Custom Silicon Breaks Nvidia's Tollbooth
The largest cloud platforms increasingly have enough AI volume to justify chips optimized for their own models, networks and power budgets. That shifts value toward ASIC designers, semiconductor IP, foundries, advanced packaging, high-bandwidth memory and data-center connectivity. NVIDIA remains central to training, but custom accelerators can erode its pricing power in predictable inference workloads and reduce customers' dependence on CUDA. The market is therefore repricing the AI supply chain from a single dominant GPU vendor toward a broader custom-silicon ecosystem.
How this story could end...
The endings this story genuinely supports, written from the same evidence as the story updates. The value ranges are illustrative examples of how the affected companies as a group could re-rate — not forecasts, targets or advice.
Mapping out the endings…
How the two sides have traded
The bull and bear baskets for this story, equal-weighted and indexed so the two lines start together.
Each scenario's basket — 3 months
Both sides on one chart — 10 stocks on the bull side, 10 stocks on the bear side. Neither side is emphasised.
Each line tracks the 10 companies most exposed to this story, so a broad basket of loosely related names does not flatten it. The full list below is unchanged.
Jul 20, 2026 – Oct 19, 2026 · 66 sessions
Equal-weighted performance of the 10 most exposed stocks over the last 3 months.
Most affected stocks
Bull case
Global Unichip helps cloud and technology customers design complex ASICs and bring them onto TSMC manufacturing processes.
Andes supplies RISC-V processor IP that can be embedded in custom accelerators as control, management and security cores.
Lumentum supplies optical components that support the high-bandwidth interconnects required to scale clusters of custom accelerators.
Microsoft's Maia accelerator gives Azure another lever to control AI capacity, cost and supplier concentration.
MediaTek is extending its advanced-chip design expertise into cloud ASIC programs, giving hyperscalers another custom-silicon partner.
IBIDEN produces advanced IC package substrates required for large, high-density AI accelerators and chiplet-based processors.
A broader range of complex AI accelerators expands demand for Advantest's high-end semiconductor testing systems.
Growing numbers of accelerator designs increase demand for ASE's advanced packaging, testing and heterogeneous-integration services.
Marvell supplies custom compute, electro-optics and connectivity technology used in bespoke cloud AI systems.
Scale-out custom accelerators increase demand for optical components and transceivers connecting servers and data-center clusters.
Bear case
DigitalOcean lacks the workload scale to amortize leading-edge proprietary accelerators, potentially widening its cost disadvantage against hyperscalers.
CoreWeave's Nvidia-heavy GPU cloud could face weaker relative economics if hyperscalers shift large inference workloads onto proprietary accelerators.
Nebius's GPU-heavy AI cloud could face rental-price and asset-utilization pressure if major customers shift predictable workloads to proprietary accelerators.
The shift toward hyperscaler-designed accelerators further marginalizes Intel's data-center compute franchise unless its foundry captures the manufacturing work.
OVHcloud may face greater infrastructure-cost pressure because it cannot spread custom-chip development across AI volumes comparable with the largest cloud platforms.
Custom chips intensify competition for the second-source accelerator opportunity that AMD is targeting with its Instinct platform.
Wider adoption of internally designed accelerators threatens NVIDIA's share and exceptional margins in standardized inference workloads.
Dell risks losing AI-server share where large cloud customers bypass branded GPU platforms in favor of internally designed systems built by ODMs.
HPE's merchant-accelerator server opportunity may narrow as the largest cloud buyers deploy proprietary silicon through customized supply chains.
A shift from standardized merchant-GPU systems toward bespoke hyperscaler hardware could pressure Supermicro's differentiation and margins.
Reporting behind this story
Looking at what moved this story…