Summary of Key Points
The exponential demand for computing power driven by the AI revolution is triggering an unprecedented wave of hardware inflation across the entire supply chain: from GPUs and CPUs to HBM memory and optical modules, even the screws in servers, almost every component is seeing price increases and shortages. This is due to a combination of three factors: a surge in demand (due to the training and inference needs of large models), a lagging supply (long production cycles and high technical barriers), and panic-driven hoarding (large companies securing production capacity and distributors speculating on prices). Upstream core manufacturers are reaping huge profits, with internet giants bearing the initial costs, which are ultimately passed on to consumers. However, this situation also poses risks such as stifled innovation, a fragmented supply chain, and potential overcapacity in the future.
1. How Wild Is the Price Surge? Prices Are Rising for Everything from Chips to Screws
The increase in AI hardware prices is not a localized phenomenon but affects every link in the supply chain:
- CPU Prices: From Stable to Explosive
In the past, CPU supply was abundant, but AI servers require more cores (for example, an 8-GPU server needs 2 high-end CPUs), leading to a sharp increase in demand. Intel raised prices by 12% across its entire range in March this year, followed by AMD with increases of 10-15% in April; the price of high-end CPUs in the spot market has risen by more than 40%, with some models experiencing daily price fluctuations.
- HBM Memory: Even More Profitable than Luxury Goods
Each AI GPU requires 5-6 HBM chips, and HBM consumes 3-4 times the number of wafers as regular memory. Samsung and SK Hynix have shifted 70% of their production capacity to HBM, causing prices to soar from $80 at the beginning of last year to $700 today (an eightfold increase), with profit margins exceeding 70%. By 2026, all of their HBM production capacity will be sold out.
- Optical Modules: Delivery Times Extended to 2028
There is a global shortage of 60% of the core optical chips for 1.6T optical modules, which are monopolized by American and Japanese manufacturers, resulting in delivery times of 18-24 months. The price of indium phosphide substrates (used to make these chips) has increased from 8,000 yuan to 25,000 yuan, a 250% rise.
- Small Parts: A Missing Component Can Stop the Entire Production Process
Prices for high-speed connectors have risen by 200%, and liquid cooling fittings by 150%; even capacitors and inductors are in short supply. A foundry stated, "Apart from the chassis, every component is in short supply. If a small chip that costs a few dollars is unavailable, a server worth hundreds of thousands of dollars cannot be delivered."
2. Why Are Prices Rising So Much? Wild Demand, Slow Expansion, and a Vicious Cycle of Hoarding
This is not just a simple cycle of demand and supply; rather, it's a complex web of interlocking issues:
- Exponential Demand: Everyone Underestimated It
The parameters of large models are getting larger (GPT-5 uses 500,000 GPUs, compared to 50,000 for GPT-4), and the demand for inference is 10-20 times that of training. For example, an AI application with 1 billion daily active users requires ten times the computing power needed for training. Demand is growing like a snowball, while supply growth is linear (it takes 3 years to build a new wafer factory, and it takes another 18 months to improve chip yield).
- Expansion Cannot Keep Up
Advanced wafer factories need $20 billion in investment to start mass production in 3 years; new chip manufacturers may take 3-5 years to enter the market; indium phosphide substrate production capacity can only be increased by more than 18 months. The most critical parts are the hardest to expand.
- Panic Hoarding Exacerbates Shortages
Large companies (Microsoft, Google, domestic firms like BAT) are securing production capacity for the next 1-2 years, and distributors (such as those in Huaqiangbei) are speculating on prices. Even small customers are stocking up in advance, creating a cycle of "rising prices → hoarding → greater shortages → further price increases."
3. Who Makes Money and Who Loses? Upstream Manufacturers Profit, Consumers Bear the Cost
This inflation is a redistribution of wealth:
- Winners: Core Upstream Manufacturers
NVIDIA's data center business generates annual revenue of $150 billion with a gross margin of 75%; Samsung and SK Hynix' combined profits from storage exceed $400 billion (higher than the GDP of many countries); the net profit of optical module manufacturer Accelink increased by 200% in half a year. As long as there are bottlenecks in production capacity, manufacturers can set prices arbitrarily.
- First to Pay: Internet Giants
Microsoft's capital expenditure is expected to increase by 75% in 2026 (to $80 billion), but due to price hikes, the same amount of money will only buy 60% of the required computing power. This is a "arms race"; not investing means falling behind, so they have no choice but to bear the higher costs.
- Final Payers: Everyone
Cloud services (AWS, Alibaba Cloud) are raising prices by 30%, and AI application APIs are no longer seeing price reductions. Mobile phones and computers are also seeing price increases due to rising memory costs. Anyone using cloud services, engaging in AI applications, or buying new devices is contributing to this inflation.
4. Three More Dangerous Risks Than Rising Prices
The price surge is just the surface; the underlying issues are more severe:
- Stifling Small and Medium-Sized Innovations
The high cost of computing power is a critical barrier for startups, with monthly bills in the hundreds of thousands of dollars being a lifeline or death sentence. Many AI startups have already gone out of business in 2026, turning AI innovation into a closed-game for giants and stifling the vitality of the ecosystem.
- Fragmented Supply Chain
The US is sanctioning Chinese chip companies while demanding that South Korea share storage profits, exacerbating global distrust in the supply chain. China is accelerating its efforts to localize production (with companies like Yangtze Memory and ChangXin expanding capacity), leading to a shift from a global to a regional semiconductor landscape, resulting in higher costs.
- Potential Overcapacity by 2028
All manufacturers are currently expanding production at full speed, and capacity will be released in 2027-2028. If AI applications do not grow as expected, demand may not keep up, leading to a semiconductor recession similar to the 70% drop in memory prices in 2019.
Conclusion: The "Affordable Era" of AI Is Far from Here
The AI revolution, like the rail and power revolutions, will inevitably experience hardware shortages and price increases in its early stages. However, this time the scale is larger and the cycle longer. We are paying the entry fee for this revolution, but whether it will be worth it in the long run depends on whether AI applications can truly create value. For now, AI services will not be free; they will only become more expensive—this is a reality we must confront.