TrendForce raises memory market forecast to $1.28 trillion on AI agent demand

TrendForce analysts have revised their 2027 memory market forecast, raising it from $842.7 billion to $1.28 trillion. The 52% jump is attributed to explosive demand for memory from AI agents, which need huge volumes for computation and caching.

Why the estimate rose by $437 billion

The main reason is the expansion of AI agents. TrendForce noted a sharp rise in memory demand for inference, training and caching of large language models. As companies deploy OpenAI's Operator, Anthropic's Claude, Google's Gemini and their own in-house solutions at scale, the load on memory grows exponentially.

In January, the analysts expected gradual market growth of 5–7% a year. But data from memory makers showed that demand was running almost a year ahead of the forecast, a signal to recalculate everything. As a result, the figure rose dramatically.

Historically, the memory market has been conservative: manufacturers plan capacity several years ahead, and forecast revisions are rare. A 52% revision during a long-term planning period is practically a revolution in this segment.

For context: if the market really does grow 52%, it would be the largest increase in the history of the semiconductor industry in 50 years. The chip market usually grows 5–15% a year in good years and falls 20–30% in crisis years. A 52% jump in one year signals that the AI era is not just hype but a fundamental shift in demand.

What eats memory in AI agents

An AI agent is not a single model but a whole system. These are the main consumers of memory:

• Context caching: LLMs with a 128K-token window keep the entire conversation in memory for fast responses
• GPU memory: HBM3E and GDDR6X for inference on accelerators (NVIDIA H200, AMD MI300X)
• Model parameters: 70-billion-parameter models need up to 300 GB of memory for the weights alone
• Fast cache: SSD and RAM for instant access to data instead of network requests
• Agent state: when a browser agent runs for hours, its internal state grows

When a dozen such systems run on a single server at the same time, memory requirements grow geometrically: not linearly, but quadratically.

Who wins

Samsung, SK Hynix, Micron and Kioxia see this forecast as a historic opportunity. If the memory market really grows 52%, that means roughly $437 billion in additional revenue. That money can fund new fabs, equipment upgrades and higher output.

Producers of HBM (High Bandwidth Memory), specialized memory for AI accelerators, stand to gain the most. Demand for HBM already exceeds supply, and prices are rising. Whoever wins contracts from NVIDIA, Intel and AMD is buying the future.

Where the forecasts could go wrong

TrendForce rarely misses in its top-line forecasts, but the memory market is full of surprises. Data compression, such as LoRA-style approaches or new model architectures, may reduce memory requirements. Or, if cloud providers switch to more efficient caching strategies, demand will slow. In addition, if a geopolitical conflict or chip export sanctions occur (as with Huawei), demand may drop sharply.

What this means

The memory market is moving from an era of steady growth to an era of exponential growth. PCs and smartphones grew it 5–10% a year: boring, predictable, and often even a decline. AI agents are doubling demand within months and changing the investment climate.

That is good for shareholders of memory makers and bad for their customers, cloud providers and corporations, since memory gets more expensive. It is also risky in the long term: if demand for AI agents slows or a technological breakthrough in data compression occurs, overproduction and a price crash will follow.

But for the next 2–3 years, TrendForce is confident: memory will be red-hot. If the forecast comes true, it will be the fastest jump in the memory market's history.