Invest1 publisher2 min readPublished
Intel's CEO puts the fix for customers' 95% reliance on TSMC on the customers who chose it
Lip-Bu Tan told a Splunk keynote in Denver that relying on one Taiwan-based company for 95% of AI chip production is very risky. The one 14A commitment he can point to, Terafab, ships its first product in mid-2028.
The Investor · Invest desk
What happened
- Intel chief executive Lip-Bu Tan told a Splunk-hosted keynote in Denver that relying 95% on one company, especially one based in Taiwan, is very risky, and said orders must be spread among competitors.
- Counterpoint put TSMC above 70% of global foundry revenue in the first quarter, far ahead of Samsung Electronics, China's SMIC and Intel.
- Intel had yield problems on its 1.4-nanometer-class 14A process and agreed in April to supply it to the Terafab project led by Tesla and SpaceX, whose first product is due in mid-2028.
- Tan said CPU demand continues to surge and predicted the current shortage will persist as the number of AI agents used for inference grows into the trillions.
- Intel hired Lee Seok-hee, the former chief executive of SK hynix, in June to run advanced packaging and back-end technology development and manufacturing in its foundry division.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- decision The engineering and schedule cost of adding a second leading-edge source falls on Nvidia, AMD and Apple, which currently route nearly all AI chip production to TSMC; better Intel yields still leave the decision with them.
- constraint Intel is courting external foundry customers on the same capital and capacity its own CPU line is short of, at a product where Tan says half of demand goes unserved.
- cost Substrate volume is being locked up by prepayment, so a diversifying buyer commits cash before wafers arrive.
- precedent With the same 95% figure already circulating in US politics, where a designer places leading-edge orders becomes a decision it may have to defend in Washington as well as in engineering review.
Tan's case is a demand-side case. The fix he described sits in other companies' procurement departments. He was candid about why the supply side is hard: "You have to manage yield, defects, cycle time and variability thoroughly to make output predictable," he said [8]. A foundry also needs the right intellectual property and support for electronic design automation, because customers arrive with different requirements, and in his ordering, "Most important, you have to provide advanced packaging as well" [26]. He said the business demands heavy capital spending [23]. On why Intel is in it at all: "The ability to do not just design but also manufacturing and advanced packaging is very important" [7].
The 95% in his sentence and the number Counterpoint measures count different things. Counterpoint's covers global foundry revenue in the first quarter, and on that basis under 30% of industry revenue is split among Samsung, SMIC, Intel and everyone else [6]. The 95% describes AI chip production. It also has a political history: in an interview before the presidential election, Trump said chip companies "stole our business, 95% of it, and now they are in Taiwan" [18].
Against the argument, the commitment Intel has on the record arrives in mid-2028, roughly 21 months after the Denver keynote [11].
One route runs through packaging. "Everything is moving more and more toward a system approach and packaging, and that is the future," Tan said [20]. Intel is positioning EMIB, which links chips through silicon bridges so that several operate as one, against TSMC's CoWoS [21]. If the binding constraint moves to packaging, a customer can buy Intel's back end without moving wafer production, and I'd expect packaging to be the likelier route to first external volume. The other runs through Washington, where the report says Tan's remarks align with the Trump administration's push on foundry volume [19].
What would show buyers actually paying for diversification is a second named 14A customer with a product date before 2028. Or a prepayment reaching Intel of the kind Tan described. "One of the biggest challenges is the substrate," he said, adding that "Japanese and Taiwanese companies are securing enough volume through prepayments" [12]. On progress so far, Tan said: "The good news is that things have improved over the past 18 months" [22].
What to watch
- A second named 14A customer, especially one with a product date earlier than mid-2028.
- Whether prepayments for substrate volume start reaching Intel's supply chain.
- Whether Intel and SK hynix cooperate on an HBM base die, which the report flags as industry speculation.