TSMC's Record High Signals AI Chip Demand Isn't Peaking

TSMC Just Broke a Record. The AI Demand Story Is Far From Over.

Taiwan Semiconductor Manufacturing Company hit a record stock price this week. Shares, up 61% year to date, closed at $467.84 on October 7 before edging lower overnight. The move matters less for what it says about one company and more for what it signals about the entire architecture of artificial intelligence spending right now.

Citi analyst Atif Malik raised earnings estimates for TSMC through 2028 in a note published Wednesday. He projects more than 40% year-over-year revenue growth for the company through next year. That forecast rests on three converging trends: continued AI compute demand, the emerging shift toward agentic AI systems, and a new networking cycle around co-packaged optics, or CPO.

The last of those is the detail most traders are missing.

Why Co-Packaged Optics Changes Everything

Co-packaged optics moves the network components that carry data between chips onto the same substrate as the processor itself. The technology reduces power consumption and latency at a scale that traditional networking can’t match. For AI clusters — the massive groups of chips working together on inference and training — that reduction compounds across thousands of interconnections.

TSMC is positioned to manufacture the advanced packaging infrastructure that CPO requires. That means the company isn’t just benefiting from more AI chips being built. It’s benefiting from a fundamental redesign of how those chips communicate.

Malik flagged this as a structural upside factor that extends well beyond the current generation of GPU-based AI workloads. That’s important context for anyone trying to pin down when the capex cycle peaks.

The Customer Base Is Widening, Not Concentrating

A common worry about the AI boom has been that it depends on a handful of hyperscalers pulling the trigger on spending. TSMC’s customer guidance complicates that narrative.

Nvidia, TSMC’s largest customer, has guided to a doubling of its order intake and over 70% year-over-year revenue growth in fiscal year 2028. AMD expects an 80% compound annual growth rate in its data center AI business. Broadcom projects its AI semiconductor revenue will nearly double in the same period.

These are not isolated companies chasing a single use case. They’re building competing architectures — custom silicon, alternative chip designs, different approaches to AI networking and inference. Each one is a TSMC customer. Each one requires advanced node capacity.

Malik noted that TSMC’s AI revenue growth could nearly double next year specifically because most of its AI semiconductor customers are reporting accelerating growth themselves. That’s a cascading effect, not a single buyer spike.

The Node Transition Is a Supply Tightener

Here’s another detail the market may be underweighting. TSMC’s next-generation process nodes are coming online in a sequence that should keep supply constrained.

AI compute workloads are moving from TSMC’s N3 node to N2 starting in the second half of 2026. Then N2 and A16 begin contributing to revenue in 2028. Transitioning production between advanced nodes is rarely smooth. Yield ramp-up, requalification, and packaging changes create bottlenecks — especially when multiple customers are making the switch simultaneously.

Tight wafer foundry supply, as Malik put it, supports sustained demand growth. In practice, that means TSMC can maintain pricing power and capacity allocation advantages for longer than a typical node transition would suggest. Competitors without access to equivalent advanced manufacturing can’t easily close the gap during that window.

What This Means for the Capex Cycle

The bigger question is whether this spending can sustain. Traditional semiconductor cycles run 18 to 36 months. AI capex has so far defied that rhythm, driven by a combination of competitive urgency among cloud providers and the physical constraints of advanced lithography.

TSMC itself is likely to increase capital expenditure over the coming years, Malik said, supported by the revenue outlook. That means the company is preparing for demand that exceeds current capacity, not demand that is already saturating it.

The numbers are specific enough to ground the argument. A 50% quarterly revenue jump isn’t a sideways trendline pretending to be momentum. A 40%+ growth projection into 2027 isn’t a best-case scenario — it’s the base forecast from a major bank covering the stock. Nvidia’s own guidance to double order intake sets a floor, not a ceiling, on what the ecosystem can absorb.

Who Wins and Who Loses

TSMC wins clearly. It captures margin and volume from a widening set of customers across multiple design architectures. Its foundry business benefits regardless of which AI chip winner emerges.

Nvidia and AMD win too, though with a caveat. Their growth depends on TSMC delivering enough advanced capacity at the right nodes. If the company stumbles on N2 yield or CPO integration, the downstream customers feel it first.

Traditional foundry competitors — Samsung, GlobalFoundries — lose relevance if they can’t match the pace of node transitions. Intel’s foundry ambitions face the same structural pressure.

Infrastructure players benefit. Power supply manufacturers, cooling system providers, and component suppliers tied to co-packaged optics all gain from a cycle that isn’t about raw chip count alone but about denser, more integrated systems.

What Comes Next

TSMC reports earnings on October 15. The consensus upgrades Malik flagged haven’t fully priced in yet, which is why the stock still has room. The key numbers to watch are management commentary on N2 ramp timelines, any mention of CPO adoption schedules, and guidance on whether capex plans are accelerating beyond market expectations.

The AI semiconductor market is projected to double next year. TSMC’s position at the center of that expansion makes it the closest public proxy for whether the cycle has genuine longevity or is running on momentum that will fade by early next year.

So far, the data suggests the former. The record price isn’t the peak. It’s a checkpoint.