Anyone looking to invest in AI gaming startups runs into the same problem: almost every pitch now has “AI” in it, and most demos look impressive for three minutes. This guide sets out the questions that separate an AI game startup with a real business from one with a good video. We use our own company, Metashift Labs, as the worked example, with the figures we have already published, so you can see what the answers look like in practice.
What does the game tech investment market look like right now?
Capital for games has been scarce. Konvoy’s Q2 2025 gaming industry report counted $193M across 60 venture deals in that quarter, down 62% year on year, after $373M across 77 deals in Q1 2025. Aream & Co. and InvestGame, summarised by GameDev Reports on 16 January 2026, called 2025 the lowest year for private gaming investment in five years.
The AI money went somewhere else. Drake Star’s Global Gaming Report 2025, published 12 January 2026, notes that the largest AI rounds touching games went to world-model companies, which build general models rather than games. The same report shows private gaming rounds recovering through the year, from 105 in Q2 to 137 in Q4 2025.
For an investor this means two things. Game studios that raise now are being tested harder than they were a few years ago. And “AI in games” covers very different businesses, from model labs to studios shipping a product to players. The framework below is for the second kind.
How do you evaluate an AI gaming company? Start with real usage, not demos
A demo shows what a system can do once, under good conditions. Usage data shows what happens when thousands of strangers use it for weeks. Three numbers do most of the work.
Retention. Day-one retention (D1) tells you whether a player who installed the game came back the next day. It is the first sign that the core loop works without the pitch deck around it. Ask for it by market, because Tier 1 markets (the US, Western Europe and similar) are the hardest and most valuable audience.
Cost per install (CPI). CPI shows what it costs to buy a player through advertising. A low CPI together with a good click-through rate (CTR) means the idea is easy to explain in an ad, which matters more than most founders admit.
Hours played. Total hours and the distribution of hours show depth. A game where a few people play a lot and most leave quickly has a different future from one where a broad group keeps coming back.
Our worked example is the month-long Steam playtest of Dominus Automa in August 2026:
| Metric | Result |
|---|---|
| Players | 12,500+ |
| Hours played | ~127,000 |
| D1 retention, Tier 1 | above 36% |
| CPI | under $0.60 |
| Ad CTR | above 4% |
| Steam wishlists | 5,500 → 16,500 |
The figures that interest us most are about how people played: 43% of players used offline mode, and around 60% of fights happened without the player present. We break this down in our playtest data on heroes playing alone. A playtest is not a launch, and an investor should treat it as evidence of demand, not a forecast.
“PC used to raise money on an idea. We raised it on proof.”
Rafał Bieniek, CEO, Metashift Labs, Strefa Inwestorów Demo Day, 15 October 2026
Who pays for the AI? Unit economics in a live game
A traditional game pays for content once and serves it many times. A game where AI runs during play pays for inference every time a player uses it. If those costs grow faster than revenue per player, more players make the business worse.
The questions to ask:
- What does one AI interaction cost, and how many does an active player trigger per day? The answer should be a number, not “it’s getting cheaper”.
- Which part of the business model covers that cost? One-off purchases fit poorly with ongoing compute. A recurring payment fits better.
- What happens to margins if usage doubles? Heavy users should be the best customers, not the most expensive ones.
In Dominus Automa, a message from a player’s hero costs a fraction of a cent. The business model is built around that running cost: a monthly subscription covering server and AI model costs, a player trading market, cosmetics and quality-of-life items, and seasonal servers every three months with a battle pass. The hero and its personality are described in more detail in our guide to agentic AI in games.
What can a competitor not copy?
Model access is not a moat. Any studio can call the same APIs, and prompts are easy to reproduce. Look for assets that take time to build and are hard to see from the outside.
- Server-side logic. In a server-authoritative game, the rules, the simulation and the AI behaviour live on servers the competitor cannot inspect. Our multiplayer server peaked at 500+ concurrent users during testing.
- Data. Every session produces data about how players and agents behave. During our August playtest players ran 22,779 offline sessions, and we analyse that history to decide what to change next.
- Community. A player base that cares is slow to build and hard to poach. Our playtest survey of 270 players came back 91% positive, and 80+ people support the project on Patreon.
There is also a product question underneath: is the AI a feature bolted onto a familiar game, or does the game only work because of it? We explain why agents that play for the player differ from smarter NPCs in AI NPCs vs AI player agents.
Has the team shipped before?
Games are hard to finish. AI makes some parts faster, but it does not remove the work of balancing, live operations, community management and marketing. A team that has shipped and run games before knows where projects stall.
Metashift Labs is a team of 8. Before raising outside money we funded the work from our earlier games business. In January 2026 we closed a pre-seed round of 1M PLN from 11 business angels from the games industry and Google, chosen for experience as much as capital.
Speed of learning vs speed of writing code
AI-assisted production is now common. The Google Cloud and Harris Poll survey of 18 August 2025 (press release) found 90% of game developers using AI in their workflows. So “we use AI to build faster” says little on its own. The useful question is whether that speed turns into faster learning: more tests, shorter feedback loops, decisions made on data.
Our own example: by our estimate, three programmers working with AI assistants wrote 273,000+ lines of code in 7.5 months, work we estimate would take 7 to 9 years at traditional industry rates. That is our internal estimate, not an independent benchmark, and lines of code are a weak measure of value. What matters more is what the speed bought us: automated tests now cover about 40% of the code, and an agent reads server logs and proposes fixes, so some bugs that took weeks now take hours. Playtest analysis now takes hours.
What risks should an investor price in?
An honest pitch names its risks. These are the ones we see for AI game companies, including ourselves.
Player sentiment toward generative AI. The GDC 2026 State of the Game Industry survey, published January 2026, found 52% of developers think generative AI is harming the industry. Among players, Quantic Foundry’s survey of 18 December 2025 (report) found 85% view generative AI in games negatively. A studio needs a clear answer on where AI appears and why. Ours is to keep AI in the process and in the hero’s behaviour, while the visible art is made by artists.
The review bottleneck. Generating code is fast; checking it is not. Senior developers and the CTO become the limit on how much can ship safely.
“If code gets written 11 times faster, someone still has to read it — and humans don’t read 11 times faster.”
Rafał Bieniek, CEO, Metashift Labs, Strefa Inwestorów Demo Day
Platform rules. Stores and platforms set their own policies on AI content and on how games reach players, and they can change them. A game that depends on one storefront carries that exposure.
Generated content quality and IP. Anything generated at runtime can go wrong in front of a player, and ownership questions around AI output are not fully settled.
How to follow Metashift Labs
We are building agentic AI for games, and Dominus Automa is the game that tests it with real players; you can read how Dominus Automa works as an automated MMORPG. We open investment rounds from time to time. If you want to hear when the next one opens, join the investor waitlist.