What this is

NVIDIA has opened two of its foundational technologies to outside parties: NVLink, the high-speed chip-to-chip interconnect, and NVHBM, high-bandwidth memory purpose-built for AI training. Previously used only inside NVIDIA's own GPUs, both are now available for third-party custom AI chips — what the industry calls XPUs (Custom Accelerators).

To put it plainly: NVIDIA is letting others drive on its "highway," but under its rules. For cloud hyperscalers (Google, Amazon) and Chinese AI chip companies developing their own silicon, the road is shorter — but the tollbooth belongs to NVIDIA.

Why open it now? The backdrop is that AI models keep growing, and no single chip can hold them — workloads must be split across hundreds or thousands of chips working in concert. How fast those chips "talk" to each other, and whether memory feeds them data fast enough, determines whether compute is wasted. The NVLink + NVHBM combination is one of the key reasons NVIDIA GPUs have led for so long. Extending this path to third parties is, in essence, widening the ecosystem moat.

Industry view

Positive voices are concentrated on the custom chip vendor side: tapping into a mature NVLink ecosystem saves substantial in-house R&D time and eases mass production. For established players like Google TPU and Amazon Trainium — whose interconnects have long been a weak point — this is a chance to close the gap.

But we are more focused on the risks. There are at least two layers worth flagging.

First, deeper ecosystem lock-in. Once you build around NVLink and NVHBM, system design, debugging, and iteration all revolve around NVIDIA. Pricing power and the upgrade cadence are no longer in your hands. Today's partner can become tomorrow's harvested customer.

Second, geopolitical variables. NVHBM's supply chain involves advanced packaging and high-bandwidth memory fabrication. For Chinese vendors, "open" is open — but whether supply is stable and whether export controls will intervene remains unclear. The door is open today; whether it stays open tomorrow is not up to those standing outside.

Impact on regular people

  • For enterprise IT: Compute procurement will get more complex. Over the next two to three years, cloud pricing may diverge sharply depending on whether NVHBM is in play — model selection will require re-bidding.
  • For individual careers: Demand stays tight for roles in AI chips, interconnect architecture, and advanced packaging. Engineers with deep hardware knowledge are gaining bargaining power; pure software backgrounds will need to upskill.
  • For the consumer market: End users won't feel anything in the short term. But as training-side costs fall, further drops in large-model API prices are highly likely — AI applications will keep getting cheaper.