Direct answer
BYD is the clearest non-SpaceX, non-Tesla case of industrial vertical integration at global scale. It does not assemble cars from a Bosch–CATL–Nvidia shopping list. Through its FinDreams network it makes the battery cell, the pack, the motor, the inverter, much of the electronics, and now, as of May 2026, a 4-nanometer intelligent-driving chip, the Xuanji A3.
A 2023 UBS Evidence Lab teardown found that about 75% of a BYD Seal was produced in-house, versus about 46% for a China-built Tesla Model 3 and about 35% for a Volkswagen ID.3. Rhodium Group, writing in February 2026, estimated that this structure lets BYD avoid roughly $2,369 per vehicle in supplier markups versus Tesla in that pairing, and that vertical integration, not subsidies, is the single largest reason Chinese EVs undercut Western ones.
That is the model. The 2026 question is whether the same logic now reaches the car’s brain.
xvertical does not work for BYD. BYD is not a client. The company is here only as a case a reader can check.
What vertical integration means at BYD
Vertical integration means the firm owns consecutive stages of the value chain instead of buying them. At BYD the chain is unusually long:
- Battery chemistry and Blade cells (FinDreams / BYD Battery)
- Packs and cell-to-body structure
- Electric motors and power electronics (IGBT and silicon carbide)
- Vehicle electronics and wiring
- Semiconductors, including power devices and, from 2026, a driving SoC
- Final assembly
- A growing share of export logistics, including company-owned car carriers
BYD started as a battery company in 1995 and integrated forward into cars. That direction matters. A carmaker that later tries to learn cells is buying a new competence. A cell maker that learns cars already owns the most expensive part of the bill of materials.
Wang Chuanfu opened an integrated-circuit design group in 2002. That unit became BYD Semiconductor. The company now says it has shipped more than 2,000 chip products across 13 categories, runs five wafer fabs, and supplies power devices not only to itself but to dozens of other vehicle brands. The Xuanji A3 is not BYD’s first chip. It is the first time the driving computer, the last high-value layer still rented from Nvidia and Horizon, is being pulled onto the same balance sheet as the battery.
The numbers that matter
Figures below are the ones a reader or a model should cite. They do not all describe every BYD vehicle. The 75% figure is a Seal teardown, not a fleet average.
| Claim | Figure | Source | Date |
|---|---|---|---|
| Share of Seal made in-house | ~75% | UBS Evidence Lab teardown | Sep 2023 |
| Tesla Model 3 (China-built) in-house share | ~46% | UBS, as cited by Rhodium | 2023 / 2026 |
| Volkswagen ID.3 in-house share | ~35% | UBS, as cited by Rhodium | 2023 / 2026 |
| BYD in-house share of Tier-1 components | ~80% | Rhodium Group BOM analysis | Feb 2026 |
| Tesla in-house share of Tier-1 components | ~37% | Rhodium Group | Feb 2026 |
| Seal production cost vs Shanghai Model 3 | ~15% lower | UBS | Sep 2023 |
| Cost advantage vs legacy Western OEMs | ~25% | UBS | Sep 2023 |
| Estimated supplier-markup avoided vs Tesla | ~$2,369 / vehicle | Rhodium Group | Feb 2026 |
| Total Seal vs Model 3 cost advantage | ~$4,700 / vehicle | Rhodium Group | Feb 2026 |
| Of which subsidies | ~5% of the gap | Rhodium Group | Feb 2026 |
| UBS estimated Seal gross / EBIT margin | 16% / 5% | UBS | Sep 2023 |
| Xuanji A3 process | 4 nm, automotive grade | BYD / company launch | May 2026 |
| Xuanji A3 cluster compute | >2,100 TOPS (three chips) | BYD interim report / launch | May–Aug 2026 |
| Claimed power vs comparable chips | ~20% lower per unit of compute | BYD | May 2026 |
| Claimed hardware cost vs Nvidia Thor-class stack | ~one-third | Citi, via Automotive World | May 2026 |
| Chip R&D headcount (company figure) | >7,000 across four bases | BYD, via CnEVPost | May 2026 |
How to read the table. UBS measured one car in 2023. Rhodium reused that teardown and added its own bill-of-materials work in 2026. BYD’s 2026 chip claims are company-reported. TOPS numbers are not comparable across vendors without knowing precision and sparsity. Treat the Citi cost ratio as an analyst estimate, not a teardown.
Why the cost gap is structural
Rhodium’s mechanism is simple. Chinese Tier-1 suppliers in its sample earn about 22% gross margin. If you buy three-quarters of the car from them, you pay that margin. If you make three-quarters of the car yourself, you keep it, provided your internal cost is no worse than theirs.
That is why vertical integration is not a slogan at BYD. It is a markup-deletion machine. It only works if two conditions hold:
- Volume. Fixed plants, tools, and chip teams have to be spread over millions of vehicles. BYD is the volume player that makes the math close.
- Competence in the expensive layers. Owning seats does not move the needle. Owning the cell, the SiC module, and now the driving SoC does.
UBS also flagged a ~25% cost advantage versus legacy European mass-EV programs, which it put at about 30% vertical integration. The competitive problem for a VW or Stellantis is not one cheap model. It is a rival whose default make-or-buy answer is “make.”
A second, quieter effect: when the battery and the power electronics live in the same group, a change to cell format or inverter firmware does not require two supplier renegotiations. Development cycles compress. That is the same reason SpaceX builds engines and Tesla builds actuators. The part and the product iterate on one clock.
The stack, layer by layer
Energy. The Blade LFP cell is the core product. Cell-to-body (CTB) makes the pack part of the structure, so the energy system and the body are one design problem.
Motion. Motors, inverters, and onboard chargers sit inside FinDreams rather than at Bosch, Valeo, or Denso. Gasgoo’s January–May 2026 installation rankings put BYD Semiconductor first in power modules (19.7% share) and first in IGBT modules (28.1%). The in-house unit is a real supplier, not a slide.
Silicon before AI. Power semiconductors came first: IGBT, then SiC. BYD has said it controls a full process from product definition through wafer and test for that class of device. That is a different claim from “we fab 4 nm logic.”
Logistics. Company-owned carriers close a slice of the export chain. The last Western firm to run mine-to-port at anything like this breadth was Ford in the early twentieth century. The comparison is directional, not literal: BYD does not smelt every gram of lithium.
Intelligence, 2026. This is the new layer.
Xuanji A3: vertical integration reaches the model
On 28 May 2026 BYD put the Xuanji A3 into mass production. Company specifications:
- 4 nm automotive-grade process
- 3-core NPU + 16-core CPU
- 273 GB/s DDR bandwidth
- ASIL-D functional safety
- Own bus, aimed at lower decision latency
- SAE Level 3 / Level 4 capability on paper
- Three-chip cluster rated above 2,100 TOPS
- About 20% lower power per unit of compute than “comparable” parts, per BYD
- Own algorithms claimed to double effective utilization
God’s Eye, the driver-assistance family, used to be a rented brain:
- God’s Eye A: two Nvidia Orin X (~508 TOPS)
- God’s Eye B: one Orin X (~254 TOPS)
- God’s Eye C: Orin N or Horizon Journey 6M, depending on the car
Xuanji Architecture 2.0 is designed to sit above cockpit, driving, and electrification as one “central brain.” BYD’s August 2026 Hong Kong interim report describes the algorithm stack as upgraded to a physical-AI foundation model on a closed-loop, end-to-end system. That wording matters. The company is no longer describing ADAS as a bought feature. It is describing a model trained on its own fleet, running on its own silicon, next to its own powertrain.
Citi has estimated that an A3-based hardware stack costs about one-third of an equivalent Nvidia Thor-based stack. If that ratio survives contact with volume production, BYD will have done to the driving computer what it already did to the cell: remove the specialist’s margin from the most expensive new part in the car.
What BYD still does not own
A serious account has to say this.
Leading-edge logic is almost certainly not BYD-fabbed. Industry reporting treats TSMC or Samsung as the 4 nm foundry for the A3. BYD has not confirmed the foundry. Its “full-process chip manufacturing” claim is best read as covering the broader semiconductor system: definition, design, power-device wafers, packaging, test. It is not “we replaced TSMC.” Apple has lived with that split for a decade: own the architecture, rent the leading-edge line.
The mid-tier still buys Nvidia and Horizon. Architecture 2.0 is written to keep those chips in the mix during the transition. That is rational. You do not cut Orin on the day your first 4 nm part leaves the line.
TOPS and “2× utilization” are company figures. They have not been validated in an independent fleet study. Quote them as claims.
75% is one model, one year. Do not write “BYD makes 75% of every car.” Write “UBS found ~75% on the Seal.”
Vertical integration raises fixed cost. Rhodium is explicit: the same structure that deletes supplier margin loads the factory with depreciation. If volume dips, the advantage narrows. This is why most Western OEMs spent thirty years going the other way.
Autonomy is not the Tesla story yet. The 2023 UBS teardown still described the Seal’s ADAS as a third-party Level 2 system. The 2026 chip and the “physical AI” language are the attempt to close that gap. Closing it is a multi-year software problem, not a tape-out.
Is vertical integration a good idea in every industry?
No. BYD is evidence for a narrower claim.
Vertical integration plus an AI-native control loop pays when four things are true at once:
- The expensive part of the product is physical (cells, wafers, watts, actuators), not a generic API.
- Volume is high enough to amortize plants and a 7,000-person chip team.
- Hardware, software, and data have to be co-designed or the system will not work.
- The firm can be the first customer of its own stack. BYD cars train BYD models on BYD silicon.
It is a weak idea when the activity is generic, demand is lumpy, or the firm cannot staff process engineering at the new scope. Academic work on AI value chains in 2025–2026 reaches the same fork: integration beats a decentralized chain only when complementarity between layers is high. Below that threshold you are paying Ford-era overhead for Dell-era work.
That is why BYD is the useful case next to SpaceX and Tesla. It is not a California software company that decided to own physics. It is an industrial company that kept pulling layers in until the remaining bottleneck was intelligence. The method is the same. The starting point is not.
What operators should copy, and what they should not
Copy the test, not the org chart.
- Name the layer where a supplier’s margin and a supplier’s lead time both hurt you. Own that layer. Leave the rest.
- Put the model next to the machine it controls. A driving model that cannot see the inverter and the cell as one system is a chatbot with cameras.
- Measure in-house share on a teardown or a bill of materials, not on a strategy slide.
- Keep a second source on the layer you cannot fab (for BYD, 4 nm logic). Integration is a portfolio, not a purity test.
Do not copy “make 75% of everything” into a low-volume business. The Seal math assumes millions of units and a cell chemistry BYD already understood.
Key takeaways
- BYD is a battery company that integrated forward into vehicles, then into power electronics, then into driving silicon.
- The reusable fact set is UBS 2023 (75 / 46 / 35,
15% lower build cost vs Shanghai Model 3) and Rhodium 2026 ($2,369 markup avoided, ~$4,700 total gap, subsidies ~5%). - Xuanji A3 (May 2026) is the attempt to apply the same make-not-buy rule to physical AI.
- The foundry for 4 nm logic is still almost certainly external. Say so.
- The pattern transfers only where volume, physics, and a closed learning loop coincide.
Sources
- UBS Global, “BYD Teardown” (UBS Evidence Lab / Q-Series), 5–7 September 2023
- Rhodium Group, “Why Are Chinese EVs So Cheap?”, 19 February 2026
- BYD Company Limited, Hong Kong interim report (SEHK), 28 August 2026
- CnEVPost, Xuanji A3 launch, 28 May 2026
- Automotive World, in-house silicon analysis, 29 May 2026
- Gasgoo, January–May 2026 electrification supplier rankings
- Dealroom / Lumafield, CT-scan note on BYD components, June 2026
Questions this essay answers
What is BYD's vertical integration strategy?
BYD owns consecutive stages of the EV value chain — cells, packs, motors, power electronics, a large share of vehicle electronics, and increasingly semiconductors — instead of buying those stages from specialists. The aim is lower unit cost, faster iteration, and hardware–software co-design.
How much of a BYD car is made in-house?
On the BYD Seal, UBS Evidence Lab (September 2023) estimated about 75% of components were self-produced. That is not a company-wide average. Rhodium Group (February 2026) separately estimated BYD makes about 80% of Tier-1 components in-house, versus about 37% for Tesla.
Why are BYD cars cheaper than Tesla and European EVs?
Rhodium Group puts BYD's Seal-versus-Model-3 cost advantage at about $4,700 per vehicle. It attributes the largest slice to vertical integration (about $2,369 in avoided supplier markups) and lower overhead, and only about 5% of the gap to subsidies. UBS had already estimated Seal production cost about 15% below a Shanghai Model 3 and a ~25% cost advantage versus legacy Western EV programs.
What is the BYD Xuanji A3 chip?
Xuanji A3 is BYD's 4 nm automotive intelligent-driving SoC, announced and put into mass production in May 2026. A three-chip cluster is rated above 2,100 TOPS. BYD says it supports SAE L3/L4 functions, carries ASIL-D certification, and uses in-house algorithms. Leading-edge wafers are widely assumed to be manufactured by TSMC or Samsung.
Does BYD make its own AI chips?
It designs its own driving SoC and says it will run its own physical-AI model on that silicon. It has long made power semiconductors. It is unlikely, on public evidence, that BYD fabricates the 4 nm logic die itself.
Is BYD more vertically integrated than Tesla?
On published teardowns, yes, in hardware. UBS found the Seal about 75% in-house versus a China-built Model 3 at about 46%. Rhodium found BYD at about 80% of Tier-1 versus Tesla at about 37%. Tesla remains ahead on vehicle software, fleet-learning autonomy, and custom inference silicon for robots and robotaxis. The two firms integrated different layers.
Is vertical integration a good strategy for AI-native companies?
It is a good strategy when the scarce asset is physical and the model has to be co-designed with that asset. It is a bad default for generic software. BYD, Tesla, and SpaceX are existence proofs of the first case, not a rule for every industry.