Valuation vertigo and flashbacks to 2000
The Philadelphia Semiconductor Index, better known as the SOX, hit 12,000 points on 13 May 2026. Year-to-date, the index is up roughly 69%. The top ten performers in the Nasdaq have more than doubled year-to-date!
SanDisk is up 558% in 2026 alone. Yours truly sat through that earlier episode, watching Qualcomm surge 2,600% in 1999 before giving most of it back, and the price moves of the last few weeks produce a very specific kind of déjà vu, somewhere between professional fascination and mild nausea.
This is not a lazy comparison. The SOX’s recent 30+% surge over 13 trading days was the largest such move since 2002. According to BTIG’s chief market technician, the last and only time the index made a similar move into a new all-time high was March 2000. Nvidia trades at roughly 33 times EV/EBITDA on a trailing basis, while its forward EV/EBITDA sits closer to 42 times, well above its 20-year median.
ASML, the Dutch lithography company that we have watched grow up from a small challenger to Canon and Nikon in the mid-90s to the monopolist it is today, commands around 35 times forward earnings, having doubled off its 52-week low. The semiconductor slice of the S&P 500 has added approximately $3.8 trillion in market capitalisation over six weeks.
None of this means the cycle is over. But it does mean the margin for error, never particularly generous in this industry, is now approximately the width of a 2-nanometre gate. Ideally, we don’t mention the C word in relation to Tech stocks (we don’t want to jinx it), but we have to ask ourselves the question whether the chip industry is ripe for a crash.
It is genuinely different this time … for the most part
To be fair to the bulls, and one should at least try, there is a real demand story underpinning this rally, not just a narrative. The primary engine is the build-out of AI data centres by the hyperscalers. Meta, Alphabet, Microsoft, and Amazon collectively told investors in late April 2026 that they plan to spend roughly $700 billion on capital expenditure this year alone. That is nearly double their 2025 spend. Amazon’s number is $200 billion. Alphabet is guiding $175 to $185 billion. Meta just raised its full-year guidance to $125 to $145 billion, with CEO Mark Zuckerberg citing “higher component costs, particularly memory pricing.” Even Bank of America, not typically a source of alarm, has projected that hyperscaler capex will surpass $1 trillion in 2027.
That kind of spending lands everywhere in the chip ecosystem. Nvidia is the most obvious beneficiary, but Applied Materials, Lam Research, Broadcom, and ARM are all riding the same wave. Micron’s revenue is expected to hit $107 billion in fiscal year 2026, up from $15.5 billion in 2023, which, as numbers go, is genuinely extraordinary. Cloud memory gross margins at Micron are running around 66%. SK Hynix posted operating profit of 37.6 trillion won in the first quarter of 2026, an operating margin of 72%, which, remarkably, exceeds Nvidia’s own 65% margin.
The traditional chip segments are suffering
Beyond AI, though, the traditional chip end-markets show a more complicated picture. The reallocation of wafer capacity toward high-bandwidth memory has created acute shortages in consumer-grade DRAM and NAND. IDC projects smartphone shipments down 12.9% in 2026 and the PC market contracting 11.3%. Analog Devices and Texas Instruments, two reliable barometers of industrial and automotive chip demand, have signalled cautious outlooks. Qualcomm and MediaTek are navigating a smartphone market that is essentially being starved of memory supply, not because demand has collapsed, but because Samsung, SK Hynix and Micron have redirected their fabs toward the hyperscalers. The consumer electronics market is not exactly booming. It is being quietly cannibilised by the AI trade.
The cycle that AI ate
For three decades, the semiconductor industry moved with the grace and predictability of a wrecking ball, swinging from boom to bust on the back of DRAM oversupply, inventory bloat and the inevitable production cuts that followed. That rhythm should, by any historical reckoning, have reasserted itself by now. It has not.
The reason is SK Hynix and Samsung. Both companies have diverted up to 40% of their advanced wafer capacity toward HBM production. Goldman Sachs lifted its 2026 DRAM supply-demand gap forecast in April to 4.9%, calling it the most severe shortage in 15 years. Both SK Hynix and Samsung have publicly warned of “significant shortages” persisting through at least 2027. DRAM prices have roughly doubled since early 2025. Suppliers are carrying only two to three weeks of stock. The familiar oversupply-driven bust, which had already been delayed by the AI demand surge of 2023 and 2024, now looks like it cannot arrive until new mega-fabs come online, probably not before 2027.
The uncomfortable question is whether AI has genuinely disrupted the cycle or merely deferred it. Put differently, would we already be in a significant memory downturn without the HBM bid from the hyperscalers? Almost certainly, yes. The consumer electronics recovery has been tepid, automotive chip demand has softened, and the DRAM glut that was building in early 2023 was resolved primarily by the HBM rotation, not by a recovery in underlying unit volumes.
The cycle has not been abolished, but it has been redirected. When HBM capacity eventually catches up with demand, whether in 2027 or 2028, the traditional question of what happens to commodity memory pricing will resurface with considerable force.
The ROI problem at the centre of everything
Here is the thing about $700 billion in annual capital expenditure: someone eventually has to earn it back. The hyperscalers are aware of this, naturally, but the math is getting harder to ignore. Bank of America calculates that hyperscaler capex now consumes 94% of operating cash flows after dividends and buybacks. Amazon is projected to be in negative free cash flow territory in 2026. Alphabet’s free cash flow is forecast to decline roughly 90% to $8.2 billion, down from $73.3 billion in 2025. Microsoft’s free cash flow is expected to slide 28% before recovering in 2027.
To bridge the gap, the big five hyperscalers raised $108 billion in bonds in 2025. JPMorgan projects $1.5 trillion in tech debt issuance over the coming years. Alphabet’s long-term debt quadrupled in 2025 to $46.5 billion. These are still strong balance sheets, but the direction of travel is notable. AI services generate approximately $25 billion in direct revenue today, or roughly 4% of what is being spent on the infrastructure that runs them.
The bears will tell you that is the entire problem in a single number. The bulls will tell you cloud computing looked exactly the same in 2008 and became a multi-trillion-dollar business. Both arguments are reasonable.
We believe the important thing for semiconductor stocks is not who’s right or wrong in that debate, but the near-term behaviour of capex. If even one or two of the major hyperscalers pauses, revises, or signals scepticism about its spending trajectory, the demand picture for Nvidia, Samsung, SK Hynix and the broader semiconductor equipment sector changes materially and quickly.
What the expanding semiconductor valuations are telling us
Let’s run the numbers properly. Nvidia’s TTM EV/EBITDA sits around 33 times, against a 10-year median of 42 times and a 5-year forward mean closer to 40 times. Its forward P/E is roughly 24 times, which looks, by current Nvidia standards, almost reasonable. The sector median EV/EBITDA for semiconductors is around 27.5 times, meaning the index, taken as a whole, is trading at a meaningful premium to its own long-run average. At the 2000 peak, the SOX traded at nearly 8 times book value. Today, it sits at approximately 5 to 6 times, which is elevated, but not historically extraordinary in nominal terms.
The bear case on multiples is straightforward. If earnings growth decelerates, as it inevitably must when the hyperscaler build-out normalises, and if multiples compress even partially toward their 20-year medians, the drawdown from here is meaningful. A reversion to long-run EV/EBITDA medians, without any decline in underlying earnings, still implies a 20% to 30% de-rating across the sector. A reversion combined with earnings misses, the kind of scenario where capex guidance is cut and inventory builds, implies something considerably more uncomfortable, in our view.
The bull case, to be clear, is also coherent. If AI demand continues to compound at current rates, if the transition from training infrastructure to inference deployment creates a second wave of hardware demand, and if the hyperscalers’ return on investment eventually validates the spending, then current Nvidia earnings multiples are genuinely undemanding.
ASML’s near-monopoly on EUV lithography, with a backlog of EUR38.8 billion and a 2030 revenue target of EUR44 to EUR60 billion, represents the kind of structural moat that does justify a sustained premium. TSMC, sitting at the intersection of every AI hardware narrative on the planet, is not going away. These are not dot com companies with no revenue and a logo.
But the SOX is 60% above its 200-day moving average. Historically, that level of technical extension resolves through either a prolonged consolidation or a sharp correction. It rarely just keeps going.
Five things that could set off the C word
The first risk, and most immediate in our view, is an earnings miss or capex guidance cut from one of the hyperscalers. The market has priced in $700 billion in 2026 spending and $1 trillion in 2027. Any credible signal that those numbers are being revisited, whether due to weaker-than-expected AI monetisation, margin pressure or simple budget discipline, would land on chip stocks like a stone through a window. Meta’s free cash flow dropped to $1.2 billion in Q1 2026, down from $26 billion a year earlier. That is the kind of number that eventually attracts Board attention.
The second risk is a DRAM price reversal. The current shortage is real, but it is partly a function of constrained supply from the HBM rotation. When Samsung’s P4L and SK Hynix’s M15X come online in 2027, the capacity picture changes. A normalisation of HBM yields, combined with new capacity, could flip the memory market from shortage to surplus faster than the current consensus expects. History, as noted, suggests this is not a hypothetical.
Third, there is geopolitical risk concentrated at a single point. TSMC manufactures the overwhelming majority of leading-edge chips, and the Taiwan Strait is not getting calmer. A serious escalation, or even a sustained spike in geopolitical tension that causes customers to seek to diversify supply chains, would create the kind of uncertainty that reprices the whole sector. Already, Apple is diverting some manufacturing away from TSMC towards Intel. M-series chips for now, but that could expand to A-series chips down the line.
Fourth, a broader market risk-off driven by sticky inflation, unexpected rate rises, or a credit event could drain liquidity from the highest-multiple corners of the equity market. Semiconductor stocks, with their 5 to 6 times book multiples and parabolic recent performance, would be among the first affected.
Fifth, the AI efficiency narrative. Every time a more capable model ships at a fraction of the compute cost of its predecessor, the investment thesis for ever-expanding hardware spend is subtly undermined. DeepSeek’s emergence earlier in 2025 already produced a brief, sharp de-rating. If a new entrant, or one of the existing hyperscalers, publicly pivots toward a more capital-efficient AI architecture, the demand assumptions embedded in current valuations get re-evaluated.
It may not be a dot com bubble, but it is definitely some kind of bubble
The dot com comparison is not a prediction. In 1999, the companies driving the rally were largely burning cash on customer acquisition and promises. Today’s semiconductor leaders are generating real, staggering earnings from genuine demand. The difference is significant and that needs to be kept in mind.
But the SOX up 65% in five months, the top Nasdaq names outperforming the dot com peak, and $700 billion in spending that still only generates $25 billion in AI-specific revenue? That is not a normal market by any means. It is a very good story running very fast, and good stories running very fast have a well-documented historical tendency to at least in part return to the mean.
Yours truly has seen enough chip cycles to know that the moment everyone agrees the cycle has been permanently disrupted is usually a reliable signal to check if the exits are clear.
