AI Demand Is Turning Foundry Capacity Into Pricing Power

The August 14 development

The new signal is pricing, not another demand forecast

For most of the AI infrastructure boom, semiconductor scarcity has been described through waiting lists, capacity expansion, utilization, advanced-packaging constraints and management forecasts. SMIC’s latest update adds a more commercial signal: it says it has raised prices for its most sought-after capacity, with increases negotiated after the first quarter being applied to wafers processed in the third quarter.

That matters because utilization and price answer different questions.

Utilization tells us how hard the factory is running. A foundry can report utilization above 90% and still lack enough bargaining power to change customer economics if contracts are fixed, competitors have spare capacity or demand is expected to fade.

Pricing tells us whether scarcity is changing the transaction. When customers accept higher prices for constrained capacity, the shortage is no longer only a production statistic. It is affecting the terms under which chips are manufactured.

This should still be interpreted narrowly. Reuters reports that SMIC’s shipment growth was driven mainly by AI-fuelled demand for chips other than CPUs and GPUs, mostly from China-based customers, as well as some earlier-than-expected orders. The story is therefore not “SMIC is suddenly supplying the world’s leading AI accelerators.” It is that AI investment is pulling on a much wider semiconductor supply chain.

What SMIC actually reported

SMIC’s second-quarter operating picture was already tight before the price discussion became the August 14 headline.

SMIC also added 8,000 wafers per month of 12-inch capacity in the second quarter. Management expects third-quarter revenue to rise another 2% to 4% sequentially while shipments continue increasing.

The profit number needs a separate caveat. Profit attributable to shareholders rose to about $479.2 million, but management said the increase was also helped by a one-time gain from a subsidiary. That makes revenue, shipment volume, utilization and realized pricing more useful for understanding the supply-demand balance than the headline profit jump by itself.

Another important distinction: the reported 5.7% increase is the change in Q2 average wafer selling price. It is not a disclosed blanket percentage increase for every Q3 customer or process. The company says it negotiated higher pricing for sought-after capacity; it has not said every wafer is being repriced by the same amount.

Why 93.7% utilization changes the economics of a foundry

A semiconductor fab is not like a software service that can add capacity by starting another virtual machine. Output depends on installed equipment, cleanroom space, process recipes, cycle times, maintenance schedules, yield, materials and customer qualification.

At lower utilization, a foundry can often absorb additional orders by using idle tools or filling open production slots. As utilization moves into the 90s, the remaining capacity is less flexible. The factory still needs maintenance. Different products use different tool combinations. Some process steps become bottlenecks before the entire fab reaches a theoretical maximum.

This creates a nonlinear effect. Moving from 70% to 75% utilization can be operationally manageable. Moving from roughly 94% toward the practical limit can force harder choices about allocation, lead times and which products deserve scarce tool time.

That is where pricing power can emerge.

None of this guarantees that prices will keep rising. New capacity, weaker end demand, customer inventory corrections or competitive supply can reverse the balance. But SMIC’s Q3 pricing statement is evidence that the current balance has moved far enough toward the foundry for at least some customers to pay more.

The less obvious AI effect

AI tightness can reach chips that are not AI accelerators

The public AI hardware conversation is dominated by leading GPUs, HBM and advanced packaging. That can make it seem as though AI demand only matters at the most advanced logic nodes.

Real AI systems consume a much wider bill of materials.

Data centers need power-management chips, networking components, controllers, optical and connectivity silicon, storage-related devices and many other semiconductors that do not resemble a flagship accelerator. Servers and networking equipment also contain large numbers of support chips produced on mature or specialty processes.

There is a second spillover mechanism. When advanced international foundries are heavily loaded by leading-edge AI products, other customers can shift less advanced work toward foundries with available capacity. Reuters reported earlier this year that SMIC was seeing some foreign customers move orders back to China as AI demand strained capacity elsewhere.

That means AI can affect a foundry even when the specific wafer being fabricated is not an AI GPU.

This is why SMIC’s own wording matters. Reuters reports that the Q2 shipment increase was mainly driven by AI-fuelled demand for chips other than CPUs and GPUs. That is a broader and more interesting supply-chain signal than a simple “AI GPU demand is booming” headline.

A higher wafer price is not the same as a higher finished-GPU price

A silicon wafer is one layer of the cost stack. A finished accelerator can also require expensive memory, advanced packaging, substrates, interposers, test, assembly, networking and board-level components. Yield at several stages can change how much usable output reaches customers.

That matters for two reasons.

First, a foundry price increase does not map one-for-one into the selling price of a finished chip. A chip designer may absorb some of the increase, negotiate it elsewhere in the supply chain or pass some portion to customers. The answer depends on contracts, product margins and competitive conditions.

Second, the most important bottleneck may sit somewhere else. For leading AI accelerators, advanced packaging and HBM availability can constrain shipments even when logic wafers are available. AIUpdateWatch’s analysis AI Chips Are Becoming Packaging Systems explains why the economic unit increasingly extends beyond the compute die itself.

The supply chain therefore has several prices that should not be collapsed into one:

• wafer fabrication price — what the foundry charges to manufacture the silicon;
• memory price — including HBM or other DRAM components;
• packaging and test cost — increasingly important for chiplet-based AI systems;
• board and system cost — networking, power delivery, cooling and server integration;
• accelerator selling price — the chip vendor’s commercial price to customers;
• cloud compute price — the downstream price of using that hardware as a service.

SMIC’s August 14 signal belongs at the first layer. It can influence the layers above it, but it does not determine them by itself.

New capacity does not solve scarcity immediately—and it carries a cost

SMIC is responding in the expected way: adding capacity and accelerating new production lines. The company’s monthly capacity rose 1.7% quarter over quarter, and it added 8,000 wafers per month of 12-inch capacity during Q2.

But fab expansion has a time lag. Installing equipment is only the beginning. New tools need qualification, process tuning, yield learning and customer approval before they contribute high-quality saleable output at scale.

Expansion also changes the income statement before it necessarily fixes the bottleneck. SMIC said first-half amortization reached $2.3 billion and expects roughly $5 billion for the full year, about 30% higher year over year. First-half capital spending was $3.4 billion.

This is the basic foundry investment cycle:

This is why high utilization is not automatically the same as permanently higher margins. A foundry can benefit from stronger prices while simultaneously absorbing the cost of a major expansion program.

Why SMIC and TSMC should not be treated as interchangeable foundries

It is tempting to read every foundry statistic as if all wafer capacity were fungible. It is not.

SMIC is the only Chinese foundry able to mass-produce 7-nanometer logic such as CPUs and GPUs, according to Reuters. But its overall business spans a broad range of process technologies and products. China accounted for about 90% of Q2 revenue.

TSMC’s current mix is much more concentrated at the advanced edge. In its second-quarter 2026 results filed with the U.S. Securities and Exchange Commission, TSMC said technologies at 7 nanometers and below accounted for 77% of wafer revenue, including 30% from 3nm, 33% from 5nm, 11% from 7nm and 3% from 2nm.

That difference matters. A percentage point of utilization at a mature-node line does not represent the same capability, capital intensity, customer set or economic value as a percentage point at a leading-edge line. Capacity cannot simply be moved from one node to another on demand.

So the correct conclusion is not “SMIC is now as constrained as TSMC” or “SMIC pricing proves the same thing as leading-edge accelerator shortages.” The useful conclusion is that tightness is broad enough to give SMIC pricing leverage within its own capacity mix, while leading-edge foundry and packaging constraints remain separate parts of the global AI supply chain.

Practical implications

What chip buyers should take from the shift

For procurement teams, the practical lesson is to track the bottleneck that applies to the exact product being purchased. A mature-node controller, a 7nm compute device and a 2nm accelerator package can all be affected by the AI buildout, but through different supply chains.

AIUpdateWatch’s N3 and CoWoS explainer covers the leading-edge fabrication and packaging side of that problem. SMIC’s latest pricing adds evidence that the pressure is also visible beyond that narrow frontier.

What to watch next

The most informative next data will show whether this is a short period of tightness or a more durable change in foundry economics.

• Q3 realized pricing: whether higher negotiated prices show up in another increase in average wafer selling price.
• Utilization: whether SMIC stays above 90% as new production lines ramp.
• Product mix: whether AI-linked non-CPU/GPU demand remains the main shipment driver or broadens further.
• New capacity: how quickly added 12-inch capacity becomes qualified saleable output.
• Amortization and margins: whether pricing power is enough to offset the rising fixed cost of expansion.
• Order migration: whether customers continue moving mature and specialty-node work toward foundries with available capacity.
• Downstream pass-through: whether higher wafer economics become visible in component or system prices, rather than being absorbed by suppliers.

If utilization remains high while prices and shipments continue rising, the evidence for durable scarcity becomes stronger. If new capacity ramps quickly and pricing normalizes, the August increase will look more like a cyclical response to a temporary mismatch.

Either way, the August 14 update is useful because it moves the AI supply-chain discussion one step closer to observable economics. Demand is not only filling factories. In parts of the market, it is changing what customers pay for access to them.

Sources and related reading

Current reporting and primary financial evidence

• Reuters — SMIC increases prices as AI-related demand tightens semiconductor supply, August 14, 2026
• Reuters — SMIC Q2 profit, revenue and demand context, August 13, 2026
• Reuters — order migration toward Chinese foundry capacity, May 15, 2026
• Reuters — SMIC capacity expansion and depreciation pressure, February 11, 2026
• TSMC — Q2 2026 results filed with the U.S. SEC, used only for advanced-node mix comparison

Related AIUpdateWatch coverage

• August 14, 2026 Daily AI Report
• AI Chips Are Becoming Packaging Systems
• What Do N3 and CoWoS Supply Constraints Mean for AI Chip Buyers?
• AI Markets