The screen economy to 2040: how cheap video generation will reshape the $411bn market for games, film and streaming

01 · MoneyWhat the screen costs: $581bn of consumer spending in 2025

In 2025 people paid about $581bn for games, streaming, pay-TV and cinema tickets. The largest segment is games: $201.6bn by Newzoo’s count 5 M. Streaming took second place for the first time: subscriptions and purchases of online video (excluding advertising) brought in $176bn and overtook pay-TV ($170bn, down 4% on the year) 6 M. Box office is $33.6bn 7 H, only 6% of that sum. Without pay-TV, games, streaming and cinema come to $411bn.

To this sum one should add the advertising that pays for free video. In 2025 YouTube received more than $60bn from advertising and subscriptions combined 8 H, more than Netflix ($45.2bn) 9 H. By PwC’s count, global advertising in media and entertainment exceeded $1trn for the first time in 2025, and the whole media and entertainment industry, together with internet access, is valued at roughly $3.5trn 10 M.

The other side of the ledger is production spending. By Ampere Analysis’s estimate, in 2026 companies will spend $255bn on films, series and broadcasting rights, 2% more than a year earlier; $101bn of that comes from streaming services, which in 2025 overtook commercial broadcasters for the first time 11 H. By our estimate (nominally +0.4–2% a year against inflation of 2–3%), in real terms this spending has not grown for several years L. This is the budget line on which generation has the greatest effect, through animation, effects, post-production and localisation; it does not make sports rights cheaper.

02 · Twenty yearsGames grew sevenfold and Netflix 66-fold (in nominal dollars), while the box office is still below its 2019 level

Twenty years ago the games market was valued at $27bn 12 L. In 2025 it is $201.6bn: 7.5 times more in nominal dollars and 4.5 times more adjusted for inflation 5,13. This is an order of magnitude: the estimates for 2005 and 2025 were made using different methodologies L. Over the same period Netflix turned from a DVD-by-post rental business into a company with revenue of $45.2bn, 66 times more than in 2005 ($0.68bn), and roughly 40 times more in 2025 prices 9 H.

The box office took a different path. The record was $42.3bn in 2019; in 2020 takings fell to $11.8bn, and in 2025 they recovered only to $33.6bn 14,7 H. Adjusted for inflation, that is 37% less than in 2019. By our estimate (takings of $8.87bn at an average price of $11.31), about 0.78bn tickets were sold in the US in 2025, 43–45% fewer than in 2005 14,15 L. In nominal dollars, revenue is held up by price: the average ticket rose from $6.41 to $11.31 15 M; in 2025 prices, US takings fell by almost 40%. The cinema has given way to the home: in 2025 Americans spent $62.2bn on home viewing, 7–8 times more than on tickets (depending on whose box-office data are used), and more than 92% of that sum went on subscriptions 16 M.

On American television sets, streaming has almost caught up with the rest of TV: in August 2026 it accounted for 49.6% of US television usage, of which 15.8% was YouTube and only 7.7% Netflix 17 H (the latest data under Nielsen’s old methodology). Netflix, the largest subscription streamer by revenue, gets half the viewing time of the platform where users make the video themselves.

03 · ProfitWho makes money: Netflix has a 29.5% margin, the other streamers have either only just turned profitable or still lose money, studios earn 5% to 20%

Netflix earned $13.3bn in operating profit on $45.2bn of revenue in 2025, a margin of 29.5% 9 H. For 2026 the company forecasts 31.5% (the forecast includes about $275m of costs for the failed Warner Bros. deal), and in the third quarter of 2026 it reported 34.6%, which management attributed mainly to the timing of costs 18 H. Disney, WBD and Paramount have only recently turned profitable: Disney's streaming business lost $3.4bn in fiscal 2022 and earned $1.3bn in 2025 19 H. Comcast's Peacock still lost about $1.1bn in 2025, and over 2020–2025 it spent roughly $10.5bn more than it earned 20 M.

Studios that sell films and series earn very different returns. Disney's content sales and licensing division made $0.39bn of profit on $8.49bn of revenue, or 4.6% 19 H. Warner Bros. Discovery's studios made $2.55bn of adjusted EBITDA on $12.6bn, or 20% 21 H. Comcast's studios (Universal) made $1.1bn of adjusted EBITDA on $11.3bn, 21.7% less than a year earlier because marketing costs rose; the box office provided only 14% of their revenue, and Comcast's theme parks earned almost three times as much as the studios 20 H.

The main point of this chart is the largest item. Content amortisation eats 36% of Netflix's revenue. But the company's cash spending on content has barely changed since 2021 ($17.5bn → $17.7bn in 2025), while revenue has grown by 52%; cash content spending as a share of revenue fell from 59% to 39% 9 H. The margin grew because revenue grew faster than content costs, thanks to price and scale; advertising brought in more than $1.5bn in 2025, about 3% of revenue (the company expects ~$2.7bn for 2026) 18 H. If video generation cut content costs by 20%, that would be ~$3.5bn a year, about 8% of revenue. Morgan Stanley estimates possible savings in film and TV production of “up to 30%”, and about 10% for the media industry as a whole, with no clear time horizon 22 M.

The industry's answer to fixed costs is consolidation. In December 2025 Netflix agreed to buy the Warner Bros. studios and HBO Max for $82.7bn 23 H, but in February 2026 it declined to outbid Paramount Skydance ($31 a share for the whole company; estimates range from $81bn to $110bn including debt). On 6 October 2026 Paramount completed its takeover of WBD 24 M. The games publisher EA has been a private company since 4 August 2026, bought out by the PIF fund, Silver Lake and Affinity Partners 25 H.

04 · Cost of a minuteFrom $2m in a blockbuster to a few dollars in a generator

To compare films, series and video generation, it helps to count the cost of one finished minute. The most expensive are feature-length blockbusters and Pixar animation: about $1.9m per minute for Avatar: Fire and Ash and $2.45m for Toy Story 5 ($250m for 102 minutes) 1,26 M. The most expensive series, The Rings of Power (~$58m per episode) and Citadel, cost $0.8–1.3m per minute; House of the Dragon (~$20m per episode) costs about $0.3m 27 M. A typical episode of a TV anime, by an unconfirmed estimate in the Japanese press, costs 20–30m yen, or about $5,000–8,000 per minute L; the most expensive projects reach 300m yen (~$1.9m) per episode 28 M.

The cheapest documented finished product is not an API price list but a Chinese turnkey AI drama. Some studios in Hangzhou, according to Caixin, make a 100-minute drama for less than 20,000 yuan (~$3,000), that is, for under $30 a minute; these are selected examples, without promotion costs 29 M. An American example with people involved: a 3-minute animated episode built from 164 generated clips, of which 41 reached the final cut, took two people two days and cost about $950, or ~$315 a minute; for other videos the same producer cites up to $750 30 L. An advert for Kalshi shown during the NBA finals in 2025 was made in two days for $2,000 on Veo 3 (excluding the cost of airtime and the creator's fee); the 15 clips used required 300–400 generations 31 M.

The documented savings in professional production are more modest and concern individual scenes. According to Netflix, the building-collapse shot in The Eternaut was made ten times faster than with conventional visual-effects tools 32 L, and 17 minutes of the documentary series The American Experiment were made twice as fast and at half the cost 33 L. In the second season of House of David (Amazon), according to Wired, AI was used for 350–400 visual-effects shots; one sequence (episode 6), according to the series' creator in a Runway case study, was made in two weeks instead of four to five months 34 L. We found no overall figure for savings across a whole production in public sources 33 M.

The animated film Critterz, made with OpenAI tools, was planned on a budget of under $30m and nine months, instead of the usual $100–200m and three years 35 L; in May 2026 it was presented at the Cannes film market, with no release date 36 L. The company Higgsfield claims that its 95-minute film Hell Grind was made by 15 people in 14 days for roughly $500,000 37 L.

05 · TechnologyWhat video generation can do in October 2026: up to 30 seconds per pass, ten minutes in the lab

The best public models, by their developers’ claims, generate up to 30 seconds in a single pass — Alibaba’s Wan 3.0 38 M and ByteDance’s Seedance 2.5 39 M. Google’s Gemini Omni Flash builds a scene in steps of 10 seconds up to a total of 40; sound is generated together with the picture 40 H. Wan 3.0, according to its developer, accepts documents as input — a PDF or text file of up to 50 pages — so it already has a “document to video” mode; there is no independent check of how well it holds a plot 38 M.

Long coherent films so far exist only in research. On 24 September 2026 Google Research described a multi-agent system built on Gemini and Veo that produced a continuous 10-minute film 41 M. Nvidia’s LongLive model generates up to 240 seconds in a single stream 42 H. The forecasting platform FutureSearch puts at 11% the probability that by 1 January 2028 a system will appear that creates a feature-length film of big-studio quality from a text prompt 43 M.

The price per second is falling, but nobody has yet shown a profit

Over 13 months Google cut the price of a second of video for developers from $0.75 (Veo 3, July 2025) to ~$0.10 (Gemini Omni Flash, 2026) — roughly 7.5 times, although these are different models 44,2 H. At the cheapest tier the fall is from $0.40 (Veo 3 Fast) to ~$0.03 per second of 360p on the price list of Runway, which resells Google’s models 45 H. Runway has not cut its prices: its cheapest tier is $0.05 per second in both 2024 and 2026, and its flagship rose from about $0.10 to $0.12 45 M.

The main event of the year in this market is the closure of Sora. OpenAI announced it on 24 March 2026; the app was shut down in April 46 M, and the API on 24 September 47 H. OpenAI explained the decision by the need for computing resources and a change of priorities; app downloads in December 2025 fell by 32% compared with November 46 M. Other companies are growing but are not showing a profit: revenue of Kuaishou’s Kling in the second quarter of 2026 exceeded 850m yuan, more than 200% up on a year earlier 48 H (according to Kling itself, about 70% came from abroad 29 L); MiniMax earned $116.6m in revenue and made a net loss of $358m in the first half of 2026 49 M; Runway raised $315m at a valuation of $5.3bn 50 M.

Leadership changes within months. In the Artificial Analysis ranking, where people compare clips blind, the top places in October 2026 are held, within the statistical margin of error (±15 points), by Wan 3.0, Utopai X, HeyGen Video 1, MiniMax H3 and Seedance 2.5, while Google’s model is ninth or tenth 51 H.

Sources: documentation from Google, OpenAI, Alibaba, ByteDance, Lightricks, Tencent and MiniMax; Runway’s price list for resold models; The Batch. “—” means the developer does not publish the figure.

06 · ChinaA preview of the 2030s: 430,000 micro-dramas in eight months, more than 90% made with AI

The market model closest to “zero-cost” production already exists — Chinese vertical micro-dramas: episodes of one or two minutes made for the smartphone. According to China’s National Radio and Television Administration (NRTA), the micro-drama market exceeded 100bn yuan in 2025 (about $14.9bn by Global Times’ conversion), with more than 800m users 53 M. For comparison: China’s cinema box office that year was 51.8bn yuan ($7.4bn) 7 H; NRTA does not specify what counts as part of the micro-drama “market”, so the comparison is approximate.

In 2026 generation entered this market. From January to August 430,000 new micro-dramas were released — 13 times more than in the whole of 2025, and more than 90% of them were made with AI 53 M. According to the China Netcasting Services Association, about 275,000 AI series were released in the first half alone — roughly 1,519 a day 54 M. Making a short AI series, by Fortune’s examples, costs less than $3,000 before promotion costs; some studios in Hangzhou make a 100-minute drama for less than 20,000 yuan, against several hundred thousand yuan and six to eight weeks for a live-action shoot 29 M.

The result: the number of titles has soared, while the share of hits among them, by indirect estimates, remains small L. The share of AI series in the 100 most popular in January 2026 was 38% (a year earlier, 7%) 55 M, although they make up more than 90% of new releases. By BigGo Finance’s count, fewer than 1.3% of AI series in the first half of 2026 reached 50m views — the threshold regarded there as the break-even point 56 L. On 23 October 2026 “Sanxingdui: Future Past” opened in cinemas; the film company Bona presents it as China’s first fully generated feature film: more than two years of work by almost a hundred people and a budget of more than 10m yuan. Advance sales came to 65,000–80,000 yuan, and the forecast for total takings is below 100,000–150,000 yuan 57 M.

In China, 13 times as many micro-dramas came out in eight months of 2026 as in all of 2025, and, by BigGo’s estimate, fewer than 1.3% of AI series reached 50m views.

07 · AttentionSupply grows many times over; attention grows by a few per cent

The market's constraint is not cameras or screenwriters but the 24 hours in a day. Americans aged 15 and over watch television for an average of 2.6 hours a day; games and leisure computer use take 37 minutes (ten years ago, 25) 58 H. In the first half of 2026 Netflix set a record: 97bn hours viewed 3 H. But 2023 as a whole saw 183bn hours, or about 93bn per half-year (Netflix's first report spoke of “almost 100bn”): over three years, growth of between zero and 4%. Netflix's revenue growth (+15.8% in 2025) is growth in revenue per hour viewed (prices and advertising), not in the number of hours (+2%) 9 H.

Music has already gone through zero production cost. At the end of 2025 streaming services held about 253m tracks, with ~106,000 added every day; 88% of tracks had no more than a thousand plays in a year, and almost half had between 0 and 10 59 M. In June 2026 fully generated tracks exceeded half of new uploads on Deezer — almost 90,000 a day on average — but drew only 1–3% of plays; by Deezer's estimate, up to 85% of plays of such tracks in 2025 were fraudulent 60 H. Meanwhile recorded-music revenue has grown for an eleventh year in a row — $31.7bn in 2025 61 H. The money has come back, but it is very unevenly distributed: Spotify paid out $11bn in 2025, and ~13,800 artists received more than $100,000 — about 1,400 more than a year earlier 62 M, while 88% of tracks have no more than a thousand plays.

Games show the same pattern. By the estimate of analyst Sulka Haro, the top 1% of games on Steam take ~94% of revenue 63 M. Netflix shows the other side: in the second half of 2024 no single title drew more than 1% of viewing, and almost half of viewing of Netflix's own films and series in the first half of 2025 went to titles released in 2023 or earlier 3 H. The catalogue ages slowly: new work has to compete with everything already made.

08 · BooksWill most novels be adapted for the screen? Technically yes, publicly no

In the US alone, new books received 4.17m ISBNs in 2025 64 M — 32.5% more than a year earlier. Self-publishing accounted for almost all of the increase (+38.7%; traditional publishers +6.6%), and 85% of the numbers were obtained by the authors themselves; whether this is real growth or a change in counting, Bowker does not say. Fiction accounts for 807,000 of these numbers (an ISBN is assigned to each format separately, so the number of titles is smaller). In 2010 Google counted 129.9m distinct book editions in all of printing history 65 M.

Demand for adaptations is documented too. In a study by Frontier Economics commissioned by the UK's Publishers Association, British films made from books in 2007–2016 earned 53% more at the worldwide box office than films based on original screenplays; about half of the 20 highest-grossing British films of that period (52%, by the publishers’ count) were based on published works, books and comics 66 H. This shows that adaptations of well-known works earned more, but not that any book would do as well: studios already pick the most successful titles, and the study was commissioned by an industry body. A shortage of space in studios' schedules is a plausible but unproven explanation of why adaptations are rare: while a minute of a series costs hundreds of thousands of dollars (section 04), only a tiny share of published books gets adapted.

A precedent that has already happened: audiobooks

A synthetic voice made the audio version almost free. In May 2025 Audible had more than 50,000 books narrated by Amazon's Virtual Voice service 67 M, and Audible itself offered publishers more than a hundred synthetic voices and a beta translation. The result a year on, according to a survey of publishers by the APA: AI-narrated books brought in 0.03% of US audiobook market revenue ($2.43bn; revenue from self-published titles may not be included), only 16% of listeners had ever tried them, and willingness to try fell from 70% to 61% 68 M.

This is the closest documented precedent, if a weak one: the publishers’ survey probably does not cover self-published Virtual Voice titles. Supply grew by tens of thousands of titles, while measured revenue from AI narration is 0.03% L. Whether this carries over to screen adaptations remains to be seen.

What a book is worth as data and as rights

The price of a book as training data has already been named. In 2024 HarperCollins offered $5,000 a title for a three-year licence to train models, split equally with the author (participation subject to the author's consent) 69 M. In Bartz — the authors' lawsuit over training on books from pirate libraries — Anthropic agreed to pay $1.5bn, about $3,000 for each of ~482,000 books; the court gave final approval to the settlement on 20 July 2026 70 H. The right to a screen adaptation is a separate right, and generation does not make the rights-holder's permission unnecessary.

Demand for “their own versions” of other people's stories is documented too. As of 25 October 2026 the fan-fiction archive AO3 holds 18.8m works (mostly short) in more than 83,000 fandoms and has almost 12m registered users 71 H. These are text adaptations that people make with no budget at all. Video generation will let this audience make their own video versions. Fan fiction exists in a legal grey area: rightsholders mostly tolerate it because it is non-commercial. A public video based on someone else's novel would probably not keep that status (our reading, not legal advice).

The calculator's arithmetic gives three conclusions. First: under its assumptions, production cost ceases to be the main barrier L. With hybrid production, a film based on a novel would cost about $280,000 in 2026 and ~$120,000 in 2030 — not tens of millions, as now. Second: only the top pays back. Even among the 10,000 most popular novels, if they are given 5% of Netflix-scale viewing, at 2030 prices roughly one adaptation in five pays back, and at 2026 prices one in twelve. Third: adapting everything is unrealistic. Films based on all new US fiction in a single year (807,000 ISBNs is an upper bound, since each format gets its own number) would cost ~$99bn at 2030 prices, or 39% of global content spending in 2026 (~$226bn at 2026 prices), and the median title would draw about 1,700 hours of viewing, or ~$400 of revenue.

09 · Home studioWhat you can generate on your own computer, now and by 2040

Short answer: short scenes are possible today; a feature-length film on a single graphics card means months of night-time work in 2026 and, if the fall in computing costs continues, weeks in the early 2030s. The weakest link is not the model but video memory, hardware prices and the human time spent editing.

What already works at home (October 2026)

• Wan 2.2 (Alibaba, open Apache 2.0 licence): the 5bn-parameter model produces a five-second 720p clip in under nine minutes on a single consumer graphics card; the smaller Wan 2.1 needs about 8.2 GB of video memory and creates five seconds of 480p in ~4 minutes on an RTX 4090 without optimisation 72 H.
• LTX-2 (Lightricks): video and sound in a single model, open weights. For a workflow on LTX-2.3, Nvidia recommends at least 16 GB of video memory and 64 GB of RAM 73 H; version 2.5 (August 2026) shoots several shots in a row and outputs up to 4K at 50 frames per second 74 M. The licence is free for companies with revenue under $10m 75 H.
• HunyuanVideo 1.5 (Tencent): 8.3bn parameters, from 14 GB of video memory; the accelerated version produces a 480p clip in 75 seconds on an RTX 4090. But the licence does not apply in the EU, the UK and South Korea 76 H.
• FramePack: minute-long clips on a card with 6 GB, at the cost of speed: about 1.5–2.5 seconds of computation per frame on an RTX 4090 77 H.
• Near-real-time generation already exists, so far at low resolution: Self Forcing produces ~10 frames per second at 480p on an RTX 4090 78 H, and Nvidia's LongLive 2.0 produces 45.7 frames per second on a server-grade GB200 in a simplified two-step variant 42 H.

The rest of the pipeline is also local. The Qwen3-Next language model has a context of 262,000 tokens, enough for a whole novel in one request 79 H; the open Chatterbox voices clone a voice from a short sample in 23 languages 80 H; ACE-Step 1.5 writes a song in under 10 seconds on an RTX 3090 and fits into 4 GB 81 H. What is missing is a system that keeps characters and plot consistent across hundreds of scenes. ViMax, the best-known open “novel to video” agent, still calls cloud models 82 H, and training a custom character model (LoRA) for Wan 2.2 takes 4–10 hours on an RTX 4090 83 L.

The bottleneck in 2026 is memory, not electricity

A shortage of memory, which manufacturers have redirected to data-centre chips, has reversed the usual trend of falling hardware prices. The RTX 5090 launched in January 2025 at a recommended price of $1,999 84 H and in September–October 2026 cost $4,300–6,000 85 L. Nvidia has raised the price of the DGX Spark with 128 GB of memory twice, from $3,999 to $4,699 in February and to $6,950 on 2 October 2026, an increase of ~75% in a year 86 M. According to the Steam survey (September 2026), 37% of computers have 16 GB of video memory or more and 1.9% have 32 GB or more; of these, 0.5% are a new 64 GB group, probably computers with shared memory 87 H.

Electricity is a minor cost. A 575 W card running for 9 minutes uses 0.086 kWh, which is 1.5 cents at the average US household tariff in 2025 88 H. In the long run, the cost of computing halved every 2.46 years (Epoch AI, 2006–2021) 52 H, but that series ended before the current shortage. Now software optimisation is doing the work: distillation cut HunyuanVideo's time on the RTX 4090 by 75% 76 H, and the NVFP4 format on RTX 50 cards gives, according to Nvidia, up to a threefold speed-up and up to 60% memory savings 89 M.

10 · GamesAI makes games easier to release but does not add player time

Games are the largest on-screen entertainment segment by consumer spending (excluding advertising). The market in 2025: $201.6bn (+9.1%), above $200bn for the first time; mobile games made $113.3bn, more than consoles and PC combined 5 M. Newzoo expects $213.9bn in 2026 (GTA VI is scheduled for release on 19 November) and $234.4bn in 2028, which is 5.1% average annual growth 95,96 M.

The flagship gets more expensive, while releasing games gets easier

The budgets of big games have grown several-fold. In 2004–2009, developing a flagship game cost roughly $30–100m (Halo 2: ~$40m; GTA IV: over $100m) L. Sony documents disclosed in the FTC case against Microsoft put The Last of Us Part II at ~$220m and Horizon Forbidden West at ~$212m 97 M. Activision court filings put the development of Call of Duty: Black Ops Cold War over its whole life cycle, including post-launch updates and excluding marketing, at ~$700m 98 M. A leak from Insomniac suggests ~$300m for Spider-Man 2 99 L.

At the same time, the barrier to entry has collapsed. In 2025 about 20,000 new games were released on Steam 100 M, against 1,713 in 2014. Almost half of them had 10 or fewer reviews as of 12 December 2025 (SteamDB); by Gamalytic's estimate on an incomplete 2025 sample, 40% did not earn even $100 100,101 M; by Sulka Haro's estimate, the top 1% of games take ~94% of estimated revenue 63 M. The share of new games tagged “generative AI used”, according to a retelling of Haro's census, rose from ~11% in 2024 to ~20% in 2025 and ~31% in 2026 (to mid-July); the share of such games in sales is 10–27%. Developers apply the tag themselves, and Valve's rules changed in January 2026 63 M.

What AI does and what developers think

Developers use AI widely, but attitudes are worsening. According to a Google Cloud and Harris Poll survey commissioned by Google (615 developers in five countries, 2025), 90% use some AI at work 102 H. According to the GDC 2026 survey (over 2,300 respondents), 36% of industry professionals use generative AI and 52% think it harms the industry, against 30% in 2025 and 18% in 2024 103 H. This comes against a backdrop of layoffs: by Amir Satvat's count, 2024 was the peak year for job cuts (15,631), and the forecast for 2026 is 14,259 104 M.

The line between game and film is blurring from the side of world models. Google DeepMind's Genie 3 (August 2025) generates interactive 720p worlds at 24 frames per second that stay coherent for several minutes 105 H; in January 2026 Google opened Project Genie to paying subscribers in the US, so far limited to 60 seconds per world 106 M.

Another fact matters more for the industry's economics: 64% of PC gaming time in 2025 went to games released before 2019 107 M. Players do not switch to new games just because there are more of them. Games have the same problem as film: the supply of titles is growing much faster than the market (Steam releases are 11.7 times higher than in 2014, while the market grew about 2.2 times over 2015–2025), and players stay in old games.

11 · Labour and lawWhat unions, courts and regulators have already decided

Unions: rules to 2030

In 2023 Hollywood went through two strikes, by writers (148 days) and by actors (118 days), and AI was one of the main bargaining issues 108 M. The writers' 2023 contract bans AI from writing or rewriting script material and does not treat generated text as source material 109 H. The 2026 cycle, according to the available sources, passed without strikes: the writers' guild (WGA, until 1 May 2030) and the actors' union (SAG-AFTRA, until 30 June 2030) ratified four-year deals; the Directors Guild (DGA) concluded a four-year deal, but we have not confirmed the date and result of the vote with a reliable source 110 M. Under the new agreements, studios may replace a human with a synthetic performer only after notifying actors and negotiating; writers gained the right to know when their texts are licensed for model training 110 M. The next big test is 2030.

US courts: fully generated work has no author

In January 2025 the US Copyright Office concluded that fully generated material is not protected by copyright, and that a prompt to a model does not in itself give sufficient control over the result 111 H. On 2 March 2026 the US Supreme Court declined to hear Thaler v. Perlmutter, so the rule stands 112 H.

In the US, a film made entirely by a model is hard to protect from copying; in other jurisdictions the rules differ. So studios and platforms will probably have to keep a noticeable share of human creative work, not for ethical reasons but to hold ownership of the product L. Disney, Universal and Warner Bros. Discovery are suing Midjourney over the generation of their characters; the cases have been consolidated and, as of July 2026, are at the discovery stage 113 M.

Regulators: labelling

China has required visible and hidden (metadata) labels on generated content since 1 September 2025 114 H. In the EU, Article 50 of the AI Act obliges deepfakes to be disclosed from 2 August 2026, and providers of generative systems to machine-label synthetic video (for systems already on the market before 2 August 2026, there is a grace period until 2 December 2026 under the Digital Omnibus package; sources differ on the status of its adoption); the requirements are softer for evidently artistic works 115,116 M.

Platforms: Disney's licensing deal with OpenAI fell apart in under four months

In December 2025 Disney and OpenAI announced a deal: Disney would invest $1bn in OpenAI shares, and Sora would receive a three-year licence for more than 200 characters from Disney, Marvel, Pixar and Star Wars, subject to the final contracts being agreed 117,118 M. On 24 March 2026 OpenAI announced the closure of Sora, both the app and the API 46 M; the Disney deal was terminated before the money arrived 119 M. The Sora 2 API was switched off on 24 September 2026 47 H. Sora no longer offers licensed Disney characters; we found no other official services with such a licence.

12 · 2040Three scenarios and a calculator: how big the market will be in 2040

Consensus forecasts end in 2028–2030. Newzoo expects the games market to reach $234.4bn in 2028 (5.1% a year) 95 M; PwC, which counts more broadly, including in-game advertising, says “almost $300bn” in 2029 120 M. Ampere forecasts that streaming subscription revenue will exceed $200bn in 2030, 29% more than in 2025 ($157.1bn) 121 M. PwC puts the box office at $39.5bn in 2030 and the whole media and entertainment industry at $4.2trn 10 M. We found no forecasts from these firms for 2035–2040.

So what follows is arithmetic, not a forecast. Up to 2030 the calculator follows the consensus; after that it follows your assumptions. The main new parameter is the share of viewing that moves to personal generation: video made for a single viewer on request. The second is the price of such an hour compared with an hour of streaming. If personal generation is cheaper, part of the money simply disappears from the market: the viewer gets the same thing for less.

The three scenarios lead to three conclusions. First, the market in 2040 will be bigger than it is now: 1.7–1.9 times in nominal dollars, but only 1.1–1.3 times in 2025 dollars (at 2.5% inflation a year); this follows from the growth rates we assume and is not a forecast L. Second, games remain the largest segment as long as their market grows no more slowly than video: for them, generation is a development tool, not a replacement for the game. Third, if people watch as much as before, the cheaper personal generation is, the less money there is in video overall. The financial gain from generation goes to the viewer as a cheaper hour, not to the industry as higher revenue.

13 · DistributionWho wins and who loses

If production gets cheaper but attention does not, value moves to what stays scarce: access to the viewer, recognisable stories and compute. Work that can be described as “shoot, paint over, dub” gets cheaper first.

The “What is happening” column is our assessment L. Sources are in the relevant sections above: Nielsen, Netflix, Frontier Economics, Kuaishou, OpenAI, invideo, Google, FilmLA, BLS via FRED, Gower Street, SteamDB, Luminate, the guilds.

Employment in the industry is already falling, although the causes most likely lie outside AI: the 2023 strikes, the correction in streaming budgets after the peak of the “series boom” (599 scripted series in 2022, 516 in 2023) and the shift of shoots out of California; the contribution of AI cannot be measured separately. In 2025 there were 19,694 location shoot days in Greater Los Angeles (excluding sound stages), 16% fewer than a year earlier 122 H. In September 2026 the US film and sound recording industry employed 328,500 people, 28% fewer than at the peak in November 2022 123 H. A survey of 300 executives by CVL Economics, carried out in late 2023 for animators' unions, forecast that by 2026 generative AI would affect ~204,000 jobs in the US entertainment industry, ~118,500 of them in film, TV and animation 124 M. Nobody has measured whether that forecast came true.

14 · Honesty and limitationsWhere this model may be wrong

15 · QuestionsFrequently asked questions

16 · MoreRead also

17 · SourcesSources and methodology

Each number in the text leads to its source. Confidence labels: H is a primary source (company report, regulator, price list), M is secondary or cross-checked against several sources, L is an estimate, a company claim or a single unconfirmed source. Conversion to dollars uses the average 2025 exchange rate; conversion to 2025 prices uses the CPI-U index (BLS). All sums are nominal US dollars unless stated otherwise.

1. Wikipedia — Avatar: Fire and Ash (budget per Variety). en.wikipedia.org M
2. Google AI for Developers — Gemini API pricing (Gemini Omni 1.1 Flash, ~$0.10 per second of 720p). ai.google.dev H
3. Netflix — What We Watched: the first half of 2026 (and earlier engagement reports). about.netflix.com H
4. US Bureau of Labor Statistics — Consumer Expenditures in 2024. bls.gov H
5. PocketGamer.biz — Global games revenue surpasses $200bn for the first time (Newzoo Global Games Market Report). pocketgamer.biz M
6. Light Reading — Omdia: online video subscriptions reached 2.24bn in 2025; online video revenue $176bn vs pay-TV $170bn. lightreading.com M
7. Gower Street Analytics — Highest grossing December since 2019; $33.6bn global total for 2025. gower.st H
8. Alphabet — Q4 and fiscal year 2025 results. s206.q4cdn.com H
9. Netflix, Inc. — Form 10-K for fiscal year 2025. sec.gov H
10. Arise TV — PwC Global Entertainment & Media Outlook 2026–2030: revenue to hit $4.2trn by 2030. arise.tv M
11. Ampere Analysis — Global content investment to hit $255bn in 2026 (press release, 12 Jan 2026). cdn.www.ampereanalysis.com H
12. PwC Global Entertainment & Media Outlook 2006–2010, as quoted in a Playlogic filing (SEC). sec.gov L
13. US BLS CPI-U annual averages (via FRED, CPIAUCNS). fred.stlouisfed.org H
14. Motion Picture Association — THEME Report 2021 (global and US box office, admissions). motionpictures.org H
15. The Numbers — Domestic theatrical market summary (tickets, average price). the-numbers.com M
16. ChannelNews — DEG Year-End 2025: home entertainment spending now far larger than cinemas. channelnews.com.au M
17. Nielsen — The Gauge, August 2026. nielsen.com H
18. Netflix — Q3 2026 shareholder letter (8-K, exhibit 99.1). sec.gov H
19. The Walt Disney Company — Q4 and full-year fiscal 2025 earnings. thewaltdisneycompany.com H
20. Comcast Corporation — Form 10-K for 2025 (Studios, Theme Parks, Peacock). sec.gov H
21. Warner Bros. Discovery — fourth-quarter and full-year 2025 results (Studios and Streaming segments). s201.q4cdn.com H
22. Morgan Stanley Research — How AI benefits and threatens the entertainment industry. morganstanley.com M
23. Netflix — Form 10-K 2025: agreement to acquire Warner Bros. ($27.75 per share). sec.gov H
24. Paramount Skydance and Warner Bros. Discovery — announcement of the merger closing date. s201.q4cdn.com M
25. Electronic Arts — EA announces completion of acquisition (4 Aug 2026). ea.com H
26. Wikipedia — Toy Story 5 (budget per Variety). en.wikipedia.org M
27. Wikipedia — List of most expensive television series (press-reported budgets). en.wikipedia.org M
28. Anime Corner — The costs of anime: prices of episodes have skyrocketed. animecorner.me M
29. Fortune — AI entertainment in China (7 Oct 2026). fortune.com M
30. invideo — How many AI video generations does it take to produce a film (vendor case). invideo.io L
31. eWeek — Kalshi's AI-generated NBA Finals ad. eweek.com M
32. TechRadar — Netflix uses generative AI VFX in a show for the first time. techradar.com M
33. Fortune — Netflix made 17 minutes of a documentary twice as fast and at half the cost (16 Jul 2026). fortune.com M
34. Runway — How House of David used Runway (customer case study); Cord Cutters News citing Wired. runway.com M
35. Implicator — OpenAI backs Critterz, a Cannes-bound AI feature on a sub-$30m budget (citing WSJ). implicator.ai M
36. Deadline — OpenAI-produced animated family film Critterz heads to Cannes market (May 2026). deadline.com L
37. News on AIR — New film made entirely using AI premiered at Cannes (company claim). newsonair.gov.in L
38. OrcaRouter — Wan 3.0 release details (compiled from Alibaba launch). orcarouter.ai M
39. TechNode — ByteDance launches Seedance 2.5 video generation model (31 Jul 2026). technode.com M
40. Google AI for Developers — Gemini Omni guide (extension to 40 s, resolutions). ai.google.dev H
41. Google Research — Automating coherent long-form video generation (24 Sep 2026). research.google M
42. arXiv 2509.22622 — LongLive: real-time interactive long video generation (NVIDIA). arxiv.org H
43. FutureSearch — In early 2028, will an AI be able to generate a full high-quality movie to a prompt?. futuresearch.ai M
44. Google Developers Blog — Veo 3 and Veo 3 Fast: new pricing (8 Sep 2025). developers.googleblog.com H
45. Runway — API pricing (incl. resale prices of Google, Alibaba, ByteDance, Kling and MiniMax models). docs.dev.runwayml.com H
46. Engadget — OpenAI is shutting down its Sora video generation app (24 Mar 2026). engadget.com M
47. OpenAI — Sora 2 model page: Sora 2 and the Videos API shut down on 24 Sep 2026. developers.openai.com H
48. Kuaishou Technology — Q2 2026 results (Kling AI revenue over RMB850m). webull.com H
49. Macau Business — MiniMax announces first-half 2026 financial results. macaubusiness.com M
50. Bloomberg — AI video startup Runway valued at $5.3 billion (10 Feb 2026). bloomberg.com M
51. Artificial Analysis — Text-to-video leaderboard. artificialanalysis.ai H
52. Epoch AI — Trends in GPU price-performance (2022). epoch.ai H
53. Global Times — NRTA at a State Council Information Office briefing: micro-drama market over 100bn yuan; 430,000 titles in Jan–Aug 2026. globaltimes.cn M
54. Seoul Economic Daily — China churns out 1,519 AI short dramas a day (22 Oct 2026). en.sedaily.com M
55. The Next Web — China's micro-drama boom, AI and state funding (1 May 2026). thenextweb.com M
56. BigGo Finance — AI short dramas: share reaching 50m views in H1 2026. finance.biggo.com L
57. Hong Kong Free Press (AFP) — China's first AI film hits cinemas but ticket sales lag (23 Oct 2026). hongkongfp.com M
58. US BLS — American Time Use Survey, 2025 results. bls.gov H
59. Hypebot — Luminate 2025 year-end music report: streaming saturation. hypebot.com M
60. Deezer newsroom — AI music exceeds 50% of daily uploads (July 2026). newsroom-deezer.com H
61. IFPI — Global Music Report 2026: recorded music revenues grow 6.4%. ifpi.org H
62. Hypebot — Spotify Loud & Clear 2026: $11bn paid out. hypebot.com M
63. Sulka Haro — Three years of AI on Steam (13 Jul 2026). fragwyz.substack.com M
64. Publishers Weekly (Bowker data) — Book output topped 4 million in 2025. publishersweekly.com H
65. Google Books Search blog — Books of the world, stand up and be counted (5 Aug 2010). booksearch.blogspot.com M
66. UK Publishers Association — Frontier Economics study: publishing's contribution to film, TV and theatre (2018). publishers.org.uk H
67. TechCrunch — Audible is expanding its AI-narrated audiobook library (13 May 2025). techcrunch.com M
68. Infodocket — Audio Publishers Association: audiobook sales up 9% to $2.43bn. infodocket.com M
69. Publishers Weekly — Publishing's AI reckoning (Sept 2026). publishersweekly.com M
70. Association of American Publishers — Court grants final approval of the Bartz v. Anthropic settlement. publishers.org H
71. Archive of Our Own — homepage counters (25 Oct 2026). archiveofourown.org H
72. Wan-Video — Wan2.1 and Wan2.2 repositories (GitHub). github.com H
73. NVIDIA — Video generation guide for RTX (LTX-2.3 workflow requirements). nvidia.com M
74. ComfyUI blog — LTX-2.5 day-0 support (12 Aug 2026). blog.comfy.org M
75. Lightricks — LTX-2.5 model card and community licence (Hugging Face). huggingface.co H
76. Tencent Hunyuan — HunyuanVideo-1.5 repository and licence (GitHub). github.com H
77. lllyasviel — FramePack repository (GitHub). github.com H
78. Self Forcing — project page. self-forcing.github.io H
79. Qwen — Qwen3-Next-80B-A3B-Instruct model card (Hugging Face). huggingface.co H
80. Resemble AI — Chatterbox repository (GitHub). github.com H
81. ACE-Step 1.5 repository (GitHub). github.com H
82. HKUDS — ViMax repository (GitHub). github.com H
83. Apatero — How long it takes to train a Wan 2.2 video LoRA. apatero.com L
84. NVIDIA — GeForce RTX 50 Series announcement (CES 2025). nvidianews.nvidia.com H
85. GPUPoet — RTX 5090 price tracker, October 2026. gpupoet.com L
86. The Register — Nvidia raises DGX Spark price (2 Oct 2026). theregister.com M
87. Valve — Steam Hardware & Software Survey, September 2026. store.steampowered.com H
88. US EIA — Electric Power Annual, average retail price of electricity (Table 2.4). eia.gov H
89. NVIDIA Blog — RTX AI Garage at CES 2026: open models for video generation. blogs.nvidia.com H
90. Law of Ukraine On Copyright and Related Rights No. 2811-IX (Arts. 12, 22, 28). zakon.rada.gov.ua H
91. 17 U.S.C. § 106 — Exclusive rights in copyrighted works (Cornell LII). law.cornell.edu H
92. 17 U.S.C. § 107 — Fair use (Cornell LII). law.cornell.edu H
93. Directive 2001/29/EC (InfoSoc), Art. 5(2)(b) — private copying (EUR-Lex). eur-lex.europa.eu M
94. ChatGPT Is Eating the World — Sora switches from opt-out to opt-in for rightsholders (Oct 2025). chatgptiseatingtheworld.com M
95. WN Hub — Newzoo: 2026 global games market forecast; GameDev Reports — 2028 forecast. wnhub.io M
96. Take-Two Interactive — Grand Theft Auto VI release date and price announcement. ir.take2games.com H
97. MP1st — The Last of Us Part II budget was $220m, Horizon Forbidden West $212m (FTC v. Microsoft documents). mp1st.com M
98. Video Games Chronicle — Court docs reveal Call of Duty games cost up to $700m to develop. videogameschronicle.com M
99. Aftermath — Big-budget game development is not sustainable (Insomniac leak). aftermath.site L
100. How To Market A Game — How many games were released on Steam in 2025 (SteamDB and other trackers). howtomarketagame.com M
101. MP1st — 40% of Steam games in 2025 earned less than $100 (Gamalytic). mp1st.com M
102. Google Cloud — 90% of games developers already using AI in workflows (18 Aug 2025). googlecloudpresscorner.com H
103. GDC — 2026 State of the Game Industry. gdconf.com H
104. GamesBeat — Game industry layoffs projected to hit 14,259 in 2026 (Amir Satvat). gamesbeat.com M
105. Google DeepMind — Genie 3: a new frontier for world models. deepmind.google H
106. 9to5Google — Google opens Project Genie (29 Jan 2026). 9to5google.com M
107. GameDev Reports — Newzoo: gaming market surpassed $200bn. gamedevreports.substack.com M
108. Wikipedia — 2023 Hollywood labor disputes. en.wikipedia.org M
109. Writers Guild of America — Summary of the 2023 WGA MBA. wgacontract2023.org H
110. Jackson Lewis — AI stays centre stage in entertainment industry collective bargaining (2026). jacksonlewis.com M
111. US Copyright Office — Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan 2025). copyright.gov H
112. Supreme Court of the United States — docket 25-449, Thaler v. Perlmutter (cert. denied 2 Mar 2026). supremecourt.gov H
113. Graphic Policy — Disney, Warner Bros., Universal and Midjourney discovery dispute (May 2026). graphicpolicy.com M
114. Cyberspace Administration of China — Measures for labelling AI-generated synthetic content. cac.gov.cn H
115. EU AI Act, Article 50 — transparency obligations. artificialintelligenceact.eu M
116. Gibson Dunn — EU AI Act omnibus agreement: postponed deadlines and other changes. gibsondunn.com M
117. The Walt Disney Company and OpenAI — agreement announcement (11 Dec 2025). thewaltdisneycompany.com H
118. NBC New York — Disney and OpenAI deal: Sora to generate videos with Disney characters (Dec 2025). nbcnewyork.com M
119. Mobile World Live — OpenAI ends Disney deal after shuttering Sora (25 Mar 2026). mobileworldlive.com M
120. PwC Global Entertainment & Media Outlook 2025–2029 (press release). en.prnasia.com M
121. Advanced Television — Ampere: streaming revenue surpasses $150bn in 2025, over $200bn by 2030. advanced-television.com M
122. FilmLA — Q4 2025 on-location production report. filmla.com H
123. US BLS via FRED — All employees, motion picture and sound recording industries (CES5051200001). fred.stlouisfed.org H
124. CVL Economics — Future Unscripted: the impact of generative AI on entertainment industry jobs (Jan 2024). animationguild.org M
125. eMarketer — US time spent with media, H2 2026 update. emarketer.com M

Read this briefing in Ukrainian →