Video summary

How Kimi K3 Is Reshaping AI Investing

Main summary

Key takeaways

Finance

Market & Sector Context

  • Technology/public markets rebounding

    • Nasdaq 100 up ~1%, with tech stocks up ~2%.
    • Strength led by compute names including NVIDIA and memory stocks.

  • Bear-market pressure in chips (temporary)

    • Chip stocks “fell into a bear market” on Friday.
    • This was tied to investor questions around China’s “Kimi k3 / Chemie k three” model release.

  • Key debate driving volatility

    • Whether more efficient Chinese AI models mean less hardware demand.
    • Whether China is closing the AI gap faster than expected.
    • Where AI value accrues across the stack (models vs. inference vs. hardware).

Key AI Investing Thesis (Venture / Private Markets)

  • China & open-weight models closing the gap with frontier models

    • Kimi k3 characterized as similar to earlier “breakthrough momentum” (the “DeepSeek moment” reference).
    • Claims that China and open-weight models are closing the gap with frontier models.

  • “Leapfrogging” dynamic (platform/product competitiveness)

    • Ongoing improvements can cause firms that don’t out-innovate to be leapfrogged.
    • Competitive differentiation increasingly shifts to data, performance, and customer resonance.

Framework / Portfolio Lens: Vertical AI vs. AI Infrastructure

The guest uses a stack-style framework:

  • AI Infrastructure (systems for building/training/running models)

    • First wave: scalable model building
    • Next wave: deploy quickly, fastly, cheaply
    • Agents as a deployment-efficiency mechanism
    • Infrastructure includes:
      • Tooling around models (“reasoning layer” + tools)
      • Connecting to external data and apps
      • Using memory to improve outcomes
    • Sector sub-areas mentioned:
      • Agent infrastructure, including agent harnesses and an identity layer

  • Vertical AI

    • AI applications tailored to specific industries/use cases (contrasted with infrastructure).

Pricing, Unit Economics, and Demand Substitution (Explicit Numbers)

  • Inference cost decline

    • Highlighted as cost down ~95% (inference infrastructure context).

  • Core argument: lower inference/model costs may not reduce total spend

    • Instead, it may shift workloads
      • Frontier workloads (expensive/highest reasoning) move to cheaper models
      • Customers still achieve similar required compute outcomes at lower cost
    • Emphasis: maintain latency and performance while reducing cost.

  • Kimi k3 token pricing (unit economics)

    • $3 per input token
    • $15 per output token

  • Caution / balance

    • “Open-weight” does not mean “free to run.”
    • Real compute costs and infrastructure needs remain.
    • Model pricing may not map 1:1 to long-term value/revenue capture.

Valuation Discussion: Model Economics vs. Big Tech Budgets

  • A “fundamentals” argument was referenced:

    • If a $20B Chinese startup can run a 2.8T (2,800,000,000,000 parameter) model with certain economics,
    • why are investors funding/“signing” OpenAI and Anthropic at valuation levels described as “evaluation of a trillion” (wording implying very high valuations).

  • Response / uncertainty

    • No definitive conclusion yet (“don’t know yet / nobody knows yet”).
    • Even if benchmarked economics look similar, OpenAI and Anthropic still generate billions in revenue, including:
      • First-party businesses
      • Third-party businesses
    • Open-weight models: not necessarily monetized as cheaply as they’re run.

Capital Markets Signals (IPO Window / Investor Preferences)

  • Recent high-profile IPO signals mentioned:

    • SpaceX
    • SK Hynix issuing ADRs (noted as happening in New York City)

  • Interpretation

    • Investors still want exposure to AI buildout—especially AI infrastructure.
    • The IPO reopening narrative may be true, but investors appear to prefer companies that are:
      • Long-standing
      • Have billions in revenue
      • “Hardly speculative” compared with early-stage tech

Tickers / Instruments / Assets / Sectors Mentioned

  • NVIDIA (equity; “compute names”)
  • Nasdaq 100 (index)
  • SK Hynix ADRs (depository receipts; issuer mentioned)
  • SpaceX (company; IPO mentioned)
  • Sectors / concepts (non-ticker)
    • Technology sector
    • Chips / semiconductors
    • Memory stocks
    • AI infrastructure
    • Open-weight AI models
    • Frontier models
    • Venture growth investments
    • Inference infrastructure layer
    • Agents / agent infrastructure
    • Identity layer

Key Presenters / Sources (Named)

  • Don (Donohoe) — Crosslink Capital partner (focus: mid-stage venture; vertical AI and AI infrastructure)
  • Crosslink Capital (source/organization referenced)
  • Bessemer (career background referenced; not a presenter)
  • Greylock — founder/host Mohamed El-Erian referenced as being on the show (per subtitles)
  • OpenAI (referenced)
  • Anthropic (referenced)
  • Kimi k3 / “Chemie k three” (referenced; attributed to “China/moonshot” in subtitles)

Disclosures / Disclaimers

  • No explicit “not financial advice” or formal disclaimer text was included in the provided subtitles.

Original video