The Misty AI Frontier and How to Spot Billion-Dollar Companies Before Everyone Else — Elad Gil
In a Nutshell
Elad Gil argues AI is in a consensus phase with compute/memory constraints keeping labs like OpenAI, Anthropic, and Google in an oligopoly, enabling rapid revenue scaling to $30B run rates; AI startups should exit within 12-18 months via acquisitions by hyperscalers amid 90-95% failure rates in tech cycles. Durable AI companies feature deep workflow integration, proprietary data, and sensitivity to underlying model improvements. Investment prioritizes massive markets first (e.g., early bets on Perplexity, Harvey, Anduril), aggressive distribution, and contrarian signals like "why now?" shifts, with personal tips on longevity basics (sleep, exercise, vitamin D, creatine) and cautious experiments (rapamycin, ibogaine).
These notes were generated by AI and may contain inaccuracies.
There are moments when being contrarian is smart, and moments when consensus is smartest. Right now, consensus is right. People overthink it, like doing hardware instead of buying more AI.
XAI got an option to purchase Cursor. Scale was partially taken by Meta. Various deals over the last year or two. Meta aggressively bid on AI talent, rational given tens of billions spent on compute. Normally, one company IPOs, enriching employees; some stay focused, others get distracted with passion projects, politics, startups, or check out.
Meta offers and matching by other tech companies gave 50 to a few hundred people across Silicon Valley a personal IPO, dramatically increasing pay packages. Unusual; similar to crypto early holders in 2017. Implications: some pursue big science, AI for science, personal quests, or quiet quit and chase vices. Like Austin's Dellionaires post-IPO.
All labs (OpenAI, Anthropic, Google, xAI) train giant models using clusters of hundreds of thousands to millions of NVIDIA chips, Hynix/Samsung memory, building data centers. Output is a flat file encapsulating humanity's internet knowledge, logic, reasoning—like DNA's few billion base pairs specifying the body and mind.
Constraints shift yearly: currently memory (Korean companies, 2-year bottleneck due to lowered capacity). Creates ceiling on model scale, inference. Everyone constrained, preventing one lab pulling far ahead despite scale laws. Labs stay close (OpenAI, Anthropic, Google) for next 2 years. Google constrained by Samsung/Micron memory. Some make own chips: Google TPUs, Amazon Trainium. Past constraint: packaging.
Forecasting problems easier than solutions; memory needs fabs, CapEx cycle. Companies under-invested, now catching up. Demand keeps growing due to impact.
Revenue scaling: Chart shows time to $1B, $10B, $100B revenue. ADP ~30 years; Anthropic/OpenAI ~1 year; Google 4 years. Generations scale faster. OpenAI/Anthropic ~$30B run rate (0.1% US GDP); AI to 0.5% GDP. At $100B, 1-2% GDP each. Excludes Azure/GCP/AWS cloud AI revenue.
Pass on Lyft Series C due to market structure (winner-take-all vs. oligopoly). In AI, predicted oligopoly aligned with cloud 3 years ago (OpenAI, Anthropic, Google durable; Meta, xAI others). No monopoly unless one pulls ahead in capabilities; compute constrains this.
Founders running successful AI companies should all take a cold, hard look at exiting in the next 12 to 18 months, which might be a value maximizing moment for outcomes.
Every tech cycle, 90-95% companies bust (auto: dozens to few; internet: 1500-2000 IPOs 1998-2000, ~12-24 survive). SaaS, mobile, crypto same. Ask: Is your AI company one of the durable dozen in 10 years, or commoditize/sell now at peak (6-12 month window, watch growth second derivative).
Bullish on AI transformation; handful companies vital long-term.
Core labs durable: OpenAI, Anthropic, Google (barring disaster).
Application companies (Harvey legal, Adept/Bridge health, Decagon/Sierra customer success):
- Does better underlying model make your product dramatically better?
- Deep/broad products, integrated, embedded in workflows (change management key, not just tech).
- Proprietary data capture/use (data moats overstated but sometimes useful; system-of-record view).
Unprecedented: market caps 100B-$3T; 1% dilution = $30B acquisition power.
Buyers:
- Big labs/hyperscalers/giant tech: Apple, Amazon, Google, Oracle, Samsung, Tesla, SpaceX, Snowflake, Databricks, Stripe, Coinbase (vertical-specific).
- Vertical incumbents (Thomson Reuters for legal/accounting).
- Merge competitors (e.g., X.com/PayPal; Uber/Lyft rumor) to stop pricing wars, combine to win vs. incumbents.
If the primary purpose is to win the market, you're already fighting big incumbents, so why make it even harder?
Biology and math background contributed to investing thinking. Math helped with technical or algorithmic CS understanding, fluency with numbers and data, and logical step-by-step thinking from pure math proofs. Proofs involve logical sequences, intuitive leaps, then fleshing out reasoning.
No single moment of realizing skill in investing; hard on self and second-guesses even now. Organically continued because getting into strong companies like Airbnb and Stripe allowed persistence.
Key to early success: moved to Silicon Valley in 2000, then San Francisco. Being in industry clusters is crucial—Bay Area for tech/AI (91% of global private AI market cap in 10x10 area), New York secondary, then drops off. Defense tech: Southern California (SpaceX, Anduril, Irvine, Orange County, El Segundo). Fintech/crypto: New York.
Shreya's team unicorn analysis: traditionally half US private market cap in Bay Area; with AI, 91% there.
Great networks from running startup, Google tenure. Airbnb: helped 8-person team with Series A strategy and intros, offered investment. Stripe: sold early API companies, emailed Patrick after hearing great things, went on walks, then invited to invest. Twitter at ~90 people. Organic: founders wanted involvement because liked talking to smart people, working on business problems, technology translation.
If you want money, ask for advice. If you want advice, ask for money. Offering advice often leads to investment opportunities.
High Growth Handbook published ~7 years ago.
Still agrees 90%: investments market first, team strength second. Exceptions: exceptional people (e.g., led Perplexity first round—Arvind from OpenAI pinged on LinkedIn pre-AI hype, met biweekly brainstorming, built prototypes fast) or very early. Anduril: Google shut down Maven (defense project), saw startup opportunity in unpopular defense tech; met Trey Stevens (co-founder, Founders Fund) at lunch.
Funded OpenAI, Harvey (early legal), early AI because only ones in important market.
AI always promised; math degree included theoretical CS, early neural nets. Google first AI company (then called machine learning). Key milestones:
- 2012: AlexNet proved scaling works.
- 2017: Google transformer architecture (basis for everything; GPT = Generative Pre-trained Transformer).
- ~2020: GPT-3 huge step from GPT-2, generalizable via API, scaling laws.
Extrapolated importance from capability steps, playing with tech, scaling law papers.
Math background helped ingest/understand, but also broad info ingestion, friends discussing (though toy-like), personally playing (GANs art, tried hiring for Midjourney-like).
Pre: MLOps teams for data/pipelines, custom models per use case, internal services—huge pain. Now: API call (few lines code), generalizable (not specialized like spell correction), embeds internet knowledge, advanced reasoning, no MLOps/hosting needed.
Using AI for 20-year angel analysis: intros, consistent good sources.
Elad: uploads founder photos, prompts models to predict founder quality from micro features (crow's feet = genuine smiles/humor, furrowed brow). Works well; explains features, personality (e.g., wry humor, holds back then chimes in). Mimics human cold reading.
Practice smiling, people.
Retrospectives: early scored companies on grid, roughly correct but randomness (e.g., dead companies sell for billions). Current market: trillions chasing same prize distorts.
Forward-focused like Marc Andreessen; little past reflection. Regrets not more aggressive on winners (power law: handful matter).
Analysis: 100 companies drove 90+% returns 2000-2024; 10 drove 80%. Hit one of 10 or bad investor.
Focus on early decisions in investing. Discussion on moving from purely angel investing to involving other investors via SPVs (special purpose vehicles). SPVs invest in a single company, unlike VC funds that raise e.g. $100M and invest across many, taking 20% carry on upside and management fees, with LPs getting 80%.
SPVs offer simplicity but high reputational risk: if it goes to zero, it hurts reputation more than a fund with diversified losers.
Early SPVs included Instacart, early Stripe, Coinbase. Criteria: outsized companies with downside protection for decent outcomes even if not maximal success. Terrified of losing other people's money; fine losing own.
Advice to scouts (e.g., Sequoia program): treat as track record for professional investing career; be fiduciary, avoid treating as free money or lottery.
Pseudonymous example not named, but emphasis on diligence for track record. Early stage: spend more time on market than most (who focus only on team). Amazing teams crushed by bad markets; okay teams succeed in good markets. Index heavily on market via customer calls, intuition (e.g., defense important, nobody doing it).
Avoid science projects (quantum, hard tech) due to science/market/capitalization risks. SPACs at market peak took hard tech public, returned capital to funds, saved industry (Chamath Palihapitiya credited seriously).
Late stage: models show 2-3x returns due to IRR underwriting (25% IRR). Skill: distinguish 0.5x, 2-3x, or 10x. Subset where growth continues via market dynamic, core insight, market share—not math, but 1-2 key beliefs.
If it's three things, it's too complicated. It's probably not going to work. If it's no things, then it doesn't make much sense. Usually there's one or two things that are really the core insights.
- Coinbase: index on crypto growth; trades all main cryptos, takes transaction cut, effectively basket of cryptos.
- Stripe: index on e-commerce growth (now more complex).
- Anduril: machine vision/drones/AI for defense important; cost-plus model vs. hardware margins; checklist of 4-5 defense tech factors.
SPV stages: Stripe first at ~8 people, followed on until personal money exhausted, first SPV ~Series C. Instacart etc. similar C/D range.
Enormous diligence: meet CFO multiple times, financials/models, customers, exec team, cash reconciliations/audits (unique to his fund). But collapses to 1-2 core questions on continued growth.
Avoid wasting time on irrelevant details (e.g., secondary cohort on tiny product); respect founder/team time as ex-entrepreneur.
2011 post on VC questions to startups outdated; VC changed: used to seed-Series D/E then IPO; now 20-year private timelines. 4-year stock vest from 1970s (IPOs in 4 years); Google took 6 years (seen as slow then).
VC shifted: growth stage was formerly public markets; Sarbanes-Oxley, private capital extended timelines, VCs now do growth.
Growth questions financially driven: core business extrapolation + ancillary options. Most companies grow on one thing for first decade (e.g., Google search/ads; later Cloud/YouTube/Waymo; SpaceX launch to Starlink).
High Growth Handbook (2018, subtitle: Scaling Startups from 10 to 10,000 People): tactical guide, not cover-to-cover; jump to chapters (M&A, hiring etc.); interviews with top practitioners.
New book: zero-to-one tactical (hire first 5, handle acquisition, raise first round).
Chapter 2 on boards: take better board member over slightly higher valuation; write board member job spec (esp. independents).
Early boards often investors (who demand seats). Board member = co-founder you couldn't hire (Reid Hoffman); committed helper.
Boards as portfolio: strategic direction, recruiting, product, customers. Proactive composition, not reactive (investors + industry seat).
Analogy: co-founder = work spouse; board = in-laws (see often, long-term, up to decade; can't fire investors contractually).
Valuation is temporary. But control is forever. (Naval Ravikant)
Control matters: board can fire CEO. Choose helpful, non-destructive people; maybe worse price for right fit.
Resources: talk to founders who've added board members; write job spec (what optimizing for); use search firms, angels, get to know investors.
Example: Holler added Sue Wagner (BlackRock co-founder; boards: Apple, BlackRock, Swiss Re) after angel investment and positive collaboration.
2018 Y Combinator blog interview on High Growth Handbook: big companies (tens/hundreds billions market cap) took aggressive distribution (e.g., Google/Firefox, Facebook ads on names in Europe). Stories omitted from TED Talks.
Founder origin stories revisionist: 'ever since 3, dreamed of accounting software'; didn't collect pins at 5 for Pinterest; didn't dream AGI at 4 for OpenAI. Product really matters; sometimes great product just wins.
Great product opens up new distribution channels or people buy it despite lack of distribution or relationships. Top companies have enormously good products plus amazing distribution engines, sometimes built into the product like Cursor or Windsurf via product-led growth where developers discover, use, and spread via word of mouth. Often includes aggressive sales and marketing. At Google, spent hundreds of millions yearly distributing search via browser toolbar installed with other software like Adobe or malware detectors—paid companies to cross-download it. Browsers like Netscape and Internet Explorer lacked instant search bars then.
Facebook bought ads against people's names in Europe (common self-searches like "Tim Ferriss") landing on sign-up flow to create network liquidity in lagging markets. TikTok/ByteDance spent billions on ads for distribution to build network for AI algorithms and content creators. Enterprise example: Snowflake spent billions on salespeople, compensation, channel partnerships. Occasionally companies win purely on superior sales/marketing/distribution, not product—bummer for technologists who believe best product always wins. Sometimes brand or early distribution edge prevails.
Post-PayPal eBay sale, dogma was fraud kills payment companies. As investor, start with beliefs, invest or start company, but it may fail despite interest—then succeed years later. Why? Technology improved, regulatory change, market shift. Example: Harvey in legal—traditionally awful selling to law firms, but AI shifts from seats/software/SaaS to selling work product or units of labor (human labor equivalents, cognition, work/labor hours). Harvey augments lawyers via knowledge corpus and collaborative tooling, opening market.
Debate: YC view—need 10x more founders for 10x big companies (founder-limited). Alternate: world market-limited—only open markets allow big companies. AI opens tons of closed markets via capabilities and CEOs asking "What's my AI story?"—unprecedented openness. AI companies not seeing explosive growth have something broken; markets so open for unprecedented ramps like OpenAI and Anthropic to tens of billions (percentages of GDP).
90% agree market first, team second—but excellent team in terrible market fails. Signals: Google shutting down Maven creates startup opportunity. Frameworks: Why now?—what shifted? Regulatory (e.g., Samsara fleet management via in-cab driver monitoring regs for sleep prevention). Tech shifts like AI plugging into all enterprise data/email/white-collar work/code/language (consumer too). Robotics different—sub-markets with hardware small relatively, no instant runway unless new.
Incumbency shifts: Inphysical competes with Hashi (bought by IBM, slows down). Fast-growing markets like Coinbase/crypto token proliferation. Common: big enough TAM (total addressable market)—avoid fake TAMs (e.g., "facilitating $30T global e-commerce, 0.1% = $300B" when really SMB website optimization). Coca-Cola example: redefined from soda (50% share vs Pepsi) to share of liquid (0.5%), led to Dasani, other drinks—reconceptualize market to expand ambition.
Past doubts on ROI for capex spend payback—likely off. Now, time to be consensus (buy more AI), not contrarian (e.g., hardware). Overthinking complicates.
Mentoring smart MIT engineering grad with angel hits raising fund without top access: start early-stage by helping founders—smart people self-aggregate, follow in, build track record, raise bigger funds, go later-stage. Works for AI or anything; no need for energy investing.
From Vinod Khosla: market entry strategy differs from disruption. Examples:
- Toys evolving: Instagram, Twitter—initial use case differs from today, consumer-centric.
- SpaceX/Starlink: entered with launch, shifted to satellites/Starlink (major driver now) via cost advantage.
Collapsed to: X, technical papers/journals (AI papers dropped due to competition; biology too though not investing), talking to people (20min with smart person > exhaustive search), using models (OpenAI, Claude, Perplexity, Gemini) for research—different uses per model. Gemini strong for travel tips/activities via Google corpus (rankings, breakdowns, multi-dimension scoring).
Deep dive ADHD/ASD (autism spectrum): autism diagnoses from 1 in thousands (30-40yrs ago) to 3% now—not older parents (minor driver), environment, but diagnostic criteria shift + school/doctor incentives (drugs). Used 2-3 models + primary literature/summary charts; prompts for output to double-check. Findings: maternal age impact > paternal (odd emphasis on paternal—societal/political?). For friend (mid-late 30s woman dating late 40s/early 50s man) worried re kids: ~10% increase per 5-10yrs incremental parental age (maternal stronger), but population frequency low so minor; main driver diagnostics. NJ schools: 60% autism diagnoses not clinical, just teacher say-so. Models aggregate trials/data, provide sources/summaries—prompts clean/check data.
Three types:
- Deep dive (interesting/investment-relevant): triangulate smartest via papers, asks—who's best?
- Smart people aggregate; polymaths with polymaths.
- Referrals or recurrent experts (e.g., Noam Shazeer via early Google AI; Kristen CEO BioAge for longevity—truth-seeking PhD/postdoc bioinformatics/aging). From Quantified Self Meetup 2008 (12 people, Kevin Kelly's house, Excel tracking) to armies of tens of thousands biohackers talking longevity.
Elad Gil's approach collapses into basics: sleep well, exercise a lot, eat well. He's considering a rapamycin pulse but waiting for real drugs targeting longevity or specific systems. Examples include eye drops for aging-related lens muscle weakening affecting focus, neurosensory aging, cosmetics of aging. He funded a clinical trial at Stanford for cosmetics of aging. Peptides are viewed as health and cosmetic applications like 5-HKCU (likely 5-amino-1MQ) and melatonin.
Short list of helpful supplements: vitamin D, creatine.
Tim is conservative, experimented earlier with Dexcom continuous glucose monitors (2008-2009, first-generation, unpleasant, rare for non-type 1 diabetics). Avoided questionable gene therapy like folistatin abroad. Heuristic: no biological free lunch to avoid pitfalls.
Current experiments:
- Different forms of ketone esters and salts for cerebral vasculature, given family history of Alzheimer's, Parkinson's, ApoE3-3 risk.
- Obacetrapib (watch, not ready).
- Rapamycin (interesting with asterisks; immunosuppressant, be careful).
Potential experiment: Norwegian 4x4 interval training + rapamycin pulsing to measure volumetric changes in hippocampus and other areas.
Basics: creatine, vitamin D. Monitor if on omeprazole (inhibits magnesium absorption). Urolithin A interesting, data mounting for mitochondrial health. Intermittent fasting, occasional 3-7 day fasting or fast mimicking diet for autophagy and mitophagy (input from Dr. Dominic D'Agostino).
Analogy: rebooting a computer/laptop fixes issues. Equivalents? Stellate ganglion block. Ibogaine for opiate addicts (flood dosing under medical supervision, high doses, fatal cardiac risk mitigated by magnesium; creates window without physical withdrawal). Historical credit: Howard Lotsoff and wife.
Ibogaine shows brain age reversal on MRIs (pre/post for veterans with traumatic brain injury; Nolan Williams lab; possibly via glial derived neurotrophic factor or BDNF). Potential for other applications.
Cautious on general anesthesia (opted for local in recent head surgery; risks memory recall, personality destabilization; poorly understood mechanisms; overused for billing). Many common meds have poorly understood mechanisms despite tolerance data.
GLP-1 agonists can systemically improve impulse control (e.g., stop smoking, reduce drinking).
Most interesting reboot: Ibogaine (beyond dopaminergic; shows possibilities). Bullish on brain stimulation and bioelectric medicine for psychiatric disorders and performance enhancement (non-invasive or implants; outpatient procedures; strong why now factors).
Spends time with neuroscientists, reads neuroscience/optical material (AI tools aid learning). Photobiomodulation (PBM) via eyes or transcranially (caution).
Many beliefs will be wrong amid big change/uncertainty; half of predictions may fail or exceed; fun due to unpredictability (perfectly predictive future boring; ties to free will).
New exercise: 10-year plan across life dimensions (despite life's unpredictability). Shifts ambition scope, near-term actions. Rejects AGI defeatism (e.g., AGI in 2 years means no plans); instead, plan and adjust amid change for interesting opportunities.
Elad Gil: X @EladGil, eladgil.com, Substack blog.eladgil.com.
Be a bit kinder than is necessary to others, but also to yourself.
Keep Tim Ferriss in your library
Save the videos and channels worth coming back to, and find them again in one place.
