The design for Tesla's next-generation AI inference processor, "AI5," tailored to Samsung's manufacturing process, has reached tape-out. James Kim, a principal engineer at Samsung Foundry, wrote on LinkedIn that the chip would be manufactured using the "latest 2nm process" at the company's Taylor, Texas fab and incorporated into a new Tesla product. The post was deleted after it was reported on, and Samsung has declined to comment on matters involving its customers. This means the update, while significant and revealed by a party involved in manufacturing, is not an official Samsung announcement.
Tape-out does not mean mass production has begun. Tesla previously stated that it completed the final chip design for AI5 in April 2026, and this latest development corresponds to the stage where that design has been converted into physical data conforming to Samsung's transistors, wiring, and design rules. From here, photomasks must be produced, wafers run, and working prototype chips obtained before performance and manufacturability can actually be verified.
From April's Design Completion to Samsung-Specific Mask Data
On April 15, Elon Musk posted a message celebrating "AI5's tape-out" to Tesla's AI chip design team. In the Q1 2026 shareholder deck released on April 22, Tesla likewise stated that it had "completed the final chip design for the next-generation AI5 inference processor in April." However, that public explanation at the time did not explicitly confirm that the Samsung-specific mask data had been finalized.
In ASIC development, a logic design that defines functionality is converted into the actual placement of transistors and metal wiring. Timing and power consumption are then analyzed. Wiring congestion is resolved, and compliance with manufacturing rules is checked before final data—such as GDSII—is handed off to the foundry. Even under the same AI5 name, the Samsung and TSMC versions must go through this process separately.
What is new about this latest post is that the physical implementation of the Samsung version has completed this verification process and is now ready to be handed off for actual chip fabrication. That said, Tesla's official plan still targets 2027 for AI5 production. Kim's phrase "soon to be incorporated" does not specify a concrete date, and there is no basis to interpret it as meaning the chip will appear in products within a matter of weeks.
Running the Same AI Software on Two Physical Versions
Tesla's manufacturing allocation has shifted within just four months. In July 2025, Musk explained that Samsung would manufacture AI4, TSMC would handle AI5, and Samsung's Taylor fab would take on AI6. However, in November of the same year, the explanation was updated to say that both AI5 and AI6 would be produced by both Samsung and TSMC. Because the two foundries differ in how they convert designs into physical form, this results in slightly different versions of AI5—but Tesla's goal is for its AI software to run identically on both.
By using two foundries, Tesla can secure capacity across two leading-edge processes and spread out the risk of ramp-up issues. Even if yield improvement lags at one foundry, there remains the possibility of supply from the other. In exchange, Tesla must verify two separate physical implementations—from placement and routing through timing, power, and testing—and maintain an environment where the same trained model performs equivalently on both. This increases both development costs and verification workload.
Tesla is willing to bear this cost because it plans to deploy AI5 at scale. In its Q4 2025 shareholder deck, Tesla set a goal of boosting AI5 to 50x the performance of AI4, citing a 10x increase in compute, a 9x increase in memory capacity, and an additional 5x improvement from hardware-accelerated quantization and softmax operations. Exactly which metrics combine to produce this "50x" figure has not been disclosed in detail, and for now it remains a design target set by Tesla.
SF2 Performance Targets and Taylor's Production Facilities
Samsung's first-generation 2nm process, "SF2," unveiled in 2023, was said to offer 12% higher performance, 25% better power efficiency, and a 5% smaller chip area compared to the 3nm SF3 process. The plan for 2nm mass production called for expanding from mobile applications in 2025 to HPC in 2026 and automotive in 2027. The development timeline for AI5 overlaps with Samsung's 2nm mass-production roadmap as it expands from HPC to automotive applications.
These figures represent Samsung's own stated targets comparing the SF2 and SF3 process generations—they are not performance differences specific to AI5 itself. Furthermore, Kim's post only mentioned the "latest 2nm process" without naming any specific derivative process such as SF2P+, SF2T, or SF2Z. Applying a generic process generation name from the outside risks getting both the performance and the production timeline wrong.
Samsung's Q1 2026 earnings report indicated that its leading-edge process lines are expected to reach full operation in Q2, along with a policy of increasing the number of major 2nm customers. In the second half of the year, the company plans to expand second-generation 2nm production for mobile applications. However, this earnings material did not name AI5 specifically, nor did it state that the leading-edge line in question is located at Taylor. A gap remains between the individual's post and the company's official disclosures that still requires confirmation.
The Taylor facility is part of an advanced semiconductor hub for which the U.S. government is providing Samsung with direct support of up to $4.745 billion. Samsung is investing more than $37 billion in central Texas, building two advanced logic fabs and an R&D fab at Taylor. Materials from the U.S. National Institute of Standards and Technology cite mass production of 2nm chips as an objective, while stating that the planned facilities are expected to become operational by 2030. AI5's tape-out has now connected this long-term plan to a concrete customer product.
Prototyping and Parity Verification Remain Before 2027
The detailed production plan for AI5 still does not perfectly align with the publicly disclosed roadmaps of either foundry. Musk had earlier explained that the TSMC version of AI5 would be manufactured in Taiwan, and later said it would be made in Arizona instead. Meanwhile, according to publicly released plans for TSMC's Arizona facility from the U.S. government, Fab 1 covers 4nm and 5nm, Fab 2 covers 3nm with production starting in 2028, and Fab 3—which will handle N2 and A16—is scheduled to come online by the late 2020s.
If the published timeline holds, it cannot be confirmed that early AI5 supply in 2027 will be covered by N2 production in Arizona. There remains the possibility that the production site, node, or schedule could change going forward. At minimum, one cannot conclude that "if the Samsung version is 2nm, the TSMC version must also be 2nm." Neither Tesla nor TSMC has disclosed the process node name for AI5, and it remains possible that the same software environment could be achieved across different nodes.
The Samsung version will now proceed to photomask and prototype wafer fabrication, post-packaging functional testing, Tesla's qualification process, and mass-production yield improvement. If the chip is to be used in vehicles, reliability testing covering temperature and vibration will also be required. If both foundry versions are used in parallel, it will also be necessary to confirm that output, latency, and power consumption for the same AI model fall within acceptable product ranges across both.
Tesla's AI5, on the Samsung manufacturing route, has now arrived at the threshold just before mask fabrication. What will change the assessment going forward is not repetition of the word "tape-out," but whether Samsung can produce working prototype chips, Tesla can complete qualification, and the 2027 production plan can be made concrete down to production volumes and the products it will ship in.