The Qualcomm Xiaomi Partnership Renewed the Same Week Xiaomi Shipped Its Own Chip

Read the May 20 press release closely and the marketing gloss burns off fast. Volume increasing each year. Multiple generations. Xiaomi among the first to adopt the next Snapdragon 8-series. To a normal reader that is boilerplate. To anyone who has seen a wafer allocation agreement, it is capacity planning and TSMC reservation schedules bleeding into a public statement. Qualcomm and Xiaomi just renewed a relationship that already runs 15 years deep, and they did it the same week Xiaomi was telling the world it had built its own chip. The Qualcomm Xiaomi partnership renewal got filed everywhere as routine. It is not.

Fifty billion yuan over a decade, roughly $6.9 billion, with 13.5 billion already burned by April: the XRing O1 is not a render or a roadmap slide. It is a production-intent SoC, and Xiaomi is committing to it while simultaneously locking Snapdragon across phones, cars, glasses, watches, tablets, and smart home gear, and pouring money into the one chip that goes straight at Snapdragon in the segment where margins actually live, the flagship handset. Both tracks, on purpose. The engineering constraints on each side are the whole explanation.

Fabbed on TSMC’s N3E, second-gen 3nm, the O1 is Xiaomi’s first in-house application processor built for the flagship tier, and the first wave of coverage got one thing wrong about it: the CPU is not custom. It is a licensed ARM Cortex-X925 cluster, the same prime core MediaTek leans on, with an on-die NPU and Xiaomi’s own ISP for the computational-photography pipeline every Chinese flagship lives or dies on. Competent silicon, until you reach connectivity. The O1 integrates no modem at all; it leans on an external MediaTek T800 for 5G, and that single gap is the whole story.

Qualcomm’s Snapdragon X80, the modem riding inside every 2025 flagship, is what people underestimate when they squint at the O1: 15-plus years of accumulated RF IP showing up as sub-6GHz and mmWave 5G, a 10-gigabit theoretical downlink when carrier aggregation cooperates, AI-assisted link adaptation, Wi-Fi 7, standards-based 5G NTN satellite messaging. You do not solve that with a sharp team and a fast tapeout. It is an RF IP portfolio, and underneath it sits carrier certification across 200-plus bands and years of field-validation data that nobody compresses. Apple already ran this experiment for everyone to watch: a decade and billions spent, the C1 finally shipping in the iPhone 16e earlier this year, and even that one sub-6GHz only with mmWave punted down the road. A company with Apple’s resources lands a constrained part after ten years, so Xiaomi’s modem clock runs in the same brutal units. The O1 can sit right there on the bench and the deal still holds.

Two tracks, split by segment, the Samsung-and-Exynos move: the O1 drops into mid-range and upper-mid devices where the modem bar falls and cost-shaving is the whole game, while Snapdragon stays bolted into the premium flagships where benchmark scores and camera ISP and NPU horsepower and global certification are all non-negotiable. That volume increasing each year line tells you Xiaomi’s own forecasts do not show the O1 eating the premium tier inside the life of this agreement. Du coup, a sober read of the modem gap, written into a contract.

Two Oryon Prime cores and six Oryon Performance cores, no efficiency cluster at all: the Snapdragon 8 Elite that anchors this whole arrangement landed in October 2024 with Qualcomm finally walking away from the old three-tier layout. Custom ARM v9 instead of off-the-shelf Cortex, a deliberate bet on sustained performance over idle microwatts, roughly 45% single-core uplift over the 8 Gen 3. The Adreno 830 does hardware ray tracing in real silicon now, not the software emulation everyone hand-waved about two years back.

Around 45 TOPS, INT4 through BF16, a VLIW execution model with dedicated tensor units: the Hexagon NPU is the part that actually decides this thing, and Qualcomm has shown Llama 3 8B running at real-time token rates straight on it. Paired with the X80, the whole package sits on TSMC N3E. What Xiaomi really wants is the next one. Qualcomm pulled its Summit up to September 23-25 this year from the usual October slot, so the next premium 8-series gets a September reveal with devices in Q4 and early 2026, almost certainly a denser 3nm node and next-gen Oryon cores, with the Hexagon climbing past 45 TOPS. Launch-partner status is not a vanity badge. It means Xiaomi ships the newest silicon before rivals, and in the Chinese premium market that drives ASP through benchmark wins and the camera-and-AI features nobody else has shipped.

Xiaomi Auto put the SU7 on the road in March 2024, cockpit and connectivity riding Qualcomm’s Snapdragon Digital Chassis, and the renewal carries that automotive thread forward instead of treating it as filler. The Ride side runs ADAS and autonomous compute on the Ride Flex SoC, a heterogeneous CPU-GPU-DSP block with dedicated ASIL-D safety-island processors, the top rung under ISO 26262, fusing the full sensor suite from camera through LiDAR and scaling from L2+ to L4. The piece I would watch harder than most is BMW’s Neue Klasse, an ongoing collaboration with Qualcomm’s Ride platform in the ADAS stack for BMW’s next-gen EV architecture. Europe’s premium segment is exactly where Qualcomm has been weak against its grip on Chinese OEMs, so a BMW win lands heavier than the spec sheet reads. The automotive design-win pipeline has crossed roughly $45 billion in announced wins, contracted revenue spread across programs that build for the better part of a decade.

The XR2+ Gen 2 is the quietest line in the contract, and the one that hands Xiaomi a finished AR stack it would take years to clone: roughly 8x the AI compute and 2.5x the GPU of the first XR2, eye and hand tracking, sub-20ms motion-to-photon, 4.3K per eye at 90Hz. The Meta Quest 3 runs the non-plus version. Xiaomi’s AR/VR line is tiny next to phones, but it is a growth vector where the silicon-and-software handoff is the entire value.

Hexagon’s real moat is the software stack underneath it, the thing nobody benchmarks the O1 against. On-device inference is where Qualcomm’s edge actually lives, and the NPU is half of it at most. The AI Stack runs application down to silicon: the AI Hub handles optimization and deployment, AI Engine Direct gives low-level graph access, the Neural Processing SDK runs the framework-agnostic runtime, and at the bottom Hexagon, Adreno and the Oryon CPU split the work. Tooling maturity compounds quietly while everyone argues about TOPS numbers, and that is the pattern I keep returning to whenever Qualcomm rolls out new compute.

Developers actually touch the AI Hub: you hand it a PyTorch or ONNX model and get back a device-optimized build, quantized and rebuilt for Hexagon, profiled across targets. Over 100 pre-optimized models sit in the catalog as of mid-year, and that ecosystem is what rivals keep underestimating. It is why the O1’s raw NPU figure matters less than it looks. Nobody has put that NPU up against Hexagon on generative workloads in public, and until it proves out on real inference rather than the easier CV tasks, Snapdragon keeps the premium tier on merit. The quantization math is unforgiving and it runs in Qualcomm’s favor: INT4 drops Llama 3 8B from about 16GB at FP16 to roughly 4GB, which fits a 12GB LPDDR5X budget and pushes north of 20 tokens a second mostly on the NPU, at a cost of maybe 1 to 3% degradation on MMLU and HellaSwag against the FP16 baseline. A tolerable tax for a real LLM inside a phone. Whatever Xiaomi markets as an AI feature, the camera pipeline most of all, rides on that NPU throughput, and Qualcomm has two years of production data on the workload that Xiaomi is still gathering.

Apple is why Qualcomm signed when it did. Apple’s modem revenue has been Qualcomm’s load-bearing wall and its loudest bear case for years, and the C1 in the iPhone 16e is the first crack in it. The handoff runs slow but one direction: one SKU this year, then more iPhone models across 2026 and 2027, then a projected full in-house transition around 2028, with Qualcomm already guiding for the Apple modem line to shrink. Against that, Xiaomi moved around 170 million phones in 2024, third behind Samsung and Apple, maybe a fifth to a third of it in the premium tier Snapdragon owns. A contractual growth commitment on that tier swaps part of the Apple annuity for a Xiaomi one. Automotive does something different in kind: a BMW-type win converts to revenue three to five years out and then pays for seven to ten years of production, a backlog that builds slowly and predictably, the floor a handset-cyclical business never used to have.

The Dimensity 9400 is the only Android premium part genuinely pushing, announced late last year on the same TSMC N3E as the 8 Elite, a licensed Cortex-X925 cluster with an Immortalis-G925 GPU. It trades real blows on CPU and it is a good chip. Where it falls short is not one TOPS figure, it is modem depth, mmWave and the heavier carrier-aggregation modes especially, plus the maturity of the tooling around Hexagon. MediaTek has taken real share in the Chinese mid-range without cracking the ultra-premium tier, and this contract bolts that door shut regardless, because Snapdragon is now locked into Xiaomi’s flagships no matter how good the next Dimensity gets.

Nobody on the marketing side mentions the foundry knot. The 8 Elite and the XRing O1 are fighting for the same TSMC N3-family wafers, Xiaomi’s homegrown program and its Qualcomm supply queued at the same fab behind Apple, AMD, MediaTek, and half the industry. N3 is constrained and booked solid. Qualcomm’s scale as a fabless buyer buys it leverage with TSMC that Xiaomi’s young silicon team cannot touch, and the multi-year volume commitment quietly obliges Qualcomm to keep enough N3 and N2 capacity reserved to deliver on it.

The H20 made the geopolitics impossible to ignore this week. Washington effectively banned Nvidia’s China-compliant accelerator in April, the one engineered to slip under the export thresholds, by demanding a license nobody expected to clear. This week the same chip got waved back through, Nvidia citing U.S. assurances the licenses will come and shipments will resume soon. A ban in April, a reversal by mid-July, nothing between them but lobbying and politics. Qualcomm and Xiaomi signed on May 20 inside exactly that volatility, Qualcomm’s heavy China exposure and its real slice of QCT revenue fully in view, and they went ahead anyway. Handset SoCs do not trip the compute thresholds that catch Nvidia’s accelerators; they do not clear the bar that triggers review. For now. As Hexagon climbs toward and maybe past 100 TOPS down the line, that line moves, and someone at BIS is already drawing it in pencil.

The moat is really one thing with smaller things stacked behind it. The modem first, 15-plus years and global certification an OEM cannot reproduce in a three-to-five-year sprint. Then the Hexagon-and-AI-Hub stack, where rivals keep closing the hardware gap but never the tooling. The ASIL-D Ride silicon needs validation years deep, and XR2+ Gen 2 has no real performance-per-watt rival in standalone headsets. The soft spots are easier to find. Oryon is strong but not untouchable, with the Dimensity 9400 right there on CPU and GPU and Apple Silicon still ahead outright. MediaTek owns real mid-range share against the 7-series. And the Snapdragon X Elite PC push stays an ugly fight against entrenched x86 for a good while yet.

I am a declared Snapdragon Insider, so weigh my read with that bias on the table. Even discounting for it, the structural logic is hard to argue with. Qualcomm is not pretending Xiaomi will never build a good chip. It cannot stop that and it is not trying to. What it has done is set a contractual floor under Xiaomi revenue that holds through the exact window where the XRing line is too green to threaten the flagship tier. By the time an O2 or O3 grows a real modem, Qualcomm will have shipped two more 8-series generations and dug its automotive and XR revenue into ground that does not care what silicon Xiaomi draws up in-house.

July 30 brings Q3 FY2025 earnings, the first print that shows how the market read all this. Two things I will watch: whether the QCT handset guidance carries Xiaomi fingerprints, and how carefully they word the sentence about Apple modem dollars sliding away. The automotive number past $45 billion is a formality at this point.

What actually nags at me is dumber than any of the strategy. It is the wafer accounting. The O1 and the Snapdragon orders sit in one TSMC queue behind Apple, both funded by Xiaomi, one of them built to make the other unnecessary someday. Qualcomm bought years with this deal. Years are not a win and nobody in the room pretends otherwise, which is why the thing runs exactly as long as it had to and not a month more.

Related reading: the flagships this partnership keeps shipping, from the Xiaomi 15 Ultra’s global 5G-A and AI-on-everything launch back to the Leica-tuned Xiaomi 14 Ultra a year before it.