Co-founder of Color Genomics, Color Genomics
Elad Gil: How to Spot a Billion-Dollar Startup Before the Rest of the World
What if the world’s most connected tech investor handed you his mental playbook? Elad Gil, an investor behind Airbnb, Stripe, Coinbase and Anduril, flips conventional wisdom on its head and prioritizes market opportunities over founders. Elad decodes why innovation has clustered geographically throughout history, from Renaissance Florence to Silicon Valley, where today 25% of global tech wealth is created. We get into why he believes AI is dramatically under-hyped and still under-appreciated, why remote work hampers innovation, and the self-inflicted wounds that he's seen kill most startups....
About Elad Gil
Elad Gil, co-founder of Color Genomics and a multi-stage investor, appeared on two podcasts in early 2024 where he discussed his views on artificial intelligence and startup markets. On the Knowledge Project podcast, Gil stated that AI is "dramatically underhyped" because most enterprises have not yet adopted it, and he predicted that within a few years the industry will be "selling units of cognition" — effectively renting AI-driven labor equivalents. He also said that if an AI company is not seeing explosive growth quickly, "something's fundamentally broken." In a separate appearance, Gil argued that generative AI has shifted the business model from selling software seats to selling "human labor equivalents," citing the example of Harvey AI in the legal sector. He noted that foundation models have "instantly plugged into a massive set of markets" including all white-collar work and code. Gil also said that while there are times to be contrarian, the current moment favors consensus, adding that investors should "maybe just buy more AI." He attributed the sudden openness of previously closed markets to both AI's capabilities and the fact that "every CEO is asking themselves, 'What's my AI story?'"
Transcript (42 segments)
I
Interviewer32:02
And like what comes to mind for me are different aspects of AI. So you have from going all the way up the stack you have electricity, you have compute, you have LLMs, you have data. Where do you see the bottlenecks being? Where's the biggest bang for the buck? Like what's preventing this from going faster?
E
Elad Gil32:24
You know, it's a really interesting question and I think there are people who are better versed than I am in it because there's this ongoing question of when does scaling run out for which of those things. When do we not have enough data to generate the next versions of models, or do we just use synthetic data and it will not be sufficient, or how big of a training cluster can you actually get to economically? How do you fine-tune or post-train a model and at what point does that not yield as many results? That said, each one of these things has its own scaling curves. Each one of these seems to still be working quite well. And if you look at a lot of the new reasoning stuff that OpenAI and others have been working on, Google has been working on some stuff as well, when you talk to people who work on that, they feel that there's still enormous scaling left for that, because those are just brand new things that just rolled out. These reasoning engines have their own big curve to climb as well. So I think we are going to see two or three curves sort of simultaneously continue to inflect.
I
Interviewer33:15
Is this the first real revolution where incumbents have an advantage? And I say that because data costs money, compute costs money, power costs money.
E
Elad Gil33:27
Yeah.
I
Interviewer33:27
And it sort of favors the Google's, the Microsofts, the people with a ton of capital.
E
Elad Gil33:34
Yeah. I think in general every technology wave has a differential split of outcome for incumbents versus startups. So the internet was 80% startup value. It was Google, Amazon, Meta, all these companies we now know and love. Then the mobile revolution was probably 80% incumbent value or 90%. Mobile search was Google, mobile CRM was Salesforce, and the things that emerged as startups took advantage of unique characteristics new to the phone, like GPS and cameras, so you had Uber and Instagram. Crypto was 100% startup value. So you go wave by wave and ask what are the characteristics that make something better or worse. In self-driving, the two winners in the west seemed to be Tesla and Google through Waymo, both incumbents, which is a bit underdiscussed because we had two dozen self-driving companies.
I
Interviewer34:58
Wouldn't that make sense though because they have the most data? Tesla acquires so much data every day, and the way they have set up full self-driving, my understanding is it has gotten really good in the last six months. One of the reasons is they stopped coding and started feeding the data into AI and having the AI generate the next version effectively.
E
Elad Gil35:19
Yeah. A lot of the early self-driving systems were people writing a lot of edge case heuristics, but they moved a lot of these systems over to end-to-end deep learning. So this modern wave of AI has really taken over the self-driving world in a strong way that has helped these things accelerate.
I
Interviewer35:40
So I think all that's true. I guess it's more a question of when does that sort of scale matter and why wasn't anybody able to partner effectively with an existing automotive company?
E
Elad Gil35:53
It really depends on the layer you are talking about. I think there is going to be enormous value for both incumbents and startups. On the incumbent side, the foundation model companies are either paired up with or driven by incumbents. OpenAI is partnered with Microsoft, Google has its own efforts, Amazon partnered with Anthropic, Facebook has Llama. I wrote a blog post maybe two or three years ago about the long-term market structure for that layer, and it felt like it had to be an oligopoly because capital is so important. Eventually you are talking about billions, tens of billions of dollars, and not that many people can afford it. The financial incentive for the cloud businesses is their clouds. Microsoft's Azure quarter was $28 billion, with 10-15% lift from AI. So the biggest funders of AI besides sovereign wealth have been clouds because they have a financial incentive. That helped lock in this oligopoly structure early.
I
Interviewer37:58
Yeah. And that's the difference.
E
Elad Gil38:00
And I guess the optimism there is that I can go use the full scale of AWS or Azure or Google and just rent time. So I don't need to make the capital investments.
I
Interviewer38:09
Right. So the optimism is you can compete with them now because you are just competing on ideas.
E
Elad Gil38:15
Well, you could have done that either way. You didn't have to take money from them because they are happy to be a customer.
I
Interviewer38:19
That's what I'm saying. You have access to infrastructure.
E
Elad Gil38:28
Yeah. Everything moved to third-party clouds that you can run on. So that is enabling. But at least for these language models, they are increasingly just a mode due to capital scale.
I
Interviewer38:41
Do you think that we just end up with three or four and they are all pretty much equivalent?
E
Elad Gil38:46
I think you can imagine two worlds. World one: an asymptote where things flatline because you can only scale a cluster so much and have so much data, in which case things converge closely over time. World two: if one model is far enough ahead in capabilities, it can help build the next model faster, creating a strong positive feedback loop. That could lead to a liftoff scenario where the model effectively creates its next version.
I
Interviewer39:52
And at that point you have an advantage that is expanding at the velocity at which you are creating the next model. GPT10 would be so much more capable than 9 that it can build 11 faster and smarter.
E
Elad Gil40:09
It really comes down to what proportion of the model building task is eventually done by AI itself.
I
Interviewer40:31
What do you think of Facebook? They have spent 50 or 60 billion and basically given it away to society.
E
Elad Gil40:38
Yeah, I have been super impressed by what they have done with Llama. I think open source is incredibly important.
I
Interviewer40:45
Why is open source important?
E
Elad Gil40:48
It levels the playing field for different uses of the technology and makes it globally available. It also allows you to remove things you may not want, because it is open weights and open source. If you are worried about a specific political bias or cultural outlook, imposing your own values on everyone is a form of cultural imperialism. Open source gives you leeway to retrain a model to reflect your country's norms.
I
Interviewer41:46
As an investor, what is the ROI on a $600 billion open source model? How do you think through what Facebook is trying to accomplish?
E
Elad Gil42:02
I don't know how Meta specifically is thinking about it. In general, there have been many times where open source has been strategically important. IBM funded Linux in the 90s as a counterbalance to Microsoft. Apple and Google funded open source browsers. So AI might follow a similar pattern. I had wondered earlier if Amazon would fund open source AI because they didn't have a horse in the race, but Meta ended up doing it.
I
Interviewer43:37
How would you think about the big players and who is best positioned for the next two to three years in AI?
E
Elad Gil43:50
It is hard because AI is the only market where the more I learn, the less I know. Things change every six months. There is a handful of companies doing well: Google, Meta, OpenAI, Microsoft, Anthropic, xAI, Mistral. The question is how the market evolves and whether it consolidates.
I
Interviewer44:38
How do you think about regulation around AI?
E
Elad Gil44:42
There are three or four forms of AI safety that people conflate. First is digital safety, about hate speech and free speech, which is less concerning. Second is physical safety, like using AI to create a virus or derail a train, but those protocols are already widely available. Third is existential safety, where AI becomes self-aware and destroys us. People mix these up and say we should shut everything down. I think society has become very risk-averse and safety-centric, and that has unintended consequences. For example, children are kept in booster seats six years longer than crash data suggests is necessary. This safety culture also affects AI.
I
Interviewer48:00
There is one in Ottawa where they have crossing guards everywhere near schools. So kids can't walk to school on their own.
E
Elad Gil48:20
Yeah. That kind of safety undermines agency. Over the last 15 years, we have moved towards fragility and microaggressions, and now we are removing independence and risk-taking. I think that has negative downstream implications.
I
Interviewer49:26
You are one of the most successful investors people have probably never heard of. You have said that most companies die from self-inflicted wounds, not competition. What are the most common self-inflicted wounds?
E
Elad Gil49:42
For early companies, founders fighting or running out of money before product-market fit. Later, getting too competitor-centric instead of customer-centric can hurt. For example, pharmaceutical distributors used to compete aggressively on market share, eroding margins until one stopped and margins improved. So focusing on your own customers is better.
I
Interviewer51:49
Scaling a company often means scaling the CEO. What have you learned about how successful CEOs scale themselves?
E
Elad Gil52:00
First, they figure out who else they need and trust them. They avoid reinventing things that are standard, like sales. They hire people who are complementary and more effective than they are. A common failure is when a founder promotes their operations lieutenant as CEO, who lacks the product vision, leading to decay. Satya Nadella at Microsoft is a good example of a CEO with a founder mindset. Also, CEOs should ignore conventional wisdom about org structure and build a team that reflects their own needs, like Jensen Huang with 40 direct reports.
I
Interviewer54:53
Is that the problem with business leadership books that are written about a particular person's style, and people try to implement it but it is not genuine?
E
Elad Gil55:06
Yes. In generic large companies, org structures might be interchangeable, but for founder-led companies, you need a structure that fits the CEO. We lived through a decade where many unproductive workplace trends like 'bring your whole self to work' took focus away from the mission. Brian Armstrong and Toby Lutke were early in pushing back, saying the workplace is not a family but a team focused on performance.
I
Interviewer57:10
And the first person to speak out against that was Brian Armstrong. Was that the moment we started to go back to founder mode?
E
Elad Gil57:33
It took time. Brian was very brave and got canceled for it. But his essay and Toby's essay made a big difference. The idea that a workplace is a team, not a family, is important. You don't tolerate someone showing up drunk at work like you might with an uncle.
I
Interviewer58:56
You are around a lot of outlier CEOs. What common patterns have you seen?
E
Elad Gil59:14
There are two or three archetypes. First, hyperfocused people who don't get distracted by side investments or press. Travis Kalanick is an example. Second, polymaths with broad interests that they pursue alongside running the company, like Elon Musk, Patrick Collison, or Brian Armstrong. They often have many interests from a young age and go deep on each. Third, some people achieve success largely through product-market fit and grow into the role, driven by a different utility curve.