Summary of "Sam Altman’s MisAI-lignment of OpenAI’s Financial Reality"

Key topics

Debate between investing in AI infrastructure (hyperscalers and hardware) versus the application/SaaS layer.
Recent Anthropic compute deal with Google + Broadcom seen as validation of strong demand for AI infrastructure and an indicator of continued capex for AI.
Large private companies and potential IPOs (OpenAI, Anthropic, SpaceX) reshaping institutional and retail capital flows.

Tickers / companies / assets / instruments mentioned

NVIDIA (NVDA) — hardware/custom silicon leader
Broadcom (AVGO)
Google / Alphabet (GOOGL)
Microsoft (MSFT)
Apple (AAPL)
Meta (META)
Amazon (AMZN)
Tesla (TSLA)
SpaceX (private; confidential IPO filing discussed)
OpenAI (private)
Anthropic (private)
Intelligent Alpha ETF — ticker GPT (Deep Water Asset Management product)
General references: “hyperscalers,” TPUs, GCP, compute tokens, capex, IPOs

Key numbers, timelines, and performance notes

Anthropic
- Run-rate revenue cited jumping from $9 billion (end of 2025) to $30 billion over ~4 months; used to explain increased compute needs and the Google + Broadcom deal.
Broadcom
- Up ~5% on the day of recording but ~25% below December all‑time highs.
NVIDIA
- Management comments implied revenue growth >40% next year vs Wall Street ~30%; NVDA fell ~8% over a 3‑day span while the NASDAQ fell ~4%.
- Street/analyst expectations referenced: large EPS growth (examples: ~70% this year, ~35% next) and gross margins discussed (~71.5% rising toward ~74%).
- NVDA trading roughly ~16x next‑year earnings (discussion).
Hyperscaler capex
- Hyperscalers guided ~60% capex growth for calendar 2026 (per speakers).
- Street expects ~10% capex growth for calendar 2027; guests expect closer to ~20% (above street but well below 60%).
Amazon capex
- Street guide cited ~ $191 billion for calendar 2026 (with ~$30–40B related to normal retail/fulfillment).
SpaceX IPO
- Confidential filing discussed; potential valuation mention up to ~$2 trillion (speaker figure). Small float expected; initial raise mentioned ~$75B (could be upsized).
OpenAI
- Not currently profitable, reportedly high cash burn (speaker referenced “$3 billion a month” contextually). Prior large private fundraising noted; a $120B figure was referenced (likely conflating commitments vs immediate cash).
Tesla deliveries
- March quarter growth: street expected 8%, actual delivered ~6% (small miss).
Apple foldable
- Market reaction: down ~5% then 2.5% after company response. Projected niche demand (~5% of iPhone sales); possible fall 2027 timing discussed.
Token pricing / compute cost dynamics
- Some commentary claims token price deflation of ~90% per year historically.
- Contrasting view from Deep Water: hyperscaler cost declines (raw compute) are one track, but price for advanced/high‑quality tokens (what funds and leading models pay) has not declined similarly and may be rising.

Methodologies, frameworks, and investment process

Capex math
- Start with hyperscaler capex guidance vs free cash flow constraints; be skeptical of models requiring sustained very high capex without FCF support.
Revenue concentration risk
- Example: NVIDIA derives ~50% of revenue from top hyperscalers — examine impact if those customers slow.
Growth vs valuation tradeoff
- Compare expected revenue/EPS growth rates to current multiples (example: NVDA growth vs ~16x forward).
Differentiate infrastructure vs application layer
- As compute commoditizes, value may shift toward application/SaaS tiers.
Token pricing analysis
- Separate hyperscaler raw compute cost trajectory from the price paid for premium tokens/advanced models.
Human‑in‑the‑loop quant strategies
- Combine machine-driven models with human oversight (example: Intelligent Alpha ETF, ticker GPT).
Practitioner approach
- Asset managers should build/operate AI tools themselves rather than treating AI as a black box.
Position sizing and psychology
- Reduce exposure when market psychology is uncertain (example: trimming NVIDIA to a half position rather than exiting).

Portfolio construction / positioning (Deep Water)

Current positioning
- Still owns NVIDIA but trimmed (reduced to “half of a position”).
- Broadcom on the watchlist — periodically owned in the past.
- Does not own Microsoft (had not owned it).
- Recently added to Apple position (belief in Apple’s opportunity in personalized AI/privacy).
- Launched Intelligent Alpha ETF (GPT) — concentrated, machine‑heavy strategy with human oversight; YTD performance tracking ahead of NASDAQ per commentary.

Risk management and cautions

Key risks
- Capex and free cash flow constraints of hyperscalers limiting upside surprises.
- Commoditization of compute/token pricing shifting value to application layer; infrastructure players may lack permanent moats without new customers.
- Revenue concentration (e.g., NVIDIA’s reliance on a few hyperscalers).
- Large private raises and IPOs reallocating capital and creating investor competition (SpaceX / OpenAI).
- High burn rates of private AI companies and uncertain paths to profitability.
- Geopolitical risk (e.g., Apple / Taiwan exposure).
- Execution risk on product launches (Microsoft Copilot, Apple Intelligence, foldables).
Psychological/market risks
- Expectations and narratives can change quickly (e.g., hyperscaler commentary causing large day moves); investor psychology can drive multiple compression despite good fundamentals.

Explicit recommendations / practical stances

No formal buy/sell commands recorded; practical guidance included:
- Be selective: own AI leaders but actively manage position sizes (example: trimming NVIDIA).
- Consider application layer/SaaS exposure if compute commoditizes.
- Monitor hyperscaler capex guidance closely — it materially affects hardware vendors’ multiples and growth expectations.
- Build and use AI tools internally as an investment firm (practice and stay hands‑on).

Notable commentary / quotes (paraphrased)

Anthropic deal = strong demand signal; Anthropic “can’t keep up” with capacity needs.

The AI cycle is still early — speakers called it the “second inning,” despite rapid developments.

Two tracks for token pricing/costs: raw compute cost vs price for premium tokens/advanced models — the latter may not deflate quickly.

SpaceX didn’t strictly “need” to IPO now but being public makes raising very large future capital cheaper and could open ownership to retail/Tesla shareholders.

Sources / presenters / references

Presenters: Dan Nathan (Okay Computer podcast) and Gan Monster (managing partner, Deep Water Asset Management).
Other referenced names: Ben Thompson (Stratechery), Sarah Friar (CFO referenced re OpenAI), The Information (reporting referenced), Doug (Deep Water team member, Intelligent Alpha lead).