Coverage: US + HK / China cross-listed
Report date: 30 May 2026
Data source: yfinance MCP (live quotes, technicals, earnings dates) + public disclosures
Author: Equity research desk (internal)
The AI compute stack is two and a half years into the buildout cycle that began with the Hopper ramp in late 2023. The cycle is not yet rolling over, but the dispersion inside it has widened materially. Three things have changed since the start of 2026:
Net positioning view: the sector is extended (SMH +54% above 200-DMA, SOXX +69%) and several names sit at or within 1.5% of all-time highs (AVGO, MU, CRDO, QCOM, ALAB at <2% off). We see the asymmetry as long the bottlenecks (memory, packaging, custom-silicon enablers), short the most-extended/most-crowded extrapolations — with two specific China longs as portfolio diversifiers.
| Layer | 2024 spend | 2026E spend | 2027E spend | 24–27 CAGR |
|---|---|---|---|---|
| AI accelerators (GPU + ASIC) | ~$95B | ~$220B | ~$285B | ~44% |
| HBM memory | ~$18B | ~$58B | ~$85B | ~67% |
| AI networking (InfiniBand + 800G/1.6T Ethernet) | ~$14B | ~$36B | ~$50B | ~53% |
| Advanced packaging (CoWoS / SoIC) | ~$8B | ~$24B | ~$36B | ~65% |
| AI compute total | ~$135B | ~$338B | ~$456B | ~50% |
Source: triangulated from sell-side TAM frameworks (Morgan Stanley, BofA, Bernstein) cross-checked against TSM CoWoS capacity disclosures and Hynix/Samsung HBM guidance. Numbers should be treated as directional — the variance across sell-side estimates is roughly ±25% by 2027.
The shape of the curve has shifted. The 2024–25 spend was front-loaded into GPU silicon and InfiniBand. The 2026–27 mix tilts toward memory, packaging, and Ethernet as hyperscalers rebalance toward total-cost-of-ownership and standards-based scale.
Three indicators we watch:
We are in mid-cycle, not late-cycle. But the easy beats are behind us — the next leg of outperformance has to come from operational execution and mix, not raw demand surprise.
| Layer | Top players | Concentration |
|---|---|---|
| Merchant GPU | NVDA, AMD | ~95% NVDA share of training |
| Custom AI silicon | AVGO, MRVL (+ TSM as fab) | AVGO ~60% of merchant ASIC |
| Foundry | TSM, Samsung, SMIC | TSM ~95% of leading-edge AI |
| Memory / HBM | Hynix, Samsung, MU | Hynix ~50%, Samsung ~30%, MU ~20% |
| Litho | ASML | ~100% of EUV |
| Etch / dep / metrology | AMAT, LRCX, KLAC, TEL | Top 4 ≈ 75% |
| AI networking | NVDA (InfiniBand), AVGO/ANET (Ethernet) | NVDA ~70% IB, ANET ~30% high-end Ethernet |
| Optical interconnect | CRDO, ALAB, MRVL, CIEN | Highly fragmented at retimer/AEC layer |
The structure is highly concentrated at the top of the stack (one foundry, one litho vendor, two GPU vendors, three HBM vendors). This is the source of both the cycle's pricing power and its tail risk.
| Ticker | Last | Day % | Mkt cap | 52wH | %off 52wH | %vs 200-DMA |
|---|---|---|---|---|---|---|
| GPUs / accelerators | ||||||
| NVDA | $211.14 | -1.5% | $5.1T | $235.74 | -10.4% | +12.5% |
| AMD | $516.10 | -0.4% | $842B | $518.09 | -0.4% | +117.2% |
| Custom-silicon enablers | ||||||
| AVGO | $446.77 | +4.7% | $2.1T | $446.77 | 0.0% | +26.8% |
| MRVL | $205.00 | +0.1% | $179B | $208.26 | -1.6% | +108.2% |
| Foundry & litho | ||||||
| TSM | $418.45 | -1.5% | $2.2T | $424.86 | -1.5% | +30.2% |
| ASML | $1,612.76 | +0.4% | $622B | $1,632.90 | -1.2% | +34.9% |
| Equipment | ||||||
| AMAT | $450.06 | +0.1% | $357B | $454.89 | -1.1% | +53.8% |
| LRCX | $318.18 | +0.1% | $398B | $322.68 | -1.4% | +65.0% |
| KLAC | $1,921.71 | -0.3% | $251B | $2,011.39 | -4.5% | +42.5% |
| Memory / HBM | ||||||
| MU | $971.00 | +5.1% | $1,095B | $971.00 | 0.0% | +186.7% |
| Hynix (000660.KS) | KRW 2,333,000 | +1.9% | KRW 1,656T | — | — | — |
| Samsung (005930.KS) | KRW 317,000 | +5.8% | KRW 2,082T | — | — | — |
| Networking / optical | ||||||
| ANET | $159.47 | +2.7% | $201B | $177.73 | -10.3% | +13.8% |
| CRDO | $236.03 | +6.2% | $44B | $236.03 | 0.0% | +62.5% |
| ALAB | $342.85 | -1.8% | $59B | $349.17 | -1.8% | +97.2% |
| CIEN | $580.23 | +1.8% | $82B | — | — | — |
| Diversified / cyclical | ||||||
| INTC | $114.68 | -5.1% | $576B | $129.44 | -11.4% | +133.3% |
| QCOM | $251.02 | +3.2% | $265B | $251.02 | 0.0% | +54.5% |
| HK / China AI compute | ||||||
| SMIC (0981.HK) | HKD 81.60 | -7.5% | HKD 654B | HKD 91.05 | -10.4% | +19.9% |
| Hua Hong (1347.HK) | HKD 161.30 | -5.1% | HKD 280B | HKD 170.00 | -5.1% | +90.1% |
| BYD Electronic (0285.HK) | HKD 29.18 | -2.1% | HKD 66B | HKD 44.10 | -33.8% | -14.9% |
| Lenovo (0992.HK) | HKD 24.00 | +21.9% | HKD 298B | HKD 24.00 | 0.0% | +124.5% |
| Sunny Optical (2382.HK) | HKD 83.80 | +13.8% | HKD 90B | HKD 90.90 | -7.8% | +21.8% |
| Cambricon (688256.SS) | CNY 1,310 | -5.9% | CNY 823B | — | — | — |
| Hygon (688041.SS) | CNY 293.90 | -6.5% | CNY 683B | — | — | — |
| ETFs | ||||||
| SMH | $598.93 | -0.2% | — | $602.14 | -0.5% | +54.1% |
| SOXX | $569.08 | -0.1% | — | $570.09 | -0.2% | +69.2% |
Source: yfinance MCP, intraday 30 May 2026.
The first thing to note: eight names in this universe are within 2% of their 52-week high (AVGO, AMD, AMAT, LRCX, MU, CRDO, QCOM, Lenovo). Memory and Lenovo are the most extended off their 200-DMA. This is not a base — it is a market that is asking a lot of execution.
NVDA is now a $5.1T market cap with the stock approximately 10% off its all-time high and +12.5% above its 200-DMA. The technical setup is the calmest it has been in 18 months — the 50-DMA and 200-DMA are converging, and the August earnings (27 Aug 2026) will be the next major catalyst. Datacenter revenue grew +63% YoY in the most recent quarter (consensus). The de-rating risk is no longer about demand — it is about share-of-incremental-spend as ASIC programs scale at AVGO and a third hyperscaler.
AMD has broken out — the stock is up +366% from its 52-week low and +117% above its 200-DMA. The catalyst was MI400 winning real share at MSFT, ORCL, and Meta inference workloads. The risk: at $842B mcap, AMD is now priced as a credible #2 in training, not just inference. Any slip on MI400 ramps in 2H26 will be punitive.
AVGO is the highest-conviction "winning" structural story in the sector. The stock is 0% off its 52-week high and the 4 June 2026 earnings is the binary catalyst. Sell-side is modelling AI revenue of ~$28B in FY26 (vs. ~$12B in FY24); we think the print will come in closer to $32-34B with a third hyperscaler customer (likely Apple or OpenAI) acknowledged for 2027. Risk: the print is a binary one, and a positioning-driven gap-fill on a "good but not great" set could take 10-15% out of the stock.
MRVL has rallied +240% from its 52-week low. The core debate is whether AWS Trainium volumes (the dominant MRVL ASIC win) decelerate as Amazon mixes to in-house designs. Earnings 28 Aug.
TSM at 30% above 200-DMA is the cleanest structural exposure to AI compute outside NVDA. CoWoS capacity is the bottleneck — every incremental kwpm of capacity flows ~$1.5-2.0B of revenue. The next earnings (15 Jul 2026) will guide on 2027 CoWoS capacity, which is the single most important number in the AI hardware stack.
ASML has had a quiet but powerful 2026: +35% above 200-DMA, ~1% off the 52-week high. EUV demand from TSM N2/A16, plus High-NA EUV at Intel and Samsung, plus a re-acceleration of memory wafer adds is driving the story. Earnings 15 Jul. The China headwind is still real but increasingly priced in (~20% of revenue, declining).
MU at $971 (mcap $1.1T) and +187% above 200-DMA is the most extended liquid name in the sector. HBM3e/HBM4 pricing is up materially, and 2026 contract pricing is locked. The print on 25 Jun 2026 is the critical event — consensus has moved up sharply, and the bar for a "clean beat" is now very high. We see the risk-reward as asymmetric to the downside at these levels.
Hynix (000660.KS) at KRW 2.33M is the cleaner, less-extended way to play HBM — 50% supply share, the structurally cheapest valuation in the trio, but lower ADR liquidity.
ANET has been the lagging, most discounted name in the AI-networking bucket — up +85% from its 52-week low but -10% off its 52-week high, with the stock just +14% above its 200-DMA. The Ethernet-vs-InfiniBand story is finally inflecting in Ethernet's favour as Meta, Google, and Microsoft all commit to 800G / 1.6T Ethernet for inference clusters. Earnings 5 Aug.
CRDO at $236 — 0% off its 52-week high, mcap $44B — is priced for AEC volumes to scale 3-4x. Next print: 2 Jun 2026.
ALAB has gone vertical: +299% off its 52-week low, +97% above its 200-DMA. The core debate is whether the rack-scale AI architecture (UALink, PCIe Gen6 retimers) makes ALAB a structural winner or a temporary beneficiary of the GB200/300 product cycle.
The sector has bifurcated.
| Bucket | NTM EV/EBITDA | NTM P/E | NTM Sales |
|---|---|---|---|
| Merchant GPU (NVDA, AMD avg.) | ~28x | ~36x | ~16x |
| Custom-silicon (AVGO) | ~26x | ~33x | ~17x |
| Foundry (TSM) | ~14x | ~22x | ~9x |
| Litho (ASML) | ~24x | ~32x | ~11x |
| Equipment (AMAT/LRCX/KLAC) | ~19x | ~24x | ~7x |
| Memory (MU) | ~9x | ~12x | ~6x |
| Networking (ANET) | ~32x | ~42x | ~13x |
| China leading-edge (SMIC) | ~20x | ~85x | ~9x |
| China mature-node (Hua Hong) | ~14x | ~40x | ~5x |
Indicative ranges; multiples calculated from current market cap and consensus FY+1 estimates from public sources. Cite primary sources before publication.
The dispersion is the story. Memory trades at single-digit EV/EBITDA on peak earnings, while networking and custom-silicon trade at >30x P/E on still-expanding earnings. The market is paying a steep multiple for visibility and discounting cyclicality even at peak.
SMH trades at ~28x NTM P/E vs. SPX at ~22x — a ~27% premium that is roughly in line with the 5-year average but well above the 10-year average of ~15%. The premium is justified by ~22% LTM earnings growth vs. SPX ~12%, but it is fully valued.
We present 5 longs, 3 shorts, 1 pair below. Each idea is sized for a 4–8 week swing, stops indicated, and earnings catalysts mapped.
| # | Side | Ticker | Last | Stop | Target | Catalyst | Conviction |
|---|---|---|---|---|---|---|---|
| 1 | Long | TSM | $418.45 | $390 | $475 | 16 Jul 2026 earnings | High |
| 2 | Long | ANET | $159.47 | $148 | $185 | 5 Aug 2026 earnings | High |
| 3 | Long | Hua Hong (1347.HK) | HKD 161.30 | HKD 142 | HKD 195 | China 2H stimulus, Aug pre-announce | Medium-High |
| 4 | Long | Hynix (000660.KS) | KRW 2,333k | KRW 2,150k | KRW 2,650k | HBM4 contract Q3 reprice | Medium |
| 5 | Long | AVGO | $446.77 | $415 | $510 | 4 Jun 2026 earnings | Medium (binary into print) |
| 6 | Short | MU | $971.00 | $1,065 | $830 | 25 Jun 2026 earnings | High |
| 7 | Short | Lenovo (0992.HK) | HKD 24.00 | HKD 26.50 | HKD 20.50 | Post-earnings exhaustion | Medium |
| 8 | Short | INTC | $114.68 | $124 | $98 | 24 Jul 2026 earnings | Medium |
| 9 | Pair | Long QCOM / Short AMD | n/a | -7% pair | +10% pair | 30 Jul / 5 Aug earnings | Medium |
Thesis (one line): the foundry sitting on the only relevant CoWoS capacity in the world is, somehow, the cheapest leading-edge AI name on a multiples basis.
Why now:
Risks: geopolitical (Taiwan), customer concentration with NVDA, Korean / Japanese capacity ramps eroding pricing power.
Thesis: ANET is the cheapest US-listed name with a clean 800G/1.6T Ethernet thesis; still -10% off its 52-week high while every other AI-networking name is at or near highs.
Why now:
Risks: AVGO/MRVL ASIC programs absorb Ethernet share; hyperscaler in-source switching displaces merchant.
Thesis: the mature-node Chinese foundry is a second-derivative play on AI server bill-of-materials — every AI server doubles 28nm/40nm content (PMIC, MCU, analog) and Hua Hong has the only at-scale, sanctioned 28nm capacity.
Why now:
Risks: China AI policy disappointment; SMIC takes 28nm share back as capacity ramps.
Thesis: the cleanest, cheapest HBM exposure — ~50% global supply share, locked-in 2026 contract pricing, and a structurally tighter HBM4 contract reprice in Q3 2026.
Why now:
Risks: Samsung HBM4 qualification; Chinese memory entrants (CXMT, YMTC); Korean won swings.
Thesis: AVGO prints on 4 June and we expect AI revenue beat + a third hyperscaler customer disclosure. This is a binary, sized smaller.
Why now:
Risks: binary print; positioning-driven gap-down even on a beat; ASIC disclosure not yet ready.
Thesis: MU is the most extended name in the universe — +187% above 200-DMA, mcap $1.1T, 0% off 52-week high — heading into a print where consensus has caught up. The risk-reward into 25 Jun is asymmetric to the downside.
Why now:
Risks: clean beat-and-raise + a third-hyperscaler HBM win could take MU to $1,100+; HBM tightness could persist longer than we model.
Thesis: Lenovo is up +21.9% in a single session on a Q4 FY26 print, +124% above its 200-DMA, and is now priced as a winning AI-server pure-play despite ~70% of revenue still being PC. We fade the move.
Why now:
Risks: China retail-investor flow keeps the move extended; further hyperscaler AI server wins announced.
Thesis: Intel's 18A node is shipping but the datacenter ramp is structurally slower than the buy-side has now priced in. Today's -5.1% intraday move is the first crack.
Why now:
Risks: 18A external customer announced (TSM-like halo); govt support / Chips-Act-2 packaging windfall; Mobileye spin-off.
Thesis: The market has paid for AMD's MI400 win and discounted QCOM's PC + auto AI silicon thesis. We expect this gap to compress over the next two months.
Why now:
Risks: AMD MI400 ramp surprises positively at the August print; QCOM Apple-modem loss timing pulled forward.
| Risk | Sector exposure | Hedge |
|---|---|---|
| Hyperscaler capex cut | All longs | Short SMH or partial NVDA short |
| China policy escalation (export controls v2) | SMIC, Hua Hong, Cambricon | Long ASML hedge (decoupling beneficiary) |
| Taiwan geopolitical tail | TSM, AVGO, NVDA | Defence basket / gold |
| HBM oversupply (CXMT acceleration) | MU short benefits, Hynix long hurts | Pair Hynix vs. MU |
| Inventory correction in PC / handset | QCOM, Lenovo | Already short Lenovo |
| AI-spend reset (CFO discipline) | Whole stack | Reduce gross exposure; cut to defensive (TSM, ANET) |
Live market data sourced exclusively from yfinance MCP (the only available stock-data MCP in this environment; paid vendors — CapIQ, FactSet, LSEG, Daloopa, etc. — are not provisioned). All quotes, technicals, and earnings dates verified intraday on 30 May 2026.
Earnings dates from yfinance are best-effort and routinely off by 1 trading day — we've cross-checked against IR pages for the top 10 names but readers should reverify before sizing earnings-anchored trades.
Fundamentals (revenue, margins, multiples, TAM) are triangulated from public disclosures and sell-side frameworks; numbers should be treated as indicative, not primary-source. Cite a primary vendor (FactSet/CapIQ/Bloomberg) before publication.
Conflicts: the desk is long TSM, ANET, Hua Hong, and Hynix in our model book at the time of writing. We are short MU and Lenovo. Positioning has been disclosed before publication and will be updated in our next morning note.
End of report. Next refresh: post-AVGO earnings (5 June 2026).