Is the semiconductor industry headed for an almighty crash? There are reasons to believe a bust, if it comes, will be less vertiginous than the boom Oct 15th 2026|5 min read For more than four decades the semiconductor industry has been governed by two laws as dependable as any in physics. The first is that the number of transistors, the tiny switches etched onto a chip, keeps rising. The second is that every boom ends in a bust. The first, which makes chips more powerful, is in rude health. Leading processors for artificial intelligence contain more than 200bn transistors, and the industry reckons it can cram in a trillion by the end of the decade. The second is stirring unease about how long the good times will last. The industry is used to booms, but not one like this. AI has created an insatiable appetite for the chips used to train and run AI models. In the first half of 2026 semiconductor sales grew faster, year on year, than at any time since the 1980s (see chart). This month TSMC, the world’s biggest contract chipmaker, reported its sixth consecutive quarter of record profits; Samsung Electronics, a South Korean chip heavyweight, reported its fourth. Yet investors are growing jittery. Despite blowout results and rosy forecasts, an index of leading semiconductor firms has fallen by over 10% since October 6th. Their fears were not helped when on October 12th Hock Tan, boss of Broadcom, a big chip designer, likened many investors in the AI trade to gamblers in a casino who “have no clue” what they are investing in. Nearly all of the industry’s expected growth hinges on chips for AI, making it a giant bet on the technology. If that bet sours—and the “supercycle”, as some have taken to calling it, comes to an end—will chipmakers suffer a bust as vertiginous as the boom? There are reasons to think not. Even as the semiconductor industry has grown, manufacturing has become far more concentrated. Most chip designers, such as Nvidia, whose graphics-processing units have powered the AI boom, do not make their own silicon. Nor do software firms such as Google and OpenAI, which have also taken to designing AI processors. Instead they rely on foundries—and one in particular. In the early 2000s some 25 firms could make the most advanced processors. Today four can, and TSMC makes nearly all of the high-end ones. Memory chips, which are also vital for AI, have gone the same way. More than 20 firms made advanced memory chips in 1990. Now only three—Samsung, SK Hynix and Micron—make almost all of them. China has produced new entrants in both processors and memory chips, but without access to the most advanced chipmaking tools, they remain behind the frontier. Consolidation has allowed a few dominant chipmakers to set the pace of expansion and dictate the terms on which they sell. They are investing heavily to capitalise on the boom. TSMC’s capital expenditure is expected to hit roughly $62bn this year, more than one and a half times as much as in 2025. Next year analysts expect it to be around $76bn. The memory-makers have also boosted their investments. Even so, their outlays are modest in comparison with the windfall from AI. As a share of revenue, TSMC’s capital spending is expected to fall from 44% in 2023 to 36% this year. Moreover, investments are being paid for from the cash that chipmakers are generating, rather than borrowing. This year TSMC’s capacity expansions will use up a little more than half of its operating cashflow, and Samsung’s barely a fifth. The debts of these chipmakers look modest next to their swelling cash piles. Manufacturers are enjoying a “very good pricing environment”, according to Bill Wiseman of McKinsey, a consultancy, and have every reason to pace their expansion to preserve it. DRAM, a type of memory used in nearly all electronics, costs about six times what it did in early 2023. TSMC has not cut the price of its three-nanometre manufacturing process, which it launched in late 2022, even though a process normally gets cheaper after a couple of years. The speed of capacity expansion has also been impeded by the fact that many new chipmaking facilities are being built in America, where construction takes “about twice as long”, according to Mr Wiseman. Consolidation has also given chipmakers more bargaining power, shifting who will bear the burden of any slowdown. Until recently, memory was sold mostly on the spot market or through short-term contracts. When demand slackened, customers were not obliged to keep buying and prices therefore collapsed. Now, with capacity scarce and only a few suppliers to choose from, buyers are signing up years in advance, and on the sellers’ terms. Nvidia has committed more than $250bn to securing supply and components until January 2029. In September Micron said that “take-or-pay” agreements, under which customers must buy the agreed volumes or pay anyway, had locked in 35% of its revenue through to 2030, with $32bn-worth of commitments, most of which involve cash deposits. Such prepayments help fund new factories. They also shift at least part of the risk of a bust onto buyers. An end to the AI boom would still bring chipmakers plenty of pain. If spending stalls, prices will drop, profits will shrink and stratospheric stock prices will come back down to earth. Customers will look to renegotiate contracts that were signed at the top of the market. The second law of the semiconductor industry still applies. This time, however, customers will share the hangover. ■ 四十多年來，半導體產業一直受制於兩條如同物理學規律般可靠的定律。第一條是：刻在晶片上的微型開關——電晶體——的數量會持續增加。第二條是：每一次繁榮最終都會以衰退告終。第一條定律讓晶片變得更加強大，目前依然蓬勃發展；用於人工智慧的頂級處理器包含超過 2,000 億個電晶體，而業界估計到了本世紀末，晶片上能塞入一兆個電晶體。至於第二條定律，則引發了大家對於榮景還能持續多久的不安。 半導體產業對榮景早已習以為常，但從未見過像這次一樣的情形。AI 對於用來訓練和運行 AI 模型的晶片產生了無窮無盡的需求。2026 年上半年，半導體的銷售額年增率創下了自 1980 年代以來的最快紀錄（見圖表）。本月，全球最大的晶片代工製造商台積電，公布了連續第六個季度創紀錄的利潤；韓國晶片巨頭三星電子則公布了連續第四個季度創紀錄的利潤。然而，投資者卻變得越來越不安。儘管業績表現亮眼、前景看好，但自 10 月 6 日以來，領先的半導體企業指數已下跌了超過 10%。大型晶片設計公司博通（Broadcom）的大老闆陳福陽（Hock Tan）在 10 月 12 日時，將許多 AI 交易的投資者比作賭場裡「完全不知道」自己在投資什麼的賭客，這讓他們的擔憂雪上加霜。 該產業幾乎所有的預期增長都取決於用於 AI 的晶片，這使其成為對該技術的一場豪賭。如果這場豪賭落空——而且正如某些人所稱的「超級週期」走向終結——晶片製造商是否會遭受與繁榮時期一樣劇烈的崩盤？我們有理由認為不會。 即使半導體產業不斷成長，製造業也變得更加集中。大多數晶片設計公司，例如以圖形處理器推動 AI 榮景的輝達（Nvidia），並不自己製造矽晶片。像 Google 和 OpenAI 這樣也開始設計 AI 處理器的軟體公司同樣不自己製造。相反地，他們依賴晶圓代工廠——尤其是其中特定的一家。在 2000 年代初期，大約有 25 家公司能夠製造最先進的處理器；如今只剩下四家，而台積電幾乎承包了所有高階晶片的製造。 對 AI 同樣至關重要的記憶體晶片，也走上了相同的道路。1990 年有超過 20 家公司生產先進記憶體晶片，現在只剩下三家——三星、SK 海力士和美光——包辦了幾乎所有的產量。中國在處理器和記憶體晶片領域都出現了新進業者，但在無法取得最先進晶片製造工具的情況下，他們依然落後於技術前線。 產業整合讓少數幾家佔據主導地位的晶片製造商能夠掌控擴產的節奏，並主導銷售條款。他們正在大量投資以利用這次的榮景。台積電今年的資本支出預計將達到約 620 億美元，是 2025 年的一倍半以上；分析師預計明年的資本支出將達到 760 億美元左右。記憶體製造商也增加了他們的投資。 即便如此，與 AI 帶來的意外之財相比，他們的支出其實相當保守。作為營業收入的比重，台積電的資本支出預計將從 2023 年的 44% 降至今年的 36%。此外，這些投資是用晶片製造商自己產生的現金來支付，而不是透過借款。今年，台積電的產能擴張將消耗其營運現金流的半數多一點，而三星則僅占五分之一左右。與這些晶片製造商日益膨脹的現金儲備相比，他們的債務顯得微不足道。 諮詢公司麥肯錫的比爾·威斯曼（Bill Wiseman）表示，製造商正享受著「非常良好的定價環境」，並且有充足的理由來控制擴展速度以維持這種環境。DRAM 是一種幾乎用於所有電子產品的記憶體，其價格大約是 2023 年初的六倍。台積電並未調降其於 2022 年底推出的 3 奈米製程價格，儘管製程通常在推出幾年後會變得便宜。威斯曼先生指出，許多新的晶片製造設施正建於美國，而那裡的施工時間「大約需要兩倍長」，這也阻礙了產能擴張的速度。 產業整合也給了晶片製造商更多的議價能力，從而改變了誰將承擔任何放緩所帶來的負擔。直到最近，記憶體主要還是在現貨市場或透過短期合約進行銷售。當需求放緩時，客戶沒有義務繼續購買，價格因此崩盤。現在，由於產能緊缺且只有少數供應商可供選擇，買家提前數年簽約，並且是按照賣方的條件簽訂。輝達已承諾出資超過 2,500 億美元，以確保到 2029 年 1 月為止的供應和零部件。美光在 9 月表示，「照價付款」（take-or-pay）協議——即客戶必須購買約定的數量，否則無論如何都得付款——已經鎖定了其到 2030 年前 35% 的收入，涉及價值 320 億美元的承諾，其中大部分包含現金押金。這類預付款有助於資助新工廠的建設，同時也將崩盤的部分風險轉嫁給了買家。 AI 榮景的終結仍會給晶片製造商帶來巨大的痛苦。如果支出停滯，價格將會下跌，利潤將會縮水，天文數字般的股價也將回到現實。客戶將尋求重新談判在市場頂峰時期簽署的合約。半導體產業的第二條定律依然適用。然而，這一次，客戶將會共同承擔這場宿醉。 ■ --Hosting provided by SoundOn