The Kimi K3 Repricing: Is Any AI Stock Actually Undervalued — And What Trade To Make If Open Source Reaches Parity by Year-End

Daily Stock & Crypto Analysis  ·  @dailyanalysts  ·  July 23, 2026 (all prices at U.S. close, ~7:09pm ET)
Highest-conviction finding: The market read the Kimi K3 open-weight shock and Alphabet's capex hike as bearish for the entire AI complex — that read is half-wrong. "Open source catching up" is deflationary for closed-model pricing rents but powerfully inflationary for compute volume and it is a gift to the one company that owns the whole stack. Alphabet (GOOGL) at $317.69, down 7.1% on the day, is the single most mispriced large-cap AI name today. If you believe open models reach parity by year-end, you want to own the toll-collectors (compute + memory) and the vertically integrated hyperscaler that pays no model rent — not sell them.

1. What actually happened (the facts)

2. Real-time price board (July 23, 2026 close)

TickerPriceDay %Fwd P/ENote
GOOGL$317.69−7.13%29.4Cloud +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%28.3Explicitly leaning into open models (no own frontier model needed)
META$606.10−3.36%26.3Reports 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
Source: Finnhub real-time quotes, retrieved 2026-07-23 23:09 UTC. *Forward P/E shown for NVDA is the consensus non-GAAP estimate (FY2027 ending Jan 2027 EPS ~$8.16–8.99 per Zacks/Street → ~24–26x at $208.76). Note: Finnhub's raw 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.

3. The core logic: what "open source on par" actually does

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:

  1. Open weights are NOT free to serve. R&D is a fixed cost; inference (COGS) scales with revenue. Kimi K3 still costs $3/$15 per Mtok, and per Stratechery it burns more reasoning tokens than Sol, muting its price advantage. Tokens are not a commodity — intelligence is, and the cheapest producer of a correct answer wins.
  2. Commoditized models expand the compute market, they don't shrink it. Patrick Moorhead (Moor Insights): LLMs like K3 will "accelerate and grow the inference market faster than without." This is Jevons' paradox — cheaper intelligence → far more usage. SemiAnalysis: Anthropic's ARR went $9B→$44B while inference margins went 38%→70%; demand is "compounding, not linear."
  3. Value is shifting to: compute (NVDA/TSMC), memory (the tightest constraint, up 6x in a year), hyperscalers/inference providers, and integrated players who own distribution. It shifts away from anyone whose entire moat is a proprietary closed model sold at a premium.
  4. China's strategy is "commoditize your complement." Xi explicitly tied open weights to AI "moving into the physical world" (robotics). Open models are a distribution weapon, not a profit center — which is precisely why they keep the pressure on closed-model pricing but do nothing to hurt the compute layer.

4. So which AI stock is undervalued? Alphabet.

HIGH CONVICTION  LONG GOOGL
Entry zone: $305–320 (current)  |  Target: $380 (1–3 months), $420+ (6–12 months)  |  Invalidation: sustained daily close below $290 (would confirm the capex-panic / cloud-deceleration thesis)  |  Timeframe: 1–3 months, with a 6–12 month leg.

Why it's the best-hedged AI name for a parity world — and why today's 7% drop is a gift:

5. The cleanest expression of the thesis: own the toll-collectors

HIGH CONVICTION  LONG NVDA
Entry zone: $195–210  |  Target: $250 (1–3 months)  |  Invalidation: daily close below $185  |  Timeframe: 1–3 months.

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.

SPECULATIVE  LONG MU (Micron)
Entry zone: $940–990  |  Target: $1,150 (1–3 months)  |  Invalidation: daily close below $870  |  Timeframe: 1–3 months.

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.

5b. The sub-sectors: chipmakers, supply chain, data-center power & 光通信 (optical)

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.

Chipmakers & silicon

NamePriceDay %Fwd P/ERead
NVDA$208.76−1.6%~24–26xAccelerator 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%~27xFoundry 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.

Data-center power, cooling & electrical (the physical bottleneck)

NamePriceDay %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.

光通信 / optical interconnect & networking

NamePriceDay %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.

SPECULATIVE  Optical / power basket idea — if you want leverage to the capex itself rather than the spenders: a basket of GEV + VRT + COHR/LITE. Entry on any broad-market pullback; these ran today, so chase is the risk — prefer adding into red days. Invalidation: a hyperscaler capex guide-down (none in sight — guides are going up).

6. What NOT to do

Do not sell the compute/hyperscaler complex on the Kimi headline. The instinct to treat "China caught up" as a reason to dump NVDA/GOOGL/MU is the DeepSeek-January-2025 mistake repeating. The names genuinely at risk in a parity world are closed-model rent extractors and application-layer software with no cost advantage and no model of their own — not the toll-collectors. Be cautious paying premium multiples for pure-play SaaS whose only pitch is "we wrap GPT/Claude"; commoditized models compress their differentiation, not Google's or Nvidia's.

7. Three-scenario framework (next 1–3 months)

ScenarioProb.TriggerGOOGL / NVDA path
Bull45%MSFT/META/AMZN print strong cloud + AI-revenue next week; oil rolls back under $90; capex fear fades as backlog is re-appreciatedGOOGL retraces to $360–380; NVDA to $240–250
Base40%Mixed megacap prints; capex-vs-FCF debate persists; oil elevated but stableGOOGL chops $310–345; NVDA $205–225; accumulate on dips
Bear15%Oil spikes further on Iran escalation; a megacap guides down cloud; broad de-risking of AI capexGOOGL tests $290 (invalidation); NVDA tests $185; step back, re-enter lower

8. Bottom line

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.