AGI Has Arrived. Here's How I'm Investing In It.

Big-picture market thesis (AGI/agentic AI impact)

• The speaker argues that the shift from AI chat “prompts” to AI “agents”—systems that can be given a goal and run tools for long periods—is a major economic change.
• They claim this shift should affect:
  • Compute demand
  • Data-center hardware mix
  • Supply-chain pricing power / margins
• Market framing:
  • “Whether or not this is true AGI… doesn’t matter to the market.”
  • What matters is whether AI is producing useful work autonomously today.

Method / investment framework implied

• Instead of betting on a single “winner,” the speaker suggests owning critical layers of the agentic AI stack.
• The focus is on layers with only 2–3 key suppliers per layer.
• Rationale:
  • “It’s easy to own most of every market,” reducing the need to predict one winner.

Key timeline / catalysts mentioned

• ~2 weeks prior (from video timing): Nvidia earnings call Q&A where Jensen Huang says tracking AGI milestones is “senseless.”
• 11 days later: OpenAI announces Astra; the speaker claims it marked “AGI has arrived.”
• Next ~days:
  • Astra demonstrations
  • Nvidia agrees to buy Hugging Face for $12.9B (per subtitles)
• Claimed framing:
  • “From ChatGPT to 01 to Astra in 4 years.”

---

“GPT6 Astra” and benchmarks (relevant to market narrative)

What Astra is (capability claims)

• The speaker frames Astra as moving from chat to agentic work by using:
  • Web search
  • Writing/running code
  • Access to software/spreadsheets
  • Ability to use a computer via screen/mouse/keyboard
• Reported workflow:
  • Breaks goals into steps
  • Selects tools
  • Checks answers
  • Iterates until done
• Hardware/agent loop implication:
  • Agents can run continuously for long durations (examples: half an hour to 5 hours, and up to 5 days).

Quantitative performance claims (benchmarks)

• PCB design / electronics workflow: lays out PCB under 3 minutes (including routing/trace details)
• 3D reconstruction accuracy: 96% vs prior best 84%
• Scientific workflow: 65% vs 53% (12 points better than “Anthropic’s Fable 5.1”)
• Legal document checking: finds 4 errors in 41 financial documents
• Coding / other AGI evaluations:
  • “Humanity’s last exam” score: 57%
  • Astra coding score: 3 points worse than Fable 5
  • Costs: “less than half as much as Fable 5 to run”

AGI definition discussion (investment relevance)

• The speaker suggests benchmarks may not directly measure “intelligence.”
• Market signal emphasized instead:
  • Agents requiring less supervision
  • Doing more productive work

---

Explicit investing “winners” / stock picks by stack layer (tickers included)

Note: The subtitles are promotional; the channel name appears as “Tickerol U.” The speaker says “which stocks I’m buying as a result,” and then lists multiple companies/tickers.

1) CPUs / compute supply constraint (pricing power)

• AMD
  • Mentioned in a context involving “Helios racks” and CPU-to-GPU ratios (e.g., “18 epic CPUs for every 72 GPUs”).
• ARM
  • Mentions ARM cores/royalties and growth in data center royalties (more than doubled YoY).
  • Claims ARM server chip “AGI CPU” has >$2B customer demand, with Meta Platforms as a lead partner.
  • Risk/caution: ARM now has wafer/manufacturing costs, potentially pressuring margins vs pure licensing.
• Nvidia
  • Mentions custom CPU efforts (“Vera”):
    • First 88-core Vera CPUs reportedly headed to Oracle, SpaceX, Anthropic, OpenAI (for “Grock/Claude/ChatGPT” per subtitles).

2) Memory (HBM/DRAM) for long-running agent jobs

• SK hynix (SKHY)
  • Claims: “making around half of the world’s HBM”
  • Reported shares:
    • 25% of global DRAM
    • 18.5% of global NAND
• Micron (MU)
  • Claims: 18% of HBM market; 23% of DRAM
  • “Revenue more than tripled YoY”
  • Earnings call end of September
  • Framed as a leading indicator “since Astro was announced.”
• Valuation claim:
  • Both memory names supposedly trade at forward P/E under 8, despite big YTD gains, because Wall Street treats memory demand as cyclical.
• Thesis:
  • Agentic AI may make memory demand more permanent, strengthening the memory cycle.

3) Interconnects (memory-to-compute and inside-rack data movement)

• Astera Labs (ALAB)
  • Mentioned as a controller for shared-memory access
  • “Revenues doubled YoY”
• Crato (CRDO) (ticker appears in subtitles)
  • Copper cables inside racks; “biggest product line”
  • Growth claims:
    • More than doubling in 2025
    • More than tripling in 2026
  • Company revenue up 115% YoY
  • Guidance: >85% growth
• Tower Semiconductor (TEM)
  • Silicon photonics unit:
    • Run-rate $180M → $680M in the last year
    • Expects crossing $1B this quarter
    • $1.3B silicon photonics contracts signed for 2027
• Optical components:
  • Lumentum (LIIT) (ticker appears as “Liit T” in subtitles)
    • Reported revenues: “more than doubled YoY” for co-packaged optics
    • Expected to “more than double again”
  • Coherent (CH)
    • Nvidia invested $2B into each of Coherent and Lumentum
    • Purchase commitments mentioned earlier in the year

4) Inference compute / specialized architectures

• Cerebras (CBRS)
  • Architecture claim: keeps data on-chip (wafer-as-chip concept)
  • Speed claim: serves open models “up to five times faster than Nvidia’s B200s”
  • Financial metrics reported:
    • Backlog: $25.4B
    • Revenue: $900M this year
    • Backlog multiple: 29x current business
    • Scheduled deliveries: ~20% by mid-2028
    • Implied delivery rate: about $2.8B/year (~3x this year’s revenue)
  • Risks/cautions:
    • Backlog concentration: “most backlog is a single contract with OpenAI”
    • Still unprofitable:
      • Net loss $450M on $180M revenue
      • Loss ~$2.98/share
      • >85% of loss attributed to stock-based compensation triggered by IPO
    • Cash: $8.6B
• The speaker characterizes Cerebras as an “early bet” on agentic AI inference speed.

---

Additional market/performance context (numbers mentioned)

Stock performance comparisons (historical)

• Microsoft
  • “$10,000 invested at start of the internet era” → “over $2 million today”
  • About 10x the S&P 500 return over the same period.
• Nvidia
  • Investing while ChatGPT kicked off the AI era (less than 4 years ago)
  • Claimed outcome: “around $150K,” “doubling every single year.”

AI compute demand growth

• Jensen comment: an AI agent uses 15–100x the compute of a human using the same model.
• Google token processing claim: 300x increase in tokens processed vs 2 years ago.
• Tokens scale analogy:
  • 3.2 quadrillion tokens per day ≈ reading a billion books/day (speaker analogy).

---

Disclosures / disclaimers

• No explicit “not financial advice” or regulatory disclaimer is included in the subtitles shown.
• Promotional disclosure:
  • VCX by Fundrise is described as the sponsor (“VCX is the public ticker for private tech”).
  • Access is provided via a link.
• The speaker is named: Alex.

---

Presenters / sources mentioned (at end)

• Alex (speaker; electrical engineer and AI researcher at MIT; “spent eight years” per subtitles)
• Jensen Huang / Jensen Hong (Nvidia CEO; referenced via earnings call remarks)
• Ethan Mollick (Wharton professor; provided emails/calendar/writing to Astra)
• Shane Le (credited with popularizing “artificial general intelligence” in 2002)
• UC San Diego (study referenced re GPT-4.5 / “human 73% of the time”)
• Anthropic (models referenced: “Fable 5.1,” “Fable 5,” “Opus 5”)
• OpenAI (Astra/GPT6; Astra demonstrations and prior testing referenced)
• PlayCo (game studio mentioned)
• Lagora (legal tech firm mentioned)
• Meta Platforms (lead partner referenced for ARM “AGI CPU” demand)
• VCX / Fundrise (sponsor)