Matan Grinberg, CEO of Factory and former string theorist, explores the shifting landscape of AI ROI, resource allocation, and the return of the polymath. He argues that the industry is moving from a period of 'token maxing' debauchery to a sober 'hangover' phase where enterprises demand clear business value and ROI. Grinberg details his journey from theoretical physics to founding an AI company, emphasizing the need for high-agency talent and the strategic decoupling of AI models from applications.
## [00:00] Intro
Harry Stebbings introduces Matan Grinberg, CEO of Factory, who transitioned from a 12-year career in string theory to software development. Grinberg posits that the future of the AI industry is defined by a race to commoditize competitors and that value accrual is highly time-dependent. He emphasizes that the age of the polymath has returned, where elite teams will be treated like professional athletes.
> *The age of the polymath is back. [00:45]*
> *The world going forward there is going to be nothing that no one can build. [00:00]*
## [01:22] Will AI actually increase GDP?
Grinberg expresses strong confidence that AI will drive meaningful GDP growth beyond the historical 2% average, though the effects will take time to permeate the economy. He explains that AI allows individuals to solve problems faster, forcing companies to choose between increasing output or operating more efficiently with fewer staff. This shift requires a fundamental adjustment in how organizations allocate human and technical resources.
> *We will see tremendous growth from these tools. I think it takes time to permeate through. [01:53]*
> *Everyone is now going to be able to solve more problems with the same number of people. [02:18]*
## [02:41] Smaller teams or bigger ambitions?
The conversation shifts to the future of engineering talent, specifically the concept of 'load-bearing individuals' or high-leverage employees whose removal would cause an organization to collapse. Grinberg suggests that AI tools act as a force multiplier for these individuals, widening the gap between those who can effectively use leverage and those who cannot.
> *Those who know how to use leverage will be able to have even more impact. [04:35]*
## [05:05] The resource allocation problem: tokens, dollars, people
Grinberg predicts that the next 24 months will see C-suite executives focusing intensely on the resource allocation problem involving tokens, dollars, and headcount. He advises leaders to prioritize their core competencies and judge success based on business outcomes like revenue rather than vanity engineering metrics like features shipped.
> *This resource allocation problem of token... is going to be the thing that over the next 24 months every C-suite is going to be thinking about. [05:08]*
> *Finally coming back to what matters in the first place. Like what are the business metrics that we want to move the needle on. [06:32]*
## [06:49] Kirkland's $500M AI bet and the build vs buy question
Harry and Matan discuss Kirkland & Ellis's $500 million investment to build internal AI tools, which Grinberg views as a potential strategic error since AI is not their core competency. He argues that such massive internal spends often lead to the realization that specialized vendors are more efficient, ultimately validating the difficulty of the problem.
> *Kirkland spending half a billion dollars to build their own AI tools... building AI technology is not a core competency of that firm. [07:14]*
## [10:01] Models, apps and infra: who gets commoditised?
Grinberg describes the current friction between model providers, application developers, and infrastructure firms, where each sector is actively trying to commoditize the others to capture more market value. He notes that value accrual is a time-dependent phenomenon, shifting based on who holds the most pricing power and leverage in the ecosystem.
> *everyone is trying to commoditize the people that are not them. [11:05]*
> *The reality is value acral is a time dependent phenomenon. [10:40]*
## [11:58] The bear case against Factory
Factory maintains a model-agnostic stance to provide customers with the best balance of price and performance across providers like OpenAI and Anthropic. Grinberg admits the primary risk to this strategy is if a single model provider achieves a significant, sustained lead over all competitors, creating a dangerous global monopoly.
> *The bare case against factory is if one model provider gets significantly better than all of the others. [12:05]*
## [13:57] The rise of open-source models
Enterprises are increasingly looking toward open-source models to manage ballooning token costs and annual budgets that are exhausted prematurely. Grinberg notes that 80% to 90% of tasks currently performed by frontier models could be handled by open-source alternatives, which serve as a vital counterbalance for less complex tasks.
> *so many of the tasks that we're doing we don't need the very frontier to do it. [14:47]*
> *there's kind of an ego thing where oh no no the work that I'm doing only a frontier model could handle. [15:15]*
## [17:08] The AI spending hangover
Grinberg describes the current state of AI adoption as a 'hangover' phase where companies are finally reviewing the massive bills accumulated during a period of unchecked usage. He predicts a healthy short-term contraction in frontier model usage as businesses prioritize actual ROI over novelty and implement strict resource allocation.
> *Phase three is the hangover where you go and look at the bill and it's like, 'Oh my god, we are spending so much. I have no idea what the ROI is.' [17:08]*
## [19:32] Token spend as a % of dev salary
Harry Stebbings questions whether token spend will eventually exceed headcount costs. Grinberg predicts that within three years, the median token spend per individual will be on the same order of magnitude as their salary, particularly for roles that gain massive leverage from AI 'droids.'
> *I would say order of magnitude. It'll probably be comparable to salary. [22:03]*
## [24:14] Factory's controversial culture: sales and engineering as one team
Matan Grinberg critiques the 'Silicon Valley fallacy' that research is the pinnacle of achievement while sales is secondary. At Factory, engineers and sales staff are fully integrated, sharing ownership of both features and closed deals to ensure the entire customer journey is treated as the product.
> *The product at factory is the entire journey from the very first time they hear our name till their 10th renewal. [25:33]*
> *If you don't have a good sales and marketing team... the second gravity returns, all of your muscles will be atrophied. [26:55]*
## [27:30] Why agency matters more than credentials
While venture capitalists often use elite credentials as a crutch, Grinberg argues they can be an 'anti-signal' if the individual lacks true agency. He prefers candidates who have demonstrated high agency by building things independently and taking end-to-end ownership of business outcomes.
> *What have you built? How have you taken ownership and agency of things end to end? [29:49]*
> *In a world where we desperately seek certainty we look for validators... that serves as a good crutch. [29:28]*
## [32:28] The age of the polymath is back
Grinberg argues that AI tools are ushering in a new era of polymaths by allowing individuals to reach the 'frontier' of multiple disciplines quickly. This shift favors individuals who can think in systems and manage uncertainty while pushing boundaries in both engineering and marketing simultaneously.
> *The age of the polymath is back. [32:28]*
> *These tools can get you up to speed to the frontier... way faster than ever before. [33:24]*
## [35:06] What we'll look back on in disbelief
Grinberg identifies writing release notes and documentation as tasks that will soon be considered a waste of expensive human engineering time. He suggests AI will soon equalize the advantage of high-quality documentation, allowing organizations to redirect human talent toward higher-value differentiation.
> *It's crazy that people used to spend hours of time writing release notes or like writing documentation. [35:24]*
## [39:25] Why the company is called Factory
Using a Tesla factory metaphor, Grinberg explains that the future of software development involves engineers designing the 'assembly lines' rather than writing individual lines of code. Humans act as architects of the scaffolding and safeguards that produce the software.
> *They're kind of like building the scaffolding around this factory that produces their software. [40:18]*
> *Engineers that build the software... they're going to have engineers that build the factories that build their software. [39:30]*
## [40:18] Labour displacement and the problems AI will finally solve
Grinberg acknowledges short-term economic shocks but remains optimistic about long-term employment. He argues that by lowering the cost of development, the market can reallocate human talent to solve a much broader range of global issues, such as dementia research, that were previously too expensive to tackle.
> *Very few of those problems that can be solved with software are we currently solving with software. [41:00]*
> *If we have more engineers who are going and solving more problems in the world, that is a net good. [41:16]*
## [44:21] Are we in an AI bubble?
Despite concerns about an infrastructure bubble, Matan identifies human behavior change as the most significant bottleneck for AI adoption. Successful enterprise integration requires navigating cultural shifts and the complexities of change management within established corporate structures.
> *The biggest bottleneck by far working with all these organizations is the human side of it. It's just like behavior change. [44:58]*
## [45:51] Lessons from selling to enterprises
Matan reflects on his transition from theoretical physics to enterprise sales, noting that success comes from genuine curiosity about a client's bureaucratic nightmares. He emphasizes that one should never try to 'sell' but rather understand if a solution can actually help the client's specific problems.
> *You should never try to sell something. You should always try to understand their problems. [46:42]*
> *People love talking about their problems and they love talking about all of the bureaucratic nightmares. [47:17]*
## [47:46] From string theory to Factory: the origin story
Matan recounts his childhood obsession with math and his drive to become a string theorist at Princeton and Berkeley. However, he experienced an existential crisis during his PhD, realizing he was pursuing the field because it was hard rather than for personal fulfillment.
> *I've just been doing this because it's hard and because someone said I couldn't do it. [49:12]*
> *I asked my dad what the hardest math was. He said string theory... I was like, okay, I'm going to be a string theorist. [48:44]*
## [50:46] Discovering code that writes itself
After exploring computer science at Berkeley, Matan became 'nerd sniped' by program synthesis—the concept of code creating itself. He realized that the most significant problems in this space would be solved in industry rather than academia, leading him to start a company.
> *It just completely nerd sniped me because the idea here is... code with the explicit purpose of creating itself. [51:03]*
## [52:30] The cold email and 3-hour walk with Sequoia
Matan reached out to a Sequoia investor who shared his physics background. Their initial meeting turned into a three-hour walk where the investor gave Matan a blunt ultimatum: drop out of his PhD immediately to either join Elon Musk's Twitter or start his own company.
> *You absolutely need to drop out of your PhD and you should either join Twitter right now... or you should start a company. [53:48]*
## [55:30] Dropping out and the $1M check
Within 72 hours of building a demo with his co-founder Eno, Matan withdrew from his PhD and pitched the Sequoia partnership. Despite a 'shitty deck,' Sequoia offered a $1M check for a 20% stake, a deal Matan accepted because they believed in him when no one else did.
> *No one else would have believed in me except him... trust and loyalty and like belief to me that matters so much more. [57:38]*
> *Drop out of your PhD and send me a screenshot. [55:16]*
## [1:01:19] Does Ivanka Trump add value as an investor?
Matan addresses skepticism regarding celebrity investors, asserting that Ivanka Trump provides significant tangible value through her intelligence and network. He notes that she and her firm, Affinity, earned their place on the cap table through active support and investor relations.
> *She is genuinely so kind, so intelligent, and like people just in throughout tech... really love her. [61:52]*
## [1:02:39] How the coding market matures
Matan suggests that the market will eventually mature into a state where AI models are decoupled from the specific applications they power. This separation is necessary to prevent misaligned incentives where model providers might otherwise 'token max' for profit rather than efficiency.
> *What is necessary for the best outcome for the consumers is going to be models that are separate from the applications. [63:01]*
## [1:07:45] The coming security danger zone
As AI-generated code grows exponentially, Matan warns that security efforts are not keeping pace, creating a 'danger zone.' He emphasizes that adversarial behavior using AI tools is still in its early stages and will become a critical market focus as stakes rise.
> *Code generated is growing exponentially. The security efforts aren't growing in kind. [68:17]*
## [1:08:50] Should US startups use Chinese models?
Matan addresses concerns regarding US startups using Chinese open-source models, specifically the fear of 'trigger words' for adversarial behavior. He stresses the importance of data exfiltration defenses and expresses a desire for the US to reclaim superiority in frontier open-source models.
> *I think it's pretty embarrassing that we don't have frontier open models in the United States. [70:33]*
## [1:11:43] Data centres and the public backlash
The conversation shifts to the public backlash against data center development. Matan argues that the United States' federalist structure acts as a 'petri dish' where states allowing data centers will see job growth and prosperity while others fall behind.
> *It's like we have little petri dishes to test out and see how things work. [72:31]*
## [1:14:22] Selling without forward deployed engineers
Matan critiques the use of service-heavy FTE models to sell AI products. He argues that if a company requires a heavy services component to make their software work, the product itself is fundamentally flawed and lacks true product-market fit.
> *If we need FTEES to make the product work, we have a [ __ ] product. [75:15]*
## [1:15:32] Grindslop, sleep and treating teams like athletes
Matan rejects 'grind slop' culture—focusing on hours worked rather than output. He advocates for treating elite engineering teams like professional athletes, prioritizing cognitive recovery and sleep to ensure high-quality decision-making and leverage.
> *Imagine trying to measure who won a basketball game by who sweat the most. [76:12]*
> *The work that we do is like might require like really deep thought... if you didn't sleep well like you're not going to make as good of a decision. [78:02]*
## [1:20:32] Anthropic vs OpenAI
When asked to choose between OpenAI and Anthropic for an IPO investment, Matan selects Anthropic based on corporate stability. He notes that OpenAI has suffered from significantly more internal turbulence and chaotic events, which negatively impacts its expected value.
> *Past is an indicator of the future and like there's just been more like random chaotic turbulent events at OpenAI. [81:06]*
## [1:21:19] Did Dario do AI a disservice?
Matan critiques AI leaders like Dario Amodei who claim AI will replace all human labor, calling the rhetoric a fundraising tactic. He argues these claims are designed to convince investors that a single company will eventually capture the entire capitalist economy.
> *The best way to convince people to do that is to say all of capitalism is gone. [82:00]*
> *Incentive is driving the outcome and the incentive is I want to raise a lot of money. [82:54]*
## [1:23:53] What he's changed his mind on
Matan shares his shift in perspective from a 'winner-take-all' view to expecting a multi-polar market with at least four frontier companies. He identifies legacy firms like EY as surprising leaders in AI adoption, moving faster than some startups due to their 'scars' from the cloud transition.
> *The bad case for humanity is when there's one that's really really good. [84:14]*
> *They are so agent native. It's crazy. They're one of our largest customers. [83:11]*
## Entities
- **Matan Grinberg** (person): CEO and co-founder of Factory, former string theorist.
- **Harry Stebbings** (person): Host of 20VC and venture capitalist.
- **Factory** (organization): AI company focused on software development automation and agents.
- **Sequoia Capital** (organization): Venture capital firm that led Factory's seed round.
- **OpenAI** (organization): Leading frontier AI model provider.
- **Anthropic** (organization): AI safety and research company, creator of Claude.
- **Ivanka Trump** (person): Strategic investor in Factory via her firm Affinity.
- **EY** (organization): Big Four accounting firm noted for aggressive AI adoption.
- **Uber** (organization): Company cited for implementing individual AI token budgets.
- **Kirkland & Ellis** (organization): Law firm that invested $500M in internal AI tools.
- **Juan Maldacena** (person): Renowned physicist at Princeton whom Matan worked with.
- **Dario Amodei** (person): CEO of Anthropic.
