KEY POINTS
- Semiconductor names ripped higher Wednesday, with AMD climbing 6.7%, Broadcom 5.1% and Micron 8.5%, the single best chip-sector session since the Iran ceasefire bounce.
- The Meta-Broadcom partnership announced April 14 locks in one gigawatt of custom MTIA chips on a 2-nanometer process and extends through 2029, validating the custom silicon thesis.
- Watch whether Nvidia's May 28 earnings print holds guidance despite hyperscaler custom-chip ramp plans.
AMD jumped 6.7%, Broadcom added 5.1% and Micron gained 8.5% on Wednesday, a session that dragged the Nasdaq Composite to a record 24,657.57 close and put semiconductors back at the top of every institutional watchlist into Thursday's open. The move was not about any single headline. It was the market pricing in a structural shift that became impossible to ignore after last week's Meta-Broadcom deal, and after a Powell speech that removed one of the last macro overhangs from the growth trade.
The immediate catalyst was cleaner-than-expected order commentary from the April 22 batch of chip-sector earnings. Hyperscaler customers are not cutting AI spend. They are reshuffling it. Some dollars are moving from general-purpose Nvidia GPUs into custom application-specific integrated circuits designed in partnership with Broadcom, Marvell and Alchip. Other dollars are flowing into high-bandwidth memory, where Micron and SK Hynix remain supply-constrained through the end of 2026.
The Chip Supply Picture
The most important chart on a chip trader's screen right now is not AMD's price action. It is Broadcom's multi-year backlog, which implies roughly $90 billion of committed AI revenue by 2029 even before this week's new commitments are added. Meta's April 14 announcement with Broadcom extends the existing MTIA relationship through 2029, commits to one gigawatt of deployed custom silicon starting this year, and sets a path to multiple gigawatts by 2027. Meta disclosed the deal explicitly as a reduction in Nvidia dependence, not as a replacement of it.
That distinction matters. Nvidia still commands an estimated 85% to 92% share of the merchant AI accelerator market, and the Meta deal does not materially dent that share on a 2026 basis. What it does is put a ceiling on Nvidia's pricing power in the 2027 and beyond budgets that hyperscalers are building right now. For every custom chip Meta, Alphabet, Amazon and Microsoft design, the incremental GPU dollar has to justify itself against a better-specified internal alternative. That is a subtle but real re-rating of Nvidia's long-run gross margin.
Micron's 8.5% ripper on Wednesday told the other half of the story. High-bandwidth memory pricing is still tightening. KeyBanc raised Micron's target to $185 on the view that HBM3e and HBM4 capacity will remain sold out through the end of 2026, and Morgan Stanley followed with a positive note on DRAM pricing. Micron is the cleanest way to play the custom silicon ramp, because every MTIA, Trainium and TPU chip needs roughly the same amount of HBM per compute unit regardless of whose logic die sits in the middle.
What the AMD and Broadcom Moves Imply
AMD's 6.7% surge carried a different message. The stock has been stuck in a range since January because its MI300X ramp has moved slower than bulls hoped, and because hyperscaler custom silicon is as much a threat to AMD as it is to Nvidia. Wednesday's move coincided with a Wells Fargo note arguing that the MI350 platform, due in the fourth quarter, will be the first AMD product that closes the software gap with Nvidia's CUDA stack thanks to improved PyTorch and vLLM support.
Broadcom's 5.1% rally was less about the Meta deal directly, which the Street had already priced in, and more about the implied read-through for Alphabet's TPU roadmap and Amazon's Trainium 3 schedule. Wall Street is increasingly comfortable underwriting Broadcom's AI revenue at 60% gross margins, which is roughly 1,000 basis points above where the merchant chip business has historically run.
The one note of caution came from Intel, which lagged the rally with a 1.2% gain. The Street is still waiting for evidence that Gaudi 4 can win a meaningful inference socket inside a top-four hyperscaler, and without that confirmation the turnaround narrative stays on hold. CEO Lip-Bu Tan has promised more foundry customer names by the Q2 print, which is the first real test of his strategic plan.
Taiwan Semiconductor added a quieter but structurally important signal to the session. The foundry's April 17 Q1 report showed AI accelerator revenue running at roughly 35% of total sales, up from 28% in the prior quarter, with the three-nanometer node already effectively sold out through the end of 2026 and two-nanometer pre-orders tracking well above management's original capacity plan. That data point sits under every other chip trade in the market right now. If TSMC cannot deliver the wafers, it does not matter whose logo is on the package. For every 100 basis points of TSMC capex upside, the merchant and custom chip ecosystem expands together, and both SMH and SOXX benefit.
What Traders Should Watch
The technical setup into Thursday is cleaner than it has been in weeks. The Philadelphia Semiconductor Index broke above its February high on volume, and futures are pointing to a modestly higher open with VIX holding near 14. Nvidia earnings on May 28 remain the single largest event risk for the group, and the options market is pricing in an implied move of roughly 8% on the print.
Three dates matter over the next five weeks. April 29 brings Alphabet, Meta and Amazon, and any of the three confirming expanded custom silicon commitments will keep Broadcom bid. May 6 is the Advanced Micro Devices earnings print, where MI350 guidance will drive the next leg of the AMD debate. May 28 is Nvidia, where the question shifts from whether Blackwell is on track to whether Rubin Ultra pricing holds up against a wave of 2-nanometer custom chips ramping into 2027.
For now, the tape says one thing clearly. Traders do not believe the hyperscaler capex cycle has peaked, and they are willing to pay a premium for the specific suppliers that sit between the hyperscaler and the foundry. That is a narrower group than the market was chasing in 2024, and it is the reason this rally has legs that the 2023 version did not.
