| Ticker | Price | Day % | Fwd P/E | Note |
|---|---|---|---|---|
| GOOGL | $317.69 | −7.13% | 29.4 | Cloud +82%; $514B backlog; 52wk range $187–409 |
| NVDA | $208.76 | −1.56% | ~24–26* | Rev +71% YoY; PEG ~0.35; 52wk high $236.54 |
| MU | $990.21 | +3.20% | — | Only green AI name; Musk thanked for memory allocation |
| AVGO | $392.47 | −1.09% | — | Custom-silicon / inference ASIC beneficiary |
| TSM | $415.58 | −1.34% | — | N3 utilization >100% H2'26; prints the whole complex |
| MSFT | $381.58 | −2.24% | 28.3 | Explicitly leaning into open models (no own frontier model needed) |
| META | $606.10 | −3.36% | 26.3 | Reports next week; open-model consumer of choice |
| AMZN | $233.66 | −4.57% | 32.9 | >$200B capex; reports next week |
| TSLA | $319.69 | −14.52% | — | Earnings/FCF miss; capex balloon |
| BABA | $114.06 | −2.14% | — | Qwen "open-source leader" narrative under threat from K3 |
pe_forward field is unreliable and overstates — it returned 42.3x for NVDA and 80.4x for AVGO, which conflict with Street consensus; those raw values were not used.
Two of the sharpest primary reads on this — Ben Thompson's "Who's Afraid of Chinese Models?" and SemiAnalysis's "AI Value Capture" — converge on a conclusion the tape ignored today:
Why it's the best-hedged AI name for a parity world — and why today's 7% drop is a gift:
Nvidia is the purest "open source expands inference" trade. Revenue is growing ~71% YoY yet the stock trades at a forward P/E of only ~24–26x on consensus FY2027 EPS of ~$8.16–8.99 (Zacks/Street) — a PEG of roughly 0.35, extraordinarily cheap for a ~70%-grower. (An earlier version of this note cited 42x from a raw data-vendor field; that figure was wrong and overstated the multiple — the correction makes NVDA cheaper, not dearer.) FY2028 consensus EPS of ~$10.90 puts it at ~19x two years out. SemiAnalysis's key insight: Nvidia is deliberately underpricing (acting as "the central bank of AI") — it has un-used pricing power, especially on Rubin-era SOCAMM memory. The Amkor $1.5B packaging deal signed today confirms it is racing to add capacity, not defending demand. Every open model that ships enlarges the addressable compute base that runs on Nvidia.
Memory is the single tightest constraint in the entire AI supply chain — DRAM prices are up 6x in a year, fabs are above 90% utilization, and MU was the only green AI name today (+3.2%) after Musk publicly thanked it for a "significant" memory allocation. Relative strength on a −$797B Mag-7 day is a tell. Speculative because it is the most cyclical link — a memory-cycle rollover is the risk.
The single most important tell of the day: while the mega-cap AI spenders (GOOGL −7.1%, TSLA −14.5%, AMZN −4.6%) were crushed, the picks-and-shovels that receive that capex were mostly green on a −$797B Mag-7 day. That is textbook rotation: Alphabet raising 2026 capex to $195–205B and Amazon to >$200B is a direct revenue line for the power, cooling, optical and networking suppliers. "AI capex is someone else's revenue." NVDA's FY2027 data-center revenue is already running $75B/quarter with hyperscaler capex budgets raised to ~$725B for 2026.
| Name | Price | Day % | Fwd P/E | Read |
|---|---|---|---|---|
| NVDA | $208.76 | −1.6% | ~24–26x | Accelerator monopoly; Vera Rubin ships H2'26; cheapest fwd multiple since 2019 |
| AVGO (Broadcom) | $392.47 | −1.1% | ~high-30s* | Custom-ASIC king; AI rev +100%+; the #1 hedge if hyperscalers move off Nvidia |
| MRVL (Marvell) | $209.32 | −0.8% | — | Custom-ASIC #2; rev +34%; 52wk $61→$330 shows how violent this trade is |
| TSM | $415.58 | −1.3% | ~27x | Foundry monopoly; N3 utilization >100% H2'26; prints the entire complex |
| ASML | $1,803 | +0.1% | — | EUV monopoly; the toll-collector on the toll-collectors; green today |
| LRCX (Lam) | $319.78 | +0.2% | — | WFE/etch; memory-capex leverage (HBM/DRAM buildout) |
If open source commoditizes the model but explodes inference volume, Broadcom and Marvell are the leveraged winners — every hyperscaler custom ASIC (Google TPU, Amazon Trainium, Meta MTIA) is a Broadcom/Marvell design win, and cheap open models make self-hosting on custom silicon more attractive, not less. TSMC and ASML win no matter whose chip wins.
| Name | Price | Day % | Read |
|---|---|---|---|
| GEV (GE Vernova) | $1,031.19 | +4.7% | Gas turbines — the scarcest item in the buildout; up 4.7% on a red tape day |
| VRT (Vertiv) | $304.04 | +1.0% | Power + liquid cooling inside the rack; direct capex passthrough |
| ETN (Eaton) | $415.13 | +2.0% | Electrical distribution / switchgear; long-cycle backlog |
| VST (Vistra) | $168.98 | +1.3% | Merchant power (IPP); data-center PPAs |
| CEG (Constellation) | $275.60 | +0.3% | Largest US nuclear fleet; 24/7 clean baseload for AI |
Power is the true binding constraint — US data-center demand is forecast to roughly double to ~66 GW by 2027 (Goldman). This basket is the least "commoditizable" part of the AI stack: an open-weight model doesn't reduce the megawatts needed to serve it — arguably it increases them (Jevons). GEV and VRT are the cleanest expressions; both showed relative strength today.
| Name | Price | Day % | Read |
|---|---|---|---|
| COHR (Coherent) | $313.22 | +0.3% | Optical transceiver + laser leader; intraday high $327.78 |
| LITE (Lumentum) | $833.64 | +0.5% | Optical components duopoly with COHR; intraday high $897 |
| FN (Fabrinet) | $517.73 | +0.8% | Nvidia's optical-module assembly partner |
| CIEN (Ciena) | $407.53 | +2.6% | DCI / long-haul optical; strong tape today |
| ANET (Arista) | $176.61 | +1.0% | Back-end AI Ethernet switching (watch Nvidia bundling risk) |
光通信 (optical) is the highest-beta way to play scaling AI clusters. As GPU clusters scale, interconnect bandwidth scales super-linearly — Nvidia is the single largest buyer of optical modules, with 80%+ of 2026 1.6T demand tied to it. The dominant module makers are the Chinese "易中天" trio — 中际旭创 Innolight (SZ 300308), 新易盛 Eoptolink (SZ 300502), 天孚通信 (SZ 300394) — which sit at all-time highs; US/global exposure is via COHR, LITE (components), FN (assembly) and CIEN. Key risk to name plainly: co-packaged optics (CPO) and active copper cabling (AEC, favored inside GB200/Rubin racks) could disrupt the pluggable-transceiver supply chain over 2027+ — so treat optical as a higher-risk, higher-torque sleeve, not a core hold.
| Scenario | Prob. | Trigger | GOOGL / NVDA path |
|---|---|---|---|
| Bull | 45% | MSFT/META/AMZN print strong cloud + AI-revenue next week; oil rolls back under $90; capex fear fades as backlog is re-appreciated | GOOGL retraces to $360–380; NVDA to $240–250 |
| Base | 40% | Mixed megacap prints; capex-vs-FCF debate persists; oil elevated but stable | GOOGL chops $310–345; NVDA $205–225; accumulate on dips |
| Bear | 15% | Oil spikes further on Iran escalation; a megacap guides down cloud; broad de-risking of AI capex | GOOGL tests $290 (invalidation); NVDA tests $185; step back, re-enter lower |
The honest conclusion the tape missed: open source reaching parity is bullish for compute, memory, and the integrated hyperscaler, and bearish only for closed-model pricing power. The best single trade for that thesis is long GOOGL into the post-earnings dip (owns the model, the chip, the cloud, and the distribution — the ultimate "commoditize your complement" beneficiary on the U.S. side), paired with long NVDA as the inference-volume toll and a speculative long in MU on the memory squeeze. Scale entries — oil above $100 and next week's megacap earnings are the near-term swing factors.
Key sources: Alphabet Q2 2026 earnings (CNBC) · Kimi K3 (CNBC) · Stratechery: Who's Afraid of Chinese Models? · SemiAnalysis: AI Value Capture · Counterpoint: Open Intelligence · Business Insider: Kimi K3 pricing. Prices via Finnhub, timestamped 2026-07-23 23:09 UTC.
This is analysis and opinion for informational purposes, not personalized investment advice. Do your own diligence and size positions to your own risk tolerance.