0

TSMC’s leading-edge chips and advanced packaging are spoken for

In sum – what we know:

• Sold out into 2027 – TSMC’s 3 nm wafer starts and CoWoS packaging are effectively fully booked, with packaging the real chokepoint rather than bare silicon.
• The Fabless Four win – Nvidia, AMD, Apple, and Broadcom secured early allocations, with Nvidia alone reserving roughly 60% of CoWoS capacity and Apple over half of initial 2 nm.
• Startups and mobile lose – New AI hardware entrants face long lead times, while Qualcomm and MediaTek get pushed off 3 nm as TSMC reprofiles capacity toward data-center chips.

There’s no single bottleneck when it comes to AI. A few years ago, perhaps there was a bottleneck related to talent. Then there was data. Then memory. Now, there’s another chokepoint that sits inside a handful of fabs in Taiwan. TSMC’s leading-edge lines — 5 nm, 3 nm, and 2 nm — along with its advanced packaging capacity are effectively sold out for AI workloads well into 2027. There’s no slack left, and the companies that didn’t book early are finding that out.

That scarcity is concentrating power and profit into a small group of dominant chip designers and memory vendors, the kind that signed multi-year contracts and prepaid for capacity before the rest of the industry understood what was coming. Nvidia is the obvious winner. AMD, Apple, and Broadcom aren’t far behind. Meanwhile, system companies are now quietly exploring secondary manufacturing at Intel and Samsung — partly as insurance against TSMC’s wait times, partly as a hedge against the geopolitical risk of having so much of the world’s compute tied to a single island.

And the squeeze runs the other way too. Late-arriving AI competitors, PC chipmakers, and smartphone vendors are being pushed off the leading nodes entirely as TSMC reprofiles its capacity toward the customers that pay the most. 

The 3 nm and packaging bottleneck

TSMC’s advanced-node utilization has reportedly reached the highest level since the pandemic-era shortages of 2021–2022. When a foundry is running its bleeding-edge fabs that close to flat-out, there’s no room left for anyone who shows up late. N3 wafer starts are said to be fully booked through at least Q2 2027, which means new high-end AI designs simply can’t get a meaningful slot until then. Lead times for custom AI chips have stretched to roughly 12 months — effectively pushing new entrants’ launches back by six months to a year.

The reason for the squeeze is straightforward. AI accelerators, host CPUs, and networking silicon are projected to consume roughly 60% of all N3 output in 2026, rising to about 86% in 2027. That nearly displaces smartphone and PC chips from 3 nm altogether, forcing consumer SoCs to either stay on older nodes or go elsewhere. TSMC prioritizes AI customers for reasons that aren’t complicated. Those chips are physically larger, command higher average selling prices, and come with multi-year purchase commitments, compared with the slower-growing and more price-sensitive consumer markets.

But wafer starts aren’t actually the tightest constraint. The real bottleneck is advanced packaging — CoWoS and SoIC specifically — which overrides bare wafer capacity. You can stamp out the silicon, but if you can’t package it into a finished accelerator, you don’t have a product to sell. This is why TSMC’s own CoWoS lines, along with key OSATs, have become the central chokepoint in the whole AI hardware chain. Designers have increasingly steered packaging work toward partners like Powertech precisely because proximity to TSMC’s CoWoS lines and a proven track record translate into more predictable schedules.

The next node is already going the same way. TSMC’s 2 nm process began mass production in late 2025, and is heavily pre-allocated. Apple has reportedly secured more than 50% of the initial 2 nm capacity, with the largest AI players placing big, long-dated orders for what’s left. That locks smaller competitors out of the next leading node before they’ve had a chance to bid for it.

All of this underlines just how far ahead TSMC currently sits. Its record profits and pricing leverage are a direct reflection of a process lead that Intel and Samsung can’t yet challenge at the absolute bleeding edge — not because they lack ambition, but because their yields aren’t there. Scarcity, in other words, has handed TSMC even more power than its technology alone would.

Winners: Nvidia, AMD, and the memory ecosystem

Nvidia is the clearest beneficiary of the entire situation, and it’s not especially close. The company controls roughly 80% of the global AI accelerator market, and long-term preferential access to TSMC capacity amplifies that dominance rather than just protecting it. TSMC disclosures and analyst estimates suggest Nvidia has booked over 800,000 wafers for calendar 2026 across its Blackwell and successor architectures. On the packaging side, the figures are even more striking — Nvidia is said to have reserved somewhere around 60% of TSMC’s total CoWoS capacity by some estimates.

What that buys Nvidia is pricing power on its accelerators, the freedom to time its product cycles without worrying about being supply-constrained relative to anyone else, and a moat against the second-tier vendors who simply can’t secure comparable capacity or timelines. The crunch, in effect, reinforces the lead Nvidia already has.

AMD, Apple, and Broadcom round out the group that analysts have taken to calling the “Fabless Four,” and all three are successfully bypassing the bottleneck thanks to early allocations. AMD’s MI-series accelerators and high-performance CPUs ride TSMC’s N5, N4, and N3 lines. Apple’s control of early 2 nm capacity secures its high-end mobile and PC silicon and underpins its on-device AI plans. Broadcom leans on TSMC for high-end networking and custom cloud silicon. When TSMC beats its guidance, the explanation is usually that these names — plus the big cloud buyers — are pulling more chips than expected. They’re treated as direct demand proxies for a reason.

The hyperscalers belong in this group, though with an asterisk. Nvidia, AMD, and the hyperscalers’ custom-silicon teams have collectively placed over US$12 billion in advanced foundry orders for delivery beyond mid-2027, ensuring they’ll have silicon even as the crunch deepens. They’re winners to the extent they reserved early. But they’re still constrained by the same packaging bottleneck as everyone else, and any late-planned project still faces the same 18–24-month lead times — being a giant doesn’t fully exempt you.

Memory is having an excellent run too. The AI build-out that TSMC’s GPUs enable directly lifts DRAM vendors, because high-bandwidth memory and large DDR stacks are mandatory for AI servers. Micron, SK Hynix, and Samsung’s memory division are all riding booming demand and rising prices tied to AI server configurations. Server OEMs, power and thermal component makers, and high-speed networking vendors benefit from the same wave. And on the packaging side, OSATs with the right capabilities have become quiet winners — Powertech is frequently cited for its ability to absorb runs when TSMC’s own lines are full, while ASE and other advanced packaging houses pick up business as designers try to avoid single-point failure.

The throughline here is consolidation. Locked-in TSMC supply functions as a near-insurmountable moat for the incumbents, and second-tier AI chip vendors without equivalent manufacturing priority are on the wrong side of it.

Losers: Startups, mobile chips, and lagging foundries

If you’re a new AI hardware startup, the math is brutal. Companies like Groq and Cerebras are facing the same 18–24-month delays as everyone else, which severely compresses their market-entry windows. With 3 nm sold out into 2027, a startup can’t scale to volume quickly even with a genuinely strong architecture — and by the time it launches, it’s likely competing against Nvidia or AMD products that already have several stable generations and deep software ecosystems behind them. TSMC scarcity, in practice, acts as a barrier to entry that favors the incumbents it’s already serving.

Smartphone and PC SoCs are getting pushed off 3 nm as that capacity gets reprofiled for AI. Qualcomm and MediaTek face real displacement pressure here, since TSMC is choosing data-center silicon over mobile flagships. The pain is sharper still for smartphone-skewed RF suppliers, who benefit little from the AI boom while also dealing with weaker consumer demand and a lack of access to leading nodes that have been repurposed away from them. It’s a bad combination — squeezed on supply and demand at the same time.

Intel sits in a genuinely awkward spot. It’s trying to claw back process leadership with its “five nodes in four years” push, yet it remains a major TSMC customer for key tiles in its consumer and server products. So it’s both competing with and dependent on the very foundry that’s reinforcing Nvidia’s and Apple’s lead — a frenemy relationship in the most literal sense. Its own foundry division can’t yet absorb high-volume AI overflow, which leaves it watching the boom from an uncomfortable distance. 

To be fair, not everyone shut out of the bleeding edge is in trouble. Non-AI logic producers and peripheral chipmakers can still reliably secure manufacturing on mature nodes, which aren’t subject to the same crunch. Plenty of perfectly good chips don’t need 3 nm. For those producers, the AI scramble at the leading edge is mostly somebody else’s problem.

Winners: Intel and Samsung as fallback options

Here’s where the story gets more interesting, because the same two companies losing at the bleeding edge are starting to win at the margins. AI resource scarcity and Taiwan-based geopolitical risk are driving major chip designers to court second-source foundries in a way they weren’t a couple of years ago. When TSMC is at 94% utilization with 3 nm booked through Q2 2027, even a foundry that lags on yield starts to look like a reasonable Plan B. The supply constraints, as one industry analysis put it, are handing Intel and Samsung unexpected advantages simply because they’re the only other manufacturers capable of advanced logic at scale.

The most concrete evidence is Samsung’s recent design wins. It has secured Tesla’s AI5 and AI6 chip programs, a prominent AI win landed outside Taiwan. More notably, Samsung has reportedly entered Nvidia’s data-center supply chain in a diversification capacity — which, if accurate, means even Nvidia is hedging its manufacturing and packaging partners rather than betting everything on TSMC. 

Intel, meanwhile, is aggressively promoting its Foundry Services to external customers, leaning hard on its upcoming 18A process and its advanced packaging technologies — Foveros and EMIB — as future pillars of an AI strategy. Coverage of TSMC’s constraints repeatedly notes that Intel’s foundry announcements carry added weight in this climate, simply because customers want a credible alternative on the table. The catch is timing. Analysts point out that customers needing capacity right now tend to look elsewhere, toward suppliers that integrate seamlessly with their existing flows and have proven AI packaging track records. Intel may evolve into a strong alternative; at present, it doesn’t fit the immediate needs of most AI buyers.

As TSMC concentrates on data-center AI chips, demand for less-sophisticated CPUs and smartphone-class parts that don’t need the absolute bleeding edge is carrying over to Intel on the PC side and to Samsung as well. That’s not glamorous work, but it’s volume, and it gives both companies a real role in the broader AI computing stack even when they’re not making the flagship accelerators.

It’s worth being clear about what this courting actually is, though. So far it’s strategic hedging rather than wholesale migration. Big buyers are signing pilot or secondary programs — Tesla at Samsung, exploratory IFS engagements — to guarantee some capacity outside TSMC. Smartphone and PC makers pushed off N3 are evaluating Intel and Samsung for future nodes where slightly lower performance is acceptable. Some companies are splitting designs, keeping the most performance-sensitive parts at TSMC while moving supporting chips or future variants elsewhere.

And the underlying constraint hasn’t gone away. Neither Intel nor Samsung currently matches TSMC on high-end logic yields, packaging track record, or the kind of seamless design-ecosystem integration that major AI buyers expect at scale. For mission-critical, top-performance accelerators, TSMC remains the default. Most of the courting to date has produced partial diversification, not a real exodus.

The future of AI compute

A huge share of global AI compute now depends on a single company in a single geography, and a handful of customers — Nvidia, AMD, Apple, a few hyperscalers — have effectively pre-booked the lion’s share of the world’s cutting-edge logic and packaging. That’s a national security question as much as a business one, and governments worried about supply-chain resilience have noticed. Trillions in broader market activity hinge on whatever TSMC’s packaging output allows in any given quarter, which is a remarkable amount of leverage for one foundry to hold over the entire AI and tech trade.

The competition angle is just as live. Nvidia’s long-term TSMC contracts don’t only secure its own supply — they consolidate a hardware advantage that lets incumbents dictate product cycles without the supply limitations that constrain everyone else. Regulators may eventually start asking whether exclusive heavy foundry allocations, combined with prepayment and multi-year deals, function as a structural disadvantage for emerging open-source efforts and startup challengers, even ones with genuinely competitive designs. A great architecture doesn’t help much if you can’t get it built for two years.

The policy response is already underway. Western and allied industrial policy is heavily subsidizing Intel Foundry in the US and Samsung in Korea and the US, in what amounts to a real-world test of whether public money can build competitive resilience. The open question is whether those subsidies can accelerate Intel’s 18A roadmap to true competitiveness, or help Samsung close an 8-to-10-point yield gap fast enough to become a genuine second source for high-end AI.

For now, the honest answer is that the gap is widening, not closing. Even with billions in state subsidies aimed at decentralizing AI compute, TSMC keeps extending its immediate timeline advantage during the exact 2024–2027 window that matters most. Companies are courting Intel and Samsung in earnest, and they probably should be — but the courting is a hedge against a future that hasn’t arrived. The present still belongs to TSMC, and to the small group of customers smart enough, or rich enough, to lock up its capacity before everyone else realized how scarce it would become.