Build1 distinct publisher3 min readUpdated
Taiwan and South Korea are splitting the AI server chain because demand exceeds what either can build. For 2026 supply forecasts, the binding step is the handoff, not any single fab.
The Engineer · Build desk

Compiled by The EngineerSomething wrong?How this is made
Taiwan's statisticians have already published a volume forecast for AI hardware, and it is filed under GDP. The official projection is 11.05% growth in 2026 [10] against 8.7% last year [9]: an acceleration of about 2.35 percentage points [1], and on the agencies' own reckoning the fastest expansion in 39 years [11], which puts the last comparable year near 1987 [2]. Nomura's Jeong Woo Park ranked the island first in the world by share of AI hardware production in a May report [8], and AI exports are what carried last year's figure [9]. A statistics office does not underwrite double-digit growth unless it expects the racks to keep leaving the port.
For anyone modelling availability rather than sentiment, the interesting part is the handoffs. Korean memory is joined to TSMC logic inside Taiwan using advanced packaging [6], after which Taiwanese firms build the accelerators into racks with their cooling systems [3], drawing on a local network of power supply, packaging and thermal specialists that lets a customer assemble a chain from several vendors [19]. Two national supply bases, and the slowest station sets the delivery date. Memory is the tightest part of it: on the reported concentration, roughly a fifth of world HBM output comes from anyone other than the two Korean suppliers [3], so a buyer without an allocation there has little to fall back on. Jensen Huang's June inscription on an SK Hynix wafer, asking for more production [4], reads as the largest customer naming the step it believes is short.
The division of labour is newer than it looks. Samsung was the exclusive maker of iPhone processors until 2015, before Apple moved to TSMC and made it the primary supplier [7], and Owen Lin's account is that cooperation between the two industries was rare while competition was the norm [2]. Park now describes them as increasingly complementary [13], and Yu-Ning Chiu's summary is that no country can do all of it alone [16], while Colley Hwang of DIGITIMES prefers "frenemy" and calls Taiwan Korea's corridor to the United States [15]. Both readings hold as long as orders exceed combined capacity. Sung Soo Eric Kim, quoted by CNN, framed it as sequencing: demand is too large for the old fight, so securing supply comes first [20].
What sits underneath is unsettled. Taiwan passed South Korea on GDP per capita in 2025 for the first time in more than two decades [12], Seoul has a $576 billion AI and semiconductor investment plan in hand [18], and the reporting is explicit that the partnership has not dissolved political and economic friction [17]. For next year, that evidence supports one supply claim and no more: volumes rise through a shared pipeline whose narrowest section is held by two companies in one country. There is nothing here that supports a price call, and anyone selling you one is not working from this material.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Taiwanese and South Korean companies are supplying the global market with AI accelerators, and demand from the largest technology companies is high enough that neither side can handle the production volume alone.
Yu-Ning Chiu says no country can do all of it on its own.
Owen Lin says cooperation between the Taiwanese and South Korean industries happened very rarely, and that competition was far more widespread.
SK Hynix and Samsung together produce about 80% of the world's High Bandwidth Memory.
Sung Soo Eric Kim, quoted in the CNN reporting, said there is no time for the old-fashioned conflict, that demand is so great that the two sides must cooperate, and that securing supply comes first. The wording reaches us through a Ukrainian-language rendering of the CNN report, so it is paraphrased rather than quoted verbatim.
Until 2015 Samsung was the exclusive manufacturer of iPhone processors; Apple later began working with TSMC, which became its main chip supplier.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single translated relay, no primary documents
Everything in the cluster comes from one publisher's Ukrainian-language rendering of a CNN report. Key numbers (80% HBM share, 8.7% growth, 11.05% forecast, $576bn plan, record trade surplus) are stated without links to the Nomura report, Taiwanese statistical release, or trade data, and quotes are paraphrase rather than verbatim. The structural claims are plausible and internally consistent, but nothing here is independently corroborated inside the supplied material.
Arrangement already in production, thinly quantified
The division of labour described is not a proposal but current production practice: Korean HBM feeding TSMC advanced packaging and Taiwanese rack integration for Nvidia AI servers, with a concentrated ~80% memory duopoly and macro prints consistent with large realised AI hardware exports. Adoption is therefore high in substance, but the supplied source gives no capacity, unit, or shipment figures, so the level cannot be scored higher.
Framing outruns the supplied numbers
The cluster's framing - that the handoff between Korean memory and Taiwanese packaging sets the accelerator clock and is the binding step for 2026 supply - is stronger than the evidence offered. The source supports a concentrated, interdependent chain but provides no capacity or throughput data that would establish which step binds, and the most eye-catching figure (11.05% growth, fastest in 39 years) is an official forecast, not an outcome. Modestly overstated rather than fabricated, since the underlying structural facts are real.
Interested voices, undisclosed positions
The quoted evidence base is dominated by parties with commercial stakes in the narrative: a sell-side analyst at Nomura, an industry research firm (DIGITIMES) that sells supply-chain intelligence, a national investment programme being promoted at a government presentation, and Nvidia's CEO publicly urging a supplier to raise output. None of these positions or relationships is disclosed in the piece, and the publisher is aggregating another outlet's reporting for traffic rather than doing original verification.
Low - structure credible, specifics unverified
Confidence is limited by single-source, translated sourcing with no primary documents, several quotes that are paraphrase, and two claims that fall short of support. The structural picture (concentrated Korean HBM, Taiwanese packaging and rack integration, mutual dependence under excess demand) is coherent and consistent with the quantified items given, which keeps confidence from falling further.
build
China's accelerator swap makes Cambricon supply, not export policy, your ship-date risk1 distinct publisher
product
Samsung's July foundry price rise moves the AI crunch into your bill of materials1 distinct publisher
product
Waymo opens the trunk: a 5nm ASIC, a quadrillion ops, and a supplier list rivals can price3 distinct publishers
invest
Micron's $10B Boise lab admits fabs alone will not close the HBM gap1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
mezha.net
1 article · August 23, 2026