The Surge in Custom AI Chip Development Can't Solve Semiconductor Supply Chain Bottlenecks
Claire Weston
Nearly every major AI company is designing its own chips to cut Nvidia dependence, but design is only the first step — the handful of fabs that can build cutting-edge semiconductors remains the real chokepoint, and that isn't changing soon.
Who is designing their own chips?
Microsoft, Google, Amazon, and Meta all have custom-chip programs. OpenAI recently partnered with Broadcom to launch its first in-house inference chip.
Anthropic is reportedly in talks with Samsung; Apple already designs chips for iPhone, iPad, and Mac, and per Bloomberg is developing a separate chip for AI servers.
China's DeepSeek is also developing its own AI chip, per Reuters. This means → in-house design is no longer a few giants' play — it is an industry-wide consensus.
Why go custom?
Bernstein analyst Stacy Rasgon told Axios the core value is bargaining leverage: "I want to have something in my hand when I negotiate with Jensen Huang."
In plain terms = going custom doesn't necessarily mean replacing Nvidia entirely — it means no longer accepting whatever price Nvidia quotes.
Custom chips can also be optimized for specific workloads, potentially lowering operating costs.
Designed — but who builds them?
TSMC handles the vast majority of cutting-edge chip fabrication. Despite investing tens of billions to expand, its executives have repeatedly said demand continues to outstrip supply.
Samsung and Intel also offer foundry services, but Intel's manufacturing process has fallen behind competitors in recent years.
All these fabs depend on ASML of the Netherlands for lithography equipment — the machines that etch circuits onto chips. ASML is the only company on earth that can make the most advanced lithography tools. This means → the single hardest-to-replicate node in the entire supply chain is one company.
Is everyone chasing the same scarce resources?
Companies trying to reduce Nvidia dependence are still competing for the same pool: cutting-edge foundry capacity, advanced packaging, high-bandwidth memory (HBM), and lithography equipment.
Rasgon noted: "If you're just starting to design a chip now, you won't see silicon for three years."
In plain terms = the "independence" that custom chips offer is independence at the design layer. At the manufacturing layer, dependence is actually more concentrated.
Does Nvidia's moat still hold?
Even companies actively pursuing custom chips are still buying Nvidia GPUs in volume — because Nvidia's combined hardware, networking, and software ecosystem is extremely hard to replicate.
Nvidia projects cumulative revenue of $1 trillion from 2025 through 2027. Rasgon said that even if competitors capture only a sliver, "that could still be tens of billions of dollars."
This reflects the real logic behind the custom-chip boom: the goal is not to destroy Nvidia, but to carve even a small slice from a trillion-dollar pie — and that slice is enormous.
Where is the ultimate bottleneck?
The final constraint is not design capability. It is that only a handful of companies worldwide can manufacture the most advanced semiconductors — and that landscape is not changing soon.
TSMC dominates fabrication; ASML is irreplaceable. The hotter the custom-chip wave, the more concentrated manufacturing becomes.
This means → custom chips solve the question of "who do I buy from," not "who can build it." The real supply-chain bottleneck sits after design.
Content is for reference only, not financial advice.