miHoYo Goes All-In on AI While 85.8% of Japanese Developers Use It, With a Gap in Realized Benefits: Two Faces of the AI Arms Race#

Hello everyone, I’m Wubian. Over the past two weeks, news about generative AI in the games industry has been unusually dense: on one side, miHoYo founder and president Liu Wei said publicly at a campus-recruitment event that within two to three years miHoYo will become “a consequential member of the domestic large-model community”; on the other, a survey by Japan’s Computer Entertainment Supplier’s Association (CESA) shows that 85.8% of Japanese game developers are already actively using generative AI at work, a big jump from 51% the previous year.

These two stories look like east and west, one company and one industry, but they are really two sides of the same thing: on the investment side, companies are betting real money; on the benefit side, developers are using the tools, yet the sense of payoff has been slow to reach players. As someone who usually writes about mobile and online games and follows large-model deployment in my spare time, I find putting these two sets of facts side by side more interesting than reading either alone. All data in this article comes from public reporting and verified figures from the CESA survey; where officials have not disclosed anything I will mark it clearly as “not disclosed” and will not speculate.

1. Two News Items Within One Week, Two Sides of One Story#

First, the timeline. The two developments are ten days apart — one representing “the people spending money”, the other “the people using the tools”:

Read alone, the first item makes you think AI has fully penetrated the games industry; read alone, the second sounds like one company’s bravado. Read together, the question becomes: with investment this large and usage this widespread, how much of it ends up becoming something players can actually feel? CESA’s data happens to provide a cold footnote to that question.

2. The Investment Side: miHoYo Treats AI as a Main Quest#

2.1 From Investment to In-House Ambition: A Five-Year Timeline#

miHoYo’s AI build-out is not a September 27 slogan; it is a rhythm of continuous, verifiable moves. Laying the public record out in time order:

The 2023 node deserves a pause. Large-model technology was applied to intelligent NPC dialogue and dynamic story generation — and filed under core gameplay. This is not an office productivity tool; it is putting model capability directly into a live-service product. For a company whose strength is the content-update cadence of anime-style mobile games, NPC dialogue and story generation are extremely high-consumption components, and getting model capability into the product line means far more than publishing a few technical blog posts.

2.2 Liu Wei’s Own Words: Confidence, Direction, and Compute#

Three sentences from Liu Wei at the recruitment event basically lay out the company’s posture:

• “I have strong confidence that for the next 2 to 3 years, miHoYo will be a consequential member of the domestic large-model community” — the goal explicitly benchmarked against the top tier of domestic large-model teams;
• “I hope to do more cool things. As long as I am here, miHoYo will never lose its direction” — positioning AI as a direction personally bet on by the founder;
• “Any team that does not firmly pursue compute and scale can never make a top-tier model” — laying the groundwork for subsequent investment in compute and scale.

The third sentence carries the most information to my eye. Everyone in AI research understands that compute and data scale are the entry ticket to the large-model race; this sentence amounts to a public admission that miHoYo is taking the capital-heavy in-house-model route, not the lightweight play of “hooking into someone else’s API and building a few applications”. At the May sharing session, Liu Wei also gave a more concrete figure — AI investment of “up to 100 billion over 3 years” — adding that if it ultimately fails, “so be it; consider it a big firework”. Upper-bound framing and a lose-gracefully posture, both on the record.

2.3 What 100 Billion Means: Half the Valuation#

The figure “up to 100 billion over 3 years” needs a reference point to have any weight:

In other words, the investment ceiling miHoYo has set for its AI strategy equals roughly half the company’s total valuation. That is not a ratio games companies usually assign to a single technology direction. Of course, one emphasis: this is an upper-bound framing, not a spending commitment — the funding structure, phasing, and accounting treatment have not been disclosed. I have seen analyses book it directly as “money already decided to be spent”, and that is a misreading.

2.4 Organization and Hiring: More Honest Signals Than a Press Event#

Press events can trade on sentiment, but org charts and hiring lists do not lie. In June, miHoYo established two wholly owned subsidiaries:

Meanwhile, miHoYo’s job site currently lists more than 50 large-model-related positions, covering model training, serving, data, audio, agents, and other directions. I went through the listings; a few representative ones:

• Large-Model Inference Platform R&D Engineer — building online services, not papers;
• Large-Model Training Platform R&D Engineer — the standard position for in-house training infrastructure;
• Senior AI Storage R&D Engineer (Large Models) — building the storage layer for massive training and inference data;
• Researcher, General Audio Understanding Models — pointing toward multimodal capability in speech and sound;
• LLM Data Processing / Agent Development Engineer (data track) — data engineering and agent applications.

Training, inference, storage, data, audio, agents — this hiring list assembles the complete silhouette of a model company. One boundary to note: the product name, parameter scale, and release timing of miHoYo’s in-house large model have not been announced; all talk of “in-house model progress” so far remains at the level of statements and organizational build-out. I treat it as a construction project under way, not a finished result.

2.5 Why It Can Afford a Bet This Big: The Revenue Base#

A heavy bet presupposes money and use cases. On revenue: according to AppMagic, miHoYo placed three titles in the global top ten of the 2025 anime-style mobile game revenue chart, with Honkai: Star Railway leading at US$420 million. On use cases: the intelligent NPC dialogue and dynamic story generation mentioned above have been running for years. In other words, the company simultaneously has “the cash flow to afford a 100-billion ceiling” and “its own product lines capable of absorbing model capability” — that is the confidence behind upgrading AI from a tool to a strategy. By the way, the competitive intensity of the Chinese anime-mobile track, where Genshin Impact sits, is itself pushing companies to pour money into technology — in this industry, innovate or fall behind has never been a figure of speech.

3. The Benefit Side: CESA’s “High Adoption, Low Realized Benefit”#

3.1 The Survey’s Basics#

Switch the lens to Japan. On September 17 — the opening day of Tokyo Game Show 2026 — CESA released a preview edition of the CESA Games Industry Report 2026, which for the first time included a dedicated survey on generative AI. The data comes from the CEDEC developer questionnaire, with 1,349 valid responses, mostly from commercial game developers in Japan. The full report is not due until December. Keep in mind the “individual questionnaire” attribute; it matters when we discuss the survey’s limits.

3.2 Adoption: From 51% to 85.8% in One Year#

The core individual-level numbers:

85.8% combined with 63% daily usage says generative AI on Japanese game development floors is no longer a novelty but a daily tool, like a search engine. The genuinely striking row is the last one: with this many users, only a single-digit share of companies have issued rules — the tools are running ahead of the institutions.

3.3 The Most Striking Set of Numbers: 837, 738, 147#

To my mind the most valuable part of this survey is the benefit-perception set. Among the 1,349 responses:

Within the same questionnaire, the number of developers who believe AI helped their own production exceeds the number who believe it improved the player experience by more than a factor of five. Most respondents agreed that AI shortens production cycles and cuts costs, which is consistent with the 837 and 738 figures; but only 147 acknowledged “players actually got better games”. This is the benefit gap I keep mentioning: adoption is sky-high, the benefits stay inside the production line, and they have not reached the players.

3.4 Where It Is Used, and How Companies Govern It#

On use cases: developers use generative AI both for art asset generation and for general administrative tasks with tools like ChatGPT and Copilot — half creation, half chores, consistent with what many industries have observed.

The company-side data comes from a separate CESA survey of member companies:

Companies’ biggest expectation is improving work efficiency and productivity, followed by shortening development cycles and lowering development and operating costs; the most widespread concern is copyright and intellectual property. In terms of governance, the most common practice is manual confirmation, editing, and review of AI output, plus designating or restricting usable tools and avoiding direct adoption of raw AI output. Putting the individual-side 85.8% next to the company-side “27 not using, 26 using” reveals a clear layering: in individual hands the tool is everywhere; at the corporate decision layer it becomes a controlled resource requiring approvals at every step. Art assets get particular caution — only 12 companies dare apply it to visual assets, which is precisely where copyright risk is most concentrated.

3.5 The Boundaries of This Survey#

Before citing these numbers, the caveats must be stated — they were my biggest concern reading the preview edition:

• The sample unit is the individual developer; multiple respondents from the same company are counted separately, so the numbers cannot reflect company-wide adoption;
• Whether last year’s 51% and this year’s 85.8% were measured on the same basis is not clarified in the report, so the year-over-year conclusion needs discounting;
• The full methodology (question wording, company vs. individual share, etc.) must wait for the full report in December.

So the rigorous statement is: the direction — a large rise in adoption — is certain, but the precise magnitude of “51% to 85.8%” deserves some reserve. The 147-vs-837 gap is also subject to these caveats, but even allowing for error, the direction of the misalignment between production-side and player-side benefit perception will not change.

3.6 The Industry Friction Behind the Gap#

The cautious attitude has its reasons. Just last month, Level-5 president Akihiro Hino publicly apologized for generative AI content used at a showcase, saying experimental use had unsettled players; earlier, Crazy Taxi: World Tour also faced player pushback over AI content. With player sensitivity to AI-generated content this high, even developers who have tasted production-side gains must weigh the reputational cost before serving it up — itself part of the reason the 147 figure looks the way it does.

4. The Middle-Layer Signal: AI Asset Labeling Is Becoming a Store-Page Standard#

Between investment and benefit there is an easily overlooked middle-layer signal: transparency about AI assets is becoming an industry convention.

NC’s AION 2 Steam store page explicitly states that “some in-game assets were created with the assistance of AI tools”, specifically naming Arcana card design and some icons as AI-generated assets. The game’s schedule: early access opens September 30, global launch October 5, free to play, available on PC via Steam and the PURPLE platform. Kotaku has already listed it among this year’s Gamescom titles whose store pages disclose varying degrees of generative AI use.

Why do I care about this detail? Because it moves “what the AI did” from a fuzzy zone inside the industry to a compliance statement consumers can see directly. Combined with Level-5’s apology above and the Crazy Taxi: World Tour boycott, game companies now have only two respectable options facing AI content: label it in advance, or apologize afterward. A one-line AI asset note on a store page is degenerating from a creative manifesto into a compliance statement — and once labeling is standard, the data on how players vote with their feet becomes more honest: accept assisted-generation assets or reject them; the market will answer.

5. Two Postures, One Unanswered Question#

5.1 Path Comparison#

Setting miHoYo and the Japanese industry side by side in one table, the difference between two AI postures is plain:

One is betting money and headcount on a future of model capability; the other is using the tools to the hilt while boxing every step in policy. The former wagers that “only a top-tier model yields the next generation of products”; the latter guards against “will using AI cause trouble”.

5.2 The ROI Question: The Time Lag Is the Key Variable#

Back to the opening question: what stands between the heavy betting and the player-perceivable benefit?

CESA’s data gives a snapshot: 837 people acknowledge AI expanded creative space, 738 acknowledge technical barriers are falling, but only 147 believe the player experience improved. This misalignment points to an observation the industry keeps re-verifying: AI improving production does not mean AI improving the product. Shorter cycles and lower costs are supply-side matters; players buy the quality and experience of the content itself, with the gates of creativity, quality control, taste, and public sentiment in between.

miHoYo’s path is, in fact, a direct response to this misalignment: rather than optimizing cost and efficiency inside existing production pipelines, bet on model capability itself, so that core gameplay like NPC dialogue and story generation grows directly out of the model — MiniMax technology entering Honkai: Star Railway core gameplay in 2023 was the first landing of this idea. The problem with this path is equally obvious: the product name, parameter scale, and release timing of the in-house model are all unpublished, and no one can promise the time lag between organizational build-out and a player-perceivable experience upgrade. The 100 billion is a ceiling, the firework is a joke; the real question is whether that lag is measured in one year or three — Liu Wei’s own framing is “a consequential member of the domestic large-model community” within 2 to 3 years, which is a promise about model capability, not a promise about player experience.

Two watchpoints here: first, the December release of CESA’s full report, when the methodological basis will be completed — whether the 147 figure holds up and whether company-side data gets revised are both worth re-checking; second, the hiring progress on miHoYo’s 50+ positions and the subsidiaries’ follow-up moves — the pace of organizational expansion often reveals a strategy’s real priorities earlier than any press event.

Closing#

Pulling the two threads together: on the investment side, miHoYo has used an investment ceiling of roughly half its valuation, 50+ positions, and two subsidiaries to elevate AI from tool to company strategy; on the benefit side, Japanese developers’ 85.8% adoption proves AI is already an everywhere-tool, while 147 benefit perceptions admit it has not yet become a good game in players’ hands. In the middle layer, Steam store-page AI asset notes have started laying this whole matter out for every player to see.

Having followed AI deployment for these years, I find “arms race” an ever more fitting term here: the two sides compete on input, not output; the scoreboard hangs inside the company where players cannot see it. The day a player can say of a game, “that NPC’s reaction felt alive to me”, the race will have its first referee. Until then, both the 100 billion and the 85.8% are process numbers.

— Wubian, NTGame Editorial