Huawei Pushes Proprietary Optical Standard to Loosen Nvidia's Grip on AI Infrastructure
The Shenzhen giant's near-package optics proposal reflects a broader strategy to shape the underlying architecture of next-generation data centres - and reduce dependence on US-controlled interconnect technologies.
A New Front in the Interconnect Wars
Huawei introduced a proprietary optical communication standard on 10 September, opening a fresh competitive front in the race to define how artificial intelligence chips talk to one another. The near-package optics approach, unveiled at an industry event in Taipei, positions the Shenzhen-based firm against incumbent players Nvidia and Broadcom, both of which have spent years embedding their own interconnect technologies into hyperscale data centres from California to Frankfurt.
At Opentechwire, we've tracked the escalating importance of chip-to-chip communication over the past eighteen months. As training runs for large language models push past the thousand-accelerator mark, bandwidth and latency between processors often matter more than the raw teraflops inside any single die. Whoever sets the plumbing standards effectively controls the blueprint for AI infrastructure.
Huawei's move signals an intent to claim a seat at that table, particularly across markets where US export restrictions have severed access to Nvidia's latest H-series accelerators and the accompanying NVLink fabrics. By publishing its own specification and calling for an industry ecosystem around near-package optics, the company is betting that a coalition of Chinese equipment makers, module suppliers and cloud operators will coalesce around a non-US alternative.
What Near-Package Optics Actually Means
Near-package optics refers to placing optical transceivers within millimetres of the processor package, rather than several centimetres away on a printed circuit board. The shorter electrical path reduces signal loss, cuts power consumption and allows higher data rates - critical when a single training cluster can generate petabytes of gradient updates per hour.
Nvidia's current approach embeds copper and some optical links directly into its Blackwell and Hopper modules, paired with proprietary NVLink switches. Broadcom supplies much of the silicon photonics and switch ASICs that underpin rival topologies from Google and Meta. Both architectures assume a degree of vertical integration: the accelerator vendor or the hyperscaler dictates module specifications, and the supply chain follows.
Huawei's standard, by contrast, appears designed for horizontal adoption. The company has not disclosed full technical parameters - exact lane counts, modulation schemes or power budgets remain under wraps - but industry observers note the emphasis on interoperability. If module makers in Shenzhen, Suzhou and Wuhan can build to a common specification, the argument goes, Chinese cloud providers gain a plug-and-play ecosystem that does not depend on a single foreign gatekeeper.
Strategic Context: Export Controls and Ecosystem Lock-In
The timing is hardly coincidental. US export controls tightened in October 2022 have constrained Huawei's ability to source advanced packaging services from Taiwan and to procure high-bandwidth memory from SK hynix or Micron. The company's in-house HiSilicon unit continues to tape out AI accelerators on older process nodes, but without access to the latest CoWoS or silicon interposer capacity, those chips face a connectivity bottleneck.
An open optical standard offers a workaround. If Huawei can persuade domestic foundries and outsourced assembly and test houses to qualify near-package optics modules at scale, it effectively decouples interconnect performance from bleeding-edge packaging. The photonics themselves can be fabricated on mature 130-nanometre or 180-nanometre lines; the innovation lies in co-design with the package substrate and thermal management, not transistor density.
There is also a defensive dimension. Nvidia's NVLink and Broadcom's switching fabrics are not merely technical choices; they are moats. Once a hyperscaler commits hundreds of millions of dollars to a particular topology, migration costs - software stacks, validated topologies, sparing strategies - make switching prohibitively expensive. By establishing an alternative standard now, Huawei and its partners aim to lock in Chinese AI infrastructure before it becomes irreversibly tied to US platforms.
The Ecosystem Challenge
Publishing a standard is the easy part. Building an ecosystem is another matter entirely. Nvidia benefits from a decade of CUDA inertia, a vast network of independent software vendors and tight co-engineering with TSMC's advanced packaging group. Broadcom's photonics have been validated in Meta's production clusters and Google's tensor processing unit pods, each representing years of joint debug and yield ramp.
Huawei's near-package optics will need at least three constituencies to succeed. First, module suppliers must invest in new assembly lines and qualify components at volume. Second, cloud operators and AI labs in China - Alibaba Cloud, Tencent, ByteDance, Baidu - must commit to deploying the technology in live training clusters, accepting the risk of early-generation bugs and performance gaps. Third, the broader semiconductor supply chain - substrate makers, connector vendors, test-equipment firms - must align roadmaps, a coordination problem that typically takes eighteen to thirty-six months even under benign conditions.
China's national policy environment provides some tailwinds. The Ministry of Industry and Information Technology has made technological self-sufficiency in semiconductors a stated priority, and state-backed investment funds have channelled capital into photonics startups and packaging specialists. But mandates and subsidies do not guarantee technical success. Yield, reliability and interoperability are discovered through iteration, not decree.
Implications for the AI Hardware Landscape
If Huawei's standard gains traction, the global AI hardware market will fragment further along geopolitical lines. Chinese data centres will increasingly run on a distinct stack - domestically designed accelerators, Huawei optical interconnects, locally fabbed photonics - while North American and European clouds remain anchored to Nvidia, Broadcom and emerging challengers such as Marvell and Intel's silicon photonics group.
That fragmentation carries costs. Developers building models that span regions will face toolchain incompatibilities. Research collaborations that require shared infrastructure will need to navigate export compliance and technical mismatches. Hardware vendors hoping to serve both markets will confront divergent certification requirements and roadmap pressures.
Yet fragmentation also creates opportunity. A bifurcated ecosystem means twice as many bets on interconnect topology, twice as many approaches to thermal and power trade-offs, and twice as many teams racing to solve the latency and bandwidth challenges that will define the next generation of AI systems. Some of those bets will fail; a few may yield breakthroughs that, in a more unified world, would never have been funded.
What Comes Next
Huawei has called for industry participation in refining the near-package optics specification, but the company has not announced formal partnerships or disclosed which module suppliers have committed to production. The next twelve months will reveal whether this initiative is a genuine ecosystem play or a unilateral effort dressed in the language of openness.
For now, the standard exists primarily on paper. The real test arrives when the first clusters go live, when training runs hit unexpected bottlenecks, and when engineers from competing firms must debug interoperability issues in the field. Standards are not adopted because they are published; they are adopted because they solve problems that matter and because enough players believe the alternative is worse.
Nvidia and Broadcom are unlikely to lose sleep over a specification document. They will pay attention when Chinese hyperscalers start ordering near-package optics modules by the tens of thousands, and when Huawei's interconnect fabric begins handling production workloads at latencies and bandwidths that rival NVLink. Until then, this remains a statement of intent - an important one, but not yet a competitive threat.
