> ## Content Index
> Fetch the complete content index at: https://nnutshell.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Factory's founder bets no single AI model will win
- URL: https://nnutshell.com/factorys-founder-bets-no-single-ai-model-will-win/
- Published: 2026-10-23T18:13:23.000Z
- Updated: 2026-10-24T03:06:51.000Z
- Description: Factory founder Matan Grinberg on enterprise adoption flipping in 2026, why model independence is the company's core bet, and why code is cheap while constraints are not.
- Author: Admin
- Tags: Matan Grinberg, Andrej Karpathy, Michael Kratsios, Factory, Cursor, OpenAI, Anthropic, Developer Tools, Agentic AI, Foundation Models, #rec-2026-09-29, #Import 2026-10-25 09:54

*Recorded 29 September 2026 · 50 min*

Factory is built on the bet that no single model provider pulls far ahead of the rest. Founder Matan Grinberg admits it is exactly the statement his paycheck needs to be true. He argues the whole economy needs it too, and that a runaway winner would force the government to step in. The evidence he offers is Cursor, whose customers lost access to a model they wanted when OpenAI moved to end its contract.

The conversation, with host Alex, traces how Factory got here. It started in 2023 with a slogan that scared off candidates, "the future is IDE free". Enterprises flipped to agents only at the turn of 2026, and token usage is now up more than 10x year on year. Grinberg describes a "software factory" in which agents improve a team's software, rather than a contest over agent counts.

The tension is that he builds autonomous coding agents while hiring hundreds of people and insisting developers gain leverage rather than lose jobs. He also holds that AGI is already here and that liability, not new regulators, should govern the labs.

1. **Adoption flipped at 2026** — [▶ 0:36](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=36s&ref=nnutshell.com)  
Factory started three years ago, when the world was barely adopting GitHub Copilot and autocomplete was the frontier. Only SF AI companies were agent-pilled, while enterprises thought Copilot was moving fast. At the turn of 2025 into 2026, spurred by Andrej Karpathy tweeting about his use of coding agents, whole enterprises became open to changing behaviour. Token usage is now well over 10x year on year. Being an engineer today versus three years ago is "night and day", while 15 years ago and three years ago looked similar.
2. **The IDE-free bet** — [▶ 2:18](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=138s&ref=nnutshell.com)  
On April 17, 2023, Grinberg and his co-founder played a game: how often could they get ChatGPT to finish a full task by copy-paste, writing no code themselves? With proper context and a properly subdivided problem, quite often. Agents were then a matter of automating that subdividing, and better models would need less of it. Their first slogan was "the future is IDE free", and founding-engineer candidates dropped out over it. It became a hiring filter. He drew the trajectory from scaling laws, give or take a year.
3. **Modular, enterprise, unsubsidised** — [▶ 6:29](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=389s&ref=nnutshell.com)  
Grinberg splits the market into model labs with apps (OpenAI, Anthropic, Google, SpaceX, Cursor), services firms like the new Accentures, and Factory, which is model agnostic and built to empower developers. Developers hate black boxes, so defaults are defaults, not lock-in. Letting them tinker with models and routing makes skeptics agent native. The focus is enterprise, where the tedious work lives, such as COBOL migrations and 30 years of code. Factory won't play the subsidisation game that the labs run for solo users.
4. **Factories, not agent counts** — [▶ 9:56](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=596s&ref=nnutshell.com)  
Factory integrates with Cursor, GitHub Copilot, Claude Code and Codex. Grinberg likens it to the ship of Theseus: swap any part and it is still your software factory. The goal is software that improves itself and learns from how users interact with it, not "use our agent". He says the posting contest over who runs 1,000 agents makes little sense, even though he has 2,000\. What matters is the team, freed from migrations and micro-optimisation to chase 10x bets and deep systems problems.
5. **Capital buys people, not compute** — [▶ 12:16](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=736s&ref=nnutshell.com)  
Factory began the year with 30 people, two of them in sales, and will end it with 300\. It has opened offices in London and Sydney and is opening one in Tokyo. The money also funds research on a model-agnostic agent that beats Codex and Claude Code, and messaging that the future is "in your hands". He rejects the "no companies left" pitch. Businesses should build only their core competency and buy the rest, as Factory buys Salesforce and DocuSign rather than staffing 10 engineers to maintain clones.
6. **Automating model evaluation** — [▶ 16:18](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=978s&ref=nnutshell.com)  
New models arrive weekly, and no engineer can know them all. Factory's software factory now benchmarks each new model against a thorough suite before adding it to the routing pool. Two engineers once did that work. Now one person owns the automation and mostly checks in. The freed engineers work on problems Grinberg thinks cannot be automated: better tool use, and managing token caching when switching models. Switching mid-task breaks the cache, so when to switch is the hard question.
7. **Constraints are the alpha** — [▶ 18:47](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=1127s&ref=nnutshell.com)  
Grinberg endorses the idea that code is cheap. The valuable part has always been defining the multi-dimensional constraints a correct solution must satisfy, a constrained optimisation that explains why engineers are paid so much. Writing the code that meets them is getting easier. On autonomy, he sees no fully unattended production for enterprise customers in the next few months, because the stakes are too high. Within a couple of years, some code may have to be AI-generated to be robust, with only certified engineers allowed near it.
8. **Hire founders and agency** — [▶ 20:47](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=1247s&ref=nnutshell.com)  
Coding interviews are weak signal whether or not AI is allowed. Factory screens for high clock speed and relentless obsession with the problem, and the best proxy is having started a company. It is now faster to acquire cool companies than to confirm a single hire is great, and the acquired teams are not failures. Grinberg says high-agency people rise, including English majors. He calls writing one of the "ultimate high agency" acts, and says "I was bad at math" no longer matters.
9. **Model-agnostic router, with a confessed bias** — [▶ 29:12](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=1752s&ref=nnutshell.com)  
The router shifts work by performance, latency and cost, a free-market allocation of intelligence. Some tasks want the frontier, others cheap and fast, others slow and cheap. Engineers shouldn't care which model runs, as with the electricity behind toast, though a manual override exists. Asked about the risk of one model winning, Grinberg admits the bias: it is hard to convince someone a statement is false if their paycheck needs it true. His defence is that the whole economy needs a multi-model world, and that a runaway winner would invite government as "the monopoly to end all monopolies".
10. **Post-AGI, liability over regulators** — [▶ 32:53](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=1973s&ref=nnutshell.com)  
Grinberg says AGI is already here: show Factory to anyone four years ago and they would freak out, yet humans shrug after a week. He worries about safety discourse shaped by paycheck bias and sees fear-mongering. His rule is that if you give people something they use to do harm, you are held liable, like the host of an apartment with a lava pit at the door. Alex adds that a bank would face Senate hearings after a sandbox escape like Hugging Face's. Grinberg praises regulating closed models while exempting US open models, calling minimal regulation of the latter a "bold statement".
11. **Margins and the Cursor lesson** — [▶ 39:39](https://www.youtube.com/watch?v=IMxtXy2SnH0&t=2379s&ref=nnutshell.com)  
Factory is margin positive, which the host calls unicorn status for an AI business in 2026\. Grinberg says reselling tokens means negative margin, while solving problems doesn't. Rivals that trained their own models to fix margins are biased to route to them, and he wants no economic reason to steer customers. OpenAI ending its large Cursor contract validates the thesis: a Cursor customer can no longer use Astra, a model he calls phenomenal. His rule is that enterprises cannot allow a single point of failure. He expects the labs to mature into cloud-style partners as the pie grows.

No enterprise should depend on a single model provider, and Cursor's customers have just been shown why. OpenAI moved to end its large contract with Cursor after Cursor joined SpaceX. A Cursor customer can no longer use Astra, a model Grinberg calls phenomenal. He treats this as the cleanest validation of Factory's founding position. Standardise on anything that is not model independent and you are at "the whims of some of these providers", whose relationships with each other keep shifting. A business cannot carry a single point of failure.

Factory's answer is a router. It moves work between models on performance, latency and cost, a free-market allocation of intelligence to problems. Some tasks need the frontier. Some need something cheap and fast. Some can run slowly and cheaply provided they finish by the end of the week. No engineer can track which model sits on the cost-performance frontier this week, so the default should be invisible, like the electricity behind your toast, with a manual override for people who want a specific model. The same logic applies to tooling. Factory plugs into Cursor, GitHub Copilot, Claude Code and Codex. Grinberg's picture is a ship of Theseus: swap any part and it is still your software factory.

The obvious objection is that Factory needs this to be true. Grinberg raises it himself. It is "very hard to convince someone that a statement is false if that statement being true is required for them to get their paycheck", and his own statement is that the future is multi-model. His defence is that the entire economy needs the same outcome. A provider far ahead of the rest would be "the monopoly to end all monopolies", and only the government could sit above that. That is an argument from incentives. It says nothing about whether the models will actually stay close. He is also certain the models are very good: AGI, he says, is already here. Nothing in the conversation says what Factory does if one model nonetheless pulls away.

The independence has a financial side. Factory is margin positive, which the host calls unicorn status for an AI business in 2026\. Grinberg's explanation is simple. Reselling tokens means negative margin. Solving people's problems doesn't. Companies that had to train their own models to repair their margins now have a reason to route to those models. He wants none. The same instinct keeps him out of the consumer market, where OpenAI, Anthropic and SpaceX subsidise usage and Factory won't compete. The focus is enterprise, where the work is tedious: COBOL migrations and 30 years of accumulated code. He expects the labs to settle into cloud-provider behaviour, partnering with competitors because the pie is growing.

The router is the end point of a longer bet that agents, not autocomplete, would be the interface. Factory started on April 17, 2023, when most enterprises were barely adopting GitHub Copilot. Grinberg and his co-founder played a game. Using ChatGPT by copy-paste, how often could they finish a whole task without writing a line themselves? With proper context and a well-subdivided problem, quite often. Autonomous agents were then a matter of automating the subdividing, and better models would need less of it. Their first slogan was "the future is IDE free". Founding-engineer candidates dropped out over it, and it became a filter for believers. He drew the trajectory from scaling laws, give or take a year.

The models alone did not flip the market. Behaviour did. At the turn of 2025 into 2026, helped by Andrej Karpathy tweeting about how he used coding agents, enterprises became open to changing how they work. Token usage is now well over 10x year on year. Grinberg's view is that what a model lacks, behaviour can supply, so the job is to meet whatever shape the model has.

That shapes what Factory sells. He does not want a bigger agent. He wants a software factory, in which a team's software learns from how users interact with it and improves itself. The contest over who runs 1,000 agents makes little sense to him. Developers hate black boxes, so defaults stay defaults and users can fiddle with models and routing. Fiddling is how skeptics become agent native. The internal proof is mundane. Every new model is benchmarked automatically before it joins the routing pool. Two engineers once did that. One now checks in on the automation. The other works on problems he thinks cannot be automated, such as tool use and managing the token cache when switching models mid-task.

If agents write the code, value shifts to what the code must satisfy. The best engineers always took an ambiguous problem and defined the multi-dimensional constraints a correct answer meets. That is "why engineers get paid so much money", and it is why Factory buys Salesforce and DocuSign rather than staffing 10 engineers to maintain clones. It also shapes hiring. Coding interviews, with or without AI, are weak signal. Factory wants high clock speed and relentless focus, and the best proxy is having started a company. It is now faster to acquire such companies than to confirm a single hire, and headcount goes from 30 to 300 in a year. Agency, not a computer science degree, is the trait he rewards, and he says an English major who treats writing as a high-agency act has it.

The least developed part of his argument concerns risk. For the next few months he keeps humans in the loop for enterprise code. Within a couple of years, some code may have to be AI-generated to be robust. On safety, he wants liability rather than new bodies: if you invite someone into an apartment with a lava pit at the door, you go to jail, and you do not call for a regulator of lava. He praises regulating closed models while leaving US open models alone. He concedes minimal regulation is "a bold statement".

What that means for the people doing the work is left to an analogy. Grinberg expects more people doing what was once called engineering who will not call themselves engineers, as with the 10,000 photos on his phone, which do not make him a photographer. Whether that reassures anyone depends on the model market staying crowded.

**The open question.** Grinberg's whole structure of routing, independence and positive margins assumes the labs stay close enough that nobody pulls away. His backstop is that the market and the government would not allow a runaway winner. That is an argument from incentives, not from evidence about model progress, and he concedes it is the claim his own paycheck requires. He gives no plan for Factory if one model does become supreme. The Cursor and OpenAI split shows model access is fragile today, but it does not show that the frontier stays crowded.

Watch the full video at [https://www.youtube.com/watch?v=IMxtXy2SnH0](https://www.youtube.com/watch?v=IMxtXy2SnH0&ref=nnutshell.com).