If you build 80% of the right game, you’re going to get like $10,000 in sales. If you build 100% of the right thing, you might get $100 million in sales.
That is Ran Mo — CEO of Proxima, creator of Suck Up!, and arguably the person with the most credibility to be an AI maximalist in gaming right now. Suck Up! was the first commercially successful game to use AI at runtime. It generated over 100 million YouTube views with zero marketing spend. Before founding Proxima in 2021, Ran led product teams at EA working on The Sims franchise and spent time at YouTube and BCG.
He has every reason to be all-in on the AI hype. He is not. And his restraint is the most instructive signal in the current landscape.
I should be transparent: I am an AI maximalist. I spent weeks sleeping two to four hours a night learning vibe coding. I believe development costs will drop 10x by the end of this year and that 2026 may be the most consequential year for AI's impact on our industry. Ran and I disagree on pace and magnitude — but the conversation we had sharpened my thinking on AI's impact on product velocity: the difference between speed and direction. AI gives you speed. It does not give you direction. And in a power-law market, direction is everything.
Here is the thesis of this piece: AI tools deliver real, measurable productivity gains on specific coding tasks. Those gains will not, in isolation, translate into more successful games. They will translate into more games — most of which will fail — while the winners pull even further ahead. The bottleneck in games was never production. It was always taste, direction, and distribution.
The Trunk and the Leaves
The most useful mental model for AI deployment in game development is actually quite simple: a tree.
The trunk is your core architecture — the game loop, the netcode, the core systems that everything else depends on. The leaves are discrete, isolated features — a grab animation with quaternion-based rotation, a UI element, a shader variant, a localization pipeline.
AI is exceptional at leaves. Ran demonstrated this during our conversation by sharing his screen — a quaternion-based object rotation system with sine-curve acceleration and slerp interpolation, built live in Unity with Claude plugged into a Rider IDE terminal. A senior engineer builds it in a day. A junior engineer needs two weeks just to learn what a quaternion is. Claude built it in under 30 minutes. That is a genuine 30x speed gain on a specific, well-scoped task.
But the trunk? Ran tested this himself — full vibe-coded Unity projects produce functional prototypes that can be architecturally catastrophic. His analogy: it's like using MS Paint to draw a square by painting every individual pixel instead of using the square tool. The result looks identical. The underlying structure is unusable the moment you need to scale. The code compiles. It runs. And then it becomes unmaintainable spaghetti when you need to iterate on systems that touch everything. "You're going to burn a million tokens and get nowhere," as Ran put it.
This reflects a current-day structural reality: AI tools are built for local optimization, and trunk-level architecture is a systemic design problem. Every tech leader I talk to is fighting this battle internally. The pressure to show AI-driven productivity gains is immense. But if you let AI-generated code infect your architectural core, you are building on sand.
At least today, building production-quality code for a complex game remains an extremely high-skill endeavor that few people in our industry can handle.
So what does Ran's actual productivity gain look like? His answer surprised me, given my maximalist leanings. On leaf-level tasks — the grab animation, the UI polish, the boilerplate — he sees 5-10x or even 30x gains. But on trunk work — re-architecting a codebase, refactoring systems — using AI can actually slow you down, because the time spent injecting context, supervising output, and auditing for silent breakage can exceed the time it would take a senior engineer to just do it themselves. The overall gain, across all work, is real but more muted than the X headlines suggest.
The strategic move is ruthless segmentation. AI on the periphery. Senior engineers on the spine.
The $100 Million Gap Between 80% and 100%
Here is where the conversation gets uncomfortable for the "AI will democratize game development" crowd.
Games are not houses. If you build 80% of a house, you can sell it for 80% of the price. Games operate on an exponential attention curve where "almost right" is commercially indistinguishable from wrong. The last 20% — the part that is pure creative judgment, directional conviction, finding the right mechanic, the right feel — is where all the value concentrates.
AI makes the first 80% cheaper and faster. It does nothing for the last 20%. Which means AI tools are making it cheaper to produce games that fail.
The data confirms the power law. The global games market generates approximately $184 billion annually. Fewer than 5% of Steam titles produce meaningful commercial returns. AI tools will dramatically increase the number of games uploaded to Steam and mobile app stores. The denominator explodes. The numerator — total player attention — stays fixed.
Ran drew the parallel to YouTube, where he worked before entering games. The barriers to entry for video content collapsed years ago. Anyone can record themselves eating breakfast and upload it tomorrow. What happened? Extreme power law distribution. The top creators captured the vast majority of attention. The long tail got nothing.
Games are heading to the same place. You must be premier television or Mr. Beast. There may not be a viable middle.
The Technically Capable Designer: Who Actually Wins
This brings us to a question most AI-and-games commentary skips entirely: if AI compresses the coding bottleneck, who is the ideal game maker in this new environment?
Ran's answer is specific: the technically capable designer. Not a 10x engineer. Not a business founder who hires coders. A designer who understands code architecture well enough to direct AI tools, read what they produce, and catch when they're building the wrong thing — but whose primary skill is creative judgment and taste.
"The limiting factor for games is not code, it's the creative output of the design," Ran said. "You can have a designer who understands code, but not very good — he would need two engineers to realize his vision two years ago. Today he might not need anyone. But he has to be somewhat technically capable."
This is the new archetype. Not the vibe coder who prompts their way to a prototype and calls it a product. Not the senior engineer who uses AI to ship 3x faster but has no design instinct. The person who sits at the intersection — who can evaluate a quaternion implementation and knows whether the mechanic it enables is worth building in the first place.
Design as Differentiator, Not Speed
Ran Mo does not use AI for art. He does not use AI for design. He does not use AI for concept work. Not even for brainstorming.
This is not Luddism from someone who doesn't understand the tools. This is a strategic decision from the founder of an AI-native game. His reasoning: in a world of AI-generated abundance, the human element of your creative output is the only differentiator. AI creative work is average by definition — it is a statistical composite of its training data. When everyone has access to the same generative tools, average becomes invisible.
"You can have the most ugly chess set and it's still chess," Ran told me. "But if you have this beautiful tile thing that isn't chess, it doesn't matter."
It’s a sharp rebuttal to the idea that AI-generated environments and world models will replace game design. The mechanics are load-bearing. A stunning procedurally generated world with no game loop is a tech demo, not a product.
And here is where Ran offered what I think is the most contrarian and underappreciated insight of our conversation — one that carries particular weight because he built the first successful AI-runtime game.
The Probabilistic Gameplay Problem
When Google announced world model capabilities, Unity stock dropped approximately 10%. The market panicked. Several VCs Ran spoke with were convinced this would change everything.
Ran's response: they're solving the wrong problem.
"People, especially people outside of games, misunderstand what games are," he said. "Games are mechanics wrapped in logic. What's really important is the mechanic. And when you build AI systems — whether it's world models or AI NPCs — they're probabilistic. If people can't predict how the probability functions, it's very difficult for them to decide what to do next."
He used a chess analogy that crystallized it for me: imagine playing chess, but the knight might do something different on the next move — you just don't know what. Would you play that game? Probably not. The reason games work is because players can form mental models of the rule system and make strategic decisions within it. Probabilistic AI breaks that contract.
This matters because Ran has more standing than almost anyone to say otherwise. Suck Up works because the comedy context tolerates unpredictability — the AI NPCs are a comedy stage, not a strategic system. But he's clear-eyed that this is the exception, not the template. "We haven't seen too many games out of Suck Up who have really gotten the stuff beyond," he acknowledged. "There's probably less there than people think there is."
The implication for studios chasing AI-in-gameplay as a differentiator: make sure you're building for a context that benefits from probabilistic behavior, not one that requires deterministic rules. Most game genres require the latter.
The Velocity Trap
This is where Ran's philosophy diverges most sharply from the prevailing Silicon Valley consensus — and from my own instincts, if I'm being honest.
Ran is deliberately lengthening his iteration cycles. Not because he is lazy. Because he learned the hard way — nearly burning out himself and his team after Suck Up — that micro-level velocity often masks macro-level misdirection.
"We would get a playtest and figure out the changes and send feedback by 8 PM, and by 8 AM the next morning we'd have the new build," he told me. "The iteration cycle was crazy fast. But then you step back after three months and you're like — we went down entirely the wrong direction."
Now he forces himself to sleep on a build for an extra day before committing to a direction. He described it through the lens of a Taoist philosophy called wu wei — the art of deliberate non-action. It sounds like mysticism until you map it to product development: in an environment where everyone is sprinting, the competitive advantage belongs to the team that pauses long enough to confirm they're sprinting toward the right destination.
"A lot of people are just kind of pedaling to go nowhere," Ran said. "And it's better just to not pedal at all in some cases."
I pushed back on this in our conversation — I think speed matters enormously right now, and the teams that figure out AI-augmented development first will have a structural advantage. But Ran’s experience reinforced something I’ve already written about: product velocity is a vector — magnitude and direction. A team moving slowly in the right direction will outperform a team moving at blazing speed in the wrong direction. AI gives you magnitude. Taste gives you direction.
Distribution Is the New Production
If creative taste is the moat, distribution is the drawbridge. And most studios still treat it like an afterthought.
Suck Up's distribution story is now well-known, but the mechanics of it deserve closer examination. "We had $0 in marketing because we were poor," Ran said, "but we had one guy who had 100,000 subscribers on YouTube play it and he got a million views in a day."
Within a week, 30 creators were playing it. Within a month, top creators with tens of millions of followers had picked it up. The game generated over 100 million YouTube views organically.
This was not luck. It was architecture. Suck Up was designed to produce entertaining content when played. The AI NPCs created a comedy stage for content creators — a canvas for drama and improvisation. The game's core loop was inherently watchable and shareable. Distribution was a design constraint, not a marketing function.
And then Proxima made a hire that signaled exactly where the industry is going: a head of marketing with approximately one million TikTok followers. Not someone from a brand marketing background at a major publisher. Someone who had built an audience themselves.
"The biggest indicator if they understand marketing today is can they build an audience themselves," Ran said. "Because until you build an audience yourself, you don't really know how to build an audience."
This is the new competency test for marketing hires in the games industry. Can you build an audience from zero? Not manage a UA budget. Not optimize CPIs. Actually create content that earns organic attention. If the answer is no, you are hiring for the old model.
The Proxima playbook — embedding distribution-native thinkers into the development cycle from day one, not bolting on marketing six months before launch — might just be the new meta. When production costs deflate, the capital you save must be redeployed toward creative risk and distribution capability. Studios that pocket the savings and ship the same game faster are simply accelerating their path to irrelevance.
The Contrarian Case — and Where It Falls Short
I want to address the strongest counterargument directly, because it is not wrong. It is just incomplete.
The argument goes like this: incumbents also get the cheap tools. A 40-person studio using AI to operate like an 80-person studio does not get squeezed — it gets leverage. It keeps its institutional advantages: project management maturity, existing player communities, platform relationships, live-ops experience. The desktop publishing revolution of the late 1980s was supposed to kill mid-tier design agencies. Instead, the agencies adapted, shed staff, and became more profitable.
This is fair. And it will be true for some mid-tier studios — the ones that restructure ruthlessly, redeploy savings into distribution and creative risk, and have the discipline to kill projects that are not working. Supercell has operated this way for years: small cells with institutional infrastructure, shared technology, and a marketing machine behind them. The model works.
But here is where the analogy breaks down. Desktop publishing produced bad newsletters. AI-enabled game development will produce playable games that compete for the same finite shelf space on Steam and the App Store. The flood is not just noise — it is functional noise. When Steam's discovery algorithms are overwhelmed by AI-generated content, they will favor quality signals: review scores, playtime depth, refund rates, developer track record. Those signals favor established studios. They also favor small teams that have found genuine product-market fit and earned organic attention. The studios they crush are the ones in the middle — too large to be nimble, too small to compete on spectacle, shipping derivative products into an ocean of content.
The Bottom Line
Every era of cost deflation in creative industries produces the same pattern. Production becomes cheap. The market floods. Attention concentrates. The winners are not the fastest producers. They are the ones with the clearest creative vision and the strongest distribution instinct.
AI does not change this pattern. It accelerates it.
The studios that survive the near future will be built around a specific kind of person and a specific kind of discipline: technically capable designers with strong creative taste, distribution thinking embedded from day one, and the wisdom to slow down when everyone else is sprinting toward the wrong destination. They will use AI ruthlessly for the leaves and protect the trunk with senior engineering judgment. They will hire marketers who have built audiences, not managed budgets. They will resist the pressure to ship fast and instead ship right.
Ran Mo nearly ran out of cash three months before Suck Up launched. One of his VCs told him they backed Proxima because the team "clung on and didn't die." That is not a glamorous endorsement. But it captures something essential about what wins in a power-law market: the conviction to stay on the right path even when the pressure to sprint in any direction is overwhelming.
Ran's parting advice was drawn from the Taoist concept of wu wei: "Sometimes it's important to slow down to speed up. Not everyone needs to be pedaling all the time."
I'm still an AI maximalist. But I left this conversation thinking less about how fast I can build and more about whether I'm building the right thing. That might be the most valuable shift the AI x gaming discourse could use right now.
