$7B, Five-Year Lock-in: ByteDance Bets on CXMT Memory for AI Inference 'Memory Autonomy'

ByteDance signed a five-year DRAM procurement agreement worth over $7 billion with CXMT. This number needs to be viewed in context: the global DRAM market size in 2024 was approximately $90 billion, and CXMT's global market share was less than 3%, but according to IC Insights estimates, its revenue in 2024 had already exceeded $3 billion. $7 billion equates to locking in more than two years' worth of CXMT's revenue upfront.

Short term: ByteDance's "hunger" for memory on the AI inference side far exceeds external imagination

ByteDance's AI layout centers on Douyin recommendations, Doubao large model inference, and Volcano Engine cloud services. All three require massive memory bandwidth and capacity.

Let's break down ByteDance's appetite for computing power. Assuming the Doubao large model processes 1 trillion tokens per day, calculated based on current mainstream MoE models (approx. 300B parameters, approx. 30B active parameters), a single inference requires about 60GB of model parameters resident in VRAM. If HBM (High Bandwidth Memory) is adopted, a single H100 card with 80GB VRAM can only run one batch. But HBM costs are extremely high, and it is monopolized by SK Hynix and Samsung.

ByteDance turning to CXMT has clear logic: use DDR5 as a "memory pool" for large-scale inference, rather than HBM.

Key Data Comparison:
- HBM3e: Approx. $12-15 per GB, bandwidth 1.2TB/s
- CXMT DDR5: Approx. $3-4 per GB, bandwidth 32GB/s (single channel)

ByteDance's AI inference infrastructure may adopt a heterogeneous architecture of "CPU+GPU+Massive DDR5." CXMT's DDR5 is plugged directly into CPU memory channels, used for storing KV Cache (Key-Value Cache) and inactive parameters, while GPUs handle core computation. This significantly reduces single-card inference costs but requires massive amounts of DDR5—which happens to be CXMT's strength.

[!tip] ByteDance's procurement scale ($7 billion/5 years, approx. $1.4 billion/year) is equivalent to more than 50% of CXMT's current capacity. This means CXMT needs to expand dedicated production lines for ByteDance, or even adjust processes.

Long term: The "orders for technology" closed loop of China's storage industry is forming

CXMT's DDR5/LPDDR5 products only entered mass production in 2023, and yield rates are still ramping up. ByteDance's order acts like a shot in the arm, providing stable cash flow so CXMT can burn money on next-generation processes (1β nm, 1γ nm).

More importantly, this order changes CXMT's customer structure. In the past, CXMT mainly relied on domestic smartphone and PC manufacturers to absorb capacity; these customers are price-sensitive with thin margins. As an internet giant, ByteDance has high requirements for memory reliability but is willing to pay a premium for stable supply. This forces CXMT to improve quality control capabilities for enterprise-grade DRAM.

Let's look at a technical detail. CXMT's DDR5 chips use a 10nm-class process (17nm), lagging one generation behind Samsung's 12nm-class process, but ByteDance's AI inference scenarios are not that demanding regarding timing. This means CXMT can cover ByteDance's needs with "mature process + large capacity," avoiding head-on competition with Samsung and SK in high-end HBM.

graph LR
A[ByteDance] -->|$7 Billion Order| B[CXMT]
B -->|Cash Flow| C[Process R&D]
C -->|1βnm/1γnm| D[Better Chips]
D -->|Lower Cost| A
D -->|Tech Spillover| E[Domestic Smartphones/PCs]

The Achilles' heel of this closed loop is: Can CXMT achieve a DDR5 yield rate above 90% within 5 years while reducing costs to parity with Samsung? If not, ByteDance's "memory autonomy" plan will turn into "buying domestic sentiment at high prices."

My judgment: This is ByteDance's "defensive investment," not the commercially optimal solution

From a purely technical perspective, ByteDance could completely buy Samsung's DDR5, which is cheaper and has more stable supply. But under geopolitical risks, ByteDance must consider supply chain security. After the US export controls against China in 2022, CXMT was almost placed on the Entity List; although it didn't happen eventually, ByteDance clearly doesn't want to repeat Huawei's mistakes.

This order is essentially ByteDance using a "prepayment + post-payment" model to help CXMT build a "ByteDance-exclusive" domestic memory production line. Short term, ByteDance's procurement costs will be 15%-20% higher than international prices. Long term, if CXMT achieves technological breakthroughs, ByteDance gains a cost advantage; if CXMT fails, ByteDance's $7 billion becomes sunk cost.

[!abstract] Core Conclusion: The $7 billion agreement between ByteDance and CXMT is essentially trading capital for time—using a 5-year window to bet that China's storage industry can catch up with international advanced levels, thereby building a hardware foundation for AI inference unaffected by geopolitics.

Finally, an industry observation

Recently reading this paper: "Memory is All You Need for Scaling LLM Inference" (arXiv:2503.19723), which mentions a viewpoint: In the inference stage, memory bandwidth is scarcer than computing power. If ByteDance can use CXMT's DDR5 to low-costly expand the memory pool, combined with its self-developed inference framework, it might bypass the HBM supply bottleneck.

My advisor asked me to try this direction, but the lab's H100s aren't enough. ByteDance's order makes me feel that domestic AI infrastructure is taking a completely different path from Silicon Valley—trading scale for efficiency, trading time for space.

One-sentence summary: ByteDance isn't betting on CXMT's today, but on how far China's storage industry chain can swim in the tide of AI inference.

Original link: https://www.ithome.com/0/981/172.htm