Huawei, Qualcomm strike multi-year patent agreement across 5G, AI

Huawei and Qualcomm Just Struck a Multi-Year Patent Deal. The AI Layer Is the Part Worth Reading.

TL;DR

Huawei and Qualcomm have reached a multi-year global patent licensing agreement spanning 5G and AI-related intellectual property — and it moves through a completely different legal channel than the US export controls that have been cutting Huawei off from advanced chips since 2019. That distinction is not a footnote. It shapes how money, technology legitimacy, and strategic leverage actually flow between the two companies. Whether the AI IP component of this deal meaningfully advances Huawei's hardware ambitions, or is mostly a floor-clearing exercise for standards compliance, remains genuinely open.

Key Takeaways

• Qualcomm confirmed a multi-year global patent licensing agreement with Huawei covering 5G and AI-related intellectual property, spanning both companies' core technology portfolios
• China has consistently represented approximately 60–65% of Qualcomm's total revenue in recent fiscal years, according to Qualcomm's SEC filings, making Huawei-adjacent licensing relationships among the most financially significant the company manages
• Huawei ranks among the top three holders of declared 5G standard-essential patent families globally, according to annual tracking by the European Patent Office and IPlytics, putting it in a position to both pay and collect royalties under FRAND obligations
• Qualcomm's licensing division (QTL) operates independently of its chip business (QCT) — meaning royalty flows from Huawei can continue even when hardware sales cannot, a structural separation that matters a great deal to both companies' negotiating positions
• US export restrictions on advanced AI chips have tightened in successive rounds since October 2023, but those rules govern hardware and software exports — patent licensing under FRAND terms operates under separate legal authority
• Huawei's Ascend 910B and 910C AI chips have been deployed by Chinese cloud providers including Alibaba Cloud and Baidu for AI training workloads, positioning them as the primary domestic alternative to restricted Nvidia H-series GPUs

What the Deal Actually Is — and What It Isn't

Let me start with the number that grounds this: Qualcomm's China-related revenue has hovered around 60–65% of the company's total for several years running. That is not a rounding error. It is the single most important fact about why this agreement exists, and why Qualcomm's legal team has maintained continuous engagement with Huawei even as the geopolitical temperature around both companies climbed.

The deal is a patent license. Qualcomm holds one of the largest portfolios of standard-essential patents (SEPs) for 5G — the foundational technology that any smartphone, base station, or connected device needs to use if it wants to participate in the standard. Under FRAND (fair, reasonable, and non-discriminatory) terms, those patents must be licensed to anyone who asks. Refusing to license a standard-essential patent is a separate legal problem from licensing it to a company on a US restricted list. This is not a loophole. It is how the standards system was designed to work, and it has produced patent revenue flows between companies that otherwise cannot do business in hardware.

Huawei is not simply a licensee here. The company holds a substantial 5G SEP portfolio of its own — among the top three globally by declared patent families, according to tracking by IPlytics and the European Patent Office. What the two sides are almost certainly doing is what they have done in prior agreements: cross-licensing, with a net royalty payment running in one direction depending on whose portfolio covers more ground. The specific financial terms of this agreement have not been disclosed.

The AI clause is newer. Both Qualcomm and Huawei have been filing patents in AI-adjacent areas — on-device inference, neural processing architectures, edge AI compression — and a multi-year agreement that covers "AI" likely sweeps in a category of patents that did not exist in meaningful numbers five years ago. What exactly falls under that umbrella, and whether it covers anything Huawei's Ascend team actually needs, is unverified.

The Hardware Wall and the IP Gate — Why Both Exist Simultaneously

Here is the thing that confuses a lot of Western readers: how can Qualcomm license patents to Huawei if Huawei is on the US Entity List?

The US Bureau of Industry and Security's Entity List restricts the *export* of controlled goods and technology to listed companies. It governs semiconductors, chip manufacturing equipment, software tools, and certain technical knowledge transfers. Patent licensing is administered separately — it falls primarily under FRAND obligations set by standards bodies, with its own legal architecture and its own carve-outs. A US company collecting royalties on a standard-essential patent from a Chinese company is not the same legal act as shipping that Chinese company a wafer fabrication tool.

This dual-track reality is not unique to the US-China dynamic. Nokia, Ericsson, and Interdigital all license 5G patents to Chinese OEMs under the same FRAND framework. The revenues from those agreements show up in European earnings reports every quarter. The story is not that this particular agreement broke through an export wall. The story is that the IP layer of the technology stack has always operated on different rules than the hardware layer — and the AI industry is about to make that distinction much more complicated.

The practical difference: Huawei can access certain AI-related IP through licensing agreements like this one. It cannot, by contrast, access TSMC's advanced nodes, Nvidia's H100s, or ASML's EUV machines. The result is a company that can claim legitimate access to certain technology families while being structurally cut off from the manufacturing capability to exploit them at leading-edge process nodes.

Huawei's AI Position Right Now

The most consequential thing happening in Huawei's AI stack is not this patent deal. It is the Ascend line — specifically the Ascend 910B and the more recent 910C, which Huawei has positioned as the primary training accelerator for Chinese AI workloads that can no longer source Nvidia H800 or H100 chips following the October 2023 and subsequent US export control expansions.

For a detailed look at how those chips have fared against Nvidia in Chinese data centers, the analysis of Huawei's Ascend chip market gains covers the deployment picture across cloud providers and enterprise customers. The short version: Ascend has captured meaningful share, primarily because Chinese hyperscalers have had no alternative for new capacity, not because it has matched Nvidia on performance-per-watt in all workloads.

Huawei's Pangu large language model family sits on top of that hardware. The company has released versions oriented at enterprise deployments — mining, finance, meteorology — rather than consumer chatbots, which is a deliberate positioning decision that sidesteps the consumer regulatory environment while maintaining enterprise credibility. DeepSeek, Alibaba's Qwen series, and ByteDance's Doubao compete in adjacent spaces; Pangu is targeting the industrial vertical rather than the general-purpose frontier race.

What Qualcomm IP potentially contributes to this picture is genuinely unclear. Qualcomm's AI patents tend to concentrate in mobile inference and edge AI — areas relevant to Huawei's device business (Mate series, tablets, wearables) more than its data center roadmap. That said, patents filed under "AI" today cover enough architectural ground that a blanket statement about what is or isn't covered would be overconfident. The companies have not specified.

What This Changes for Western Founders and Consultants

It confirms that the bifurcation of the tech stack is not uniform. Hardware export controls and IP licensing operate on different timelines, different legal regimes, and different strategic logics. If you are advising clients on China tech exposure, the question "are we licensing IP to a Chinese company?" and the question "are we selling them chips?" require entirely different compliance analyses.

Qualcomm's ability to maintain this relationship matters for the broader patent ecosystem. If the US government were to restrict patent licensing to Entity List companies — which has been discussed but not implemented — it would create a significant FRAND compliance crisis for the entire global standards system. Western companies holding 5G and emerging AI SEPs would face impossible choices between US export law compliance and international treaty obligations. Watch for this in policy discussions over the next 18 months.

For founders building on Qualcomm's Snapdragon AI stack: this deal has no direct effect on your product roadmap. But it signals that Qualcomm is actively managing its China business relationships in ways that preserve optionality. A company collecting royalties from Huawei has different incentives in Washington than a company that has written off that relationship entirely.

For consultants tracking Chinese AI tooling: the Ascend ecosystem is real and growing. Alibaba's Qwen models now run natively on Ascend infrastructure; Baidu's ERNIE deployment has expanded on Ascend clusters. The question is not whether Huawei's AI stack is functional — it is whether it can close the performance gap with leading Nvidia hardware at scale, and whether the software ecosystem (CANN, MindSpore) develops the depth that CUDA has accumulated over fifteen years. That gap is larger than the hardware gap and less discussed.

A Comparison: Where Huawei Sits in China's AI Hardware-Software Stack

Checklist: How to Assess Your China AI IP Exposure Before It Becomes a Compliance Problem

• Audit what you license vs. what you sell. Patent licenses to Chinese entities operate under different rules than hardware exports. Know which category your China relationships fall into — and which might shift.
• Map your SEP obligations. If your product uses 5G, Wi-Fi 6/7, or Bluetooth, you may owe royalties to Huawei regardless of your preference. This is not optional; it is FRAND.
• Don't assume AI patents are the same as AI chips. Licensing an algorithm patent to a Chinese entity is a different compliance question than selling that entity a chip that runs the algorithm. Do not conflate them in your legal briefs.
• Watch for FRAND-vs-export-law friction in policy. Proposals to restrict patent licensing to Entity List companies are real. If implemented, they would create conflicts with international treaty obligations. Track the Bureau of Industry and Security rulemaking calendar.
• If you're evaluating Ascend-based deployments: benchmark your actual workload, not Huawei's marketing claims. The 910B's performance on transformer-based training has been reported as competitive with the Nvidia A100 on certain tasks, but the software overhead from CANN vs. CUDA is non-trivial and workload-dependent. Neither Huawei's benchmarks nor dismissive Western comparisons are reliable.

Where This Is Heading

AI patents will become the next standards fight. 5G SEPs took a decade to sort out through international standards bodies. AI — particularly the foundational techniques behind large language models, hardware-software co-design, and neural network compression — is proceeding without that standardization layer. The companies accumulating the most AI patents now (Huawei, IBM, Samsung, Qualcomm, Baidu, and a cluster of US hyperscalers) will have significant leverage in whatever licensing regime eventually emerges. This deal is early positioning for that future.

Huawei's device business and its AI infrastructure business are on diverging trajectories. The Kirin chip revival (the Mate 60 Pro's 9000S, manufactured at SMIC's 7nm-class process) is a consumer-facing story about Shenzhen's engineering capability under constraint. The Ascend AI chip story is a data center infrastructure story, aimed at enterprise and sovereign AI buyers across China. These are different businesses with different competitive dynamics, and they get conflated in Western coverage constantly.

The software moat matters more than the hardware specs. Huawei can close hardware gaps faster than it can rebuild the equivalent of fifteen years of CUDA ecosystem development. MindSpore and CANN are functional; they are not drop-in replacements for the stack that every major AI research group in the world has built on. The question of whether Chinese AI infrastructure can develop genuine software ecosystem depth — not just functional parity but the community, the tooling, the model zoo, the debugging culture — is more consequential than any single chip announcement.

Qualcomm's China strategy will shape US tech policy debates. A company with 60-plus percent China revenue has lobbyist incentives that differ from a company that has fully written off the market. As the US legislative calendar fills with additional export control proposals and FRAND-related IP bills, Qualcomm's willingness to maintain agreements like this one will make it an active participant in those debates — on the side of preserving licensing flows.

The multi-year framing matters. Both sides agreed to a multi-year structure, which signals that each has an interest in predictability over a planning horizon. For Huawei, that means IP litigation risk reduction while it focuses engineering resources on hardware and software development. For Qualcomm, it means a revenue stream that survives further geopolitical turbulence. Neither party is betting on the current environment to improve.

FAQ

Why can Qualcomm license patents to Huawei if Huawei is on the US Entity List?

The Entity List restricts exports of controlled goods and technology under the Export Administration Regulations. Patent licensing on standard-essential patents is governed separately — primarily through FRAND obligations under international standards bodies and through commercial contract law. Refusing to license a FRAND-encumbered patent to anyone who asks is itself a legal problem. The US government has not, as of this writing, issued a general prohibition on patent licensing to Entity List companies, though the question has been debated in policy circles.

Does the "AI" portion of this deal mean Qualcomm is helping Huawei build better AI chips?

Almost certainly not in any direct sense. Patent licensing conveys the right to use patented methods — it is not a technology transfer agreement, a joint development agreement, or a chip supply arrangement. Qualcomm's AI patents mostly concentrate in mobile inference, edge processing, and on-device AI efficiency, which are relevant to Huawei's device business more than its Ascend data center line. Whether any specific Ascend design benefits from Qualcomm IP is not publicly known.

Is Huawei actually competitive in AI, or is this mostly domestic market capture?

Both. Inside China, the Ascend platform has captured real share among enterprises and cloud providers that cannot source Nvidia GPUs for new capacity. Outside China, Huawei's AI hardware has essentially no presence — export controls, reputational risk, and the CUDA ecosystem lock make international sales implausible in the near term. The Pangu LLM family is similarly positioned: it has genuine enterprise deployments in China across energy, mining, and finance verticals, but its international footprint is minimal. "Competitive" and "globally competitive" are very different claims here.

What does this mean for Qualcomm's relationship with Nvidia?

These are not directly competitive at the IP licensing level. Qualcomm's QTL division monetizes standards-based patents; Nvidia does not hold comparable SEP portfolios. At the chip level, Qualcomm's edge AI and on-device inference business could eventually intersect more directly with Nvidia's push toward agentic and edge inference, but the Huawei patent deal does not materially change that dynamic.

Should Western AI companies be worried about Huawei's AI patent portfolio?

If your product uses 5G or any technology covered by Huawei's SEP portfolio, you likely already owe royalties — or should be. That is established, not new. Huawei has been actively litigating and licensing its 5G patents in Europe and Asia. On AI-specific patents, the picture is less settled. Huawei files a large number of AI patents; which of those will prove enforceable, cover meaningful techniques, and survive challenge is genuinely unknown. Worth tracking, premature to panic over.

Is the Ascend 910C actually as good as an Nvidia H100?

Huawei has made competitive claims, and some independent assessments from Chinese AI research groups suggest the gap has narrowed for specific transformer-based training workloads. Benchmark comparisons remain limited and often come from sources with an interest in the outcome. The honest answer: the hardware gap has closed more than most Western analysts expected two years ago; the software and ecosystem gap remains large; neither side's published numbers should be taken at face value without workload-specific testing.

What's the single most important thing to watch from here?

Watch whether the US moves to restrict FRAND patent licensing to Entity List companies. If that rule changes, it creates cascading legal conflicts between US export law and international standards obligations that would affect every company holding AI or 5G SEPs — not just Qualcomm. That regulatory question is more consequential than any individual deal announcement, and it has received far less coverage than it deserves.