MediaTek and the Fragmented Compute War.Original analysisNot investment advice
In 2020, the MediaTek thesis was Huawei substitution. In 2026, the story is bigger. Compute is fragmenting across phones, edge devices, cars, hyperscalers, and national supply chains. MediaTek sits in the middle.
In 2020, the MediaTek thesis was simple. Huawei was being cut off from advanced chip manufacturing. US export controls threatened Huawei’s Kirin smartphone chips, 5G chips, and broader silicon roadmap. The obvious conclusion was that Huawei and Chinese Android OEMs would need alternative SoCs, and MediaTek stood to benefit.
That thesis was directionally right. The 2026 version is more interesting.
Huawei did not disappear. SMIC and Huawei shipped 7nm-class Kirin chips inside the Mate 60 Pro.23 But Huawei also did not fully escape the constraints, and the follow-on Kirin 9020 in the Mate 70 looked incremental rather than a major redesign.4 Qualcomm remains the premium Android incumbent. Apple has its own silicon. Xiaomi is experimenting with in-house chips. Hyperscalers want custom AI ASICs. Automakers want AI cockpits. Edge devices need on-device inference. TSMC’s leading-edge capacity is becoming more strategic every year.
The story is no longer “Huawei loses, MediaTek wins.” The story is fragmentation. MediaTek sits in the middle of it.
The correct 2026 claim is not that MediaTek wins because Huawei got blocked. The correct claim is that MediaTek benefits from compute fragmentation, but only if it keeps moving up the value chain faster than its old smartphone business slows down.
I. The 2020 thesis
In May 2020, Dylan Patel published a SemiAnalysis piece arguing that MediaTek stood to gain the most from the new US-China cold war.1 The argument was tidy. Huawei’s Kirin roadmap was under pressure from export controls. Chinese Android OEMs needed an alternative non-US SoC supplier. MediaTek’s Dimensity 5G line lined up well with sub-6 GHz 5G, which mattered more globally than mmWave. The framing was largely a smartphone SoC story.
I revisited that piece because it has aged in a way most cold war predictions did not. The direction was correct. The shape of the win turned out to be different. The 2020 framing read MediaTek as a phone-chip substitution play. The 2026 reality is a wider story about who builds compute when every region, OEM, and cloud wants more control over silicon.
If Huawei loses access to advanced self-designed chips, Chinese Android OEMs will need alternative SoC suppliers. MediaTek could benefit by moving up from lower-end Android phones into the mid-to-premium 5G market.
The piece was conservative in 2020. It assumed the export controls would hold, the Huawei recovery would be limited, and the smartphone SoC market would be the centre of the action. Two of those held. The third moved.
II. What changed by 2026
Six years later, three facts complicate the simple substitution story.
First, Huawei is not dead. TechInsights teardowns confirmed that SMIC manufactured the Kirin 9000S in the Mate 60 Pro on its 7nm-class N+2 technology, a made-in-China milestone that Reuters reported as a real breakthrough.23 Second, Huawei is not unconstrained. TechInsights described the Mate 70 Kirin 9020 as incremental rather than a major redesign, consistent with SMIC remaining behind the leading edge.4 Third, Taiwan added Huawei and SMIC to its export control list, requiring Taiwanese firms to obtain government permits before exporting to either company.5 The simple binary of 2020 has become a fractured market.
That is the new strategic environment MediaTek operates in. Not a Huawei collapse. Not a Huawei comeback. A constrained Huawei, a fragmented Chinese OEM base, a more strategic TSMC, and a global silicon roadmap pulled in many directions at once.
Huawei blocked → MediaTek opportunity
- Huawei Kirin · under export control pressure
- Chinese OEMs · need a non-US SoC supplier
- MediaTek Dimensity · the substitution candidate
Many silicon roads, no single winner
- Huawei · constrained domestic silicon with SMIC
- Qualcomm · premium Android incumbent
- Apple · internal silicon, vertical model
- MediaTek · TSMC-linked neutral platform
- Hyperscalers · custom ASIC demand
- Nvidia · AI ecosystem and partnerships
- Chinese OEMs · multi-vendor approach
- Edge AI & auto · new compute footprint
III. MediaTek is no longer just cheap Android
The old MediaTek brand was low-cost Android SoCs. That brand was useful, but it understates what the company is now trying to be. Dimensity flagships are competing in premium Android. The TSMC N2P tape-out signals leading-edge intent rather than mid-tier survival.7 Smart Edge Platforms is climbing toward half of revenue. AI accelerator ASIC projects for US hyperscale customers are on the public record.611 A Grace Blackwell Superchip was co-designed with Nvidia for the DGX Spark workstation.910
None of those pieces fits a story of MediaTek as a discount Android shop.
MediaTek’s story is not that smartphones are booming. It is that silicon content and silicon complexity are rising across more device categories.
IV. The TSMC access advantage
The most important structural difference between Huawei and MediaTek is not design talent. It is manufacturing access.
Huawei has to work around export controls and SMIC’s constrained roadmap. MediaTek can ride TSMC’s leading-edge cadence. The company has announced a flagship SoC tape-out on TSMC’s enhanced N2P process, with the first chipset expected in late 2026. MediaTek describes N2P as offering up to 18% performance improvement at the same power, around 36% power reduction at the same speed, and roughly 1.2x logic density versus N3E.7 TSMC’s own 2025 annual report places N2P and A16 volume production in the second half of 2026 and the next A14 node in 2028.8
That is a wide gap. Not because TSMC is the only foundry, but because no other foundry can offer the same combination of process maturity, packaging integration, and AI/HPC customer base at the leading edge today. MediaTek being on TSMC is not a small detail. It is a strategic position.
V. Smartphones are mature, but premium silicon still matters
This is not a smartphone boom story. Smartphones are mature and cyclical. MediaTek’s own Q1 2026 numbers showed Mobile Phone revenue down 17% QoQ and 15% YoY, with Mobile Phone at 49% of revenue.6 Counterpoint described 2025 as a year of only modest global smartphone shipment growth, with 2026 facing softness as components are pulled toward AI data center prioritisation.12
If MediaTek were still a unit-growth story, those numbers would be the headline. They are not, because the value per chip is rising. Premium silicon wins, higher ASPs, AI on-device features, better modems, and content-per-device gains start to matter more than handset shipment counts.
The smartphone market is not exploding. The value per useful chip is.
VI. Smart Edge is becoming nearly as important as phones
The segment that quietly recasts MediaTek as something other than a phone-chip company is Smart Edge Platforms. In Q1 2026 it grew 23% sequentially and 13% year over year, reaching 46% of total revenue.6 The drivers MediaTek calls out span connectivity, computing, automotive, and broader share gains, including Wi-Fi, broadband, TVs, automotive cockpit silicon, IoT, edge AI, and local inference.
You can see the identity shift in the revenue mix. Phones still pay the bills, but they no longer define the company.
Phones first
Edge to cloud platform
VII. The real surprise is AI ASICs
The most interesting line in MediaTek’s Q1 2026 commentary was not about phones or edge devices. It was the public framing of the AI accelerator ASIC program.
MediaTek said its first AI accelerator ASIC project for a US hyperscale customer was on schedule, with around US$2 billion revenue contribution expected in Q4 2026 and the project scaling to multiple billions in 2027. A second AI accelerator ASIC project is underway with mass production targeted by the end of 2027. The company also framed the cloud ASIC TAM at US$70–80 billion in 2027 and reiterated a 10 to 15 percent share target.611
This is not normal smartphone silicon. It is a serious bid to become a custom silicon execution partner for hyperscalers that want their own AI chips, but do not want to build a frontier design and supply-chain organisation from scratch.
MediaTek is not trying to beat Nvidia head-on. It is trying to become the design and integration partner behind customers who want their own AI silicon.
VIII. Why MediaTek has the right skill stack
AI accelerator ASICs need a specific cluster of capabilities. Power-efficient design. High-speed interfaces. Memory subsystem design. Die-to-die interconnect. HBM integration. SerDes. Voltage regulation. Packaging awareness. Customer-specific integration. TSMC execution discipline.
MediaTek talks about the AI ASIC effort in those exact terms. The company has discussed work on 400G SerDes, 64G die-to-die interconnect, an advanced 3.5D platform, custom HBM solutions, and integrated voltage regulators for future multi-kilowatt XPUs.6 The Nvidia GB10 Grace Blackwell Superchip, which powers the DGX Spark workstation, was co-designed with MediaTek. MediaTek contributed CPU, memory subsystem, and high-speed interface design work. GB10 pairs a Grace 20-core Arm CPU, a Blackwell GPU, and 128 GB unified memory, delivering up to 1 PFLOP of AI performance in a desktop form factor.910
That is the through-line. Phones taught MediaTek low-power design. Modems taught connectivity. Consumer SoCs taught system integration. High-volume devices taught execution discipline. AI ASICs need all of those skills, but scaled up to multi-kilowatt and hyperscale-grade reliability.
IX. The geopolitical middle position
MediaTek occupies a strange but valuable place in the geopolitics of compute. Taiwan-based. TSMC-linked. Not American. Not Chinese. Android-OEM-friendly. Globally integrated. Dependent on US-aligned semiconductor tools and TSMC, and exposed to Taiwan-related risk.
That position has been re-priced by the export control environment. Taiwan’s own export control list now includes Huawei and SMIC, with Taiwanese firms required to obtain permits before exporting to either.5 US controls continue to limit advanced semiconductor access to China at large. Huawei and SMIC remain on the constrained side of that line. TSMC continues to define what leading-edge looks like.
The strategic interpretation is uncomfortable but honest. MediaTek benefits from being neutral enough for many OEMs around the world, but its neutrality is not independence.
MediaTek is neutral, but not detached. Its neutrality works because it is plugged into TSMC, not because it is outside geopolitics.
X. Value chain map
This is a value-chain map, not investment advice. The point is to show the shape of the dependency. Compute fragmentation is a multi-vendor environment, and MediaTek’s value capture depends on how many of these adjacent boxes it can either serve, partner with, or stay neutral to.
Quick terms
- SoC
- System-on-chip. A single chip integrating CPU, GPU, memory controllers, modem, and other blocks.
- Fabless
- Chip company that designs chips but does not manufacture them.
- Foundry
- Company that manufactures chips on behalf of fabless designers.
- ASIC
- Application-specific integrated circuit. A chip designed for a particular workload.
- AI ASIC
- Custom chip designed for AI workloads, typically inference or training acceleration.
- Edge AI
- AI computation on devices near the user rather than only in cloud data centers.
- NPU
- Neural processing unit. An on-chip accelerator for neural network operations.
- SerDes
- Serializer/deserializer. The high-speed I/O block that moves data between chips.
- Die-to-die
- Communication between chiplets inside the same package.
- HBM
- High-bandwidth memory. Stacked DRAM dies sitting next to an accelerator on an advanced package.
- TSMC N2P
- An enhanced 2nm process platform from TSMC, scheduled for second-half 2026 volume production.
- Export controls
- Government rules restricting technology exports to specified entities or countries.
- Hyperscaler
- A large cloud or data-center operator, e.g. Microsoft, Google, Amazon, Meta, Oracle.
XI. What could break the thesis
A serious piece needs counterarguments. The case for MediaTek as a beneficiary of compute fragmentation has more than one honest failure mode.
- Huawei improves faster than expected. A faster recovery on the Kirin roadmap, even on SMIC’s constrained technology, would shrink the Chinese OEM substitution opportunity.
- Qualcomm defends premium Android. The top of the Android SoC market is where ASPs and margins live. Qualcomm is not a passive incumbent.
- OEMs build in-house silicon. Xiaomi and others moving more design in-house compresses MediaTek’s phone TAM at the top, even if unit volumes hold.
- Smartphone demand stays soft. A second weak year of handset shipments amplifies the Q1 2026 mobile decline and stresses the volume base.
- AI ASIC concentration risk. Multi-billion-dollar ASIC programs concentrated on a small number of hyperscale customers carry execution and customer-power risk.
- ASIC revenue is lumpy. Project-based silicon revenue does not arrive smoothly. Quarter-to-quarter volatility can swamp the operational story.
- TSMC capacity and pricing. Leading-edge wafer cost is rising. MediaTek’s gross margins on flagship and custom silicon must defend against that.
- Taiwan geopolitical risk. Any escalation around Taiwan would re-price every Taiwan-based fabless company, MediaTek included.
- China exposure under restriction. Tighter US-China rules could complicate sales to certain Chinese customers or platforms.
- Margins in custom silicon. Hyperscalers negotiate hard. Custom silicon revenue can be large at lower margin than mainstream merchant SoCs.
- Nvidia and hyperscalers keep more in-house. If they pull more integration work back in-house, MediaTek’s partner role narrows.
- Edge AI monetisation is slow. Smart Edge growth is real, but on-device AI monetisation has been slower than the hype curve suggested.
The correct claim is not that MediaTek wins the new cold war. The correct claim is that MediaTek benefits from compute fragmentation if it keeps moving up the value chain.
XII. The fragmented compute war
In 2020, the question was: who supplies Huawei if Kirin gets blocked? In 2026, the question is bigger: who supplies compute in a world where every region, device maker, cloud provider, and AI platform wants more control over silicon?
That is where MediaTek becomes interesting. It can sell flagship SoCs to Android OEMs that want an alternative to Qualcomm. It can supply connectivity and compute into edge devices. It can support automotive AI cockpits. It can use TSMC’s 2nm roadmap. It can co-design with Nvidia at the high end and build ASICs for US hyperscalers at the system level. It can benefit from China’s reluctance to depend fully on US suppliers, while still operating inside the TSMC-centred advanced manufacturing ecosystem.
That is the 2026 MediaTek story. Not cheap Android chips. Not just Huawei substitution. A neutral fabless silicon platform for a fragmented compute world.
1 Patel, D. (May 2020). Mediatek Stands to Gain the Most from the New Cold War. SemiAnalysis. Historical anchor for the substitution thesis. Used as inspiration only. No content, structure, or charts reproduced.
2 TechInsights (2023). TechInsights finds SMIC 7nm (N+2) in Huawei Mate 60 Pro. Identification of the Kirin 9000S on SMIC’s 7nm-class N+2 process inside the Mate 60 Pro, framed as a made-in-China manufacturing milestone, more advanced than SMIC 14nm but not at 5nm-class equivalence.
3 Reuters (Sep 4 2023). Teardown of Huawei’s new phone shows China’s chip breakthrough. Reuters validation of the TechInsights teardown and broader context on Huawei and SMIC at 7nm.
4 Reuters (Dec 11 2024). Huawei’s new Mate 70 phone chip shows no major redesign, TechInsights says. Coverage of the Kirin 9020 as incremental rather than a major redesign, consistent with continued SMIC 7nm-class manufacturing.
5 Associated Press (2025). Taiwan adds Huawei and SMIC to export controls list. Taiwan requiring its firms to obtain government permits before exporting to either company, with geopolitical context.
6 MediaTek (Q1 2026). Q1 2026 earnings transcript. Mobile Phone revenue down 17% QoQ and 15% YoY, Mobile Phone 49% of revenue, Smart Edge Platforms 46% of revenue with 23% QoQ and 13% YoY growth, first AI accelerator ASIC project tracking to around US$2 billion in Q4 2026 and scaling to multiple billions in 2027, second AI accelerator ASIC project targeting mass production by end of 2027, and discussion of 400G SerDes, 64G die-to-die, 3.5D platform, custom HBM, integrated voltage regulators, cloud ASIC TAM of US$70–80B in 2027 with a 10 to 15 percent share target.
7 MediaTek (2025). MediaTek develops chip utilizing TSMC’s 2nm process. Flagship SoC tape-out on TSMC N2P, first chipset expected late 2026, up to 18% performance improvement at same power, around 36% power reduction at same speed, and 1.2x logic density versus N3E.
8 TSMC (2025). 2025 Annual Report. N2P and A16 volume production scheduled for second half of 2026 and A14 for 2028, with AI/HPC and mobile roadmap context and TSMC’s advanced process leadership.
9 MediaTek (2025). NVIDIA DGX Spark features GB10 Superchip co-designed by MediaTek. GB10 Grace Blackwell Superchip with MediaTek-led CPU, memory subsystem, and high-speed interface design, paired with Nvidia’s Blackwell GPU.
10 Nvidia. DGX Spark product page. DGX Spark system with 20-core Grace Arm CPU, Blackwell GPU, 128 GB unified memory, and up to 1 PFLOP of AI performance for developer workstation use.
11 Reuters (Oct 31 2025). Taiwan’s MediaTek targets billions in AI accelerator chip revenue by 2027. Independent confirmation of MediaTek targeting billions of dollars in AI accelerator revenue by 2027 and broader data-center ASIC TAM context.
12 Reuters (Jan 12 2026). Apple leads global smartphone market with 20% share in 2025, says Counterpoint. 2025 smartphone market context, modest shipment growth, and 2026 softness driven by component pressure linked to AI data center prioritisation.
- The AI Memory Wall. Companion essay on DRAM, HBM, packaging, and semicap as the new center of computing.
- The Dry Resist War. Companion essay on why Lam’s Aether could become critical to AI-era chipmaking.
- MediaTek Investor Relations. Quarterly transcripts, segment commentary, and Smart Edge/AI ASIC framing.
- TSMC Annual Reports. Process roadmap, packaging, and AI/HPC positioning.
- TechInsights. Independent teardown analyses of Kirin and other advanced silicon.
- The AI Field Manual. Reference layer for the AI stack: hardware, memory, models, agents, safety, economics.
This is Essay No. 014. The topics: intelligence, AI, systems, knowledge, and the questions underneath the questions everyone else is asking. If you read this far and disagreed with any part of it, write to me. I read everything.