The cleanest part of the screen is also the most revealing. RegentQuant said on its May 1 snapshot that NVIDIA trades at a forward PEG of 0.33, TSMC at 0.52, Broadcom at 0.56, and AMD at 0.82 RegentQuant. In plain English, the largest, most widely owned names are not priced as if the market has run out of patience for them. They are priced as if earnings growth is still catching up to the move.

That leaves the more stretched names in a different camp. Arm Holdings sits at a forward PEG of 16.0, with a forward P/E of about 120 times against expected EPS growth of just 7.5% RegentQuant reported. KLA Corporation, Microchip Technology, and GlobalFoundries also land above a PEG of 2.0. Qualcomm and Skyworks show negative implied growth, which makes PEG itself meaningless for them. Screens are useful. They are also perfectly happy to stop making sense when the earnings line gets messy.

So the rally is bifurcated. The names carrying the index are priced on visible growth. The narrative-heavy names, and the slower-growth industrial names that always seem to be described as “strategic” when the multiple gets awkward, are where the strain shows. That is the market telling you which assumptions it trusts.

If the top of the sector looks reasonably priced, memory looks absurdly cheap. RegentQuant put Micron’s implied EPS growth for the coming year at 588.8%, Samsung’s at 458.0%, and SK hynix’s at 359.6% RegentQuant. Those are real numbers. They are also the sort of numbers that can make almost any stock look like a bargain if you stop reading too early.

The problem is not that the growth is fake. It is that the growth is coming off a trough. RegentQuant flagged that as a known flaw in the PEG screen, and the memory complex is the textbook example RegentQuant. The honest read is blunt: forward earnings are expected to recover violently, the stocks have already priced in a meaningful chunk of that recovery, and the residual upside depends on whether high-bandwidth memory keeps absorbing supply faster than DRAM capacity comes back online.

IDC said in late April that HBM has become the primary constraint in the AI accelerator supply chain, most capacity is already pre-committed through 2026, and meaningful new supply is unlikely before late 2026 IDC. SemiAnalysis added in mid-March that HBM consumes roughly three times more wafer capacity than commodity DRAM, a gap that could widen to nearly four times as the industry transitions to HBM4 this year SemiAnalysis. That supports pricing. It also limits how much volume can move if demand slips.

The forecast spread is worth keeping in view too. IDC’s late-April outlook puts 2026 semiconductor revenue at $1.29 trillion, while Deloitte projected $975 billion in early February IDC Deloitte. Those numbers are not supposed to be identical, and they are not. Still, the gap is a reminder that this market rests on a narrow set of assumptions about AI demand and memory pricing, not on some broad consensus that would survive much abuse.

This is not a setup that screams multiple compression. The valuations at the top are not stretched enough to make that the obvious first domino. The more realistic risk is that earnings estimates come down, and they come down quickly.

The first pressure point is hyperscaler capex. IDC said in late April that hyperscale capital expenditure exceeded $100 billion for the first time in the third quarter of 2025, and that the four largest hyperscalers are expected to increase capex by 70% year over year to about $600 billion in 2026 IDC. SemiAnalysis said in mid-March that Google’s 2026 capex expectations have roughly doubled versus prior assumptions, mostly because of datacenter and server spending SemiAnalysis. If any of those companies sound less enthusiastic on the next earnings call, the growth lines supporting the sector’s nicer PEG ratios will start to bend.

The second pressure point is memory pricing. Deloitte said in early February that consumer DDR4 and DDR5 prices rose about fourfold between September and November 2025 as HBM demand pulled wafer capacity away from commodity memory Deloitte. Deloitte also said further price increases of as much as 50% were likely in the first half of 2026 Deloitte. That kind of move can keep margins happy for a while. It can also set up a sharp reversal if consumer end markets weaken faster than expected.

The third pressure point is export controls, especially around China. NVIDIA’s most recent annual filing, filed in February 2026, says the company had not received licenses to ship its restricted high-end AI products to China, and that the U.S. “AI Diffusion” rule could eventually impose worldwide licensing requirements on covered export classifications NVIDIA 10-K via SEC. That is the sort of line investors like to skim past on a good day. On a bad one, it is the whole story.

TSMC adds a different kind of tension. SemiAnalysis said in mid-March that effective N3 utilization is expected to exceed 100% in the second half of 2026, and that TSMC will not be able to add enough capacity to fully meet demand for the next two years SemiAnalysis. TSMC’s board approved about $15.5 billion in capital appropriations in November 2024 to support long-term capacity plans TSMC 6-K via SEC. Scarcity helps pricing and margins. It also caps shipment upside. The foundry gets to charge more because it cannot make enough. Lovely problem, until it is the only problem.

The main mistake in reading the April rally is treating valuation as the whole debate. It is only the visible part. The more important variable is what happens to forward earnings estimates, because the valuation work looks benign only as long as those estimates stay where they are.

That is why sector concentration matters so much. Deloitte said in early February that the top 10 global chip companies had a combined market cap of $9.5 trillion as of mid-December 2025, with the top three accounting for 80% of that total Deloitte. Deloitte also said high-value AI chips now drive roughly half of semiconductor revenue while representing less than 0.2% of unit volume Deloitte. That is a remarkable setup. It means a tiny part of the market has an outsized effect on the rest of it.

The long-term story is still there. IDC projects semiconductor revenues reaching $1.75 trillion by 2030, with data center chips approaching half the total market IDC. That is plenty of room for the AI buildout to keep paying the rent. But near-term fragility does not disappear just because the secular case looks strong. The market can be right about the destination and still get the path wrong.

For now, the useful watchlist is simple. Listen for hyperscaler tone shifts, track memory pricing as DRAM supply comes back online, and keep one eye on China policy. If those assumptions stay intact, the rally has room. If they crack, the PEG ratios that look reassuring today will turn into a neat little memorial to consensus, which is the market’s favorite decoration and least reliable guide.