential for AI labs, not just cost optimization."
The precedent here matters:
- Google's TPUs have been in production since 2015, giving DeepMind consistent access to compute while competitors scrambled
- Meta's MTIA chips launched in 2023, purpose-built for recommendation and ranking workloads
- Amazon's Trainium and Inferentia chips now power significant portions of AWS AI instances
Anthropic has Amazon as an investor and cloud partner. Amazon has its own chip efforts. So why would Anthropic still need custom silicon? Because relying on a cloud provider's hardware, even a friendly one, still means being second in line. When Amazon allocates Trainium capacity, AWS revenue-generating customers come first. Anthropic knows this.
The SK Hynix angle is telling. HBM supply is viciously constrained. SK Hynix and Samsung control most global production. Nvidia buys massive quantities for H100 and H200 GPUs. If Anthropic can secure direct HBM allocation by building its own chips, it bypasses the Nvidia allocation queue entirely.
The Implication
Watch for OpenAI to announce custom silicon within six months. They can't let Anthropic and Google operate on independent timelines while they wait for Microsoft to prioritize their Azure capacity needs. The AI lab that controls its chip roadmap controls its training schedule. The lab that controls its training schedule wins the race to AGI.
For the rest of us, this means the agent economy gets built on increasingly fragmented infrastructure. Your AI agent might run on five different chip architectures depending on which lab's models you're using. Portability and interoperability just got harder. The Fourth Web needs open standards more than ever.