High gross margin in chips doesn't equate to high profits

Looking at the AI industry chain separately, the most eye-catching segment upstream is chips, especially GPUs and accelerators. The underlying logic of value capture is simple: computing power is still a scarce commodity, and the software-hardware ecosystem, development tools, and customer migration costs amplify 'first-generation leadership' into 'multi-generational leadership.' This is also why the leading GPU designer can achieve a gross margin level similar to platform-type businesses: NVIDIA disclosed in its earnings release that the FY2025 year-end GAAP gross margin guidance is approximately 70.6% (non-GAAP approximately 71%). This figure implies that from a value chain perspective, upstream chipmakers have an extremely thick profit margin per 'unit product,' and they hold greater pricing power when supply and demand are tight.

However, the 'high gross margin' in the chip segment does not necessarily equate to the 'largest profit pool.' The reason lies in two ceilings: the first is supply chain bottlenecks (advanced packaging, HBM supply, capacity allocation), and the second is customer pushback on pricing (cloud providers developing their own accelerators, diversified procurement). As such, the upstream sector captures 'scarcity rents' with high gross margins, but profit fluctuations tend to follow economic cycles and supply-demand dynamics. For investors, the core metrics for chips are whether they can 'maintain ASP and gross margin,' and whether capacity/packaging constraints prevent incremental demand from fully translating into revenue.

High-efficiency foundry, highly leveraged cloud

Moving down from chip design to manufacturing and system integration, the logic of value capture begins to diverge: one side is 'capital-intensive, reliant on yield and scale' advanced manufacturing, and the other side is 'engineering-heavy, reliant on resource allocation' server and data center systems. Taking wafer foundries as an example, the market often measures their efficiency and bargaining power through gross margin/operating margin: Taiwan Semiconductor's 2024 annual report disclosed a gross margin of 56.1% and an operating margin of 45.7%. This is extremely rare in the manufacturing industry, reflecting the technical barriers of advanced processes and customer stickiness, allowing foundries to earn more than just processing fees while having a certain degree of pricing power at advanced nodes. From a 'profit pool' perspective, although the unit gross margin of foundries is lower than chip design, their massive capital expenditures and large revenue base make their profit pools significant; at the same time, their risks resemble long-cycle capital expenditure gambles: once demand slopes flatten or competition intensifies, depreciation and utilization rates will magnify profitability elasticity on the downside.

More noteworthy is how the "system layer" is reshuffling traditional hardware profits. AI training/inference isn't just about buying a single GPU; it involves an entire rack system, networking, cooling, power, and software stack. The value capture at this layer presents two outcomes: for general OEM/ODM players, it's mostly about high volume with thin margins; but for cloud providers with software platforms and customer access points, the system layer becomes a "leverage point" to turn capital into long-term cash flow. Therefore, you see cloud giants being extremely aggressive in capital expenditure — for instance, a recent report by the Financial Times highlighted Amazon’s plan to make large-scale capital investments to advance AWS’s AI strategy, essentially locking in pricing power for the next wave of "cloud computing supply."

Who holds the customers and billing?

What ultimately determines where profits flow in the AI industry chain often lies downstream at the "billing point": whoever can turn computing power into sustainable subscriptions, usage-based billing, or embedded value-added services will capture a more stable profit pool. The cloud is the most direct example: Amazon's latest earnings release disclosed that AWS’s segment operating profit for 2025 (according to the reported figures) was $45.6 billion, compared to $39.8 billion in 2024. Meanwhile, Amazon’s Q4 2024 earnings announcement showed AWS’s 2024 segment revenue at $107.6 billion. Comparing the two, AWS’s segment operating margin appears to be above 30% (approximately 37%) — this "high operating leverage" is the essence of cloud value capture: heavy hardware depreciation, but once utilization rises and services are layered on, incremental revenue converts faster into profit.

The model layer represents another form of "profit pool competition." Large models themselves may not naturally have high gross margins because inference costs and price wars can erode profitability; however, once models are deeply integrated with workflows, data, and enterprise software, they can become "allocation rights": rather than just selling model APIs, AI gets bundled as an upcharge package into existing subscriptions like CRM, Office, customer service, development tools, etc., redistributing profits towards those who own customer relationships. This is why Microsoft emphasized revenue and operating income growth in its FY2024, pointing out that Azure’s scale is driving margin improvements (even though gross margins are structurally impacted by heavy AI infrastructure investments).

(Chip and Computing Power Series No. 32)

$NVIDIA (NVDA.US)$ $Taiwan Semiconductor (TSM.US)$ $Advanced Micro Devices (AMD.US)$ $MetaX Integrated Circuits (688802.SH)$

48K Views