Unreal Engine 5 Was Supposed to Make Games Cheaper. It Made Studios Bigger.

On July 6, 2026, Xbox CEO Asha Sharma sent a memo explaining why 3,200 people were losing their jobs. The line that got quoted everywhere was the one about returns: in a typical year, Microsoft lost 64 cents on every dollar it invested in its game studios. The line I keep coming back to is a different one. Sharma wrote that she intended to cut Xbox’s management structure from as many as fourteen layers down to five, and preferably three.

Fourteen layers of management. To ship video games.

That memo arrived at the end of a four-year stretch in which game development tooling improved more than it had in the decade before it. Unreal Engine 5 shipped in production-ready form in April 2022 with two technologies designed to delete entire categories of labor from the art pipeline. By 2026 it was everywhere. Halo walked away from Slipspace, an in-house engine with components approaching twenty-five years old, to adopt it. Most of the Western AAA slate now runs on it.

Costs went up anyway. Team sizes grew, cycles stretched past eight years, and a lot of the output shipped with frame time problems that players noticed inside the first hour. By the ASGC tracker’s count, more than 53,000 people have lost jobs in games since 2022, with another 14,259 forecast for 2026 alone.

The gap between what that tooling promised and what the industry did with it is the most useful thing in this whole downturn, and almost nobody is looking at it.

Two systems carried the pitch, and both of them were real.

Nanite is virtualized micropolygon geometry. Before it, every static mesh needed a hand-authored LOD chain: the same rock modeled at 20,000 triangles, then 5,000, then 1,200, then 300, so the renderer could swap in cheaper versions as the camera pulled away. A person built those. A person maintained them every time the art changed. Nanite streams and scales geometric detail on its own, which Epic describes as largely removing polygon count and draw call constraints.

Lumen is fully dynamic global illumination. Before it, lighting was baked. Epic’s 5.0 release notes put the change plainly: with Lumen you no longer author lightmap UVs, wait for lightmaps to bake, or place reflection captures. Anyone who worked in UE4 knows the rhythm that replaced. Build the scene, kick off a bake that could run for hours on a complex level, come back, find that the light placement no longer works, move one light, bake again. Overnight bakes were routine. Lumen made lighting interactive.

World Partition did something similar to level streaming, swapping hand-tuned streaming volumes for automatic spatial partitioning.

Then there is the price. Unreal Engine 5 is free until a title clears a million dollars gross, at which point Epic takes five percent. A studio could delete two full-time production disciplines from its pipeline at zero marginal cost until it was already successful.

None of that was marketing. The technology does what Epic said it does. I have watched artists who spent a decade fighting lightmap UVs move to Lumen and get their afternoons back.

Remove the manual LOD pipeline and the bake cycle and the labor required per unit of visual fidelity falls. Budgets compress, or fidelity rises at constant cost. That is the entire point of a tooling improvement.

The pre-UE5 baseline is unusually well documented, because Sony’s lawyers botched a redaction. In a filing submitted during the FTC’s attempt to block the Activision acquisition, black marker failed to obscure the numbers. Horizon Forbidden West cost $212 million over five years with a peak of more than 300 full-time developers. The Last of Us Part II cost roughly $220 million with a peak over 200. The UK Competition and Markets Authority’s 418-page market report put the average AAA budget five years earlier at $50 to $150 million.

Those are the numbers UE5 was supposed to bend downward. Instead Concord spent eight years in development on a budget industry estimates put between $200 and $400 million, which Sony has never confirmed, and sold roughly 25,000 copies before the servers came down after two weeks and Firewalk Studios was closed.

A polygon budget is a cost. It is also a governor. When an environment artist has 40,000 triangles for a hero prop, somebody has to decide what that prop is for. When the lighting bake takes six hours, somebody decides which lighting changes justify six hours. The constraint forces a prioritization conversation, and the conversation is worth more than the constraint.

Remove the constraint and you do not automatically bank the savings. You remove the conversation. Scope expands to fill whatever the tooling permits, because nothing in the room is saying no anymore. Then you need coordination staff to manage the expanded scope, and managers to manage the coordinators, and eventually a CEO writes a memo about fourteen layers.

I have watched this failure mode for thirty years on the Microsoft stack under a series of different names. When ORMs arrived, the pitch was that developers would stop hand-writing SQL and ship faster. At more than one client I walked into, what actually happened was that nobody wrote SQL anymore and nobody read it either, so N+1 queries multiplied quietly until somebody called me at two in the morning about a database going down under load. Autoscaling did the same thing to capacity planning. Serverless did it to cost forecasting. The constraint that used to force a decision got abstracted away, the decision stopped getting made, and the expense surfaced somewhere nobody had instrumented. That is most of why I built PLogger the way I did. Observability is not about explaining a failure after it has already cost you a weekend. It is about keeping a constraint in the room after the tooling has removed it.

Epic had no obligation to ship that constraint. The studios were supposed to build their own, and most of them never did.

The strongest evidence that UE5 is not a cost-elimination machine is what happened to the company that wrote it.

On March 24, 2026, Epic Games laid off more than 1,000 people, roughly 20 percent of its workforce, leaving about 4,000. Tim Sweeney’s memo was blunt. The downturn in Fortnite engagement that started in 2025 meant Epic was spending significantly more than it was making. The cuts came alongside more than $500 million in identified savings across contracting, marketing, and unfilled roles, and they landed weeks after Epic raised V-Bucks prices on the grounds that running Fortnite had gotten considerably more expensive.

Fortnite runs on Unreal Engine 5. Epic wrote Unreal Engine 5. If that engine converted reliably into lower operating costs, its author would have been first in line for the benefit instead of cutting a fifth of its staff to stay funded.

ARC Raiders is built on Unreal Engine 5.

Embark Studios shipped it in October 2025 with roughly 70 people after about three years of work. It became Nexon’s most successful global launch, peaked near 700,000 concurrent players across platforms, out-peaked Battlefield 6 on Steam in its second week, and passed 16 million copies by May 2026 per Nexon’s Q1 report, which came in around $959 million, up 34 percent year over year. More than half of active players have logged over 100 hours. Embark published a four-update roadmap in January 2026 and shipped every update on schedule.

Then there is the detail that reframes the whole argument. Embark did not use Lumen or Nanite. They ran a modified build of UE5 developed with NVIDIA. They kept World Partition and Runtime Virtual Texturing, which remove labor without surrendering control over what a frame costs. They used the PCG Framework and Houdini for procedural placement rather than staffing an army of artists to position rocks by hand. And they ran Unreal Insights continuously, hunting bottlenecks throughout production instead of at the end of it.

Seventy people took the engine, switched off the two features it was marketed on, kept the parts that eliminated drudgery, and never stopped measuring.

Epic’s engineers gave a talk at Unreal Fest 2025 enumerating seven root causes of hitching in UE5 titles: level streaming and inefficient static mesh use, PSO compilation, garbage collection, heavy synchronous loading, and unoptimized Blueprints among them.

Read that list as a practitioner. Not one of those is a rendering defect. Every one is an engineering discipline failure, the kind that appears when nobody owns a frame time budget. The engine makes them easy to hit, in fairness. A fresh UE5 project turns on Lumen, Nanite, Virtual Shadow Maps, reflections, and temporal anti-aliasing by default, and switching a feature off can leave hidden dependencies running and quietly eating resources. That is a real footgun, and it is also one that a team with a profiler open does not step on, which is why Embark’s game runs well and a number of far better-funded ones do not.

Sweeney’s public position has been that the responsibility sits with developers who target high-end hardware first and treat optimization as a late-stage chore. He is mostly right, and he still had to cut a thousand jobs.

The Steam top ten in January 2026: Counter-Strike 2, Dota 2, PUBG, ARC Raiders, Apex Legends, Rust, Bongo Cat, Stardew Valley, Delta Force, Grand Theft Auto V.

Two of those run on Source 2, which shipped in 2015. Rust is a Unity survival game from 2013. Stardew Valley is a farming sim one person wrote. GTA V is thirteen years old. Bongo Cat is a free application in which pressing keys makes a cartoon cat punch your taskbar, and it outranked most of what the industry shipped that year. Roblox is running about 132 million daily active users on its own engine.

Visual fidelity at the top of the charts is trending down, not up. Whatever players are optimizing for, triangle count has never been on the list.

The uncomfortable reading of this is that the tooling worked exactly as advertised and the industry could not metabolize it. Nanite deleted the LOD pipeline. Lumen deleted the bake. Both were genuine wins, handed over free until a game was already earning.

What Epic could not ship was judgment about what to do with the capacity it freed up. Studios that already had that judgment turned the tooling into margin and shipped on time. Studios that had been using budget constraints as a substitute for product decisions lost the only mechanism forcing them to choose, expanded scope until the org chart needed fourteen layers, and then found out the loop closes regardless. It just closes on the people making the games rather than the people approving the budgets.

The engine was never going to save anyone. It is a renderer. What it removed was the excuse, and a large part of this industry turned out to need the excuse.

That is worth holding onto, because the industry is about to run the experiment again with a much bigger tool.

Generative AI is the same event moved up the stack. Unreal Engine 5 removed the constraint on what an asset cost to produce. AI removes the constraint on whether a person produces it at all. And the Steam data already shows how that resolves, because the experiment has been running for three years in public.

Sulka Haro ran a census of 53,597 Steam releases between mid-2023 and mid-2026, checking every store page for Valve’s AI disclosure. Games carrying the flag went from 10.9 percent of releases in 2024 to 19.9 percent in 2025 to 30.8 percent so far in 2026. AI-flagged launches climbed from roughly 13 a month before the disclosure mandate to about 530 a month. Non-AI launches grew modestly across the same window, from around 1,030 to 1,320. Between 60 and 90 percent of Steam’s release growth is AI-flagged titles, and almost none of them earn anything. The median paid game launching in early 2026 made about $350 across its entire lifetime.

Then there is the split inside that number. AI-disclosed games are a third of new releases but only 10 to 27 percent of estimated sales. Among the flagged games that failed to find an audience, 72 percent had used AI for visuals, which is the application players look directly at. The ones that succeeded used it where nobody is staring: voice at 24 percent against 8, localization at 18 percent against 6. Their disclosure text reads differently too, full of words like auxiliary and reviewed and refined.

Same technology. The cohort that treated it as a multiplier on craft made money. The cohort that treated it as a substitute for craft produced 530 games a month into a store where roughly one percent of titles take ninety-four percent of the revenue.

GDC’s 2026 survey caught the sentiment turning over. Fifty-two percent of industry professionals now think generative AI is having a negative effect, up from 30 percent a year earlier and 18 percent the year before that, with the sharpest opposition among visual and technical artists at 64 percent. But the number underneath is the one I would watch. Upper management reports using AI at 47 percent. Individual contributors report 29. The people who approve budgets believe in the tool at nearly twice the rate of the people who would have to make it work.

That gap is where the next round of cuts gets authorized, and it is the same gap that produced fourteen layers of management. An executive looking at generative AI sees a cost line disappearing. An artist looking at the same tool sees a constraint being removed from a pipeline that already lost its last governor in 2022.

Embark is instructive here as well. They used machine learning for the ARC machines’ locomotion and AI text-to-speech to let 70 people produce at the volume they do, while their chief creative officer has said plainly that they do not use generative AI and have no goal of replacing performers. They pointed the technology at work nobody wanted and kept people on the work players actually notice. That is the same decision they made about Lumen and Nanite, applied to a different tool.

Sweeney went out of his way in the Epic layoff memo to state that the cuts were not AI-related, and I believe him. AI did not cause this downturn. What it offers is one more chance to remove a constraint without installing anything in its place, arriving at the exact moment the industry finished demonstrating what it does with that opportunity.

The tooling keeps getting better. It has never once been the variable.

Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and Chief Architect at Prana Entertainment in Las Vegas. He has spent close to thirty years shipping production software on the Microsoft and .NET stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy, and Allegiant Air. He currently builds production agentic AI systems, including MCP servers, Microsoft Agent Framework implementations, RAG pipelines, and SQL Server 2025 vector search, and maintains PLogger, an AI-powered diagnostic and observability framework. He was employee number six at Break.com, co-hosts Edge Grip Podcast, trains Brazilian jiu-jitsu in Las Vegas, and is the author of the novel The Archive of Lost Suns.