I Was Wrong. This Is a Historic Buying Opportunity.

Finance-Focused Summary (Markets, Investing, Companies, Strategy, Risks)

Market Regime / Sentiment

• The CNN Fear & Greed Index has been near “extreme greed” for ~3 weeks (the longest stretch so far this year), suggesting stretched positioning despite ongoing background risks such as war, supply-chain issues, and rising costs.
• Warren Buffett / Berkshire Hathaway is presented as a counter-signal:
  • Berkshire reportedly holds ~$400B cash, about 32% of its portfolio (described as an all-time record).
  • The implication: even with euphoric sentiment, a major institution may see value ahead.

“Where Opportunity Is” Thesis

The presenter argues that as large players resume deploying cash, opportunities will concentrate in AI infrastructure, especially:

• AI data center CPUs (not just GPUs)

A key framework correction is highlighted:

• The presenter previously underestimated CPU requirement scaling for agentic AI and became too GPU-centric.

---

Tickers / Assets / Instruments / Sectors Mentioned

Companies / Implied Tickers

• Nvidia (NVIDIA) — used for GPU comparisons (e.g., “Vera GPUs,” “Jensen Huang” numbers)
• Advanced Micro Devices (AMD) — earnings and AI CPU opportunity; deal activity with Meta
• Intel (INTC) — foundry/customer wins; earnings reaction
• Meta Platforms (META) — committed 6 GW of AMD Instinct GPUs; also framed as a potential early flagship for ARM AGI CPU
• OpenAI — referenced in connection with AMD warrant dilution
• Arm (ARM) — launched AGI CPU; royalties and supply constraints
• Apple — possible manufacturing talks; previously used Intel chips (2006–2020) then moved to M1 (per presenter)
• Google — multi-year Intel deal for custom chips
• Tesla — linked to Intel Terafab in Austin (as a customer)
• SpaceX
• XAI — linked to Intel factory activity
• Qualcomm — referenced historically as part of Arm’s ecosystem
• Sub Quadratic (SubQ 1M preview) — startup discussed for AI efficiency

Benchmark / Index

• S&P 500 (used as a growth-rate comparison)

---

Key Numbers & Claims

Sentiment & Capital Allocation

• CNN Fear & Greed: “extreme greed” for ~3 weeks
• Berkshire Hathaway
  • Cash: ~$400B
  • Cash as % of portfolio: ~32%

AI Infrastructure Spending (Earnings Backdrop)

• “Largest tech companies” announced >$700B in AI infrastructure spending this year
  • +77% YoY
  • Even with supply chains “at a standstill”

CPU vs GPU Scaling (Agentic AI Framework)

• Traditional data centers: ~1 CPU per 8 GPUs
• Agentic AI workloads are argued to be CPU-centric (tool calls, orchestration, context/memory)
• Jensen Huang estimate (as quoted by presenter):
  • 12,000 GPUs require 400,000 CPU cores
  • This implies roughly ~33:1 CPU-to-GPU ratio (cores-based)

Rack / Chip Math Presented

• Reuben GPUs per rack: 72
• Vera rack CPU configuration:
  • 256 CPUs per rack
  • 88 cores per CPU
  • → ~23,000 CPU cores per Vera rack
• Presenter correction (rack confusion):
  • For 167 GPU racks, CPU racks needed:
    • ~18 CPU racks
    • i.e., ~9:1 (GPU racks : CPU racks) (described as rack-based confusion)
• Chip-count correction:
  • Needs ~4,600 Vera CPUs per 1,200 Reuben GPUs
  • Equivalent: ~1 CPU per 2.6 GPUs

AMD Earnings & Growth

• AMD quarter revenue: $10.3B, +38% YoY
• AMD data center revenue: $5.8B, +57% YoY
• Claim: most growth is in data center CPUs (not GPUs)

Meta Commitment to AMD GPUs & Warrant Structure

• Meta committed 6 GW of AMD Instinct GPUs
  • 1 GW: fully customized MI450s for Meta’s workloads
• Warrant described:
  • AMD gave Meta a warrant for ~10% of AMD at $0.01 per share
  • Shares only if AMD hits $600
• Dilution framing:
  • Dilution could be ~20% total because OpenAI has a similar deal
• “Net” valuation framing (presenter logic):
  • If $600 trigger occurs, value “worth” about $480
  • Only ~$30 more than current level (per presenter)
• Investment implication presented:
  • Despite dilution, AMD’s long-term outlook is supported by CPU market growth.

AMD CPU Market Growth (Lisa Su cited)

• Data center CPU market “nearly triple” by 2030
  • Implied CAGR: 35%
  • ~3x faster than the S&P 500 (comparison)

Intel Foundry / Customer Narrative (April Wins)

• Intel Terafab:
  • $25B chip factory in Austin
  • Customers listed: Tesla, SpaceX, XAI
• Intel multi-year deal with Google for custom chips (internal cloud)
• Bloomberg claim (per presenter):
  • Apple in early talks with Intel and Samsung about manufacturing chips in the US
• Intel earnings:
  • Revenue: $13.6B, +7% YoY
  • EPS: 29 (presented as up from 13 cents last year)
  • Stock reaction: +24% the next day
  • Described as best day since the dot era

Arm AGI CPU Specifics & TAM Math

• Arm launched AGI CPU (first Arm-designed/sold chip)
• Specs presented:
  • Up to 136 cores per chip
  • 60 chips per rack
  • Cores per rack: just under 8,200
• CPU support math (corrected):
  • Nvidia: 1 CPU per 2.6 GPUs
    • 4,600 Vera CPUs for 12,000 Reuben GPUs
  • Arm: 1 CPU per 4 GPUs
    • Needs ~3,000 Arm CPUs
• Claimed benefit:
  • ~54% better performance
  • or ~40% fewer CPUs for the same data-center workload
• Efficiency / cost claims:
  • “Roughly double performance per watt versus Intel and AMD”
  • “Data center construction cost savings”: ~$10B per gawatt (compute), claimed by Arm
• Meta implication (as argued):
  • If Meta uses ARM CPUs instead of AMD for the 6 GW project:
    • ~$60B construction cost savings (presenter ties this to AMD/Meta payment/warrant logic)
• Arm demand & financial expectations:
  • Demand “doubled” within first 6 weeks
  • Expected $1B+ CPU sales in first year
  • Expected $15B annual chip revenue by 2031 (presenter notes current revenue < $5B)
  • Supply constraint: stock dropped ~10% due to inability to build fast enough
  • Growth claim: “quadruple” annual revenue over ~5 years

Sub Quadratic (Unverified) AI Efficiency Breakthrough

• Model: SubQ 1M preview
• Research version:
  • 12M token context window (up to 12x bigger than “most Frontier models”)
• Planned shipping version:
  • 1M tokens (matches peers)
• Cost claim:
  • >300x cheaper to run
  • Example (presenter framing): “price of a cup of coffee” vs higher compute cost elsewhere
• Funding / valuation:
  • Seed funding: $29M
  • Launch valuation: $500M
• Due-diligence caution (explicit):
  • No peer-reviewed paper
  • No public test model
  • No reliable external benchmarks
• Conditional monitoring plan:
  • If Anthropic / OpenAI / Google publish responsive work on “subquadratic attention,” claims may be validated
  • Otherwise, might be “smoke and mirrors”
• Investment implication if true:
  • Cheaper inference → more demand
  • Potential positive spillover to CPU demand (compute becomes cheaper per task)

---

Methodology / Framework

Agentic AI CPU Sizing Framework (Correction Highlighted)

1. Start with traditional ratio: ~1 CPU per 8 GPUs
2. Apply agentic AI workload assumption:
   • tool calls, sub-agents, orchestration, context management → runs on CPUs
3. Use Jensen Huang’s system-level core estimate:
   • 12,000 GPUs require 400,000 CPU cores
4. Translate cores into rack requirements using provided rack specs:
   • Reuben GPUs per rack: 72
   • Vera rack: 256 CPUs/rack, 88 cores/CPU → ~23,000 cores per Vera rack
5. Correct “rack vs chip” mistake:
   • racks can mislead; align chip/core counts with implementation needs
6. Final conclusion:
   • For AI data centers: ~4,600 Vera CPUs per 1,200 Reuben GPUs
   • Equivalent: approximately 1 CPU per 2.6 GPUs

Investing Framework Emphasis

• “Be fearful when others are greedy; greedy when others are fearful” (Buffett referenced)
• Watch where institutions deploy capital next:
  • implied watchpoints include cash deployment, buybacks, and capex winners

---

Recommendations / Cautions (As Stated)

• Presenter frames investing opportunity as CPU beneficiaries of agentic AI efficiency and infrastructure spend.
• Key uncertainties / downside risks:
  • Sub Quadratic efficiency claims are explicitly “not verified”
  • Arm growth could face limits due to supply constraints
  • AMD warrant/dilution structure is presented as a material downside/cost to the Meta deal
• No formal “buy/sell” ticker recommendations are provided in the subtitles.

---

Disclosures / Disclaimers

• No explicit “not financial advice” line appears in the provided subtitles.
• The presenter states Sub Quadratic efficiency claims are not verified and could prove either meaningless or transformative.

---

Presenters / Sources Mentioned

• Alex (presenter; described as ex–MIT electrical engineering and AI researcher)
• Warren Buffett (Berkshire meeting comments)
• Greg Ael (Berkshire’s new CEO referenced)
• Jensen Huang (numbers cited; attributed to GTC 2026 per presenter)
• Lisa Su (AMD quote cited)
• Arm CFO (expectations cited)
• Bloomberg (reported Intel/Apple manufacturing talks per presenter)
• Anthropic, OpenAI, Google (referenced as potential sources to validate Sub Quadratic “subquadratic attention” claims)