The Four Questions

For every wave, I asked the same four questions:

• What started it — the specific technical unlock, and where it came from.

• Which companies got created because of it.

• Which of those companies vanished — and why.

• Which didn’t — the ones still standing, and what they did differently.

A note before diving in: this is necessarily retrospective. Every cycle produced far more failures than the famous names suggest. These patterns don’t guarantee an outcome — they show up, repeatedly, among the companies that ended up mattering.

Cycle Zero: The Dream (1950–1956)

What started it. In 1950, Alan Turing asked whether a machine could think, proposing a test: if a hidden machine fools a human in conversation, it’s earned the label “intelligent.” Six years later, a small group at Dartmouth College — John McCarthy, Marvin Minsky, Claude Shannon, Nathaniel Rochester — coined the term “artificial intelligence” for a proposal claiming intelligence “can in principle be so precisely described that a machine can be made to simulate it.” The bet: two months, the right people in a room, real progress. It took about seventy years, and counting.

What it means. The technical ceiling was total: nothing happened that a programmer hadn’t explicitly hand-coded in advance. The idea that a machine could adjust its own behavior from raw experience — the foundation every cycle since has been built on — didn’t exist yet as a working technique. What was new in 1956 wasn’t capability. It was ambition.

• Could: play simple games (Arthur Samuel’s checkers program, IBM, 1950s — one of the first machines that improved by playing itself); prove basic math theorems.

• Couldn’t: see, hear, hold a conversation, or do anything beyond what a programmer had explicitly hand-written in advance.

The companies: there weren’t any. Every dollar came from the government — ARPA (later DARPA) funded MIT, Stanford, and Carnegie Mellon directly, betting on basic research with no product in sight. A pattern repeats later in this story at every scale: patient, product-agnostic money arrives first — government grants here, corporate labs in the 1990s, today’s frontier labs burning investor cash pre-revenue. There was no bust in Cycle Zero simply because there was nothing yet built to bust.

Cycle One: The Age of Rules — Two Booms, Two Busts (1960s–1993)

What started the wave. The first real AI boom was symbolic AI: teach a computer using rules, like handing a new employee a rulebook — “if the patient has a fever and a rash, consider measles.” Stack enough if-then rules into an “inference engine” and you got an expert system: software that mimicked one human expert’s judgment in one narrow field, well enough that the expert’s employer would pay for it. Not “a machine that thinks,” but “a machine that replicates this specific senior employee’s judgment, at a fraction of the cost.”

This wasn’t AI’s first attempt. The broader rule-based promise had already collapsed once, a decade earlier:

• The ALPAC report (1966): after years of funding, machine translation was slower and less accurate than a human translator — the panel recommended the US stop funding it.

• Perceptrons (1969): Minsky and Papert mathematically proved single-layer neural networks couldn’t solve basic problems, like telling a line drawing was one shape or two — freezing neural-network research for over a decade.

• The Lighthill Report (1973): a UK review, by a mathematician outside the field, found AI’s general-purpose promises undelivered. The UK, then the US, slashed funding within two years; researchers rebranded their work as “informatics” just to keep raising money.

It took roughly a decade to rebuild confidence — not around general intelligence, but around the narrower, provable promise above. That’s what turned into the billion-dollar boom below.

Two pools of money were doing two different jobs here, and they’re easy to conflate.

• Research stayed government-funded: DARPA kept funding MIT, Stanford, and CMU through the 1980s, and was often these companies’ single largest customer.

• Equity was investor money for the first time in AI’s history: Symbolics pursued VC aggressively from 1980; Thinking Machines raised ~$16M from private backers including CBS’s William Paley; Teknowledge and IntelliCorp raised VC, and Teknowledge later IPO’d.

• Why it matters: government money funded the customer base, investor money funded the company — exactly why defense budget cuts later hit this industry so hard.

What it means. Building one of these systems meant sitting with a human expert for months, translating their judgment into an explicit decision tree: if symptom X and test result Y, conclude Z, unless W is also true. It genuinely wasn’t faking it — the logic inside was that expert’s knowledge, written down instead of carried around in one person’s head.

• Could, for the first time: diagnose a blood infection at a level competitive with practicing physicians (MYCIN, Stanford); identify unknown chemical compounds from raw lab data (DENDRAL, also Stanford); configure a complex computer order without errors (XCON, saving Digital Equipment Corp an estimated $40M a year).

• Couldn’t: handle anything its programmers hadn’t explicitly anticipated, learn anything new once deployed, or apply what it “knew” to any other domain.

The companies.

In plain terms: the first row isn’t companies at all — MYCIN, DENDRAL, and XCON were the research systems and internal deployments that proved rule-based AI could deliver real value, before any business existed to sell it. Specialized AI hardware (“Lisp machines”) meant computers built to run one programming language, Lisp, as fast as possible — the way graphics work later needed its own GPU. Knowledge-engineering consultancies meant firms that interviewed a human expert for months, then hand-translated that judgment into thousands of formal rules — a service business built around one very slow, manual step.

Sources: Computer History Museum oral histories; Inc. Magazine’s 1995 account of Thinking Machines’ financing; Wikipedia and contemporaneous coverage of Symbolics, Lisp Machines Inc., and Thinking Machines Corp.’s bankruptcy filings; SEC filings and Finextra on Teknowledge’s 2005 sale to Intuit; Wikipedia on IntelliCorp’s 2019 wind-down; Stanford’s historical accounts of MYCIN and DENDRAL; CMU and DEC retrospectives on XCON.

Almost every company above died within a decade — and their real, measured results didn’t save them. By 1985, corporations were spending over a billion dollars a year on this technology, and it wasn’t hype: these systems delivered exactly what the “Could” list above promised, in production — Symbolics alone controlled roughly 70% of a real, growing hardware market. Two mechanisms killed almost all of it anyway, one per archetype.

Why the Lisp machine makers died:

• Their value proposition got commoditized. Symbolics’ machines were the only practical way to do serious AI research — until Intel’s 386 chip (1986), mass-produced for ordinary business computers, got cheaper and faster far quicker than any niche AI chip could. Once a $3,000 ordinary computer matched a $150,000 Lisp machine, the specialized box had no reason to exist.

• Their core customer base evaporated. Thinking Machines held on longer than most — until the Strategic Defense Initiative (“Star Wars”) scaled back its AI funding in the late 1980s and took Thinking Machines’ largest customer with it.

• Execution made a bad hand worse. Symbolics compounded a shrinking market with leadership chaos and expensive leases. Lisp Machines Inc. never had Symbolics’ capital to begin with, and folded first, in 1987.

Why the knowledge-engineering consultancies died:

• Brittleness. A system only followed the exact rules it was given; an unanticipated situation got no answer, not a wrong one.

• The knowledge-acquisition bottleneck. Turning an expert’s judgment into formal rules was inherently slow, and didn’t get faster with practice.

• The maintenance nightmare. One rule change could unpredictably interact with thousands already installed, so updates took nearly as long as building the system.

• The economics stopped working. Multiply that across dozens of clients, and lifetime cost caught up with the sales pitch — clients did the math and walked. Japan bet over $1.3 billion on the identical idea at national scale (its Fifth Generation Computer project) and got the same result.

Almost nobody survived independently, and almost nobody got bought out either — there was no Google or Microsoft yet rich enough to acquire the talent. Failure simply meant failure: shareholders lost essentially everything. What outlasted the companies were the ideas: XCON’s rule-engine logic still lives inside the “business rules engines” banks use today, and the name “Thinking Machines” was reborn, independently, as Mira Murati’s AI lab in 2025.

The Quiet Years: Machines That Learned From Data (1990s–2000s)

What started it. Out of the wreckage of rules-based AI, a humbler idea took root: instead of telling the computer the rules, show it thousands of examples and let it find the pattern. This is machine learning, and it emerged partly because the previous approach had just publicly failed — nobody wanted to fund “artificial intelligence” anymore, so the people still working on it stopped calling it that and simply shipped statistics inside ordinary products.

What it means. Instead of a programmer hand-writing “if this, then that,” the software was shown labeled examples — this email is spam, this one isn’t — and worked out its own statistical rules for telling them apart, rules no human wrote down or could fully explain. That’s still the core idea inside every AI system built since.

• Could: filter spam from millions of labelled examples; rank search results by which pages people actually clicked, rather than any human-authored rule; flag fraud from patterns in real time.

• Couldn’t: pick its own signals — a human still had to hand-choose what data to feed it. No general intelligence, just very good narrow statistics, hidden inside a bigger product.

The companies — a few well-documented examples, not a full list. Nobody called themselves an “AI company” in this era, a direct consequence of Cycle One’s collapse. The three below are the best-documented cases, not exhaustive — most large tech companies of the period ran similar experiments without a business model attached: Microsoft’s Bayesian spam-filtering research shipped inside the Office Assistant (“Clippy,” 1997); Apple built speech recognition (PlainTalk, 1993) and handwriting recognition into the Newton. None of it was marketed as “AI,” and none made money directly.

Google — no revenue model (1998–2000) No ad business at all when Kleiner Perkins and Sequoia wrote a $25M check in June 1999 — AdWords didn’t launch until October 2000. The bet: PageRank was measurably better than Yahoo’s or AltaVista’s, and the dot-com thesis was capture the audience first, monetize later — the same logic later used on DeepMind, and on Safe Superintelligence today.

PayPal — “Igor” (~2000) Fraudsters were draining accounts faster than PayPal could process transactions profitably. Max Levchin’s team built a statistical model — nicknamed after a taunting Russian fraudster — that cut fraud below 0.5% by 2001. Arguably saved the company.

No bust this time — the one clean exception in this story. The technology kept compounding quietly until it re-emerged as its own category in Cycle Two

Cycle Two: The Deep Learning Boom (2012–2017)

What started the wave. Geoffrey Hinton had explored neural networks since his 1970s PhD, through two winters almost nobody else took seriously. In 2012, he and two students, Alex Krizhevsky and Ilya Sutskever, entered the ImageNet competition with AlexNet — a deep neural network trained on GPUs, the video-game chips that turned out to be shockingly good at this math. AlexNet nearly halved the previous best error rate in a single year, proving more data, more compute, and the right architecture beat hand-engineered approaches. It was less a new idea than an old one (neural networks, decades old) finally meeting enough data and cheap hardware. Hinton joined Google the following year; he resigned in 2023 to speak freely about the risks of the field he’d just re-legitimized.

What it means. Earlier systems needed a human to hand-pick which signals mattered. A deep neural network, given enough labeled examples and compute, worked out its own signals — edges, then shapes, then whole objects, stacked in layers, with nobody defining them in advance. That’s what “deep” means: not smarter in some abstract sense, just many more stacked layers of self-taught pattern recognition than anything tried before.

• Could, for the first time: label a photo’s contents at near-human accuracy on unseen images; transcribe speech reliably enough for consumer products; and, via DeepMind’s AlphaGo (2016), beat world Go champion Lee Sedol by playing millions of practice games against itself.

• Couldn’t: understand language at all — a model that tagged a photo of a dog had zero idea what “dog” meant in a sentence. Every model was single-purpose, with no transfer between tasks.

The companies. Four distinct business models emerged, each with its own funding logic:

Put simply: frontier research labs meant a team raising money purely on scientific reputation — the closest thing here to Cycle Zero’s government-funded research, just paid for by investors. AI infrastructure for other builders meant picks-and-shovels: chips or tools bought by other AI companies, not ordinary businesses. Applied AI, enterprise SaaS meant AI wrapped into a plug-and-play tool a bank or retailer could buy and use immediately. Data infrastructure for training meant the unglamorous work every model depended on: humans, at scale, labeling raw data.

Sources: Google/DNNresearch and Google/DeepMind deal reporting; DeepMind’s ethics-board terms, AI Principles, and 2023 Google Brain merger; Intel’s Nervana and Habana Labs announcements; Uber AI Labs coverage; Salesforce’s account of the MetaMind acquisition; Richard Socher’s 2020 departure and You.com; Kensho acquisition reporting; Clarifai/Nebius coverage, May 2026; Uber’s acquisition of Mighty AI; Meta’s Scale AI investment.

What worked. The failure mode changes here: in Cycle One, failure meant bankruptcy. Here, even an ending often looked like a win.

DNNresearch → Google, $44M (2013) A bundled purchase of the AlexNet-winning techniques and the people who built them. It paid off almost immediately — the technique shipped inside Google+ Photo Search within about six months, even after the team dispersed (Sutskever to OpenAI in 2015, Krizhevsky around 2017).

DeepMind → Google, $500M–$650M (2014) No product, no revenue, raised on research reputation alone — the same “fund the team, work out the model later” logic that funded Google itself in 1999. It paid off: DeepMind’s algorithms cut Google’s data-center cooling energy ~40% within a year, then came AlphaGo (2016) and AlphaFold, whose protein-structure predictions won a 2024 Nobel Prize. By 2023 it had absorbed Google Brain; today it’s the org building Gemini — commonly called one of the best tech acquisitions ever made.

MetaMind → Salesforce (2016) Not a failure. Richard Socher became Salesforce’s Chief Scientist; MetaMind’s tech became the foundation of Salesforce Einstein. He left in 2020 to found You.com, now valued around $1.5B.

Kensho → S&P Global, $550M (2018) A natural-language tool for Wall Street analysts, backed early by Goldman Sachs and later by its own future acquirer’s peers, JPMorgan and Morgan Stanley. The largest AI acquisition of its time.

Clarifai — independent for over a decade, then absorbed differently Never acquired outright; raised a modest ~$101M and evolved past pure image recognition. In May 2026, its team and technology moved to Nebius via a team-hire-and-license deal — a reminder that “still independent” can be temporary even after ten-plus years.

What didn’t.

Nervana → Intel, $408M (2016) → quietly killed (2019–2020) An extraordinary price for a 48-person startup. Intel then paid ~$2B for rival Habana Labs, whose chip was already shipping and reportedly outperforming Nervana’s — which hadn’t shipped yet. Intel discontinued Nervana within weeks. A big price tag doesn’t guarantee the technology survives.

Geometric Intelligence → Uber (Dec 2016) Bought to seed “Uber AI Labs.” Founder Gary Marcus left after roughly three months; the lab published respected research but was wound down in 2020 amid pandemic cost-cutting, the same year Uber sold its self-driving unit. The team, not the founding thesis, is what Uber kept.

The buyers were consistently the same handful of companies with the balance sheets for it — Google (twice), Intel, Uber, Salesforce, S&P Global. The same instinct behind the $44M DNNresearch deal drove Meta to pay $14.3B for 49% of Scale AI in 2025 — roughly 1,000x bigger.

Cycle Three: The Language Boom — Transformers to ChatGPT (2017–2023)

What started the wave. In 2017, Google researchers published “Attention Is All You Need,” introducing the Transformer: an architecture that reads a whole sentence at once and learns which words matter most to which others, instead of plodding word by word. That made training on basically the whole internet’s text practical for the first time. Every major model since — GPT, Claude, Gemini, Llama — descends from this idea. Five years later, on November 30, 2022, OpenAI released ChatGPT as what was internally a “low-key research preview.” It had a million users within five days.

What it means. The leap is generalization. Cycle Two’s systems each needed a separate model trained from scratch for one task, with no transfer between them. A large language model, trained once, could be redirected to a brand-new task just by describing it in plain English — no retraining, no new code. That specific thing had never been true of any AI system before.

For the first time, a single model could:

• Draft an email, summarize a document, translate a paragraph on request.

• Write and debug code in dozens of languages; answer open-ended customer questions correctly most of the time.

• Pass professional exams (GPT-4 scored in the top 10% of bar-exam takers) and hold a coherent multi-turn conversation.

Couldn’t reliably: chain more than a couple of logical steps together; remember anything outside the current conversation; use outside tools or take real-world action on its own; or avoid confidently inventing plausible-sounding facts — “hallucination.”

The companies. Cycle Three splits into eight business models — the largest and most varied of any cycle, with nearly $90 billion invested in AI startups in 2023 alone:

Two names need a plain-language definition. Open-weight foundation model means Stability AI published the actual model file for anyone to download and run, unlike OpenAI or Anthropic’s closed, API-only models — the same openness that made it widely copied, and hard to keep monetizing. Autonomous computer-use agents meant software operating an ordinary screen the way a person would, years before that became standard in Cycle Four.

Sources: Microsoft/Inflection deal reporting (March 2024); Google/Character.AI reporting (August 2024); Amazon/Adept reporting via Semafor and CNBC; Stability AI board/CEO coverage; Midjourney and ElevenLabs revenue estimates; Runway’s 2023–2026 funding rounds; Perplexity AI’s 2024–2025 valuation history.

Too early to call a verdict on the foundation labs — OpenAI and Anthropic are both still burning enormous capital, and neither has settled into a stable, profitable business. But a few application-layer companies have now generated real, growing revenue independently for three-plus years — long enough to say why they weren’t swallowed the way Inflection or Character.AI were.

Midjourney — bootstrapped, ~$500M ARR by 2025 Founder David Holz swore off outside capital after watching VC pressure wreck his last company. Launched in Discord, July 2022, charging $10/month from day one — no free tier. 15M users and ~$200M revenue by 2023, on a 40-person team.

ElevenLabs — ~$330M ARR by end of 2025 Founders Mati Staniszewski and Piotr Dąbkowski had no AI-research pedigree — just the best voice-cloning product in the market, priced from early on.

Why they scaled so fast: both treated the product as the whole business from day one — real pricing, real revenue, minimal burn. Why the frontier labs haven’t disrupted them: image-generation taste and voice quality are a feature to a company like OpenAI or Google, not the thing they obsess over — while for these two it’s the entire product. Neither ever had a free tier for a bigger player to undercut. Runway and Perplexity, both richly valued heading into 2026 (~$5.3B and ~$20B), look like they may be following a similar script — but younger and less proven on sustained profitability.

What didn’t work — and the failures split cleanly into a few reasons.

Free product, no way to charge for it (the companion-app trap):

Inflection AI → Microsoft, ~$650M (March 2024) Mustafa Suleyman (DeepMind co-founder) built “Pi” — free, genuinely loved, with no enterprise budget to extract. Every conversation burned compute with no revenue back. Microsoft licensed the tech and hired the founders into what’s now Copilot.

Character.AI → Google, $2.7B (August 2024) Users spent 75–120 min/day in the app — more than Netflix, YouTube, or TikTok — on a free tier, plus lawsuits over teen dependency. Google licensed the tech and brought founders Noam Shazeer and Daniel De Freitas back to DeepMind. Same trap as Inflection, sharper.

The lesson: engagement that would thrill investors elsewhere becomes a liability the moment every session burns real money.

Thin layer, no protection once the model improved (the wrapper graveyard): Copy.ai, Writesonic, Rytr, Anyword, and dozens like them grew fast — until GPT-4 (March 2023) and free ChatGPT did the job better, for nothing. The same lesson as Teknowledge and IntelliCorp in 1987, compressed from years to months.

Real product, spending discipline broke first (money, not technology):

Stability AI — CEO resigns, March 2024 Giving away Stable Diffusion free while burning $8M/month against $5.4M/month revenue isn’t a business. Lightspeed’s board letter accused Mostaque of mismanagement; he resigned. Every other collapse this cycle was a technology problem — this one was a spending-discipline problem.

Real product, the technology just wasn’t ready (too early):

Adept AI → Amazon, ~$300M (June 2024) Built exactly what Cycle Four’s agents would become, just too unreliable for enterprises to trust in 2022–23. Same reverse-acquihire pattern as Inflection and Character.AI — for a team that was early, not a product that was free.

Cycle Four: Thinking Slower, Then Acting Autonomously (2024–2026, happening now)

What started the wave. Cycle Three showed a great model alone isn’t a moat — Midjourney and ElevenLabs won on product and pricing, not the smartest model. The next unlock followed that logic: instead of making the model bigger, make it do more per query, then let it act on its own.

• Reasoning models (late 2024): OpenAI’s o1 and o3 taught models to pause and generate a long internal chain of reasoning before answering — spending compute at question-time instead of only during training. On ARC-AGI, a benchmark designed to resist memorization, o3 jumped from ~32% to 87.5% in about a month.

• Agentic AI (2025–26): the frontier shifted from models that answer to models that act — give the AI a goal, let it plan its own steps, use tools, check its own work, and keep going for hours. The practical difference between a search engine and a junior employee.

What it means. Both shifts run the same mechanism: instead of one-shot answers, the model runs an internal loop — generate a step, check it, decide the next, repeat. A reasoning model runs that loop against its own logic; an agent runs it against the outside world, taking real actions and observing what happens.

• Can now: open an unfamiliar codebase and fix its own bugs with minimal supervision; review a 500-page legal production and draft a first-pass memo; operate a real software interface — clicking, typing, navigating — rather than only calling a predefined API; stay on one goal for hours.

• Still doesn’t work reliably: long, ambiguous, multi-week projects without drift; full autonomy on anything touching real money or physical risk without a human checkpoint.

What We Can Learn From This

Every wave started with a research paper — watch research, not headlines, to spot the next one early. AlexNet’s 2012 paper preceded Cycle Two by months; “Attention Is All You Need” (2017) preceded ChatGPT by five years; the o1/o3 research preceded “agentic AI” by mere months. By the time a wave is obvious, the paper that started it is old news to whoever was reading it.

Companies that stopped innovating always lost — and the lower the initial moat, the shorter the window to build a bigger one. Lisp machines’ moat was raw hardware speed, wiped out once general-purpose chips caught up. Nervana’s was a chip roadmap, made obsolete by a better rival within three years. Copy.ai and its peers’ moat was “wraps GPT-3 in a nice interface” — gone within a single model upgrade.

No revenue, or revenue on broken unit economics, means survival now depends on investor conviction, not product quality — and conviction can vanish for reasons that have nothing to do with either. Internally, when a new paper or bad quarter makes the tech’s promise look overstated (Perceptrons did this in 1969, Lighthill in 1973). Externally, for reasons unrelated to the science at all (Uber’s pandemic cost-cutting killed Uber AI Labs in 2020). Character.AI and Inflection burned out the same way on a shorter clock — record engagement, zero revenue, patience gone within about two years. The lucky ones get bought first, by someone who already knows how to monetize what they built: DNNresearch and DeepMind never had a business model; Google did.

Unit economics at scale deserves its own diligence, per use case. Edge cases and context changes quietly kill it — the knowledge-engineering consultancies’ per-client maintenance cost is the clearest example in this history. And willingness-to-pay versus cost-to-serve needs scrutiny before the fundraising, not after: Inflection and Character.AI had millions of engaged users, but nobody stress-tested early enough whether those users would ever convert to a price that covered the compute cost of serving them.

The value flowing to hyperscalers keeps compounding. The check size for buying into the next wave early grows roughly an order of magnitude bigger each decade — and it’s almost always the same five or six companies with a balance sheet large enough to write it.

The next wall is probably one of these:

• A reliability wall — right often enough to demo, not yet right enough to trust unsupervised.

• An economics wall — the cost per query of a reasoning model or long-horizon agent still dwarfs what most use cases can pay.

• A physical-world wall — everything built so far understands text, images, and code; almost nothing acts safely in physical space.

• A trust and regulation wall — one well-publicized autonomous-agent failure with money or healthcare could freeze adoption overnight, the way one report froze a whole field in 1973.

• A talent-and-capital concentration wall — if the next unlock needs ever more capital, fewer companies can even attempt it.

The Moral of the Story

So — is AI a bubble, the way Dalio and every other skeptic keeps saying? Wrong question. Every cycle in this piece was overhyped, oversold, and eventually walled off from whatever it had promised — and every single one left the technology permanently better than it found it. The bubble talk is usually right about the froth. It’s almost always wrong about the arc underneath it.

The better question is the one this whole piece has been answering: which wall this cycle is about to hit, who’s quietly working on the unlock for what comes after it, and who’s going to get bought, who’s going to get wiped out, and who’s going to compound independently when it arrives.

For nearly two years, his team did what he calls “the full press” — inviting Dylan Field to their summit, helping him run a board search, showing up for him in every way except writing a check. Field told them he’d let them know when the time was right. Then COVID hit, the market fell apart, and Field called: “Now’s the time.” Internally, George’s own team was still debating whether the market for designers was even big enough to justify the price. It didn’t matter. Two years of proximity to the business meant they’d already done the work a live deal process never gives you time to do. They said yes.

That’s the case for proactive over reactive investing, in one story.

Reactive investing is evaluating a company only once it’s already in a live, competitive process — a round being shopped to a room full of investors at the same time you are.

Proactive investing is finding a company you want to own, then finding your own right time and price to get in, independent of whether anyone else is looking at it yet.

What’s wrong with reactive:

the best companies get priced up the moment a round goes live, and you rarely have enough time in a live process to actually diligence a business the way it deserves.

• For most investors, the only real edge left in that setup is having a nuanced, contrarian view the rest of the room doesn’t share — and companies you’re right to be contrarian about are rare almost by definition, which makes reactive investing a bet you can win occasionally, not a strategy you can run at scale.

• Worse, the best growth investors already know this, which is exactly why they preempt rounds before they ever go live. So a lot of what you’re “reacting” to in a live process has already been passed on by the people who got there first.

There’s a real payoff to going proactive

• For the investor, it means never paying an auction price, having actual time to build conviction instead of renting a checklist from someone else’s data room, and getting a look at the company before it’s competing against every other term sheet in the market.

• For the founder, it means raising from someone who already understands the business in depth and has already shown up to help before being asked to — a very different experience from opening a data room to twenty investors he’s never met and hoping he picked well.

How do you actually do proactive?

It splits into two separate problems: how you spot the company, and how you get close enough to the people behind it before anyone else does.

Spotting the company - happens one of three ways.

1. Customer backward — you talk to customers and look for what they’re obsessively in love with, then work back to the company behind it. This is literally how Neil Mehta finds companies at Greenoaks: what he calls a “Jaw-Dropping Customer Experience.” Before his team ever takes a first meeting, they’ve already talked to the company’s customers and picked the product apart in granular detail — the customer’s reaction is the signal that sends them looking for the company in the first place, not something they check afterward.

2. Founder forward — you go find the people first, and this one comes down to how deep your network actually runs. Sometimes it means backing someone who’s already proven it once: Elad Gil has built much of his investing around getting close to repeat founders early, often before they’ve even decided what to build next. Sometimes it means going a layer earlier than that — knowing the right person inside an exceptional company well enough that they tell you who the standout talent is before anyone else has noticed. If you’d been close to the right person inside PayPal in 2001, you wouldn’t have needed to “discover” Reid Hoffman, Max Levchin, or the rest of what later got nicknamed the PayPal Mafia — you’d already have known exactly who they were, years before any of them started a company, because someone on the inside had already told you.

3. Third-party data — you let data surface the company before it’s visible anywhere else. Pat Grady and Alfred Lin have described on the Uncapped podcast with Jack Altman how Sequoia’s own sourcing system pulls in information “from the internet,” but also from “a bunch of other sources, some of which are paid, some of which are proprietary” — feeding real signal into the same system that tracks their founder relationships.

Getting close to them before anyone else is the harder half — a great signal doesn’t matter if the founder won’t take your call. A few edges that actually earn that call:

1. A nuanced point of view on the space — Bill Gurley spent years publicly writing about marketplaces and network effects on his blog, Above the Crowd. That reputation meant he’d effectively earned marketplace-expert status before any deal existed — he could call almost any marketplace founder in the world and get a real conversation about their business, with no live process underway and nothing being asked for. The expertise itself was the access.

2. A podcast, events, or a community — Harry Stebbings interviewed hundreds of founders and investors on 20VC for years before he ever raised a fund of his own. Those conversations weren’t just content; they were a running, years-long view into who was building what and how those businesses were actually progressing — so that by the time it mattered, he already knew exactly when to get in.

3. Genuine product depth — Bijan Sabet and his partners at Spark Capital were already avid personal users of Twitter back in 2008, well before it looked like an obvious business. Sabet has said Spark tends to invest in things the partners actually use themselves — that hands-on familiarity is what let him talk to Twitter’s founders as a fellow obsessive user, not as an outside investor pitching himself.

Finding the right time to get in.

Being close to a company doesn’t automatically tell you when to act — that’s its own separate skill, and it usually comes down to one of three moves.

1. Spot the inflection point and manufacture a round — you see the moment a business tips, before the founder has fully processed it themselves, and you go to them with a term sheet before a data room even exists. Say you’ve been talking to a founder for a while and, in one of those conversations, they mention the pilots are going unusually well and starting to convert. Because you already know the business, you also know who those pilot clients are — so you can go check with the clients yourself and confirm the contracts are actually about to close, weeks before the founder has even updated their own board deck. You show up with a term sheet on the back of that. A reactive investor, seeing the round once it’s live, has no way to catch up to that — they don’t know the company or its customers well enough to have formed that conviction in time.

2. Preempt off the calendar, not just the metrics — you track the unglamorous stuff: how much cash is left, how many months of runway remain, how long it’s been since the last round. Founders raise on a fairly predictable clock, and knowing where they sit on it lets you show up before they’ve started calling anyone else.

3. Offer additional capital into an internal round — instead of waiting for an external process, you add capital to a round the company is already doing internally with its existing investors, getting in at a price that was never set by a market at all.

This is also, honestly, why I spend the time I do on this podcast, on writing, on LinkedIn, and just being out talking to founders and other investors build a real point of view, show up consistently, and let the relationship come first, long before there's ever a round to talk about.

It's the version of proactive I'm actually trying to practice, not just write about. If any of this resonates, take a look at my thesis on marketplaces hiding inside SaaS here

In February 2000, Fortune ran a cover story asking “How Much Are Your Eyeballs Worth?” — the thesis being that whoever captured the most attention online would capture the most value. It didn’t hold: nearly every poster child of that era — Excite, Lycos, TheGlobe.com, Pets.com — went bankrupt or got acquired for a fraction of what they’d raised, because attention could be bought but loyalty couldn’t.

I see the same setup playing out in AI now, and the stakes are higher for me: AI carries far more disruption risk than any past wave, and Indian startups have nowhere near the capital cushion their American counterparts do to survive being wrong about where value actually sits.

An essay that stuck with me was “Thinking Through the Future for LLM Companies... and What This Means for B2B AI Startups” by Sarah Tavel’s, General Partner at Benchmark: She outlines how silicon valley is still underestimating the disruption risk from model players and outlines three ways for B2B AI startups to build defensiblity :

• A network effect

• proprietary or hard-to- access data

• executing fast enough to land-grab an overlooked vertical before anyone notices.

Among these “Network effects” striked a chord with m for four reasons, rooted in what I actually see in India:

• I’m already seeing early evidence of AI SaaS that are slowly turning into marketplaces — and marketplaces are exactly where network effects show up.

• Network effects they compound faster than the other two, and they’re a genuinely hard moat for a model provider to disrupt, no matter how good its next model is.

• It also shields an Indian companies far better from global competition than anything else does.

• India has enough — and more — idle supply sitting across labour, SMBs, and products, all waiting to be matched with the right demand.

  
  Let me explain what i mean by AI SaaS turning to marketplaces

Marketplace Businesses Hiding Inside SaaS

Here’s what makes AI different from every SaaS wave before it: it’s no longer just managing an internal workflow — it’s a system of action, reaching outside the company to interact directly with vendors, customers, candidates, whoever sits on the other side of a transaction. Do that across enough counterparties, and a company isn’t automating a process anymore. It’s building a network.

That’s the pattern I keep finding across Indian AI, category after category:The path looks the same every time. A company starts by automating a workflow — searches, reach-out, screens whatever the task is to help it find right talent, supplier , product etc. While doing the same for over 100 customers they gain context over 10x more suppliers which then helps them connect demand / supply much better , it follows the patter :

Automating a Workflow → Understanding Demand Side’s Intent → Understanding the Supply Side’s Intent → Matching

So what does this actually look like inside a real company? Start take example of Zip

Case study: Zip

Zip is the clearest proof of it at scale — and also the clearest example of a company that hasn’t figured out how to get paid for it yet.

I was sharing an early version of this thesis over lunch with Siddharth, an investor at Accel who previously spent time at DST Global. He pointed me to a company he’d worked with during his time at DST - that fits this pattern: Zip.

Zip has raised roughly $358 million across six rounds — Y Combinator, Tiger Global, and CRV backed it early; DST Global, BOND, Adams Street, and Alkeon joined a $190 million Series D in October 2024 that valued the company at $2.2 billion. What it does: an employee submits a purchase request, Zip routes it through approvals, generates the purchase order, and matches the invoice for payment, plugged into whatever ERP the company runs.

The number worth sitting with: Zip’s revenue went from roughly $18 million in 2023 to roughly $193 million in 2025 — about 10x in two years. Some of that is ordinary SaaS growth. A jump this large also tracks with a company that changed what it was selling, not just how much of it.

2020–2023: Zip’s pitch was simple. Employees were emailing spreadsheets to buy things; nobody knew what had been approved; finance found out about a purchase only when the invoice arrived. Zip gave everyone one form, routed it to the right approver automatically, and handed it to the accounting system — a traffic cop for paperwork that was happening anyway, charging like one too: a straightforward subscription, $300–600 per employee a year, plus a small usage fee.

2024: Zip added more steps to the same idea — vendor bidding, contract management, auto-matched invoices. A bigger slice of the process meant a bigger subscription bill, but the logic hadn’t changed: Zip was still just processing decisions the company made on its own.

2025 onward : by now, Zip had quietly built something valuable just by doing its original job for four years — visibility into what thousands of companies were buying, from whom, and at what price, across more than 7 million suppliers. In 2025, it started using that data. Its Market Intelligence Agent surfaces vendors a company has never worked with. Its Price Negotiation Agent, launched that October, compares what a company pays to what similar companies pay, and negotiates the price down on its behalf. Zip now markets these in savings terms — millions of dollars, hundreds of thousands of hours — not efficiency terms.

In the language of the chain: Zip has climbed all the way from workflow to matching. What it hasn’t done yet is capture liquidity as something it gets paid for — it still charges the way it always has, a subscription, with no take-rate, no savings-share fee, no vendor-side pricing.

That, to me, makes Zip meaningfully undervalued today. It’s built the two hardest ingredients of a marketplace — a large base of demand-side data, and a live discovery-and-negotiation layer — and simply hasn’t cracked the pricing model to capture the value of what it’s built yet. Once it does, it’s worth a lot more than it is today.

It’s not just Zip. This same pattern is playing out across categories — here’s how far it actually spreads.

There are several companies doing this across sectors and use cases :

None of these companies is doing anything unusual — each is a variation on the same pattern. Sitch runs it in dating, interviewing both people before ever proposing a match. FleetWorks runs it in logistics, replacing the phone-and-text load-matching between a broker and a truck carrier. Ralo - is your ai agent that helps you find best mortages.

All of them, in the end, are dabbling with the same question: liquidity.

Lets understand one section in detail - take hiring

Juicebox and Sorce: automating from opposite ends

Juicebox is a SaaS business - founded in 2022, having raised $116 million to date, including an $80 million round led by DST Global, with Sequoia, Coatue, and Y Combinator also on the cap table, and roughly tripling its ARR since mid-2025. Worth being precise about what it replaced: a recruiter used to search LinkedIn manually, reach out by hand, coordinate the back-and-forth, then sit with the hiring manager to calibrate on what “right” actually meant. Juicebox automates the whole loop — plain-English search instead of Boolean strings, automated outreach, a feedback loop that gets sharper with every profile you react to. It charges the way SaaS always has: a per-seat subscription, $139–199 a month, with AI features and contact credits as add-ons.

Sorce flips the identical problem to the other side and helps job seekers find jobs. Upload a résumé once, then swipe through job recommendations like a dating app — swipe right, and Sorce’s agent fills out the application and writes a tailored cover letter for you. Users have crossed 20 million swipes and a million submitted applications, with placements at SpaceX, Anduril, NVIDIA, and OpenAI.

Zoom out, and Juicebox and Sorce are fighting the same battle: AI against AI. One screens candidates faster for companies; the other gets candidates through more applications faster. Both are pure automation plays today.

My bet: both realise automation alone never creates lasting value. A candidate doesn’t want to apply to more jobs — they want the one or two that are actually great for them. To get there, both need real context on the side they don’t currently have. Push far enough, and they converge on something that looks a lot like Jack and Jill.

Jack and Jill already starts where Juicebox and Sorce are heading — the far end of the chain. Founded in early 2025, it’s raised a $20 million seed round led by Creandum, with early backers that include angels from Anthropic, Lovable, and ElevenLabs. Jack is an AI agent that works for the candidate, while Jill works for the company; the whole system learns constantly about both sides of the market and aligns incentives on both sides of the hire instead of optimising for just one.

Why this excites me as an investor

• Option value - if i enter any of these companies before they start monetising supply side, I can enter most of these companies at SaaS multiple at their current SaaS revenues, having good visibility into multiple other revenue unlocks that can come in near future. i can see 3-5 more ways they can monetise the existing context they have build while being a SaaS tool.

• Moat - depth of context on both supply / demand accumulated over years of making matches happen and learning what works for each party is deep moat that makes the product super sticky.

• Inflection point - Once you start monetising supply side you will probably see more market-pull form both ends since the value prop is significantly better now due to network effects and just like how every other marketplace and that should mean much higher growth rate.

What is yet to be solved

For these companies to actually earn a network effect instead of just hoping for one, here’s what needs to happen - and what I’d urge founders building in this space to spend real time on.

Business model

Imagine if Google billed you monthly for customers it had already connected you to. Or imagine you just bought credits to show your ad to a fixed number of people. Neither model would have scaled the way Google did. Instead, Google built an auction — and the auction gets paid whether or not the click ever converts. What sets the price isn’t a binary success or failure; it’s Quality Score: how good your product,

your landing page, your relevance actually is. A weak advertiser pays more for the same customer. A strong match pays less. Google gets paid either way, and the system gets smarter with every click.

Right now, these companies is mostly experimenting with two models, and both miss this insight:

• Pay-for-outcome — a hire made, a meeting booked — charges everyone the same rate: the same fee percentage, or the same flat price per meeting, regardless of how good the company or the product actually is at converting. A company with a weak employer brand or a vague hiring bar is far harder to place well than one that’s an obvious draw for talent, but today both pay the identical rate. There’s no Quality Score equivalent.

• Pay-for-credits (usage / subscription) — search, reach-outs, sourcing — has the mirror problem: a flat price per credit no matter how likely that credit is to actually produce a good outcome. A company that’s hard to sell into, or a candidate who’s hard to place, consumes far more of the marketplace’s real effort per attempt, and pays exactly the same as one that doesn’t.

I think there could be a much better model for these AI native marketplace thast yet to be discovered. The fix isn’t picking one of these models over the other — it’s building something closer to Google’s auction: a price that varies with the quality of whoever is asking for the match, set continuously, rather than a flat fee or a binary payout. That’s a genuinely hard pricing problem, and nobody in this category has solved it convincingly yet.

That’s a genuinely hard pricing problem, and nobody in this category has solved it convincingly yet. Three questions worth sitting with:

Pricing strategy — how do you price for match quality the way Google prices for Quality Score, instead of charging everyone the same flat rate?

Capturing Option value - A candidate who doesn’t land the job can be told exactly why — what the hiring manager was actually looking for, where the gap was — instead of silence. A supplier who loses an RFP can be told exactly why they lost — price, delivery terms, a missing certification — instead of a form rejection. That’s real option value: a second product built entirely out of the exhaust from matches that failed, monetizable in its own right, and almost nobody in this category is building it deliberately yet.

Payment flow — how do you make sure the payment physically flows through you, rather than around you?

Proprietary context

Every reach-out, every search, every successful match is a chance to learn something real about how a buyer or seller actually behaves, not just what they say they want. That context is what makes your next match better than your last one — and it’s the closest thing you have to a real moat, because a competitor can copy your interface in a weekend, but they can’t copy years of accumulated context overnight. Ask yourself:

• On the demand side — if you’re a candidate-facing agent, are you learning which kinds of roles a person actually thrives in, versus just accepts?

• On the supply side — if you’re a recruiter-facing agent, are you learning what a specific hiring manager really means by “a strong communicator,” beyond what’s written in the job spec?

Context tells you who to match. It doesn’t tell you when to pull the trigger — that’s a question of liquidity and timing.

Liquidity and tipping point

A lot of the thinking I see in this category is still pure SaaS logic: add X clients at X average contract value, and that gets you to X revenue next year. That’s the wrong frame for a marketplace. Two questions to ask instead:

• How much real context have you built on your buyers, and how much on your sellers? Based on how deep it runs on both sides, how many good matches can you realistically make happen?

• How close are you to the tipping point — where more buyers show up because you clearly have context on the right sellers for them, and more sellers show up for the mirror reason?

Get the timing right, and the last piece is making sure people keep coming back on their own.

Growth and retention loops

In the internet era, the best marketplace founders thought deliberately about growth loops and retention loops — often more than they thought about the product itself. Uber’s loop, for instance, ran on frequency: every extra driver on the road shortened the next rider’s wait time, which brought that rider back sooner, which pulled in more drivers. That loop wasn’t an accident. It was designed.

I don’t see the same depth of thinking from most AI founders today. Most default to whatever growth motion a SaaS company would run, instead of asking how AI itself can become the loop. The same gap shows up in how they think about liquidity: a SaaS founder thinks in contracts and expansion revenue; a marketplace founder has to think in matches and network density. Two questions worth sitting with:

• Growth loop — how do you design a genuine, AI-native growth loop here

• Retention loop — how do you design a genuine retention loop, one that creates real market pull instead of just another reminder notification?

What I’m Spending My Time On

Beyond what founders should focus on, here’s where I’m spending my time right now.

1. Not every market is created equal

Bill Gurley’s 2012 essay, “All Markets Are Not Created Equal: 10 Factors To Consider When Evaluating Digital Marketplaces,” is still the best single checklist for this. Most of it holds up for AI marketplaces without much modification:

• A genuinely new experience versus the status quo

• Real economic advantage for both sides

• A real opportunity for technology to add value

• High fragmentation of supply and demand

• Low friction for supplier sign-up

• A large TAM

• The ability to expand the market rather than just digitise it

• Being part of the payment flow

• True network effects

Three things change materially for AI marketplaces, though.

1. Frequency: an always-on AI manager can manufacture frequency in categories that never had it — Casa turns home services into a monthly membership; Counsel Health turns a doctor’s visit into an ongoing chat.

2. Disintermediation: Lawhive’s AI does enough of the actual legal drafting that a lawyer has a reason to stay on the platform, not just take the first client and leave.

3. And the value proposition itself can change, not just the delivery — a flat, predictable fee instead of unpredictable hourly billing is a genuinely different product, only possible because AI does enough of the real work to make the economics hold.

That changes which markets are actually worth building in, and I’m spending real time re-running Gurley’s checklist with these factors adjusted for AI. Even after all that, one more filter matters more than any of it.

2. Not every network effect is the same

Just because a business is building context on both sides of a transaction doesn’t mean it will produce a strong network effect. Some of these models will genuinely compound the way a marketplace should. Some won’t — the context might not transfer well across users, or the category might not have enough natural connective tissue between one match and the next. Telling those two apart isn’t obvious from the outside, and it needs deep, rigorous thinking rather than a pattern-match on “it has AI agents on both sides, so it must be a marketplace.” Telling them apart is where most of my own time goes right now — building the judgment and the tools to know the difference before the rest of the market does.

What’s Next

Two posts are coming next.

• One goes deep on the whitespace within this opportunity in India — where the biggest gaps still are.

• Second on new playbook on AI services - where i have been spending some time

Subscribe to get alerted when they’re out.

I am working very closley with Floworks team helping them navigating some of the challenged above and i will keep you posted on updated learnings on this space

If you’re a founder building in this space i would love to chat, reach out to me on Linkdien or via mail (sauravg@capria.vc).

If you’re an investor tinkering in this space - i am happy to share notes, please reachout

If you fundamentally believe my arguments are flawed - please write to me , i would love to hear your thoughts !

And if this was useful, share it with someone who should be thinking about it too.

In my last piece, I studied twelve legendary investors and found the pattern behind what made them lean in — spot the tailwind, pick the business model riding it, back the team built to win it. This piece asks the opposite question: using those same sharp instincts, why did some of the best investors alive say no to the exact companies that made other investors legendary?

I set a narrow bar for what counted. The investor had to already have a real, repeatable track record of great calls elsewhere — this isn’t a piece about mediocre judgment missing good companies, it’s about why excellent judgment still misfires. And for every case, I needed a dated, first-person record of what the investor actually said or wrote on the day they passed — a memo, a quote, a deposition — not a version reconstructed after everyone already knew how the story ended. Hindsight makes every miss look obvious; I wanted what looked true before the outcome existed. I built this set from investor interviews, SEC filings, court documents, and firms’ own published retrospectives, including Bessemer’s public “Anti-Portfolio,” where partners document their biggest misses in their own words — dozens of deals in total. I wasn’t collecting stories. I was looking for the same decision showing up again in a different room, a different decade, wearing different clothes. Two patterns kept recurring.

There are two completely different ways to miss a great company, and they deserve to be treated separately because they’re solved in completely different ways.

1. The first is not believing in the founder, the market, or the business model — deciding this isn’t a real business before a single number ever comes up.

2. The second is missing because of price — correctly believing in the founder, the market, and the business model, and still misjudging what it was worth, for one of a couple of very specific reasons. Belief first, then price. Underneath both, though, the same mistake keeps recurring: judging the thing sitting in front of you — the category, the price — instead of the capability underneath it that would decide what it could still become.

   

Part One - Not Believing the Founder, the Market, or the Business Model

This is the mistake that isn’t about spreadsheets at all. It’s about whether you believed the market was real, the business model actually worked, or the people in front of you could pull it off.

1. Believing the category was already won — Google, 1999

In 1998, Stanford professor David Cheriton and Sun co-founder Andy Bechtolsheim watched a ten-minute demo from Page and Brin and each wrote a $100,000 check on the spot — one made out to “Google Inc.,” a company that didn’t legally exist yet. A year later, former Amazon executive Ram Shriram brought the deal to John Doerr at Kleiner Perkins and Michael Moritz at Sequoia. On June 7, 1999, the two rival firms co-invested $25 million jointly, valuing Google at roughly $100 million, with almost no revenue on the table.

Source: Google press release, June 7, 1999; Steven Levy, In the Plex.

Cowan, at Bessemer, passed without ever hearing the pitch. And his instinct wasn’t lazy — in 1999, Yahoo held roughly 46% of search-related traffic and MSN about 40%. “Another search engine” looked, correctly, like a settled market.

The difference between Cowan and the two men who wrote the check wasn’t information. It was method. Cowan appears to have pattern-matched: a search engine already exists, so this market is spoken for. Doerr and Moritz reasoned from the constraint itself — is search quality, not distribution, the actual battleground here, and is that solvable in a way Yahoo’s and MSN’s model structurally can’t copy? Their answer was yes. Moritz, in Google’s own 1999 release: “Google should become the gold standard for search on the Internet.” Cowan reached his conclusion about the market before he ever asked that question. Doerr and Moritz, on this account, reached theirs by asking almost nothing else.

2. Letting a failed attempt cloud judgment on the whole category — Facebook, 2004

Bessemer’s Jeremy Levine was cornered in a lunch line at a corporate retreat in 2004 by a persistent Harvard undergraduate named Eduardo Saverin. Levine’s reply, per Bessemer’s own page: “Kid, haven’t you heard of Friendster? Move on. It’s over!”

Source: bvp.com/anti-portfolio.

Friendster really was suffering severe scaling failures by 2004 — slow load times and outages so persistent they became the company’s defining trait. But Levine let one company’s failure stand in for the entire category’s prospects. In thirty seconds, a specific, fixable failure at Friendster — a slow, badly scaled product — became, in his head, proof that “social network for young people” itself was a dead idea.

The same year, Accel’s Jim Breyer appears to have already been developing a first-principles thesis on social networking, before Facebook ever pitched: a category leader would need real identity, verified users, and a dense starting community before wide expansion. When Facebook’s actual pitch matched that model almost point for point, Breyer moved fast, overrode internal skepticism, and put in $12.7 million for Facebook’s May 2005 Series A — a single stake that alone returned Accel’s entire Fund IX. Levine was reasoning from Friendster’s failure. Breyer, on this account, was reasoning from his own independent model of what would have to be true for the category to actually work, and then checking Facebook against it.

Source: TechCrunch, “Accel Partners’ Extraordinary 2005 Fund IX,” November 2010.

3. Confusing the wedge for the endgame — eBay and Amazon

Cowan also passed on eBay in 1997: “Stamps? Coins? Comic books? You’ve GOT to be kidding. No-brainer pass.” He wasn’t wrong about what he saw — collectibles really were eBay’s biggest category. Benchmark’s Bruce Dunlevie and Bob Kagle looked at the same site and invested $6.7 million for a bit over a fifth of the company — a stake various retrospectives value anywhere from roughly $2.5 billion to $5 billion within two years of the 1998 IPO, still cited as one of the best-performing venture investments in Silicon Valley history. Benchmark’s read was that collectibles were the trust-building wedge for a generalizable auction mechanic, not the ceiling. Cowan’s read stopped at the coins.

Source: bvp.com/anti-portfolio; FundingUniverse, “History of Benchmark Capital.”

Amazon ran the identical mistake in public markets. Barron’s 1999 “Amazon.bomb” cover story — the stock down from $221 to $118.75, valuing the company at $19 billion against a $720 million loss — quoted retail consultant Kurt Barnard directly: “Once Wal-Mart decides to go after Amazon, there’s no contest.” The piece even noted, almost dismissively, that Bezos had already moved past books into music and drugs. Skeptics read the expansion as more ways to lose money. It was actually the tell that “bookstore” was never the real business — just the category with the lowest trust barrier to start in. Amazon fell 95% by 2001, making the skeptics look right for two years, then compounded roughly 47x from that same $19 billion mark by 2019, per Barry Ritholtz’s own retrospective count, and far beyond since.

Source: Barron’s, “Amazon.bomb,” May 31, 1999.

4. Not believing in the founders — PayPal and Apple

Cowan passed on PayPal’s Series A too: “Rookie team, regulatory nightmare.” Both halves were true — Peter Thiel and Max Levchin had never worked in financial services, and PayPal spent years fighting state-by-state money-transmitter battles serious enough to plausibly kill the company.

Source: bvp.com/anti-portfolio.

Here’s the job, I think, when you’re evaluating a founder. First, spot real obsession and first-principles thinking — that part is rare, and you can’t manufacture it in someone who doesn’t have it. Second, work out honestly what’s actually fixable through hiring and help, versus what’s a structural hole nothing patches. Third, go do the fixing, instead of just noting the gap and walking away. Cowan, on the record we have, skipped straight to a label — “rookie” — and never did any of the three.

Google is the cleanest case of someone who did all three. Doerr and Moritz weren’t blind to the fact that Brin and Page had never run a company — they just treated it as learnable, not disqualifying, and made an experienced CEO a real condition of the round. It took about two years, and real resistance from the founders, before Eric Schmidt came in as CEO in 2001. Brin himself admitted it later, without much enthusiasm: “Basically, we needed adult supervision.”

Source: Steven Levy, In the Plex.

Apple runs the same playbook a generation earlier. Don Valentine’s own 1977 Sequoia memo — the one that actually funded Apple — says “management questionable” in his own handwriting, and by his own account, “a lot of people wouldn’t invest in Apple, wouldn’t even talk to Apple, because Steve was so odd.” Valentine didn’t ignore that. By most accounts of the memo, he treated Mike Markkula’s involvement as an experienced operator as a real condition of the deal, before Sequoia’s money went in.

Source: Sequoia’s 1977 Apple investment memo, published 2026; Sequoia oral history, articles.sequoiacap.com/apple-story.

PayPal never got that kind of help from Cowan, because he never looked past “rookie” long enough to ask what was fixable. It rode eBay’s own growth instead, which bought the team enough room to learn on the job — eBay bought PayPal for $1.5 billion in 2002, four years after Cowan’s pass. “Rookie” and “odd” are true statements about a founder. They’re not verdicts. What’s actually hard to find is obsession and first-principles thinking. Skills, by comparison, you can go hire.

5. Not believing the unit economics could ever work — Instacart, Tesla, and WeWork

Sometimes the skeptics were, in the moment, simply correct — and this is the case that keeps the rest of this piece honest.

Bessemer passed on Instacart’s 2013 Series B — $37 million in annualized GMV across six cities — for being “daunted by negative gross margins.” That wasn’t a misreading; grocery delivery genuinely lost money on every order at that stage. Instacart went public in September 2023 at $10 billion, but it took a decade, and margins improved specifically because density changed the underlying economics — the objection was real, and the company outgrew it rather than disproving it.

Source: bvp.com/anti-portfolio; Instacart IPO coverage, September 2023.

Tesla is the sharper version, because the investor who passed also personally loved the product. Bessemer’s Byron Deeter test-drove a Roadster in 2006, put down his own deposit, and passed on the $40 million Series C, telling his partners: “It’s a win-win. I get a great car and some other VC pays for it!” His reason: “negative margin company.” Accurate — Tesla’s 2006 net loss was $30 million against essentially zero revenue. Tesla passed $30 billion in market cap by 2014, and what actually closed that gap wasn’t just improving car margins — it was entire businesses that didn’t exist in the 2006 thesis at all: energy storage, solar, and — in the estimation of many Wall Street analysts today — Full Self-Driving software above everything else. Deeter’s math on the car in front of him was right. It said nothing about businesses that hadn’t been proposed yet.

Source: bvp.com/anti-portfolio; Tesla Series C coverage, BusinessWire, May 31, 2006.

Sometimes, though, the flawed unit economics never get fixed. WeWork’s growth, on a chart, looked like a software company’s. What mattered was $47.2 billion in lease obligations against $17.9 billion in balance-sheet liabilities — a mismatch that had nothing to do with growth rate and everything to do with a leasing business wearing a technology company’s multiple. Scott Galloway called it in writing before the collapse: “Any equity analyst who endorses this stock above a $10 billion valuation is lying, stupid, or both.” Masayoshi Son, who backed it at $47 billion using almost the same language he’d used for Alibaba — “I fell in love with WeWork” — later admitted: “We made a failure on investing in WeWork and I’ve been admitting that several times I was foolish.”

Source: Scott Galloway, “WeWTF,” No Mercy / No Malice, August 16, 2019; SoftBank Q4 FY2019 earnings call, May 2020.

Negative margins aren’t, by themselves, a signal of anything. Sometimes they’re a stage a real business passes through on its way to scale. Sometimes they’re the entire business, permanently, no matter how the story is told. They look identical on the spreadsheet in both cases. Telling them apart is the actual skill.

Part Two — Right About the Category, Wrong About the Price

This is the more interesting mistake, because in every case here the investor correctly believed in the founder, the market, and the business model. They understood exactly what the company did. They looked at the number attached to it and concluded the number couldn’t be right — for one of two reasons. In every case, they were pricing the business as it existed. What they missed was the capability underneath it, and where that capability could still go.

1 — The same business compounds longer than the room believes

There’s a reason this catches so many smart people, and it isn’t a lack of intelligence — it’s wiring. We’re not built to sense how fast something can grow once it actually starts working, and every tech cycle compounds faster than the one before it. ChatGPT reached one million users in five days — a milestone Netflix took three and a half years to hit, Twitter took about two years for, Facebook about ten months. It crossed 100 million monthly users within two months, per a widely reported UBS analysis — the fastest-growing consumer application ever, at the time. Each cycle resets the speed limit, and intuition never gets the memo.

Source: Reuters, “ChatGPT sets record for fastest-growing user base,” February 2, 2023.

Stack a second blind spot on top: most of these businesses are also network-effects businesses, where value doesn’t grow with the number of users, it grows with the number of possible connections between them — twice the users can mean roughly four times the value, since value rides on the square of the network, not the headcount. Nobody here was bad at math. Their intuition was calibrated for a slower, more linear world, and that’s exactly what broke.

Bessemer’s Levine met Brian Chesky in January 2010, Airbnb’s first $100,000 revenue month. Chesky’s ask — a $40 million valuation, roughly 33 times annualized revenue — got the verdict “crazy.” February revenue was $200,000. March was $300,000. By April, Airbnb raised again at 1.5 times the “crazy” price — on a lower effective multiple than the one Levine turned down, because revenue had tripled faster than the price moved.

Source: bvp.com/anti-portfolio.

YouTube adds a twist: the buyer agreed with the skeptics and paid anyway. When Google bought it for $1.65 billion in 2006, Mark Cuban wrote that “only a moron” would buy it, predicting it would be “sued into oblivion” — and five months later Viacom filed a real $1 billion suit. Eric Schmidt’s own 2009 deposition records that his internal estimate of YouTube’s worth was $600–700 million, “much lower than we paid for it.” Real monetization didn’t arrive until 2010. Google didn’t disclose a single YouTube revenue figure until 2019. The first number the public ever saw — $4.04 billion, for one quarter, in Q1 2020 — was roughly two and a half times the entire purchase price, in ninety days.

Source: Schmidt deposition, 2009, reported by CBS News/CNET; Viacom v. YouTube complaint, SDNY, March 13, 2007.

Google’s own round contains the cleanest version of the blind spot, aimed at the investor himself. In his first meeting with Larry Page, John Doerr asked how big the company could get. Page said: “Ten billion.” Doerr, assuming market cap, pushed back that a realistic ceiling might be “as high as one billion dollars.” Page corrected him: “And I don’t mean market cap, I mean revenues.” The most quotable fact from Google’s Series A isn’t a story about investor foresight. It’s the investor underestimating the exponent in front of him.

Source: Steven Levy, In the Plex, excerpted by TechCrunch, April 4, 2011.

2 — A second business growing invisibly inside the “expensive” one

Amazon’s AWS launched in 2006 with $21 million in first-year revenue, invisible inside a $10.7 billion company, and wasn’t even reported as its own segment until 2015, by which point it was $7.88 billion and growing 70% a year — soon after, it was generating the majority of Amazon’s operating profit, while the retail business everyone actually argued about stayed thin-margin the whole time. Advertising ran an almost identical script a few years later: quietly started around 2012, folded for years into Amazon’s catch-all “Other” revenue line, and not broken out as its own disclosed line item until 2022 — by which point the underlying business had already been roughly a $10.1 billion revenue stream back in 2018.

Source: Amazon SEC filings, 2015 AWS segment disclosure and 2018 10-K; The Motley Fool, “Amazon Breaks Out Ad Revenue for the First Time,” February 7, 2022.

Apple’s Mac — the direct descendant of the exact product Sequoia was pricing in 1977 — is under 10% of Apple’s revenue today. iPhone, nonexistent until 2007, was just 0.5% of revenue in its own first year and 52% by 2023. Services — not a business Apple reported, or even had, in 1980 — hit 26% of revenue by 2025. Nobody pricing Apple in 1977 was pricing a subscriptions business, because it hadn’t been invented.

Source: Apple 10-K filings, FY2023–FY2025.

ARM is the same story, with the second business arriving even later. SoftBank paid $32 billion for it in 2016 — a 43% premium that sent its own stock down 11% same-day — on an Internet of Things thesis that turned out directionally right and about a decade early. The actual re-rating came from AI and data-center compute, a category that barely existed as an idea in 2016. Nvidia ran the same play in slow motion, in public view, for over a decade: founded in 1993 to build gaming graphics chips, its 2006 bet on CUDA — letting those chips run general-purpose computing — was treated by Wall Street for years as an expensive distraction, and its market cap went largely sideways for the better part of a decade while that “distraction” matured. Data-center revenue first passed gaming revenue in the quarter ended July 2020, then went from $15 billion to $47.5 billion in a single fiscal year once generative AI arrived. Nvidia’s market cap crossed $1 trillion in May 2023 and $3 trillion thirteen months later. The chips were always the same instrument. Nobody, for a decade, was pricing what they’d eventually be asked to compute.

Source: Nvidia Q2 FY2021 earnings release; Nvidia FY2023–FY2024 annual results; market-cap milestones per CNBC and Bloomberg, 2023–2024.

The uncomfortable admission underneath this mechanism is that spotting the specific hidden business ahead of time is usually a losing game. Almost nobody could have named AWS inside a bookstore in 2003, or a graphics company’s AI future in 2007 — these businesses were invisible on purpose, buried inside consolidated numbers or simply not yet invented. What you can actually assess, without needing to predict the exact unlock, is optionality — does this capability have more than one surface it could plausibly point at — and, more usefully, whether the people running the company have already shown you they know how to find a second act. Bezos had already expanded past books before AWS ever existed. Jensen Huang funded a decade-long, unpriced bet on general-purpose computing while still running the gaming business well enough to pay for it. That track record of unlocking is a real, observable signal, well before the specific unlock ever shows up in a filing. The specific unlock almost never is. Don’t price what you can see. Price what’s underneath it.

What I’m Still Figuring Out, as an Investor

I don’t have this fully solved, and I’d be wary of anyone who claims they do — every investor in this piece is demonstrably brilliant and still has a page-length list of misses sitting right next to their legendary wins. What follows is less a finished framework than a set of working beliefs, and a few open questions I haven’t resolved.

A Few Things I Believe

• What’s in front of you is rarely the whole business. It’s the same idea that explains almost every miss in this piece, just stated as a rule instead of a story. Here’s the hierarchy I keep coming back to: revenue streams sit inside products, products sit inside hard capabilities, and hard capabilities sit inside soft capabilities.

– Revenue line: just an output — the number today, and nothing about tomorrow.

– Product: the business under the revenue line — what a customer is actually buying.

– Hard capability: the business under the product — the durable, structural thing (a network, a logistics engine, a brand’s pricing power) the product happens to be running on right now.

– Soft capability: the business under all of it — the people, the culture, the org’s actual DNA, its ability to attract talent nobody else can. Google’s clearest soft capability was never Search as a product — it was the ability to pull in the best engineers on Earth, which kept manufacturing new hard capabilities (ad auctions, distributed systems, deep learning) long after any single product matured, and each of those hard capabilities kept unlocking new products and new revenue lines underneath it.

– Every layer deserves real scrutiny on its own terms. But the soft capability matters most, because products change, revenue lines change, even hard capabilities erode eventually — and the soft capability is the thing that either regenerates a new one when the old one runs out, or doesn’t.

• Play your own game, deliberately. In my last piece, I noticed something across almost every investor with more than one massive win: each had built a genuinely distinct, repeatable playbook, not just good instincts. Bill Gurley’s was marketplaces — a specific checklist for spotting idle supply in fragmented industries. Masayoshi Son’s was what I called a penetration-curve time machine - tracking a technology’s adoption curve in a leading market to predict what’s coming in a market running years behind it. Both were really the same move: a repeatable way of spotting the capability underneath a category before anyone else bothered to look. A real playbook is what lets you go deeper than the generalist in the room and catch the nuances everyone else misses.

– I’m choosing to build mine deliberately, as part of my own playbook, rather than borrowing someone else’s

– My working rule so far — still more hypothesis than proven discipline — is to stay cautious on businesses with obviously flawed unit economics unless I have real, tested conviction I can tell a Tesla from a WeWork.

– Most of the time I don’t have that conviction yet, so the honest move is to say so rather than force it, even if it means missing some companies that eventually fix their margins the way Tesla and Instacart did.

• Then stress-test, narrowly. Only after that do I switch into devil’s-advocate mode: narrow the risks down to the one or two that would actually be fatal, and decide whether to spend energy fixing them — the way Valentine pushed Jobs toward Markkula, the way Doerr and Moritz made an experienced CEO a real condition of Google’s round — or simply calibrate honestly whether they’re survivable at all, the way Cowan should have with PayPal’s regulatory fight once eBay’s growth was visible.

Few Open questions i am spending time on

That’s the shape of what I believe today — not a finished system, just one I’m actually testing against every new deal I see. What I don’t have is a clean answer for how to apply any of it in real time, before the outcome is obvious to everyone. Four questions in particular keep circling back.

• How do I actually price compounding? Every tech cycle seems to compound faster than the last one, so the instinct that correctly priced the last cycle’s winners will underprice this cycle’s, by an amount you can’t know in advance. Doerr underpriced Google’s own ceiling by roughly 10x — in the room, while writing the check.

• How do I spot Act Two before it exists? Some optionality is visible if you look — Bezos had moved past books before AWS existed. Most isn’t. Nobody could have modeled Nvidia’s AI business off its 2007 gaming numbers. I only seem to get good at telling these apart in hindsight.

• How do I price soft capability? Team and culture is where almost everything in this piece ends up mattering, and it’s the layer with the least data attached to it. Revenue and margins show up in a filing eventually. Culture mostly doesn’t, until it’s already gone wrong.

If you found this research valuable, I’d be grateful if you shared it with someone who enjoys thinking deeply about startups, investing, or company building. That’s still how most new readers discover my writing.

And if you think I’ve missed something or disagree with one of the conclusions—I’d genuinely love to hear from you. Some of the best ideas in these essays have come from conversations with readers after I hit publish - Please write me in comments / Linkedin

If you’re new here, I write deeply researched essays on venture capital and company building every few weeks. If that sounds interesting, I’d love to have you subscribe.

On one hand, AI is doing parts of your job — and doing them well. Layoffs are happening. And somewhere in the back of your mind, a question keeps surfacing: does my role exist in five years? On the other hand, you are shipping more than ever. The slow, boring parts of the job are disappearing. You are spending more time on the things that actually matter.

So what do you do with that? What does your job look like in the near future? What does a product team look like when AI does half the work?

A few weeks ago, I published a piece analysing 12 legendary investors across 6 decades - One key learning stood out: the best investors spend an incredible amount of time with technical talent to understand where the puck is going. They get close to the builders.

I took that seriously.

Over the last 4 weeks, I’ve been doing exactly that - sitting with product managers across big tech, early-stage startups, and everything in between.

Alongside that, I went deep on history. Analysing how tech roles have actually evolved across every major platform shift of the last 30 years. Who evolved. Who got left behind.

During one of my catchups with my dear friend Amrut (VC at Nexus Venture Partners) we got deep into this topic. Amrut spends a lot of time with AI builders here and, like me, believes the best way to understand where things are going is to sit in a room with the people doing the work. We decided to host a PM gathering over coffee as part of our ongoing series with technical talent. Twenty slots. 150+ registrations.

The room ended up with PMs across Google, Microsoft, Swiggy, Atlassian, Visa, 11Labs, and a range of AI-native startups. What follows is my summary of findings — and some open questions I am still sitting with.

The conventional wisdom is split in two camps

Camp one: AI lets one person do everything — generate a UI, write code, run analysis, draft copy, all in a day. You need fewer people. Product pods shrink. Headcounts drop.

Camp two: Just because you can produce something with AI does not mean it is good. A PM who generates a Figma screen is not a designer — they produced a design. Specialists survive, they just stop doing execution work and focus entirely on judgment.

Both camps are right about something. But if you look at how tech roles have actually changed across every platform shift, a clearer pattern emerges.

What history actually shows

Every major platform shift has changed roles — not eliminated them. The pattern is consistent. Roles do one of three things.

They split. The webmaster owned the entire early web — server, HTML, database. As the web scaled, one role became four: front-end developer, back-end developer, DBA, sysadmin. None of those categories existed in 1994.

They merge. When complexity moves to machines, redundant roles collapse into one. The sysadmin managed servers. The release manager handled deployments. When cloud made both a software problem, they merged into the DevOps engineer — one role, broader scope, higher pay.

They evolve. When a platform changes how you work without changing what you own, the role adapts or disappears. Java developers became Android developers. Web developers became React developers. People who did not move became irrelevant.

The webmaster became four specialised roles in under a decade. The sysadmin and release manager merged into DevOps in three years. Every platform shift reshapes roles. None have eliminated them.

What’s happening now

Three things are happening simultaneously across product organisations — and they are pulling in different directions depending on where you sit.

1. Customer-facing pods are merging

The direction is clear: engineer, designer, and PM are converging into a single product owner — one person who can take an idea from customer insight to shipped product without a handoff chain. The Forward Deployed Engineer is the most complete version of this already in the wild. An FDE combines PM, solutions engineer, customer success, and deployment into one person who owns the customer end-to-end. Microsoft has started building the same model with Forward Deployed PMs who own deep customer domains.

Three forces are driving this convergence across the rest of the industry.

PMs are covering more ground per person

As implementation gets cheaper, the traditional PM→Designer→Engineer dependency is weakening — each PM now owns more surface area with fewer dedicated engineers. Across Google, Atlassian, Swiggy, and AI-native startups, PMs are creating prototypes, generating designs, and analysing customer feedback themselves. At Major big techs and startups, PM-to-engineer ratios have risen measurably.

Bureaucracy itself is an incentive to merge

Every handoff introduces delays, context loss, and re-alignment overhead. That friction costs more than most teams realise. A senior PM at Google put it directly:

“I am comfortable being an A-grade PM, B-grade engineer, and C-grade designer if it eliminates weeks of collaboration overhead.”

This PM now builds mocks and makes minor code changes before involving designers or engineers — validating ideas in days instead of waiting weeks.

Execution work is collapsing. Strategic work is expanding.

The time PMs spent gathering information, writing updates, and running alignment meetings is being automated away. What remains — and is expanding — is judgment work: prioritisation, trade-off decisions, identifying what to build next.

• A PM at 11Labs stopped scheduling meetings with analytics teams entirely — she queries AI directly instead.

• At Visa, conversations have shifted from information sharing to decision-making because everyone arrives with AI-generated context.

• Atlassian trained roughly 70 PMs on internal operating systems that automatically summarise customer interviews, generate product updates, and answer product questions.

2. Infrastructure product teams are splitting

While customer-facing pods get leaner, the opposite is happening in infrastructure — and the logic is precise.

As AI generates more code, the cost of building features falls toward zero. But the cost of trusting what you built does not. When a model can hallucinate, degrade silently in production, or behave differently across user segments, the constraint shifts from can we build this to can we trust what we built.

This is creating a new layer of roles that did not exist two years ago: AI ops PMs managing operational reliability, evaluation engineers whose entire job is designing test suites that catch model failures, platform PMs who own internal tooling, compute PMs managing infrastructure behind model serving, ML safety PMs, developer tools PMs, and data infrastructure PMs. These are not temporary titles — they reflect genuinely new complexity that someone has to own.

Infrastructure work also compounds in a way feature work does not. A single evaluation engineer building a shared testing framework improves the reliability of every product team using it. Feature work scales linearly with headcount. Platform work scales with the number of teams building on it.

Swiggy’s AI team described the split explicitly: Product Engineers focused on customer surfaces, and Infrastructure Engineers building the shared systems — coding agents, analytics platforms, and evaluation frameworks — that make Product Engineers far more effective.

3. Product leaders must evolve

The roles are not the only thing changing. What it means to be excellent at them is changing faster.

Domain expertise is resurging

When anyone can generate a summary of payment regulations in seconds, the value is no longer in having the information. It is in knowing what it means for a specific customer, edge case, or regulatory environment. That takes years to build. AI does not compress it.

A PM at Visa was direct: AI can help someone learn about payments quickly, but it cannot replace years of understanding the edge cases and regulatory trade-offs that come up in production. Startups are already acting on this: healthcare AI companies are hiring doctors, legal AI companies are hiring lawyers. Generic PMs are losing ground. Domain specialists are gaining it.

PMs must operate more like founders

AI is not reducing headcount — it is raising expectations. The question has shifted from how many can we remove to what else can we build with the same people. The PM who creates value by coordinating execution is doing work that smaller AI-empowered teams can now do. The PM who identifies a market gap, validates it fast, and drives it to revenue is doing work that compounds.

Taste is the hardest moat

As AI makes average output easy to produce, the gap between good and excellent becomes the primary differentiator. Knowing which design feels right, which metric actually matters, which customer problem is worth solving — these require judgment built over years of observation and failure. They cannot be prompted.

T-shaped skills are now the baseline

The future product person moves fluidly from customer insight to design to shipped product without depending on a chain of specialists. The biggest winners will not be the best coders or the best PMs in isolation — they will be the people who combine product thinking, domain depth, and AI-native execution.

How hiring is changing

Hiring is a lagging indicator — but it is starting to move. The hardest problem has always been that the skills that matter most — judgment, ownership, adaptability, energy — are exactly what interviews measure worst.

One concrete change: Google now brings the hiring manager in right after the phone screen — not at the end. Previously, candidates would clear three or four rounds of generic evaluation before a hiring manager was involved. Too many people were passing the generic bar but failing once they hit the actual team. The fix: establish team context early, before investing more rounds in a candidate who may not fit regardless of raw PM ability.

The preferred signal now is asking candidates to walk through a product they have actually shipped — the trade-offs made, and what happened after launch. AI literacy is now tested directly: not whether someone uses AI, but how they use it, which tools they reach for, how they handle hallucinations. The direction is toward working sessions where candidate and interviewer solve a real problem together.

Open questions I am still sitting with

These are the things I do not have clean answers to yet. If you have a strong point of view on any of them, share it in the comments - if your argument is genuinely interesting, we would love to have you at the next session.

1. Who is better positioned to become the customer-facing product owner — an engineer who knows how systems work and has to upskill on business context and customer judgment, or a PM who knows the business but has to upskill on tech, or a designer? Each starts from a different foundation. Which matters more?

2. For customer-facing PMs increasingly operating like founders — finding new revenue streams, owning business outcomes — what is the real incentive to stay at a large company rather than just start something themselves?

3. What would most engineers actually prefer — upskill toward the product owner role, or go deeper into infrastructure? The answer shapes how companies should train and retain engineering talent.

4. Are there industries or use cases where the product pod does not shrink or merge — because design judgment and specialisation genuinely matter more and you need deep specialists? What are they?

5. How does hiring change when we are optimising for intrinsics rather than skills? How do founders, VCs, and recruiters identify the next generation of product leaders when the old proxies — pedigree, domain knowledge — matter less?

6. How will entry-level talent learn the craft when most entry-level tasks will be done by AI? As a leader, how are you thinking about attracting and training the next wave of junior product talent?

Amrut and I are continuing this series — hosting more technical folks over coffees and dinners across Bangalore. If you want to be part of the conversation, drop a comment below.

Let me also take a moment to thank , Rohan , Anjali & Sanchit for feedback on first draft.

That gap pushed me to look for primary sources: investment memos, books, recorded talks, documented deal histories. The goal was to understand not what great investors say they look for, but how they actually think.

The twelve investors below were selected on three criteria: each has multiple investments with documented returns above 50x (verifiable from public filings or acquisition announcements); each left a documented record of their decision-making; and together they span six decades of investing (1972 to present), every major stage, and fund sizes from individual angels to $100B. Any pattern that holds across that range is structural — not a product of era, firm, or personal style.

All outcomes above are sourced from public filings, acquisition announcements, or documented investment records.

The Conventional Wisdom Gets It Wrong

The standard framing is that great investors fall into two camps — market-first or founder-first. Don Valentine: “We don’t choose people. We choose markets. Give me a giant market always.” Masayoshi Son invested in Alibaba after a brief meeting with Jack Ma because of his “shining eyes” and leadership presence. These accounts suggest two completely different philosophies.

But when you look at the actual deals not the quotes, the pattern is identical across all 12.

1. They spot a massive technology tailwind that signals the right moment for a new category.

2. They identify which business model or theme is up for grabs within that tailwind.

3. They find the team best positioned to build the category-defining company.

The style difference is in the order of operations. The underlying investment logic is the same.

Even Son’s Alibaba investment — his most “founder-first” story — was preceded by a specific market thesis: China’s internet penetration in 2000 matched the US in 1995. He knew the category before he met the founder. The “shining eyes” determined which founder got the cheque, not whether he would write one.

A note before the analysis: this is necessarily retrospective. Great investors also make many failed bets, and success makes decision-making appear cleaner than it was in real time. The goal here is not to argue these frameworks guarantee outcomes, but that certain patterns appear repeatedly among investors who compounded exceptional returns over long periods.

How They Spot Tailwinds

Seven signals appear consistently across these investors’ best decisions. Each is grounded in specific, concrete data inputs.

1. Spend Time With Technical Talent

The most reliable leading indicator of the next major category is what the most capable engineers are doing voluntarily, with no commercial incentive. They reveal the next platform through what they build in their spare time, what they complain about in technical communities, and what they adopt simply because it’s better — even when rough and unfinished. The investors who exploit this do it deliberately, making recurring visits to university labs, developer communities, and technical conferences. The question they’re asking: what will be possible in three to five years that is not possible today?

John Doerr / Kleiner Perkins → KPCB internet portfolio: Sun Microsystems, Netscape, Amazon, Google

Doerr made regular visits to university computer science departments and DARPA-funded research labs throughout the 1980s and 1990s, asking a single question: what will be possible in three to five years? From those conversations, he saw networked computing being built by researchers years before any commercial product existed. The Kleiner Perkins internet thesis came from watching the most technically capable people at Stanford — not from market research. Sun, Netscape, Amazon, and Google all followed that same thesis.

Source: John Doerr, Measure What Matters (2018); Kleiner Perkins portfolio history

Elad Gil → Stripe (~$65B valuation, 2023)

Before investing in Stripe, Gil was embedded in the developer community as a practitioner who had run technical teams at Google and Twitter. He observed engineers voluntarily adopting Stripe’s API — seven lines of code versus weeks of compliance overhead — with no sales contact from Stripe. His investment came from independent observations across unconnected engineering teams, not from a pitch.

Source: Elad Gil, Invest Like the Best podcast; Lex Fridman podcast

2. Track Infrastructure Cost Curves

Every time the cost of a core technology drops by an order of magnitude, a new class of products becomes viable that wasn’t before. Valentine’s method: track the cost curve of enabling technologies — semiconductors, networking silicon, storage — and calculate the price at which a specific product becomes viable at volume. He didn’t wait for the product to exist. The opportunity opens when cost is 12–24 months from viability and closes when the category becomes obvious.

Don Valentine / Sequoia → Apple (IPO $1.78B, December 1980)

Sequoia’s 1977 investment memo classified the deal as “M. Priority” and described the market as “Home — Hobby Computers” at >$500M. Valentine’s framework: microprocessor cost had dropped to the level where a sub-$1,500 consumer computer was viable for the first time. Sequoia invested in a $600K round for ~10%. Apple IPO’d at $1.78B market cap on December 12, 1980.

Source: Sequoia 1977 Apple investment memo, published April 2026; Apple IPO, December 12, 1980

Don Valentine / Sequoia → Cisco (Sequoia invested $2.5M for ~30%, 1987)

Networking silicon had fallen below the cost point where router hardware could be sold profitably to enterprises outside the Fortune 500 for the first time. Cisco’s founders had been rejected by numerous other investors. Valentine saw the cost curve; others didn’t.

Source: Don Valentine oral history, Computer History Museum

3. Track Capability Thresholds

Technology doesn’t just get cheaper — it crosses capability thresholds that make previously poor experiences genuinely good. Botha’s approach: track hardware capability curves (camera quality, GPS accuracy, battery life, processor speed) against application categories where experience was still inadequate. The question at each step: is there an app that would be genuinely good at this capability level, but was genuinely poor at the level before? When yes, the category has just become investable.

Roelof Botha / Sequoia → Instagram (Facebook acquired for ~$1B, April 2012) The iPhone 4, released June 2010, shipped with a 5-megapixel camera — the threshold at which mobile capture quality crossed from tolerable to genuinely good. Instagram launched in October 2010, 3.5 months later. Photo sharing behaviour was already established on desktop. What changed was the hardware making mobile capture genuinely good for the first time.

Source: Sequoia Capital Crucible Moments podcast, Roelof Botha episode

4. Track Technology Penetration Across Geographies

Technology adoption follows a predictable S-curve in every market. Once you have that curve for the leading market, you can identify lagging markets running 18–36 months behind. Son’s method: if China’s internet penetration in 2000 matched the US in 1994–95, the business models proven in the US between 1995 and 2000 are likely to prove out in China over the next five years — at prices that reflect local investor uncertainty rather than confirmed outcomes. The edge is acting on that arithmetic before local consensus forms.

Masayoshi Son / SoftBank → Alibaba ($20M for ~34%, 2000; ~$150B peak value)

Son tracked internet penetration rates across geographies and calculated that China’s adoption curve was running several years behind the US. He met Jack Ma briefly. What confirmed his decision was Ma’s vision and conviction — but the category thesis existed before the meeting. SoftBank’s $20M generated returns reported at approximately $150B at peak value.

Source: QZ; Fortune; South China Morning Post; Son in multiple public interviews

5. Identify Irreversible Behaviour Shifts Already Underway

The signal isn’t whether a company has created a new behaviour. It’s earlier: is there a behavioural shift already happening — driven by new technology — that a company can systematise and own?

People do not wait for a great product before they start changing how they behave. They improvise. They do the new thing badly with inadequate tools: sharing video files by email, coordinating rides via SMS, paying each other through clunky bank transfers. The behaviour exists before the right product does.

The question to ask: is there something people are clearly trying to do, at meaningful scale, with tools that were not designed for it? If yes, the demand is real. What doesn’t yet exist is the infrastructure to make it reliable. That gap is the investment.

Roelof Botha / Sequoia → YouTube (Sequoia invested ~$11.5M; Google acquired for $1.65B, October 2006)

People were already sharing video before YouTube launched — through email attachments, crude hosting sites, and peer-to-peer networks. The behaviour was real but fragmented. YouTube didn’t create video sharing; it made an existing, proven demand dramatically better. Botha’s September 2005 investment memo recognised that broadband had crossed the threshold making video streaming viable — and that enormous, proven behaviour was waiting for the right infrastructure to own it.

Source: Sequoia YouTube investment memo, September 2005 (public via litigation); Google acquisition press release, October 2006

6. Track Underutilised Supply in Fragmented Markets

In fragmented industries, significant supply capacity sits idle not because demand is absent but because the matching layer is broken. Gurley’s marketplace checklist: identify a large, fragmented industry; measure supply utilisation; calculate why utilisation is low — matching problem, trust problem, or discovery problem? Assess whether a platform solves it in a way that compounds with scale. High fragmentation + high idle supply + a matching barrier = a marketplace worth building.

Bill Gurley / Benchmark → Uber ($11M Series A, February 2011)

Gurley published his marketplace framework openly on his blog “Above the Crowd.” Travis Kalanick read it and reached out directly. The blog didn’t generate warm introductions — it generated founders who had already self-selected against the framework.

Source: Kalanick in multiple interviews; Bill Gurley, “Above the Crowd”; Benchmark Series A reported February 2011

Bill Gurley / Benchmark → OpenTable (Series B 2000; Priceline acquired for $2.6B, June 2014)

Same framework: restaurant table inventory sitting empty because phone-based reservation created friction on both sides of the market.

Source: Gurley on OpenTable in multiple interviews; Priceline-OpenTable acquisition: WSJ and TechCrunch, June 2014

7. Track How Long a Category Has Gone Without Structural Innovation

In markets where the dominant product design hasn’t changed in a decade or more, a specific dynamic builds: customers often continue paying because alternatives are equally inadequate, not because the product is good. Complexity accumulates. Pricing opacity grows.

Customer frustration alone is not enough. The real question is: how long has the dominant design been unchanged, and has a new technological shift now made a dramatically better architecture possible? Long stagnation plus an enabling technology equals an open category.

Peter Thiel → PayPal (co-founded 1998; eBay acquired for $1.5B, October 2002)

Online payment in 1998 required merchant accounts, payment gateways, and lengthy application processes — complexity that served financial incumbents rather than merchants or consumers. The dominant infrastructure hadn’t structurally changed since the early 1990s. PayPal replaced the process with an email address and a password.

Source: eBay acquisition press release, October 2002; Thiel, Zero to One (2014)

Doug Leone / Sequoia → ServiceNow (Sequoia led $41M round, 2009; IPO June 2012 at ~$2.96B)

Enterprise IT service management was dominated by BMC Remedy, first released in 1991. By 2009, the dominant architecture was ~18 years old — built for on-premise infrastructure that cloud computing was making obsolete. ServiceNow was rebuilt from scratch for a cloud-native world. Net revenue retention above 120%. Leone argued against a $2.5B VMware acquisition offer, calling it “giving away the company.”

Source: Sequoia Capital, “The ServiceNow Story”; ServiceNow IPO prospectus June 2012

What Kind of Company Rides the Tailwind?

The same tailwind can produce structurally different types of companies. Being clear on which type shapes what you look for in the team. Five patterns appear consistently:

A — Bottleneck Created by New Infrastructure. The new platform creates a constraint that didn’t exist before. Oracle solved the data management problem created by the PC wave. Stripe and Twilio solved the payments and communications bottlenecks created by AWS.

B — Bottleneck About to Become Evident. The infrastructure is in place; the constraint hasn’t hit critical mass yet. Investing before the pain is acute gives you entry price before the category is contested. ServiceNow in 2009: cloud adoption was making on-premise IT management inadequate before most enterprises felt it acutely.

C — Act 2 of a Proven Model on New Infrastructure. A behaviour was validated on one platform; new infrastructure makes the same behaviour dramatically better on the next. YouTube was Act 2 of photo sharing (proven via Flickr on desktop) once broadband removed the buffering barrier.

D — A New Delight Made Possible by New Technology. Something genuinely new — not a migration from a prior platform, but a behaviour only possible because of the specific capabilities the new platform provides. Instagram wasn’t Flickr on a phone. The combination of persistent connectivity, always-present camera, and social graph created a sharing loop desktop couldn’t replicate.

E — A Differentiated Attempt at an Existing Space. The category exists; a new entrant has a structural insight competitors cannot replicate without rebuilding from scratch. Google’s PageRank required re-crawling and re-indexing the entire web. Altavista couldn’t flip a switch. Facebook’s real-identity architecture meant Myspace would have had to ask 100M pseudonymous users to verify themselves.

How They Source

Three mechanisms appear consistently across all 12 investors: network, media, and going deep into a domain before the category becomes obvious to others. Each is deliberate, not passive.

1. Network — The Pre-VC Career Gives You the First Edge

Most of these 12 investors didn’t enter venture capital as career investors. Their pre-VC careers — as engineers, operators, or founders — gave them networks no generalist investor could replicate.

• Don Valentine spent seven years at Fairchild Semiconductor, the nursery of Silicon Valley. Its alumni founded Intel, AMD, and National Semiconductor. Apple and Cisco entered his pipeline through those relationships, not through formal pitch processes.

• Peter Thiel had personal relationships with the entire PayPal founding team — who went on to found or fund LinkedIn, YouTube, Yelp, Palantir, and SpaceX. His $500K Facebook investment came through Sean Parker, who he knew through that network.

• Elad Gil spent years as a product leader inside Google and Twitter. When Stripe began getting adoption among developers, he heard about it from engineers he knew at other companies — not from a pitch.

• Marc Andreessen had operated at the technical frontier across two complete waves before ever writing a cheque. Founders sought his board involvement because he had done the operational work himself — a different category of trust from reputation alone.

The investors who sustained their edge built networks intentionally — not to accumulate contacts, but to maintain ongoing access to the best technical talent and deal flow in their target sectors. Masayoshi Son built government relationships in Japan that gave SoftBank a regulatory and distribution advantage in mobile and internet infrastructure. His $100B Vision Fund secured a critical $45B commitment from Saudi Arabia’s Public Investment Fund — a relationship that predated the fund.

2. Media — Publishing Creates Inbound

The mechanism has shifted across three eras. In the 1980s–90s, it was trade press: investors quoted regularly in publications that founders read became the first call. In the 2000s–10s, it became the blog — from being cited in someone else’s piece to owning a point of view directly. In the 2020s, it’s the media house.

Bill Gurley / Benchmark → Uber — Travis Kalanick found Gurley through the blog

Gurley published detailed frameworks on marketplace dynamics with enough specificity that founders could run their own business through the framework. Kalanick read it and reached out directly. The blog didn’t generate warm introductions — it generated founders who had already self-selected against the framework.

Source: Kalanick in multiple interviews; Bill Gurley, “Above the Crowd”

a16z describes itself as a media company that monetises through venture capital — it runs one of the largest technology podcast networks, publishes Future.com, and employs a full editorial team. 20VC closed a $400M third fund in October 2024, raised in approximately four months. The pattern — build a media presence with a specific point of view, convert the audience into a deal pipeline — is now a recognised playbook in venture capital.

3. Going Deep Into the Domain

Domain immersion means spending more time in a specific category than any other investor, before the category is obvious. The result is a “prepared mind” — a pre-built framework against which every incoming company is evaluated instantly. The decision is fast not because the investor moves quickly, but because the intellectual work was done before the company arrived.

When you’re known to think more carefully about a specific category than anyone else, founders route deals to you without being asked — not for your capital, but because a conversation with you produces insight they can use. The compounding is real. It explains why the best investors in a given wave — Gurley in marketplaces, Wilson in open-protocol networks, Andreessen in developer tools — see a disproportionate share of the best companies.

Jim Breyer / Accel → Facebook ($12.7M Series A, 2005; returned entire Fund IX)

Accel had been tracking social networking before Facebook pitched and had formed a specific view of what the category leader would look like: a network built on real identity, with verified users, starting from a dense community before expanding. When Facebook pitched, Breyer moved quickly against internal scepticism because the intellectual work was already done.

Source: TechCrunch, “Accel Partners’ Extraordinary 2005 Fund IX,” November 2010

How They Pick the Category Leader

Six signals for category leadership appear consistently. These are concrete data points, not impressionistic assessments of founder quality.

1. Trusted Referral

A trusted referral means three things together: someone with direct, hands-on knowledge of the product; who knows you well enough to judge what you look for; and who has your best interest at heart. All three together is what makes it signal rather than noise. David Cheriton had personally invested in Google alongside Andy Bechtolsheim and introduced both Moritz and Doerr — not because he was promoting it, but because he thought they’d regret missing it.

2. Network-Verified Signals on Competitive Position

Category leadership is confirmed not just through the company’s own metrics but through third-party signals: traffic growth data, app store rankings, API usage patterns. Botha’s YouTube investment memo shows competitive position assessed using traffic growth data before the term sheet was signed.

3. Proven Leadership in a Constrained Geography

The category leader dominates one geography before any competitor does. Geographic concentration makes competitive dynamics, unit economics, and retention visible at real scale — without the noise of premature national expansion.

• Uber: San Francisco dominance in 2010–11 proved end-to-end model economics before national expansion.

• Airbnb: New York City host density and guest repeat rates in 2010 validated the two-sided marketplace.

• OpenTable: San Francisco restaurant density was the proving ground before national rollout.

4. A Structural Differentiation That Cannot Be Quickly Replicated

The category leader has a product advantage competitors cannot close by increasing budget or headcount.

• Facebook: Real identity at the architecture level. Myspace couldn’t replicate this without asking 100M pseudonymous users to verify themselves.

• Google: PageRank required re-crawling and re-indexing the entire web with a different algorithm. Altavista couldn’t flip a switch.

• Stripe: Developer-first API design meant adoption was bottom-up through engineering teams. Incumbents’ sales-led motion couldn’t replicate this without a structural rebuild.

5. Organic Growth With Zero Paid Acquisition

Zero paid acquisition at meaningful scale is more than a cost metric — it means the product is spreading through genuine value delivery.

Marc Andreessen / a16z → GitHub ($100M Series A, July 2012; Microsoft acquired for $7.5B, October 2018)

GitHub grew from launch in 2008 to 1 million users entirely through word of mouth — no sales team, no marketing budget. Engineers added GitHub profiles to resumes. Recruiters asked for GitHub handles.

Source: TechCrunch, July 2012; Microsoft press release, June 2018

6. Network Effects as a Structural Moat

The category leader has a product that gets more valuable with each additional user. Unlike a technology advantage (which a well-funded competitor can eventually match), network effects improve over time rather than decay. Two-sided marketplaces, communication platforms, and protocol-layer products all exhibit this.

Fred Wilson / USV → Coinbase (Series B, 2013)

Wilson’s thesis: open protocols (Bitcoin, Ethereum) would eventually require a trusted consumer on-ramp. The company that built the most trusted interface to the protocol would capture significant value from the protocol’s network effects. USV invested in Coinbase’s Series B in 2013 — four years before the 2017 crypto boom.

Source: USV thesis, usv.com; Coinbase direct listing April 2021

The Replicable Playbooks

The investors who built the most durable track records operationalised their frameworks into repeatable playbooks — the same logic applied across categories and cycles. Four stand out for specificity and longevity.

Playbook 1 — Bill Gurley: Marketplace Supply Utilisation

What share of supply is idle? What’s the structural reason for the mismatch? Does the platform solve the matching problem in a way that compounds with scale?

• Uber: Idle vehicle capacity in major cities; real-time mobile matching as the fix. Led $11M Series A, February 2011.

• OpenTable: Restaurant table inventory wasted by phone-based reservation friction. Series B 2000; acquired by Priceline for $2.6B, 2014.

• Grubhub: Restaurant kitchen capacity idle during off-peak hours. Just Eat Takeaway acquisition for $7.3B, 2021.

The framework was published openly. The founders who called Gurley were the founders whose businesses already passed the first test.

Playbook 2 — Masayoshi Son: Penetration Curve Time Machine

Track technology penetration rates across geographies. Identify markets running 18–36 months behind the leading market. Invest in the proven model at the point where the outcome is statistically likely but not yet priced in by local investors.

• Yahoo Japan: Japan’s internet penetration in the mid-1990s tracked several years behind the US. SoftBank co-founded Yahoo Japan in 1996.

• Alibaba: China’s e-commerce penetration in 2000 tracking behind the US. SoftBank’s $20M for 34% generated approximately $150B at peak value.

Playbook 3 — Don Valentine: Cost Curve Economics

Map the cost decline curve of enabling infrastructure. Identify products that become viable when cost crosses specific thresholds. Invest while the cost is still falling — the window closes when cost stabilises and the category becomes obvious.

• Apple (1977): Microprocessor cost had dropped to the level making a sub-$1,500 consumer computer viable for the first time. Sequoia invested in a $600K round for ~10%; Apple IPO’d at $1.78B market cap.

• Cisco (1987): Networking silicon had fallen below the cost point where enterprise routers could be sold profitably outside the Fortune 500. Sequoia invested $2.5M for ~30%.

Playbook 4 — Elad Gil: Infrastructure Layer Cluster Investing

When a new infrastructure layer achieves developer adoption — AWS compute, iOS, Ethereum — a cluster of enabling companies must be built before the application layer can scale. Invest in the enabling companies first. Infrastructure appreciation leads application appreciation by 2–3 years.

• Stripe: Payments infrastructure for the web commerce layer being built on AWS.

• Airbnb: Trust and payment infrastructure for peer-to-peer marketplace transactions.

• Coinbase: Consumer on-ramp to the crypto protocol layer before consumer applications existed at scale.

The Through Line

The investors on this list who built 20-year track records share one property: they turned their judgement into a system. Tailwind identification is systematic. Sourcing follows the thesis. Category leader identification follows concrete, repeatable signals.

Last month I made a simple mistake. I booked two hotel rooms instead of one for my parents while they were travelling. I called to cancel. Explained everything. Put on hold. Transferred. Explained again. Told it was being processed.

I called back the next day. New agent. No record of my first call. I explained everything a third time. By then, my parents had already checked into the hotel.

Not one person knew what the previous person had said. Every interaction started from zero — with someone who had no memory of why I was there and no real stake in fixing it.

Now imagine this instead.

I call once. The same voice picks up — already aware that I accidentally booked two rooms twenty minutes ago. It cancels the right booking, confirms the refund, and messages my parents so they don’t worry at check-in. Two days later, it follows up: “Just checking — the refund should have reflected by now. How was your parents stay?”

I did not just get my problem solved. I felt taken care of. That is not customer service. That is a relationship.

— — —

ACT 2 IS AROUND THE CORNER

Act 1 is not finished — it is still scaling, still being sold, still generating returns. But it was built around a specific thesis: automation. Reduce cost. Handle repetitive workflows. Prove that machines could talk.

That thesis worked for two types of companies.

Enterprise Voice AI. India started of with AI services companies, delivering custom made voice ai agents for enterprises with a FDE model — primarily for outbound BFSI workflows: collections, loan repayments, lead qualification, insurance renewals. The economics made sense. Low stakes to try; the quality bar was easier to clear because expectations for outbound calls were already low.

India moved outbound-first for a structural reason. Inbound felt risky — a visible failure when a customer called for help. Outbound was forgiving. You were interrupting someone who did not ask to speak to you, so a slightly awkward AI was no worse than a slightly awkward human.

That window is narrowing. Regulators tightened outbound regulations in early 2025,, Connect rates have compressed sharply — spam labelling, call-blocking apps, and DND enforcement have taken a real toll. The differentiator is no longer who calls more. It is who can make the call worth answering.

In the US, more horizontal voice agent companies came up with inbound customer support being primarily the stronger wedge. Support is structurally broken — long waits, staffing shortages, inconsistent quality. Voice AI increasingly handles predictable queries well. Critically, inbound callers have already decided they want help. Customer intent is high, the quality bar is easier to clear, and the relationship opportunity is much richer.

Vertical Voice AI. The second category placed a different bet: if voice becomes good enough, entirely new experiences emerge. Language learning, healthcare navigation, companionship.

Some of these are working. Language learning is the clearest success — demonstrated measurable outcomes from conversational practice at scale. The use case fits well: structured practice, pronunciation feedback, repetition without embarrassment.

Other categories — companionship, healthcare navigation, financial guidance — feel earlier. Not because demand is missing. Because they require something harder than automation: trust, continuity, emotional nuance.

And that is precisely where Act 1 runs out of road.

Act 2 does not replace Act 1. It expands the category. It begins when Voice AI becomes dependable enough to build relationships, not just handle transactions. That unlocks consultative sales, healthcare navigation, personal assistance, education, companionship. The thread connecting all of them is the same thing missing from my hotel experience — a presence that knows you, shows up consistently, and follows through.

The question is not whether that threshold arrives. It is when. And what has to be true for it to happen, there is some evidence that this is already in motion but when would it truly it scale?

To answer that, I want to map the six dimensions of a relationship — and be honest about where Voice AI stands on each one today.

— — —

THE 6 PILLARS — A MAP TO ACT 2

These are not checkboxes. They are a map.

The pillars where AI leads are compounding. The pillars where it lags are the bottlenecks. When enough bottlenecks clear, Act 2 starts.

— — —

1. PRESENCE — AI Already Leads

Think of the last time something went wrong unexpectedly — a flight cancelled, a payment declined, a family health scare. You probably called someone. And if they picked up, that single moment likely built more trust than ten routine interactions.

For humans, there are limits: business hours, hold queues, shift changes, tired agents at the end of long days.

For Voice AI, this pillar is already crossed.

Good deployments answer instantly, operate 24/7, and handle thousands of simultaneous interactions without fatigue or mood variation. The AI that picks up at 2 a.m. while you are stressed — patient, calm, and fully attentive — creates a relationship moment that most traditional support organisations rarely even attempt.

This is a structural advantage that widens over time, not narrows. And it is the clearest signal that Act 2 is not hypothetical — at least one dimension of the relationship equation has already flipped.

— — —

2. UNDERSTANDING — The Biggest Blocker

A customer calls and says: “I’ve been dealing with this for three weeks.”

Those seven words contain frustration, exhaustion, a history of failed attempts, and sometimes an implicit threat to leave. They are not a support ticket. They are a signal that someone wants to feel heard.

Understanding has two layers, and AI’s story on each is very different.

WHAT YOU SAY:

In clean audio with clear speech and standard vocabulary, the best systems are genuinely good. But the real world is not a benchmark environment.

Noisy call centres. Thick accents. Medical or legal terminology. City names, product names, account numbers. People talking over each other. Multilingual switching mid-call.

In these conditions — which are the conditions of most real contact centres — error rates rise substantially. Turn-taking breaks often enough in production to feel noticeable. Speaker diarization fails often enough that transferred calls routinely lose context.

The honest framing: AI transcription is good enough for structured, predictable interactions. In messy, real-world speech, it is still clearly below human performance.

WHAT YOU MEAN:

This is where the gap widens.

Hearing “I’m just checking on my order” and understanding it as “I’m anxious this gift won’t arrive before the wedding” is a different cognitive act. Or recognising that “I just wanted an update” sometimes means “This is the third time I’m calling and I’m losing patience.”

Humans do this subconsciously and well. AI does it inconsistently — improving with audio-native models that preserve tone and hesitation rather than discarding it in transcription, but not yet reliable across the range of human emotional expression.

Understanding is probably the single biggest blocker to Act 2. It is where the experience most often breaks down. And it is the pillar that will determine, more than any other, when the inflection point actually arrives.

— — —

3. CAPABILITY — AI Already Leads on Structured Workflows

Unless you are a therapist or a close friend, building a relationship with someone who listens beautifully but cannot act is difficult. Empathy without action becomes frustrating quickly.

What most people underestimate: AI is not just faster than humans here. It is structurally different.

A human agent logs into a CRM, pulls the account, finds the transaction, navigates the refund workflow, and updates the record. That takes minutes, involves multiple systems, and introduces error at each step. An AI agent with proper integrations executes in parallel — no switching, no hold time, confirmation sent before the call ends.

For defined workflows — appointment booking, payment processing, insurance intake, order tracking — AI already executes faster and more reliably than most human agents.

Where it still breaks: edge cases. When a customer says “actually wait, I want to change that” and the system is two API calls deep into a transaction, recovery logic gets complicated fast. Many systems freeze, fail silently, or hand off without explanation.

The winning systems will not have the best models. They will have the deepest integrations and the best recovery logic when things go wrong.

— — —

4. CONNECTION — The Gap Is Narrowing

Two things live inside this pillar: conversational feel and emotional adaptation.

CONVERSATIONAL FEEL:

The shift from Speech to Text to Model to Voice to continuous audio-native pipelines — now in production with OpenAI’s GPT-Realtime-2 and Gemini 2.5 Flash Native Audio — removes the latency from treating transcription and synthesis as separate steps. The model hears you rather than reads you. The best systems now respond well under 300 milliseconds. At these speeds, the mechanical pause that gives AI away starts to disappear.

Voice quality has closed considerably. For short interactions, the gap that would have been obvious two years ago is not obvious today.

EMOTIONAL ADAPTATION:

This is where things get harder.

The difference between saying “I understand your concern” in a flat tone and slowing down to say “That sounds genuinely frustrating — let me fix this right now” is enormous. One is a script. The other feels like care.

Audio-native models preserve tone, pacing, and hesitation that transcription-based systems lose. Companies like Hume are building models that understand the emotional weight of what they are saying — not just how to sound appropriate, but why. These are meaningful advances.

But the important shift in the Connection story is this: sounding human is increasingly commoditised at the frontier. The next race is knowing when to be warm, when to be concise, and when to slow down. That is a harder problem than naturalness — and a more defensible one.

— — —

5. CONTINUITY — AI’s Most Underestimated Advantage

You call your doctor’s office. They already know you came in last week, what was discussed, and they ask how the medication is working — before you say a word.

That is not service. That is a relationship.

Here is what most people miss: we assume humans are good at continuity. In practice, human service organisations are structurally bad at it.

Different agents take your call each time. Context disappears at transfer. Shift changes reset institutional memory. Staff turns over. The person who handled your case last month has left — and whatever they knew about your preferences, your history, your frustration went with them. What most organisations deliver is not continuity. It is the occasional illusion of it, when you get lucky with the same agent twice.

AI has no such problems within its design constraints. The same identity shows up every time. Context from the last call is retrieved before this call begins. The tone, the persona, the standards — consistent across call 1 and call 10,000. No bad days. No shift variation.

The leading platforms now use vector retrieval to surface relevant history and trigger engines for proactive outreach — renewal nudges, post-resolution check-ins. In production.

The nuance: total memory is not the goal. Helpful memory is. A system that retrieves everything creates noise; one that retrieves the right things creates trust.

Being briefed on a file is still different from being remembered. But AI is structurally better positioned for continuity than human service organisations have ever been — and that gap widens as memory curation improves.

— — —

6. TRUST & RELIABILITY — The Hidden Foundation

Every other pillar sits on top of this one. Remove trust and the whole structure collapses.

A voice agent tells a customer their loan is approved. The customer acts on it. The agent was wrong. One mistake after six months of good interactions — and suddenly everything feels fragile.

Reliability has three dimensions, and AI’s story on each is very different.

BEHAVIOURAL CONSISTENCY:

AI already has a clear advantage. Same standards. Same patience. Same persona. No cutting corners on a Friday afternoon.

Humans are inconsistent in ways that erode trust slowly and invisibly. Software tends not to be.

FACTUAL RELIABILITY:

This remains deeply unsolved. Models still occasionally say incorrect things with genuine confidence. In high-stakes environments — healthcare, insurance, financial guidance — occasional mistakes are not acceptable, no matter how smooth the voice sounds.

The best systems increasingly say “I’m not fully sure — let me verify that” rather than confabulating a confident answer. Trust is not built because someone never makes mistakes. It is built because they behave responsibly when uncertainty appears.

The winning approach: deep integration with source-of-truth systems — CRM, EHR, policy documents, inventory — constrained generation grounded in verifiable facts. The moat is trusted data access, not model quality.

ENTERPRISE-GRADE RELIABILITY:

This is the dimension the industry under-discusses. And it is probably the biggest gap between demos and production.

Enterprises do not refuse to deploy because the AI sounds slightly robotic. They refuse because it fails unpredictably. An API times out mid-transaction. A customer changes their mind three steps into a workflow and the system cannot recover. An ambiguous request sends the orchestration layer into a loop. A retry creates a duplicate booking.

In production at scale, these failure modes compound. The systems that win enterprise are not the ones with the most impressive demos. They are the ones that handle the conversations that do not go as expected — gracefully, silently, without the customer noticing.

The breakthrough moment is quiet. It happens when users stop noticing reliability at all — when “wow, this AI actually handled it” becomes simply “of course it did.” That invisibility is when Act 2 is truly underway.

— — —

What Determines When Act 2 Arrives

The six pillars show where AI already leads, where it lags, and what the bottlenecks are. Two of them — understanding and enterprise-grade reliability — most directly throttle the transition. The question is what actually moves them.

Five things are worth watching.

1. Audio-native models hitting production quality

The move to continuous audio-native pipelines is the most consequential architectural shift currently underway. The immediate unlock is not emotional intelligence — that takes longer than the demos suggest. It is smoothness: lower latency, better interruption handling, fewer of the mechanical pauses that make conversation feel artificial.

More importantly, models that hear rather than read preserve acoustic cues — pacing, pitch, hesitation — that transcription permanently discards. That directly addresses understanding, which is the biggest bottleneck to Act 2. The architecture shift and the blocker are the same problem.

2. Vertical models cracking real-world speech

Benchmarks are clean. Production is not.

The models that matter are not the ones performing best on standardised tests. They are the ones trained on what real calls actually sound like: a customer in a moving car, switching between Hindi and English, trying to pronounce a locality name that appears in no training corpus.

A model trained on millions of real BFSI conversations understands that “ECS bounce” means something different from “technical error.” A model trained on healthcare intake knows when “a little discomfort” might need escalating. The gap between generic and vertical is wider in understanding than in any other pillar. This is the work that is least visible and most important.

3. Reliability becomes the competitive moat

The industry has been over-indexed on naturalness. The race has moved.

Sounding human is increasingly commoditised at the frontier. Companies like ElevenLabs, and Cartesia are making Voice AI more expressive — and that matters. But the enterprise deals will be won by whoever also makes deployment boring-reliable: orchestration that does not break when reality deviates from the happy path, retry logic invisible to the caller, edge-case handling that does not require human escalation.

A conversation with an AI the customer trusts completely — because it has never failed them, never given wrong information, never lost context mid-transfer — is worth more than a conversation that sounds ten percent more warm. The moat is not voice quality. It is operational reliability at scale.

4. Sesame — the company betting on relationship as the product

Most Voice AI companies optimise for task completion. Sesame is optimising for something different: relationship continuity as the product itself.

Founded by Brendan Iribe — co-founder of Oculus — and backed by Sequoia, their thesis is that an AI companion should be someone you want to return to, not just someone who handles your request. Their AI characters generated millions of minutes of conversation shortly after launch — not because users needed to, but because they wanted to.

Their research identified the precise frontier: without conversation history, their model is indistinguishable from human speech. With context, human speech still wins. They published this because it names exactly what they are building toward.

The hardware vision — smart glasses, ambient and always-on — turns switching costs from functional to emotional. That is a categorically different business than anything Act 1 produced. If it works at scale, it is proof that relationship-layer AI is a real category, not a feature.

5. Grounding and deployment reliability at scale

Whoever solves this wins enterprise. Not model quality.

The system grounded in every source of truth — CRM, EHR, policy documents, inventory — constrained to only say what it can verify. And the operational layer that does not break when reality deviates: invisible retries, graceful recovery, workflows that handle the 20% of conversations that were not in the design brief.

RAG-based grounding is already reducing hallucination rates meaningfully. The next frontier is confidence calibration at scale — systems that know what they do not know, every time.

The moat is not the model. It is trusted, reliable deployment at scale.

— — —

What’s next

Act 1 is still running. Still generating value, still scaling, still being sold. It was the right first chapter.

Act 2 is where it gets 10x better, there is some evidence this is already in motion — Indian Voice AI services companies are seeing new set of clients and use cases beyond BFSI — sales and collections, most of them are taking defensible niches than remaining a generic custom built voice agent company. As the model gets better, we would some of the use cases like companionship, therapy, coaching etc take off

The interface layer of the next decade might be a voice. Because it is more reliable, more consistent, more present, and increasingly better at understanding than the human on the other end of the line.

My hotel story was a failure of relationship. Three different agents. No memory. No continuity. No follow-through.

The company that fixes that experience — reliably, at scale, across every interaction — does not just reduce support costs.

It earns the customer. And the company that earns relationships at scale may quietly own a category that does not yet have a name.

After about two hours of chatting with founders at the mixer, I was heading out, and that’s when I bumped into this founder building his company in language learning. Once he finished his elevator pitch, I asked him, “This space is getting super competitive, right? Any thoughts on how are you planning to differentiate?”

That threw him off; he said “ Yaar pichle 2 ghante mai sabne mujhe yehi pucha hai, tum VC log ye moat ko leke itna obsessed kyu ho ? mai apko batao moat voat kuch nahi hota, bus execution se dhande chalti hai” (for the past 2 hours every single VC has asked me the same, Why is every VC obsessed about moats? There is no such thing as a moat It is the execution that matters).

I thought to myself, I haven’t even brought up moats yet; I was just asking about how you are planning to get to PMF. I can understand his frustration, but my situation was no different. A large part of the founders I met had a me-too product with little to no unique insights about the market.

I also noticed during the event that people were using the word moat with varied meanings — some had coined terms like short-term moat, Intangible moat, Soft moat, etc. — and each had their own definitions of what it meant.

In this blog, I will debug the question of “Are moats real?”. Or is it some VC-invented term? What do investors mean when they ask about your competitive strategy, and how are we seeing founders think about it?

As a founder, once you have defined where you want to get to (your vision), there are two outstanding questions — how do you plan to get there, and how do you stay there (moat)? Let’s talk about each one by one :

Section 1: How Will You Get There?

Some founders call this a short-term moat, but there is hardly any moat at the seed stage. What you need is a plan of action to firefight your way to PMF first and Critical mass / Scale second.

Executing being hard is a given; there is no way around that, but do you have a plan of action that comes from differentiated insight about the customer?

To make this more relatable, let me give you an example. Say you are hiring for a head of marketing for your startup, and you ask them, “So how would you approach reducing CAC?: and they say, “i’m gonna execute very hard. “ For sure, but do you have an initial plan of action that is insightful and differentiated from the other candidates I have spoken with? i am sure there is a reasonable chance this might change over time, but do you have a well-researched day 1 hypothesis?

We are talking about fighting your way broadly with two types of players — incumbents and startups. Let’s discuss each of them one by one.

1.1 Differentiating from other startups

The first step in differentiating from the competition is understanding where you stand in the spectrum of competitive intensity.

If you stand in low, competitive spending, you ideally want to spend 80% of your time thinking (and explaining to investors) why others do not see this. Or do they see this, but they can’t / don’t want to go after it? A vast market with a low supply of founders is not rare. In my last blog, I explored 2–3 reasons this happens. Please refer to it here.

On the other hand, if you are at the other end of the spectrum, then you want to make sure you have a very well-thought-through plan of action to compete because the more competitive intensity, the less time to figure PMF before everyone writes you off, including your customers

when we speak with founders in these categories, we hear two not-so-concrete answers :

Yes, there’s a lot of competition, but we’ll execute harder”

Our competitors are doing X, but we’re doing Y. (eg: yes, there are lot of language learning companies, and most of them are done on adults while we are focussed on kids)

I have already discussed why the first answer isn’t good. On the second one, they try to say, “Hey, we are doing Y, and they are doing X. It’s not direct competition if you think of it.” If you are in a similar subspace, you will get head-on very soon, so why will doing Y give you leverage? (Why is concentrating on kids a better strategy to get to PMF?).

For example, you could say something like

“The critical problem in English fluency is getting the confidence to speak in front of a bunch of people, which is why community learning is essential to drive outcomes; learners need to get used to speaking in front of people. Most of the players in the market are concentrated on 1:1 learning with an AI tutor. We are focused on a community platform powered by AI, and therefore, we have 10x better engagement than other players in the market, and the community will eventually become a Moat.”

The key is to have unique insights about the user that differentiate them from the rest — it’s a compounding of such insights that will help you win the fight.

1.2 How do you compete with incumbents?

Startups have always competed with incumbents on two levers: Agility and Niche focus. Incumbents can never move fast and break things or focus extensively on a niche set of uses, which is the curse of scale.

In the post-Gen AI world, incumbents are more agile than ever, and startups seem to be competing for a global set of users from day 1. This means you must pull harder on the agile lever and define the niche set of users you are going after.

There are three kinds of players we are talking about here :

Legacy players

• (eg: Banks in financial services or hospitals in healthcare) — Both founders and VCs are least worried about these guesses, especially in sectors where tech adoption has been slow; these are slow-moving mammoths — just the pace at which startups can move provides huge leverage.

Big tech

• The Open AIs and Googles of the world — at one end, VCs have been worried about Open AI and the other 5 big model players running over so-called “Thin wrappers” (startups with no proprietary data/tech innovation moat). At the other end are founders who believe there is no such thing as a thin wrapper :

• At a recent AI event in HSR, while answering a techie’s question around the same, Rohit Agarwal (Founder of Portkey) said, “Generally, AI founders get very triggered when you mention the term Thin wrapper; we used to call Freshworks a database wrapper. That is what it is; the value added was at the workflow/UI level. So having a technical IP is not necessary to create moat).”

• I couldn’t agree with him more; if you think of it all, SaaS are database wrappers — there is little to no technical moat, and they do well by understanding a user and their pain points well and building custom workflows / UI for them.

• But here is the caveat — it cannot be a generic use case for a generic audience.

• Andrew Ng, in his lecture on opportunities of AI, talked about “Fads along the way” taking the example of the Lensa AI app — which saw its revenue rise and fall quickly because it was a generic product for a generic set of users.

• So the question to ask is, am it solving a real problem for a specific set of users with the workflows that I understand better than anyone else ? Are they willing to pay for a workflow layer?

Scaled-up startups in the domain

• (eg: Practo in healthtech or Razor-pay in fintech) — This is a little tricky. These guys have been more agile than ever, and unless what you are building needs a different set of capabilities than what these guys currently have, they are bound to jump on you.

• The capability you need in AI language learning is pedagogy/ user engagement — this is clearly what Duolingo of the world has; while most startups have claimed that conversational language learning is something they are not catering to now it was an obvious market for them to tap into and Duolingo has just entered the market, even players like Unacademy pounced on the market — this doesn’t mean there won’t be any startup winners in the space, but everybody in the space is in for huge bloodshed.

• But note if you have a niche focus on, say, Indian youth trying to learn English for better employment opportunities, that will help you build more customized features around it, build a solid set of users who love you, and then expand to the next set of users — if you are thinking along the lines, please do highlight this in your pitch deck as well.

Section 2: How Will You Stay There? (Or what moat are you building ?)

The above meme shared by a founder summarises how some founders feel about MOAT — as if it’s a VC-invented term that means nothing in the real world, but that is not the case. Let me explain.

When VCs try to assess if the following business model has, they ask whether competing with your business becomes increasingly difficult once you get a critical mass.

It’s not just the VCs; you would see any seasoned second-time founders think about it quite early on.

in a recent interview, Sajith Pai (Partner at Blume) asked Ranjeet (founder of Pratilipi) what was his criteria of choosing a space/business model to build, and this is what he said :

Well, you would say all this sounds great in theory, but -

Aren’t all businesses supposed to get disrupted eventually?

It’s tempting to dismiss moats because we all know that disruption is inevitable. Why build a moat when something new could always come along? The answer is simple: while no moat is permanent, they give you a point-in-time advantage. And that advantage buys you time to innovate further or extend your lead. Moats give you breathing room to sustain your business so you’re not constantly firefighting against competitors. The best moats grow as you scale your business (think about Google’s AdWords biz).

Aren’t culture/values/ability to innovate, etc, the only moat?

This is a 40-year-old debate in corporate strategy, on which Harvard Business Review has written a great article that concludes that this debate will not end any time soon.

What makes and breaks companies is culture/value/capability to innovate, etc. (some refer to this as intangible or soft moats). Those are genuinely nonreplicable and sit at the heart of the organization. However, it’s essential to understand whether you are using these to build a business that gets increasingly difficult to compete with as it scales or if you are using them to just firefight competitors on a day-to-day basis.

Ashish Mohapatra (founder of Ofbussiness), in this podcast (in the section by should Business be Complex), explains why he doesn’t like to depend on intangible moats. He says

“While intangible moats tend to be generally hard to sustain or tend to break at scale, they can also make you overly dependent on execution intensity, which is not a place he wants to be.”

I would add that if culture/values, etc, are the only moats that bring in huge keyman-ship risk — since most of it would be a reflection of the person leading the organization (which is not a place any investor/founder wants to be in, especially at the scaled up stage)

2.1 Common Types of Moats We See These Days

Let’s break down some common types of moats and how to think about them:

Product/Experience moat:

• In a blog titled “Generative AI companies have MOATs eventually,” Brandon Glenklen argues that all GEN AI companies could eventually build moats based on iterative improvements they make to the product based on accumulated user feedback over some time; this is precisely what we discussed the section on competing with Open AI, while I agree with him mostly there are two caveats I want to call out:

• This will be significantly more problematic if you are in the high-competition zone because the time to get to PMF and build the moat layer becomes significantly lower.

• Ensure it’s not a generic use case that can be served well with the next-generation LLM or, eventually, AGI.

Distribution moat:

• This has to be one of the moats; you must think of defendable ways of acquiring consumers, given how performance marketing costs have increased. Ideally, you need more than just a distribution moat because that has now become table stakes.

Brand Moat:

• Is a large portion of your customer base likely from word-of-mouth? This is critical in industries where personal recommendations, such as healthcare or SaaS, drive product discovery. However, this may not apply to all industries. For instance, in AI-powered English language learning, users are more likely to discover products through ads rather than seeking referrals or advice from peers..

Data moat:

• The founder’s argument goes like this: We have this initial set of data, which we used to make this recommendation engine and get some customers, giving us more data. More data means better recommendations mean more customers, which again means more data, but this is hardly true. We would have seen 100+ startups in the past month or so — apart from some niche areas like healthcare, we hardly find any data moat. The following conditions should be valid for data to be a moat.

• An extensive set of Proprietary Data should significantly improve the product/service. — More data is not always better; we have observed diminishing returns in the context of AI scaling. Therefore, understanding to what level data can enhance your product for the end users is super important below is an example of a use case in which the leverage of data is low:

• The Data moat should not erode over time — as time passes, the data shouldn’t go stale. Also, new data acquisition shouldn’t get costlier (while incremental value reduces)

• The startup should have proprietary data sources/strategies that are scalable (asset substantial enough for large-scale model training), continuous (data can be resampled over the period), diverse (adequately reflect real-world scenarios), and legally compliant.

And more often than not, the 1st condition isn’t proper, let alone 2 and 3.

At an applied Gen AI level, the most significant outcomes will come from business model innovation that will leverage the unique capabilities of Gen AI.

If you’re a founder building in Gen AI and thinking about how to differentiate and scale, I’d love to hear from you. Please reach out to me on Linkedin. I end up replaying all my inbounds, or you can write to me at sauravg@capria.vc. Let’s explore how we can collaborate to build enduring companies.

Why are ideas important, when make or break lies in execution?

If you are an operator this might be the first question that pops up in your mind, Ashneer Grover once said in a public forum

“Ideas India mai 2.5 rupay per kilo bhikta hai, execution is everything”

If you end up listening to his full speech, you will realize what he is saying is “Ideas are not a MOAT in itself” which I do agree with — of course, great ideas have no value if it’s not executed well, but I am sure he would agree with me when I say that getting the where to play right would have 10x more impact on your chances of success and size of outcomes than getting the how to play right.

Recently, Toshan Tamhane — Global Chief Operating Officer at UPL, Ex-Senior Partner McKinsey explained this beautifully in an episode in the barbershop. He says :

“Great CEOs get the where to play right, I did an analysis on sports back in the day

At that time the number one player in tennis was Roger Federer and his income was $50M, and the no 10th player was Juan del Potro, his income was $15M. When you then look at the next 5 racket sports, the average income of the top 1 player is just $900k.

You see the delta you can create with hard work, commitment, and execution is 3.2x (delt between Juan and Roger) whereas picking the right areas to play is 15x (delta between Juan and the top 1 player in other sports)”

I am sure this is the case with cricket vs. other sports in India currently and the same plays out in businesses as well.

So it’s not about picking one idea or business model, but it’s about picking an idea space — a set of closely related ideas in an area that is ripe for disruption.

Y Combinator is a huge proponent of picking the right idea spaces, they say

“Over the last 10 years if you started in fintech infra, vertical SaaS for enterprise your chances of building a billion dollar business was astonishingly high whereas if you are building in consumer hardware or social networks or ad tech success rates were really thin”

So what makes a great idea space?

I think, Two big buckets of what make a great idea space is either it’s an existing large market with a low supply of founders or it’s a small but growing market that might soon be a large enough market, lets talk about the two one by one.

1. Large markets, with a low supply of founders

Wait, but why on earth will there be a large market with a low supply of founders? to understand this we need to understand Dalton’s theory of supply — demand equation. Dalton says The classic supply-demand equation applies to startups as well. Demand is ideas and Supply is founders

There are certain ideas that get a large number of founders mostly these are sexy businesses and a bunch of them are what Paul Graham would call “Made up startup ideas” here is a write-up from Paul’s blog summaries how to avoid this trap

To know more on how to spot such ideas Check out Paul’s blog here, or check out my notes on his blog here.

There are a bunch of ideas where the supply of founder are less and a large market could be hiding in these idea spaces, here are two main reasons why founders end up missing it :

Spaces that are hard to get started — Paul Graham calls this “Schlep Blindness” The biggest example is stripe — For over a decade, every hacker who’d ever had to process payments online knew how painful the experience was. Thousands of people must have known about this problem but nobody went behind this huge opportunity since they would have to make deals with banks, move money, deal with fraud, and comply with complex regulations until Stripe came and solved it.

This is one of the key reasons why venture studios like 2070 Health exist, Healthcare as an industry is hard to build in. You can not move fast and break things in healthcare, it’s a highly regulated and relationship-driven industry and thus there exists a large amount of white spaces, which we are unlocking one by one.

To know more on how to spot such ideas Check out Paul’s blog on schlep blindness here, or check out my notes on his blog here.

Space that’s unsexy: One of my favorite founders Ashish Mohapatra of ofbusiness has said across multiple public forums that Unsexy businesses have huge profit pools to be captured, he went on to define three questions to ask to find if this is an unsexy business :

One, it does not have to be in metros.

It should be something a common man can’t understand. He understands food delivery. Does he understand steel purchase?

Thirdly there has to be hard work and education in it. To lend, you have to get the customer, underwrite it (the loan), then you have to convert it, then you have to service it, then you have to collect the money. The more execution there is, the tougher it is to replicate.

Another founder who has been pretty vocal about how unsexy businesses could have large whitespaces is Adit palicha, founder of zepto

2. Small but fast-growing markets

When there is a small but fast-growing market and in my opinion, this happens when there is a fundamental change in :

• Consumer behavior creates a need or increases receptiveness of certain solutions: for example, Airbnb came out during the recession when homeowners were looking for more income or Zepto came out around COVID when people were struggling to get groceries.

• Tech/infra/regulation that makes a new solution, GTM, or cost structures possible: for example, Zynga the gaming company came out when Facebook had just let third-party developers make apps and games. YouTube came out when broadband penetration had crossed 50% in the country and Adobe Flash had made it much easier to watch videos online, similarly, a large amount of internet startup ideas suddenly became viable when Jio fundamentally changed the data infrastructure in India.

This is the reason why a large number of highly successful companies are founded during a recession or any such major events, for example, Covid has made remarkable changes in consumer behavior that make a lot more business possible.

This is also the reason Why, the “Why now” slide is important in your pitch deck, what investors are looking for is “Are there strong tailwinds in the market that this idea could ride on”

To understand the importance of timing, I highly recommend this TED talk that summarizes this pretty well.

So how have founders usually got good ideas?

I see two broad ways in which founders have reached the ideas :

• A Side project that turned into a startup: a hacker trying to create a cool toy/project or a user just trying to solve his/her pain points, Companies such as Facebook, Instagram, slack, and Spanx, all started like this. Note these folks were already on the hunt for ideas and were ready to all in once he/she saw early signs of user love. Reid Hoffman in his podcast Masters of scale has talked about this briefly, in his episode with Sara Blakely (founder of Spanx) he touched upon “how to constantly be in the hunt for ideas and how to seize it when you see early signs”

• A consulting style due diligence: In a recent episode of Barbershop Rahul Chaudhary talked about how he and his cofounder did a consulting style DD of 34 ideas filtered through 7 in-depth evaluations in 3 months, similar was the case with other founders such as Karan Bajaj, Abhiraj bhal, Shantanu Deshpande, Ashish Mohapatra etc. Yes, a structured thesis-driven approach to identifying a market does work.

Concluding thoughts

In conclusion, your choice of idea space would have a huge impact on your chances of success and the size of outcomes than how well you execute. In order to pick the right idea space, look out for 1. Large markets with a low supply of founders — these are usually in an unsexy space or hard to get started with. or 2. Small but fast-growing markets — these usually happen when there is a fundamental shift in consumer behavior, tech, infra, or regulation.

please feel free to reach out to me on LinkedIn — I mostly end up replying to all my inbounds.