Xiaomi unveiled three custom chips on Monday, and the cleanest way to understand the launch is by where each one goes. XRING O3 goes into a phone. XRING O100 accelerates a local AI model. XRING D100 goes into a car.
Then Xiaomi put all three into one 150-watt desktop prototype called the AI Cube, running 120-billion- and 3-billion-parameter models locally. It is an unusually literal demonstration of the company's Human × Car × Home strategy: one silicon programme reaching across the entire ecosystem.
It is also a demonstration of what neither Xiaomi's reporting nor our Xiaomi model can separate. Xiaomi reports a single segment called Smart EV, AI and other new initiatives. Our model drives that whole line from 104,199 units at $35,118 each. Those units are vehicles. Monday's phone silicon, AI accelerator and prototype computer have no independent volume, price or revenue driver.
The summary card above lays out the mismatch: three products, three end markets, one reported segment, and one vehicle count doing all the work in the valuation.
Three chips, three jobs
The products are not variations of the same part.
| Chip | What Xiaomi announced | First use |
|---|---|---|
| XRING O3 | 3nm mobile system-on-chip; 24B transistors; 10-core CPU; 16-core GPU; 200-TOPS NPU; LPDDR6 | Xiaomi 18 Fold and Pad 9 Pro Max in September 2026 |
| XRING O100 | 6nm AI accelerator; wafer-level vertical stacking; 1.22 TB/s bandwidth; up to 330 tokens/s claimed | Consumer devices in 2027 |
| XRING D100 | 3nm driving chip; 20-core CPU; 16-core NPU; up to 160GB unified memory; models up to 200B parameters | Commercial use planned for 2027 |
Every performance figure is a Xiaomi claim from the launch, not an independent benchmark. The distinction matters most for O3: the company claims more than 5.2 million in AnTuTu and large generation-on-generation gains, but the chip has not yet shipped in a retail device where anyone outside Xiaomi can reproduce them.
O3 is the near-term product. Reuters reports a 200,000–300,000-unit shipment target, citing an unnamed source, and says TSMC will manufacture it on 3nm. Xiaomi said on its earnings call that devices using the prior O1 had already passed one million cumulative shipments, although Reuters' source put O1-powered smartphones at about 150,000. That gap is not a contradiction: Xiaomi's disclosed figure includes phones, tablets and watches.
O100 is the more novel architecture. Xiaomi says it uses 6nm wafer-level vertical stacking and hybrid bonding at a 1.4-micrometre pitch to deliver 1.22 TB/s of bandwidth, sixteen times a conventional flagship phone, with on-device inference as fast as 330 tokens a second. No price, power budget or saleable product was announced.
D100 is the automotive part. Xiaomi calls it China's first high-compute smart-driving chip on 3nm and says it has completed validation. It disclosed the CPU, NPU and memory capacity but not total TOPS, foundry, unit cost or a vehicle programme. Commercial application is scheduled for 2027.
The AI Cube joins them together, but it is explicitly a prototype. Its 33,874 CNC-cut cooling apertures are a lovely launch detail; they are not a production plan. No retail price, release date or volume was given.
The missing line is not small enough to ignore
In the June-quarter results, Smart EV, AI and other new initiatives produced RMB24.9 billion of revenue. Smart EV contributed RMB23.9 billion, or 96%. The remaining RMB1.0 billion is every other business in the segment combined, not a disclosed AI or semiconductor revenue figure.
That is as far as the income statement lets an outside reader go. Xiaomi did disclose the spending around it:
- RMB2.38 billion of segment capex in the quarter, 65.8% of group capital expenditure.
- RMB9.23 billion of group R&D, up 18.9%, without a chip-specific split.
- A RMB2.6 billion segment operating loss, despite a 19.2% gross margin.
- About RMB0.2 billion of group free cash flow, our subtraction of RMB3.8 billion of operating cash flow and RMB3.62 billion of capex.
The business therefore spent about thirteen times group free cash flow on segment capex in the same quarter, while announcing two 3nm designs and one 6nm accelerator. That does not prove the chips caused the spend — most segment capex may be factories and vehicle tooling — and Xiaomi gives no allocation that would let anyone claim otherwise. It does show why treating AI as a rounding error is no longer adequate.
What the launch does to our valuation
Our published base case for Xiaomi is $18.59 per ADR as of August 20. Remove the entire Smart EV, AI and other-new-initiatives vertical and it falls to $13.28, a $5.31 reduction. The combined segment is therefore 28.6% of the base case.
The problem is not that the model hides its approximation. The $35,118 input is explicitly segment revenue divided by vehicle deliveries, not the average selling price of a car. The problem is that the approximation gets less representative with every non-car product Xiaomi adds.
Move the delivery slider and the model moves MiMo, O100, D100 and every future AI device with it. A quarter with 20% more cars but no change in AI investment appears to grow the whole programme 20%. A quarter with flat car deliveries and the first commercial O100 revenue appears to have no AI growth at all.
This is where most analysis quietly invents a split. We are not going to. Xiaomi has not disclosed chip revenue, chip capex, internal transfer pricing or a segment operating result below the combined line. Monday's launch is evidence that a separate AI/silicon driver will eventually be needed; it is not enough information to build one now. The model does not change today, and the reason is lack of disclosure rather than lack of economic importance.
The strategic logic is clearer than the economics
The three chips give Xiaomi control at the three places where its ecosystem competes: the application processor in the phone, local inference across devices, and the driving computer in the car. The benefits could include lower supplier dependence, tighter software-hardware optimisation and one AI stack stretching from MiMo on a handset to a driving model.
None of those benefits is yet a number. O3 reaches customers next month; O100 and D100 are 2027 products; the Cube is a demonstration. Meanwhile the smartphone gross margin is 8.5%, the combined EV and AI segment is loss-making, and four consecutive quarters of free cash flow have each stayed below RMB1 billion.
The launch makes Xiaomi more vertically integrated. It also makes the request to management more specific: stop making investors infer the economics of three silicon businesses from the number of cars delivered.
What to watch
- O3 shipments. Reuters' 200,000–300,000 target is source-reported, not company guidance. Retail shipments and independent benchmarks are the first test.
- A commercial O100 product. A device, price and memory configuration would turn the 1.22 TB/s demonstration into something that can be modelled.
- The first D100 vehicle. Xiaomi has disclosed no model, TOPS figure or unit-cost comparison with the Nvidia hardware it currently uses.
- A revenue or spending split. Chip revenue, internal transfer pricing, R&D or capex would all be useful; today Xiaomi publishes none of them.
- Segment cash conversion. The next filing needs to show whether the RMB2.38 billion quarterly capex run rate is building profitable volume or extending the loss.
Sources and provenance. Chip specifications, launch timing and AI Cube details are Xiaomi claims from its 24 August XRING briefing, carried by contemporaneous Chinese reporting; none is independently benchmarked. TSMC production, the O3 shipment target and O1 smartphone count are Reuters reporting based partly on unnamed sources. June-quarter revenue, margin, deliveries, R&D, capex and cash flow are from Xiaomi's 18 August results; free cash flow and all ratios are derived by R40. The $18.59 base case, $13.28 ex-segment case and $35,118 unit driver are R40 model outputs as of 20 August, not company guidance.
