Spotting billion-dollar AI companies early: Elad Gil on investing and the frontier

Executive overview

Most companies built during a technology wave go bust — AI is no exception. A small handful will define the next decade; the rest should consider exiting in the next 12–18 months while valuations are high.

Elad Gil argues the compute memory constraint (~2 years) is keeping all major labs roughly equal for now, but when it lifts, a single lab could pull far ahead. At the application layer, durability comes down to workflow embeddedness, proprietary data, and whether a better underlying model makes your product better or obsolete.

The core insight: being consensus — just buy more AI — is the right call right now. Most contrarian framings are overthinking a once-in-a-generation moment.

The AI compute constraint

• Memory (HBM from Korean chipmakers) is the current bottleneck — expected to last ~2 years
• Every lab is supply-constrained; no single player can buy 10x the compute of rivals
• Constraint enforces rough capability parity across OpenAI, Anthropic, Google, xAI for now
• When constraint lifts, one lab may pull dramatically ahead via scale law advantages
• Previous bottleneck was packaging; next may be power and data center construction
• Demand keeps outpacing forecasts — no fracking-style workaround has emerged

The personal IPO phenomenon

• Meta's aggressive bidding forced all major tech companies to match pay for top AI researchers
• 50–several hundred researchers experienced a simultaneous wealth event — tens to hundreds of millions per person
• Analogous to the 2017 crypto wave, but spread across Silicon Valley rather than one company
• Likely to trigger a subset shifting to passion projects, big-science bets, or disengagement

Market structure: oligopoly, not monopoly

• Core labs (OpenAI, Anthropic, Google, xAI, Meta) form a near-term oligopoly
• No single lab has pulled far enough ahead on capability to become the default for everyone
• Compute constraint acts as a ceiling that preserves competitive parity in the short run
• Application-layer companies face a different question: are you one of the ~dozen that survives?

What makes an AI application durable

• Does your product get materially better as the underlying model improves?
• How deeply are you embedded in customer workflows — change management is the real barrier, not technology
• Are you building multiple integrated products that are hard to pull out?
• Proprietary data as a moat is often overstated, but useful in system-of-record contexts
• Harvey (legal), Anduril (defense), Decagon/Sierra (customer success) cited as examples with structural depth

The exit window for AI founders

• History: ~1,500–2,000 internet companies went public in 1999–2001; ~1–2 dozen survived
• Every tech cycle — SaaS, mobile, crypto — follows the same 90–99% failure rate
• Many current AI companies face predictable headwinds: lab commoditisation, model capability leapfrog, market shifts
• Founders should watch the second derivative of growth — plateau signals the peak window
• Mega-cap buyers (Apple, Amazon, Google, Oracle, Stripe, Coinbase, Thomson Reuters) have unprecedented buying power — 1% of a $3T market cap is $30B
• Merger of private competitors is underused; x.com + PayPal cited as the archetype

How Elad Gil invests: market first

• Market beats team at most stages — great teams get crushed by terrible markets
• Framework: identify one core belief that, if true, makes the company very large
  • Coinbase: index on crypto volume growth
  • Stripe: index on e-commerce growth
  • Anduril: AI + drones will be critical for defense
• Avoid science projects; skipped most hard-tech SPACs (capitalization and science risk)
• For growth-stage: one or two questions collapse complex models — is the core durable?
• Does full diligence (CFO meetings, cash reconciliation, customer calls) but compresses it to the key questions

Spotting great markets

• Ask: why now? What has shifted — regulation, technology, competitive landscape?
• AI shifted from ML Ops pipelines to a generalised API anyone can call in a few lines of code — that unlocked every white-collar market simultaneously
• Regulatory shift example: fleet management cameras mandated by law → software wedge
• Competitive shift: Hashi acquired by IBM → Inphysical gains startup runway
• Reframe the TAM: Coca-Cola went from 50% soda share to 0.5% of all liquids — changed scope of ambition

Distribution and the untold founder stories

• Google toolbar: paid every site on the internet to bundle the client app
• Facebook: bought ads on people's own names in Europe to seed network liquidity
• ByteDance spent billions on paid distribution to build TikTok's content graph
• Snowflake: billions in enterprise salesforce and channel partnerships
• Great product opens distribution channels; great distribution sometimes wins even without the best product

Getting into deals: geography and network

• 91% of global private AI market cap is in the Bay Area — location is the single biggest edge for early-stage access
• Defense tech clusters around Southern California (SpaceX, Anduril, El Segundo)
• Early deals (Airbnb, Stripe, Coinbase) came from helping founders operationally first — investment followed naturally
• Perplexity: founder cold-messaged on LinkedIn because Elad was publicly discussing AI when almost no one else was
• Anduril: spotted the gap after Google shut down Maven; found Trey Stevens at lunch

SPVs, track records, and fiduciary mindset

• First SPVs focused on companies with massive upside AND meaningful downside protection
• Scouts who treat allocated capital as "free money" are building a bad track record for themselves
• Power law is extreme: ~10 companies drove ~80% of all technology returns over two decades — did you own one?
• Regret skews toward not investing more in the winners, not toward bad bets

Boards

• A board member at its best is a co-founder you couldn't hire
• Founders are reactive about boards; they should write a job spec as they would for any hire
• Investors who hold board seats can't be removed — choose the person, not the valuation
• Valuation is temporary; control is permanent
• Build a portfolio of board members with complementary coverage (strategy, customers, product, finance)

AI dogma worth questioning

• ROI timelines on AI capex are probably being underestimated
• The consensus view — keep buying AI — is correct; elaborate contrarian theses often miss the obvious
• Harvey disproved the dogma that selling to law firms is a bad business by shifting from software seats to units of cognitive labor

Information diet and use of models

• Primary inputs: X (Twitter), technical papers, 20-minute calls with domain experts
• Now uses multiple models in parallel for research (Gemini for travel/rankings, Claude/OpenAI/Perplexity for deep dives)
• Runs structured prompts: ask for primary literature, summary charts, and checks output across models
• Polymathic people aggregate together — spending time with smart people yields more referrals to other smart people

Longevity and health (brief)

• Most interventions collapse to: sleep, exercise, diet
• Conservative on supplementation: vitamin D, creatine, magnesium monitoring (especially if on PPIs)
• Rapamycin is interesting with significant caveats — immunosuppressant risk is real
• Skeptical of anesthesia overuse; cautious about poorly understood mechanisms
• Bullish long-term on non-invasive brain stimulation and bioelectric medicine as the next frontier