AI Chip Valuations: Two Warnings Every Investor Needs Now

The AI Hardware Rally Has a Number Problem

Semiconductors now represent 18% of the S&P 500. Let that sink in for a moment. Software stocks — historically the dominant growth sector — peaked at around 15-16% of the index at their all-time highs. Technology stocks broadly peaked near 25%. Chip valuations have never been this concentrated in the S&P 500, and they've quadrupled their historical average weighting of under 5%.

That single data point tells you two things simultaneously: there is still theoretical room for this rally to run, and the risk profile is rising with every percentage point gained. This is not a reason to panic-sell your Nvidia or dump your Micron position. It is, however, a reason to stop investing on vibes and start investing on data.

Two concrete warnings are now visible in the numbers — one from chip market structure, one from AI usage metrics at Google. Understanding both is the difference between riding this wave intelligently and getting caught holding the bag when the rotation finally turns.

Warning One: Semiconductor Concentration Risk in the S&P 500

When a single sector quadruples its historical weighting in the world's most-watched equity index, momentum traders and institutions take notice — and they pile in. That is precisely what Goldman Sachs data confirms is happening right now.

Hedge funds and mutual funds have been rotating aggressively out of software and into semiconductors throughout 2025. The specific data points are stark:

• Mutual funds are carrying their widest underweight in software excluding Microsoft since 2012
• Hedge fund exposure to software is at its lowest level since 2019
• Both hedge funds and mutual funds reduced Microsoft positions in Q2
• Semiconductor additions include ASML, Intel, Micron, AMD, and Broadcom

This is textbook momentum chasing. Institutions with quarterly performance targets do what they must: they sell the laggards and buy the leaders. In the short term, this amplifies gains. In the medium term, it creates fragile, overcrowded trades.

The historical irony is not lost here. This rotation strategy — despite feeling sophisticated — tends to produce returns that underperform a simple S&P 500 or Nasdaq 100 index fund over time. Retail investors who follow hedge fund flows six months late are often the ones absorbing the losses when the rotation reverses.

The parallel to draw is the software sector's own peak concentration. Software hit roughly 15-16% of the S&P 500 before rolling over meaningfully. Semis are already at 18% and climbing. The ceiling exists — we just don't know exactly how high it is.

Key takeaway: Concentration is not inherently a sell signal. But 18% sector weighting, with institutional momentum piling in, means you need an exit framework, not just an entry thesis.

Warning Two: Google's Token Usage Data Reveals a Decelerating Growth Rate

This is the warning that most retail investors will completely miss — and it's arguably more important than the concentration risk.

At Google I/O, the company proudly announced its token usage trajectory:

• 2024: 9.7 trillion tokens
• 2025: 480 trillion tokens
• 2026: 3,200 trillion tokens (3.2 quadrillion)

On the surface, this looks exponential. Plot those numbers and the curve shoots nearly vertical. Impressive, right?

Now apply calculus — specifically, look at the rate of change of the growth rate, what analysts call the second derivative. The growth rate went from 50x year-over-year to 6.6x year-over-year. That is an 86% collapse in the growth rate. The first derivative — the growth rate itself — is now a negatively sloping line.

To be precise: AI token usage is still growing massively in absolute terms. Nobody is saying AI adoption is reversing. But the acceleration of that growth is slowing sharply, and that deceleration matters enormously for chip demand forecasts.

Why? Because semiconductor companies — Nvidia, Micron, AMD, and the broader ecosystem — are priced for continued explosive growth in AI compute demand. That demand is ultimately driven by token usage, model training runs, and inference workloads. If the growth rate of token consumption is compressing, the demand curve for new chip capacity will eventually compress with it.

One important caveat: these are Google's numbers specifically. Claude (Anthropic) has reportedly been growing at a much faster clip, particularly through Claude Code. OpenAI's ChatGPT enterprise adoption continues to scale. The Google data is not the entire market — but it is a leading indicator worth tracking closely.

Key takeaway: The second derivative of AI token growth is negative at Google. This does not kill the hardware trade today. It tells you the trade has an expiration date, and the clock is ticking.

What Micron's Numbers Actually Tell You

Micron is the most instructive case study for understanding where the AI hardware trade is right now — both its extraordinary strength and its forward-looking complexity.

The current financials are genuinely remarkable:

• Revenue has approximately tripled while cost of goods sold rose only 20%
• EPS projections were revised from $26 to $59.16 — Wall Street was off by more than 100%
• The company is paying down $4.6 billion in debt while investing $11.7 billion in plant, property, and equipment
• Balance sheet coverage of short-term obligations remains healthy
• The weekly RSI has held above 70 for roughly seven months — overbought, but with fundamental justification

The pricing power Micron currently commands is exceptional. When revenue triples while input costs barely move, margins expand dramatically. This is what happens when demand structurally outpaces supply in a capital-intensive industry — and AI data centre build-outs have created exactly that dynamic for DRAM and NAND memory chips, particularly HBM (High Bandwidth Memory) variants critical for AI accelerators.

Micron's dual manufacturing model — using both in-house fabs and subcontractors, unlike Nvidia which outsources almost entirely to TSMC — gives it cost flexibility and supply chain resilience that pure fabless designs lack.

But here is where the forward picture gets complicated. Current analyst consensus projects:

• Next year (FY2026): +76% EPS growth — legitimately strong
• FY2027: -5.4%
• FY2028: +6.8%
• FY2029: -2.7%

The valuation case for Micron is almost entirely built on that 76% growth year. PEG ratio calculations beyond 2026 become increasingly difficult to justify at current multiples. And by 2028-2030, Chinese semiconductor self-sufficiency — driven by Huawei's chip development programmes and state investment — is likely to introduce meaningful competitive pressure on memory pricing.

Wall Street has been dramatically wrong about Micron to the downside already. They could be wrong again — this time underestimating the runway. But the structural story beyond the next 18 months becomes materially less compelling.

Key takeaway: Micron's current fundamentals are exceptional. The trade thesis gets thinner as you project beyond 2026, and the data is already telling you that.

Where Institutions Are Actually Placing Capital Right Now

Beyond the semiconductor rotation, it's worth understanding the broader institutional picture to contextualise these moves.

Companies like Berkshire Hathaway — with multi-decade investment horizons and no quarterly performance pressure — have largely sat out the AI hardware sprint. Their stock has been essentially flat for over a year. That is not a failure; it is a deliberate refusal to chase momentum at elevated valuations. Patient capital does not need to win every quarter.

Contrast that with hedge funds and mutual funds operating under annual or quarterly return benchmarks. Their incentive structure requires them to chase momentum. When semis are running and software is lagging, they rotate into semis. When the reverse happens — and it will — they'll rotate back. The retail investor who follows this rotation six to twelve months late absorbs the transition costs.

The smarter institutional play, which some long-only funds are beginning to execute, is building a barbell: maintaining semiconductor exposure while quietly accumulating software names at historically low relative valuations. Microsoft, Salesforce, and similar enterprise software companies sitting at multi-year underweights in institutional portfolios represent the asymmetric setup once the hardware trade peaks.

Key takeaway: Follow institutional positioning for context, not as a trade signal. The rotation into semis is mature. The rotation back into software has not yet begun — but the conditions for it are building.

How to Think About the AI Trade From Here

None of the warnings outlined above constitute a sell signal today. The data does not support that conclusion. What the data supports is a more precise, disciplined framework for managing AI-related positions going forward.

Here is the practical framework:

Watch for these inflection signals:

• Semiconductor S&P 500 weighting approaching 20-25% (the historical software and technology sector ceiling range)
• Token usage growth rates declining across multiple major platforms simultaneously, not just Google
• Micron or Nvidia forward EPS revisions turning negative for the first time
• IPOs of ChatGPT (OpenAI), SpaceX, or Anthropic — public market listings of the AI-native companies often mark late-stage enthusiasm peaks
• Chinese HBM and advanced GPU chip announcements reaching commercial scale

What still supports the trade:

• AI inference demand continues to require massive memory bandwidth — HBM demand is structural, not cyclical, for at least the next 12-18 months
• Hyperscaler capex (Microsoft, Google, Meta, Amazon) remains at record levels through 2025 guidance
• The software sector's eventual re-rating higher will not eliminate hardware demand — it will coexist with it
• DDR5 and HBM3E product cycles for Micron are still in early adoption phases

The AI hardware trade is not over. It is maturing. That distinction matters enormously for position sizing, time horizon, and exit planning.

Practical Conclusion: Data Over Narrative

The most dangerous thing an investor can do right now is treat the AI chip rally as an indefinitely self-sustaining trend. The most profitable thing an investor can do is hold appropriate exposure while monitoring the specific data points that will signal when conditions are changing.

Semiconductors at 18% of the S&P 500 are historically elevated — not a reason to sell, but a reason to know your exit. Google's token growth second derivative turning negative is a leading indicator worth tracking across the industry. Micron's fundamentals are exceptional today with meaningful uncertainty beyond 2026. And institutional momentum is fully invested in hardware, which means the rotation fuel is largely spent.

The investors who will navigate this transition well are not the ones who called the top earliest. They are the ones who watched the data, understood what it meant, and acted on evidence rather than emotion — in either direction.

Set your watchlist. Know your signals. Stay in the trade until the data tells you otherwise.

Frequently Asked Questions

What does semiconductor concentration at 18% of the S&P 500 actually mean for investors?

It means the sector has quadrupled its historical average weighting of under 5% in the index. For context, software peaked around 15-16% and broad technology around 25% at their respective highs. At 18%, semis are in historically unprecedented territory, which amplifies both upside momentum and downside risk if sentiment shifts. It is a signal to monitor position sizing and have a defined exit framework, not necessarily to sell immediately.

Is Google's declining token growth rate a sign that AI demand is collapsing?

No. Absolute token usage is still growing enormously — from 480 trillion in 2025 to a projected 3.2 quadrillion in 2026. What is declining is the rate of acceleration. The growth rate dropped from 50x year-over-year to 6.6x, an 86% compression in momentum. This matters because chip demand is priced for continued explosive growth. Slowing growth rates, even at high absolute levels, will eventually translate into lower demand growth for new semiconductor capacity.

Why has Micron stock performed so strongly, and is it still worth buying?

Micron's stock has surged because Wall Street dramatically underestimated its earnings growth — projecting $26 EPS while the company is now tracking toward $59. Combined with exceptional pricing power (revenue tripling while COGS rose only 20%), this produced a sustained earnings beat cycle. The near-term case remains strong with projected 76% EPS growth next year. The risk increases beyond 2026, where analyst consensus shows flat to negative growth, and Chinese competitive pressure in memory chips is likely to emerge by 2028-2030. Any investment decision should be calibrated to that specific time horizon.

What sectors could benefit when the semiconductor trade eventually rotates?

Software is the most likely beneficiary of a rotation out of semis. Mutual fund underweight in software is at its widest since 2012, and hedge fund software exposure is at its lowest since 2019. Enterprise software companies like Microsoft, Salesforce, and Adobe are sitting at historically depressed relative valuations versus hardware. When institutional investors begin rebuilding software positions — driven by improving earnings comparables and AI monetisation visibility — the rerating could be significant. The rotation has not started yet, but the conditions for it are the most favourable they have been in years.