NVIDIA Management: Custom Chips Unlikely to Shake CUDA Ecosystem, Cloud Provider Revenue Share Rises to 20%
nashnova research
Nvidia management pushed back on the custom-chip threat in a dialogue with Bernstein, noting the company's share of hyperscaler capex has climbed from ~5% pre-ChatGPT to ~20% — and that the real moat is not GPU hardware itself but the CUDA ecosystem and general-purpose programmability.
Custom chips are gaining ground — why isn't Nvidia worried?
Management offered a key data point: Nvidia's share of hyperscaler capex rose from ~5% before ChatGPT to ~20%, moving into the mid-20s range.
This means → even as Google TPUs — Google's in-house AI chips — and various custom ASICs — chips designed for a single task — keep deploying, Nvidia's slice of spending has grown, not shrunk.
The core argument: Nvidia's value comes from performance, interchangeability, and the CUDA ecosystem, not the GPU hardware itself. In plain terms = customers are buying an entire software toolchain and developer community, not just a chip.
GPU vs. custom chip — which fits where?
Management cited Elon Musk's Colossus project as an example: GPUs can handle shifting workloads and be rented out; ASICs are optimized only for narrow, fixed tasks.
This means → when AI models iterate quickly, general-purpose programmable systems matter more. Custom ASICs only win when the workload is highly predictable.
This reflects a deeper industry reality: AI is still changing fast, and most customers cannot lock in future compute needs in advance.
Data centers are getting more expensive — where are the bottlenecks?
Management noted that data-center content is heading toward ~$40 billion per gigawatt — not just chips, but CPUs, networking, LPUs — low-latency processing units — and other system-level components.
This means → Nvidia delivers a system-level solution, not a single chip, which separates it from pure chip-level custom-silicon vendors.
On supply, Nvidia remains ~70% supply-constrained in fiscal 2027. Bottlenecks span wafers, CoWoS packaging — an advanced chip-packaging technology — high-bandwidth memory, optical transceivers, land, power, and rack capacity. Qualifying a new foundry takes more than two to three years.
How will the customer mix shift?
Hyperscalers will remain ~50% of revenue long-term. The other half will grow faster, driven by emerging cloud providers, industrial, enterprise, sovereign, and AI-lab customers.
Frontier AI labs account for roughly 20% of Nvidia's revenue. These labs switch flexibly between building their own infrastructure and using cloud providers; both modes will coexist.
Management expects ~80% of models will eventually be open-source. In plain terms = open-source expands who can deploy AI, but it does not make the underlying compute free — open-source models are not the same as open-source compute.
How will earnings be returned? What to watch next?
On capital allocation, management said that once supply and infrastructure commitments are resolved, the plan is to return nearly all remaining free cash flow to shareholders through buybacks and potentially higher dividends.
This means → Nvidia sees expanding capacity as the top priority right now; shareholder returns come after the supply bottleneck eases.
The most direct test ahead: whether the custom-ASIC displacement thesis leaves a mark on cloud capex share — Nvidia's ~20% share would need to start declining before the threat is confirmed.
市场有风险，内容仅供研究参考，不构成投资建议。