Chip Stocks Race Toward Biggest Gains Since Dotcom Era on AI Demand

On Wednesday, May 27, South Korea’s benchmark Kospi index crossed 8,228 points — a single-session surge of 4.65% that pushed its year-to-date return to exactly 100%. Two chipmakers drove it there. SK Hynix, Nvidia’s primary supplier of high-bandwidth memory, climbed 9.21% in the session and crossed the $1 trillion market-capitalisation threshold for the first time in its history. Samsung Electronics followed, up 2.68%, having crossed that same mark just weeks earlier. The benchmark’s performance now approaches the Nasdaq 100’s 102% surge in 1999 — right before the dotcom bubble burst. That comparison is everywhere right now. The question is whether it illuminates or misleads. Business Standard

The chip sector’s current ascent did not emerge from nowhere. It is the downstream consequence of a capital commitment so vast it has few modern precedents. The five largest hyperscalers — Amazon, Alphabet, Microsoft, Meta, and Oracle — have collectively committed more than $600 billion in capital expenditure for 2026, a 36% jump from 2025 and more than four times what the entire publicly traded U.S. energy sector spends annually on drilling, refining, and distribution. Goldman Sachs’ Global Institute puts the cumulative AI infrastructure bill at approximately $7.6 trillion between 2026 and 2031, spanning compute, data centres, and power infrastructure. Investing.comGoldman Sachs

Chips sit at the centre of every line in that ledger. What has happened to semiconductor equities since that spending unlocked is a story about what happens when extraordinary demand meets a supply chain that is structurally incapable of responding quickly — and about how fast markets can price the gap between those two things.

Chip Stocks Post Dotcom-Scale Gains as AI Demand Reshapes Global Equity Markets

The semiconductor industry’s market capitalisation has grown more than fourfold since the AI boom began, climbing from $2.2 trillion in May 2023 to $9.4 trillion in April 2026. That expansion — compressed into three years — eclipses the pace of the late-1990s technology build-out in raw velocity. Morningstar

South Korea’s Kospi is the most vivid single expression of this run. According to Korea Exchange data, Samsung Electronics has surged 149% year-to-date, while SK Hynix has climbed 215%, serving as the primary engines of the entire index’s rally. Together, the two chipmakers account for more than 42% of the Kospi’s market capitalisation, a record concentration that has made the index a de facto leveraged bet on the global AI supply chain. Disruption BankingCNBC

The American semiconductor picture is only marginally less spectacular. The VanEck Semiconductor ETF — the sector’s clearest benchmark vehicle — rallied nearly 49% in 2025, its third straight year of gains and a continuation of a run that began with a more than 72% surge in 2023. Over the same period, AMD’s AI accelerator revenue climbed 289% and Broadcom’s accelerator revenue surged 840% between the March 2023 and March 2026 quarters. These are not the numbers of a normal cycle. CNBCMorningstar

What connects them is a single supply constraint the market has only recently begun to price properly: high-bandwidth memory, or HBM.

HBM is the specialised memory architecture required to run large AI models at scale. It is power-efficient, extraordinarily dense, and technically brutal to manufacture — each generation requiring chipmakers to stack silicon wafers with sub-micron precision at volumes that push the limits of yield. Micron confirmed on its fiscal Q1 2026 earnings call that its HBM capacity was sold out through the entirety of calendar year 2026. SK Hynix controls an estimated 57–62% of global supply. And gross margins on HBM production have reached 60–70%, dramatically higher than the margins on standard DRAM. Investors have done the arithmetic. Manufacturing Dive

The data centre revenue figures for Nvidia make the demand side equally clear. In its fiscal year 2026, the company reported data centre revenue of $193.7 billion out of total revenue of $215.9 billion — a figure that transforms the company’s historical identity as a gaming-chip maker into something more like a sovereign infrastructure provider. Of the U.S. Market Index’s 85.6% gain since May 2023, 22.76 percentage points come from semiconductor stocks, of which 12.2 points come from Nvidia alone. One company’s earnings power has become a meaningful factor in national wealth creation. Morningstar

Why This Rally Looks Different From 1999 — And Why That Still Doesn’t Mean It’s Safe

What is driving semiconductor demand in the AI era?

AI demand is fuelling semiconductor growth through a convergence of AI accelerator chips led by Nvidia’s Blackwell architecture, high-bandwidth memory needed to train and run large language models at scale, and power-dense data centre infrastructure. Nomura describes a “triple memory super-cycle spanning DRAM, HBM and SSD memory” since the third quarter of 2025, and forecasts annual revenue and earnings growth of roughly 30% for memory suppliers over the next three to five years, following an estimated seven- to eightfold profit increase in 2026. CNBC

The dotcom comparison is not wrong, but it is partial. The late-1990s run in technology equities was primarily a valuation event: companies with no revenue, no product, and sometimes no coherent business model attracted capital because connectivity was understood to be transformational. What was unclear then was when, how, and for whom that transformation would monetise. Speculative capital flooded in ahead of any earnings. The capacity being built — fibre-optic networks, server farms — was eventually used, but not before it buried the companies that built it.

The current semiconductor rally has a different texture. Earnings are real. Revenue is growing. The spending driving chip demand is not investor sentiment but committed corporate capex from companies with balance sheets thick enough to sustain it through years of uncertainty. Goldman Sachs describes the Kospi as its “highest-conviction equity market” in the region, forecasting 300% earnings growth in 2026 — the strongest annual profit expansion in any Asian market since the 1999 Asian financial crisis recovery. Goldman’s 12-month Kospi target stands at 9,000; JPMorgan’s bull case goes to 10,000. Disruption Banking

The structural argument rests on two legs. First, AI workloads are not discretionary in the way earlier technology adoption waves were. Companies that fail to build or buy AI infrastructure do not simply grow more slowly — they risk obsolescence in markets where AI-native competitors move faster and cheaper. That creates a degree of demand inelasticity that typical semiconductor cycles have lacked. Second, the mix shift in memory consumption has been profound. Servers now account for 60–70% of memory demand, up from roughly 30% before the AI boom began, according to analysts at Jefferies. That shift is permanent unless a competing architecture emerges requiring less memory — something no credible roadmap currently proposes. Fortune

Yet the word “structural” is doing a great deal of work in bullish analyst reports right now. Permanent cycles have a way of revealing themselves as long cycles. The distinction matters enormously for investors entering at today’s prices.

The Downstream Consequences: Consumer Electronics, Geopolitics, and the Price of Concentration

The gains are generating their own side effects, and they are not uniformly positive.

The reallocation of manufacturing capacity toward HBM is crowding out conventional DRAM and NAND production. The total number of wafers a fab can produce is fixed at any given point. Producing HBM displaces commodity memory. Consumer electronics — smartphones, laptops, game consoles — compete for what remains. Nintendo raised the price of the Switch 2 in part because memory input costs rose. PC manufacturers are absorbing higher bill-of-materials charges. The memory tax on everyday devices is not abstract: it is already surfacing in retail prices, and it will persist structurally so long as AI data centre construction continues at current rates. Samsung is expanding HBM capacity by 50% in 2026, partly at its $17 billion facility under construction in Taylor, Texas — but that capacity will not be fully online in time to relieve the immediate shortage.

The geopolitical dimension is pointed. Taiwan’s Taiex is being driven by TSMC, which now accounts for over 40% of that benchmark’s market capitalisation and holds roughly 68% of global foundry revenue by advanced nodes. South Korea’s Samsung and SK Hynix, between them, supply the dominant share of global HBM. The global AI buildout — a $7.6 trillion investment program made primarily by American hyperscalers — runs through a narrow geographic corridor in East Asia. Allianz Research has noted that this concentration creates a scenario where the entire AI infrastructure programme is effectively hostage to a handful of facilities in a narrow geographic corridor — a rebalancing that could, over time, be addressed by semiconductor capacity investments in Europe and the U.S., but those will not mature in time to alter the current cycle’s supply dynamics. CNBCAllianz

Bank of America estimates the total addressable market for AI data centre systems will reach over $1.2 trillion by 2030, representing a compound annual growth rate of 38%, with AI accelerators alone representing a $900 billion opportunity. Those figures provide the macro ceiling. What they don’t tell you is how the gains within that market distribute across suppliers, geographies, and time — and the current distribution is unusually narrow. Yahoo Finance

The Kospi’s daily average trading value surpassed 40 trillion won for the first time this month, reaching a record 48 trillion won — but Samsung and SK Hynix alone account for 43% of that total. An index that rises on two stocks is not diversification. It is sector exposure dressed in national-market clothing. Disruption Banking

The Case for Caution: What Chip Bears Are Actually Arguing

Not everyone finds the “structural” framing convincing — and the most substantive critique deserves a fair hearing.

The core bear argument is not that AI demand is fake. It’s that cycles have a way of reasserting themselves, and that the current valuation run is pricing in a scenario with no setbacks, no substitution effects, and no demand disappointments. Memory chip economics have been notoriously cyclical. The industry’s last sustained boom — driven by cloud computing and smartphone adoption — collapsed into one of the worst sector downturns in decades between 2022 and 2023. Those who argued that secular demand had permanently broken the cycle in 2021 were not wrong about the demand thesis; they were wrong about the timeline and the magnitude of the overshoot.

A chip expert at Harvard struck this note in a recent interview, warning that memory markets have been here before and that “this too will pass.” The observation is not contrarianism for its own sake. It’s a reminder that scarcity, in semiconductor history, reliably attracts supply — and that supply, when it arrives, tends to overshoot the demand it was chasing. Fortune

The concentration risk at the index level amplifies this concern. Goldman Sachs strategist Tim Moe’s observation to CNBC that “it’s the AI hardware theme that’s clearly what is propelling things” is candid acknowledgement that the breadth of these rallies is thin. When Samsung and SK Hynix account for 42% of the Kospi, any demand deceleration, yield disappointment, or hyperscaler capital expenditure revision doesn’t just hit two stocks — it hits the national benchmark. CNBC

There is also the architecture question. Nvidia announced in October 2025 that it intends to use LPDDR5 — a lower-power, cheaper variant of consumer DRAM — for inference GPUs by the end of 2026. If inference workloads, which are growing faster than training workloads, shift toward memory architectures that don’t require HBM’s premium engineering, the margin story changes materially. The market has priced in a world of permanent, high-margin scarcity. Scarcity, in this industry, has always eventually produced its own cure.

The honest answer to whether 2026 is 1999 again is: not quite, and not entirely different either.

The earnings are real. The demand is structural. The spending commitments are in writing and backed by corporate balance sheets that did not exist in the last millennium’s internet frenzy. The hyperscalers are not startup dreamers burning venture capital; they are the most profitable companies on earth, allocating capital with clear awareness of what pulling back would mean competitively. That backstop — genuine, cash-generating demand from entities that cannot afford to stop spending — is something the dotcom era never had.

Yet markets that price perfection are vulnerable to the imperfect. The semiconductor industry has never produced a cycle that matched its most optimistic projections from the peak. Technology always wins the long game. The companies that own the technology don’t always win along with it.

The chip rally of 2025 and 2026 is simultaneously a genuine reflection of structural change and a warning about what happens when markets are given permission to price in only the best version of that change. The gains are real. So is the distance between the price and the proof.