Samsung committed $73.24 billion to capital expenditures and research in 2026. This amount exceeds half of the projected annual operating profit for the company. The investment targets AI accelerator memory, advanced foundry nodes, and next-generation packaging. Samsung aims to lead the AI semiconductor era through this massive spending. In June 2026, the company expanded this plan with a 2,450 trillion won domestic investment framework that runs from January 2026 through December 2040. This massive capital injection targets the growing demand for agentic AI, which drives orders for high-bandwidth memory and server-grade storage.
The Mach-1 AI accelerator sits at the center of this strategy. Naver, the South Korean cloud and search giant, ordered $752 million worth of Mach-1 chips. Samsung expects to deliver between 150,000 and 200,000 units to Naver. Each unit costs approximately $3,756, or 5 million won. Naver plans to use these accelerators for AI inference in the servers that power its Naver Place map service. This contract allows Naver to reduce its reliance on expensive Nvidia AI chips. Samsung intends to use the Naver deal as a stepping stone to supply other Big Tech firms, including Microsoft and Meta.
Mach-1 technical specifications and performance
The Mach-1 functions as an AI inference accelerator based on an application specific integrated circuit (ASIC) design. It combines Samsung’s proprietary processors with low-power DRAM chips. This design differs from Nvidia accelerators, which use a combination of GPUs and high-bandwidth memory (HBM). The Mach-1 targets the transformer model and is specified to reduce data bottlenecks between the processor and memory.
| Specification | Detail |
|---|---|
| Chip Architecture | ASIC |
| Memory Type | Low-power DRAM |
| Unit Price | $3,756 (5 million won) |
| Naver Order Volume | 150,000 to 200,000 units |
| Total Contract Value | $752 million (up to 1 trillion won) |
| Efficiency Increase | 8x |
| Bottleneck Reduction | 1/8th of current levels |
Samsung claims the Mach-1 improves efficiency by eight times. The chip also reduces the bottleneck phenomenon between memory and GPU chips to one-eighth of current levels. These improvements allow for large language model inference using low-power memory instead of power-hungry HBM. The Mach-1 targets edge computing applications that require low power consumption, minimal dimensions, and low costs.
Architectural differences between Samsung and Nvidia
You already know that Nvidia dominates the AI training market, so the significance of Samsung’s shift toward inference lies in the cost-efficiency of the Mach-1. Nvidia’s current accelerators consist of GPUs and HBM. In contrast, the Mach-1 uses an ASIC design with LPDDR memory. This combination aims to solve the energy and cost problems inherent in the GPU-HBM architecture.
The Mach-1 price is one-tenth the cost of Nvidia’s hardware. While Nvidia’s Blackwell architecture targets real-time trillion-parameter models at a lower cost than the Hopper series, it still relies on high-power components. Samsung’s Mach-1 focuses on the inference side of the market. It competes with solutions like AWS Inferentia rather than high-performance processors like the AWS Trainium. The Mach-1 provides a way for companies to run AI workloads without the high price and power requirements of traditional GPUs.
The HBM roadmap and memory competition
Samsung aims to rival SK Hynix in the high-bandwidth memory market. The company is currently sampling its HBM3E 12H chips to customers. This product is the industry’s first 12-stack HBM3E DRAM. Samsung plans to start mass production of these chips in the first half of the year.
The company’s HBM roadmap shows massive growth projections. Samsung expects HBM shipments in 2026 to be 13.8 times higher than the 2023 output. By 2028, annual HBM output volumes will reach 23.1 times the 2023 level. For the sixth-generation HBM, code-named "Snowbolt," Samsung plans to apply a buffer die to the bottom layer of the stacked memory to increase efficiency.
High-bandwidth memory remains a chokepoint in the AI inference era. The demand from hyperscale data center operators, such as Microsoft, Google, Amazon, and Meta, creates a seller’s market. Samsung’s investment in HBM4 and advanced packaging aims to secure its share of this expanding market.
Foundry market share and the 2nm race
Samsung’s foundry business faces stiff competition from TSMC. In 2025, TSMC held 69.9% of the global foundry market with $122.54 billion in revenue. Samsung held 7.2% of the market, which is a decline from its 13% share in early 2024. Samsung’s foundry revenue in 2025 was $12.63 billion. This gap between Samsung and TSMC remains a primary challenge for the company’s semiconductor division.
Samsung implemented Gate-All-Around (GAA) technology at the 3nm node in June 2022. This made Samsung the first foundry to use GAA technology commercially. TSMC deferred GAA until its 2nm node. Samsung is now working on its 2nm GAA production. The company has already secured its first 2nm AI chip order from Japan’s Preferred Networks.
The transition from FinFET to GAA represents a massive architectural shift. Both Samsung and TSMC plan to integrate backside power delivery (BSPD) at their 2nm nodes in 2025. Samsung relies on its experience with 3nm GAA to gain an advantage. Will Samsung’s early lead in GAA technology translate to higher yields and more customers at the 2nm node?
AI infrastructure and the broader supply chain
The rise of agentic AI drives a massive demand for specialized hardware. This demand affects the entire semiconductor ecosystem, including manufacturing, power, and water supplies. In Korea, the government must invest 72.8 trillion won in 70 transmission lines by 2038 to prevent delays in semiconductor fab construction. KEPCO faces significant challenges in this area, as its total debt reached 210.7 trillion won as of the end of June.
In the United States, the semiconductor supply chain sees significant investment. Over $827.8 billion in private investments have been announced across 30 states since 2020. This includes the Advanced Manufacturing Investment Credit and various manufacturing grants.
China is also expanding its own chip capacity. Goldman Sachs projects that China’s supply of advanced-process wafers at 7 nanometers and below will grow by 46% per year through 2035. This growth will narrow China’s domestic gap between advanced chip supply and demand from 92% last year to 34% in 2035. However, China still trails Samsung and SK Hynix by two to three technology generations.
Strategic positioning in the AI era
Samsung positions the Mach-1 as a tool to disrupt the current AI hardware hierarchy. By combining its processor and memory expertise, the company targets the high-growth inference market. The contract with Naver proves there is significant interest in alternatives to Nvidia. This deal provides the volume needed to refine the Mach-1 technology for larger customers.
The company’s strategy involves three pillars: reach, openness, and confidence. Reach involves putting AI across all devices. Openness requires easy access to AI tools. Confidence relies on security, such as Samsung Knox, to protect data. The Mach-1 supports this vision by providing the hardware necessary for efficient, low-power AI processing.
Samsung’s $73 billion investment is an attempt to control the AI hardware stack from the memory to the logic. This vertical integration gives Samsung an advantage that pure-play foundries or memory makers lack. The company’s ability to deliver the Mach-1 and HBM4 at scale will determine if it can reclaim its lead in the AI era.
The Mach-1 is Samsung’s most aggressive attempt to decouple the AI industry from Nvidia’s pricing power.
