Most important take away

Tom Hulme argues that foundation models are commoditizing so fast they should be thought of like power stations — capital-intensive assets that depreciate over months, not years — and that the durable value in AI will accrue to cloud providers and incumbents with proprietary data and distribution, not to standalone model companies. For founders and investors, the practical lesson is to prioritize clock speed, fundamentals, and option value over heat, momentum, and structured rounds that paper over reality.

Summary

Actionable insights for investors:

• Pick a strategy and stick to it. Don’t abandon a portfolio approach because one company looks exciting — dumping outsized capital into a single name breaks the model that drove your earlier returns.
• As an angel, write small, consistent checks (e.g., $5K) rather than varying check size by conviction. Conviction-weighted sizing is the biggest mistake angels make; you cannot stack-rank your own winners ahead of time.
• Don’t follow on as an angel. Pro-rata in priced rounds pits you against VCs in a different game; Hulme’s data shows follow-on capital would have hurt his returns.
• Stop outsourcing diligence to other angels or to the presence of a “name” lead. A $5M check from a $1B fund is coffee money and creates dangerous social validation.
• Use regret minimization for secondaries: founders and early backers should take 10–20% off the table when offered. Modern SPV infrastructure (AngelList, etc.) now makes this far easier than a decade ago.
• Watch for structured rounds (convertibles to avoid pricing, >1x liquidation prefs, IPO ratchets). VCs use them to protect TVPI; founders should evaluate the whole package, not just headline price (“tell me the price, I’ll tell you the structure”).
• Run pre-mortems. Borrowed from Kahneman: explicitly list how the decision could be wrong and inoculate against it. Beware “complete conviction” — it doesn’t exist.
• The four S’s of VC: Sourcing, Selecting, Supporting, and Selling (to LPs, founders, and exec hires). Fight to stay sharp on sourcing as you mature.
• “Serendipity favors the connected.” Network value scales with Metcalfe’s law; deep networks are the real engine behind apparent luck.

Career and founder advice:

• Bullying and adversity taught Hulme empathy and a chip on his shoulder — he won’t ask anyone to do what he wouldn’t do himself. Direct, honest feedback is the most empathetic act, even when uncomfortable.
• The best founders pay attention to feedback and the market. If feedback isn’t landing, consider whether it’s your feedback that’s off.
• Great founders run the scientific method: a startup is a series of unanswered questions, and the best founders pick the right order and answer them efficiently. Perfection is the enemy of progress.
• V1 should be the minimum needed to get meaningful feedback. Don’t ship signup pages with no product description and call it product-market fit.
• “Free kills feedback.” Charge design partners and ICPs — you can’t tell if something is valued if no one pays.
• Velocity > speed (speed in a given direction). Aim activity at a specific question; running without direction is wasted motion.
• Best founder questions to ask: How did you first make money? What’s your unfair advantage? Why now? What keeps you up at night? Founders with no paranoia are a red flag — “only the paranoid survive.”
• Ideas are cheap; execution is everything. Market timing risk is real — being too early is tantamount to being wrong unless you have cockroach-mode survivability (UiPath at $500K ARR for 9 years is the model).
• Premature scaling is one of the biggest ways VCs damage companies. Excess cash raises costs, lowers clock speed, and makes the team less adaptable when iterating toward PMF.
• Naive vs. insider founders both work, but naive founders must have humility and recruit specialists. Without that, they have to learn everything from scratch.
• Watch cultural debt as carefully as technical debt — it’s more insidious, harder to reverse, and got worse through COVID-era remote bloat. The best founders make deep, not shallow, cuts and use burning-platform moments to reset culture.
• Junior team members in particular need in-person time to learn, shadow, and absorb informal feedback. Hybrid often delivers the worst of both worlds when teams don’t coordinate their in-office days.
• Nature dominates nurture more than people realize — you can help founders be better, but you can’t fundamentally change them.

Tech patterns and theses:

• Foundation models = power stations. Training is the build-out, inference is the power, and everyone is building next door with the same H100s. Meta’s open-sourced Llama with 350K H100s (~14% of global supply) and $100B+ planned spend makes the asset depreciation cycle brutally short.
• End state: cloud providers (AWS, GCP, Azure) become the cash cows; they will acqui-hire foundation model teams and give models away as utilities. The Microsoft–Inflection deal was effectively a GPU cluster purchase, not a model purchase.
• Gen AI is a sustaining innovation (à la Clay Christensen), not a disruptive one — it gets sprinkled across incumbents to cut costs and improve products. Microsoft is exceptionally well-positioned because data + distribution sit inside Word, GitHub, and Epic-adjacent workflows.
• Hulme’s rough capital-loss estimates: foundation models 90% to zero, application layer 70% to zero, ~20% of value flows to incumbents.
• In the application layer, look for proprietary data, distribution, or end-to-end enterprise readiness (Synthesia as an example). Use Sam Altman’s test: would the company be happy or devastated if models improved 100x? Invest in the “happy” ones.
• Memory and agency are the things foundation models haven’t cracked — a model with persistent personal memory and probabilistic anticipation would be genuinely differentiated.
• Be skeptical of AGI timelines that correlate with fundraising needs. Founders who don’t need to raise (Zuckerberg, Hassabis) are markedly more cautious than those who do (Altman, Amodei, Musk).
• Robotics is becoming generalizable for the first time thanks to vision + LLMs + multimodal + cheaper components — a category Hulme previously considered uninvestable.
• Defense/military tech matters more over the next 20 years; structural barriers (procurement, in-house skill requirements) favor winner-take-all dynamics (Anduril, Helsing).

Chapter Summaries

• Childhood and empathy: Bullying gave Hulme grit and empathy; honest direct feedback is the most empathetic gift.
• Three types of investors: smart-smart, passive-passive, and dangerously dumb-but-thinks-they’re-smart. Avoid the third; both of the first two are fine.
• Angel investing experiment: 27 companies, ~4.5x DPI, ~24x TVPI pre-2015 — but he couldn’t have stack-ranked his own winners, which undermines reserves strategies and self-conviction.
• Structured rounds and TVPI games: Convertibles, >1x prefs, and IPO ratchets are creeping back because VCs are incentivized to protect marks. Founders should weight structure, not just price.
• Damaging investor behavior: The biggest VC damage isn’t interference, it’s stuffing companies with too much cash and triggering premature scaling.
• Winners vs. zeros: Winners came from fundamentals founders who compounded over a decade (GoCardless); zeros came from hot, momentum-driven seeds with no liquidity escape (Fab, Massive Health).
• Angel mistakes: Outsourced diligence, social validation traps, and not spending enough time understanding how founders actually think.
• Transition to VC — the four S’s: Sourcing, Selecting, Supporting, Selling. Selling to LPs, founders, and exec hires is underrated.
• Networks and luck: “Serendipity favors the connected” — Metcalfe’s law applied to careers.
• Outcome scenario planning: Use pre-mortems, option-value framing, and acknowledge there’s no such thing as complete conviction.
• Fund returners debate: Hulme pushes back on the dogma — there are other valid strategies (PE-like, debt-like) at growth stage. Pick a lane and don’t betray it.
• Stripe Series G at COVID lows: GV invested $100M extending the $32B round when others froze with “fear of looking stupid.”
• Foundation model thesis: Models = power stations; commoditizing fast; value flows to cloud providers; consumer revenue isn’t sticky; memory and agency would be true differentiators.
• Application layer: Demand proprietary data/distribution and the “happy if model is 100x better” test.
• AI and incumbents: Microsoft’s distribution moat through Word/GitHub/Epic-style integration is formidable; expect ~20% of AI value to flow to incumbents.
• Building companies: Clock speed wins; V1 just needs to generate good feedback; charge customers; velocity beats raw speed.
• Culture debt and remote work: Culture debt is harder to reverse than tech debt; ambitious early-career people should be in person to learn from osmosis.
• Quick-fire: Climate change needs AI’s PR team; AGI timelines correlate with fundraising needs; Neuralink pitch was unforgettable; Hulme is a super-bull on Elon; biggest mind-change is robotics; nature beats nurture; GV’s biggest wins include 277x on Robinhood seed and billions on Uber.