Verified
Netflix
Experimenting- Streaming
- ~16,000 employees (2025)
- In-house foundation models
- Self-hosted LLMs on AWS
- GenAI on ~300 titles (2026)
Netflix deploys AI mainly inside its own product and studio: a recommendation foundation model and open-weight LLMs it post-trains and serves itself on AWS GPUs, generative AI tools that creative partners used on about 300 titles in 2026, and AI tools across its advertising workflow. Netflix has run machine learning in recommendations for two decades, reports no dollar return from AI, and has put Chief Product and Technology Officer Elizabeth Stone over product, engineering and data since Feb 2026.
research.netflix.comTheir approach to AI
- The platform
- Netflix builds its own AI stack: a foundation model that centralizes member preference learning, the MediaFM content model, and an AI Platform that post-trains open-weight LLMs and serves them on vLLM inside Netflix's production environment (Jul 2026).
- The models
- Netflix trains its own recommendation and media models and post-trains open-weight LLMs, CNBC reported in May 2025 that OpenAI powers its generative search, Eyeline licenses commercial video diffusion models (Oct 2025), and Greg Peters said in Oct 2025 that most generative AI uses at Netflix are bought tools.
- The reach
- Netflix's recommendation models shape the homepage for more than 325 million paid memberships (Jan 2026), and creative partners used generative AI workflows on about 300 Netflix titles in 2026 (Jul 2026).
- The catch
- Netflix has published no business return it credits to AI, only engagement lifts and single-production examples, and has reported no count of employees using AI tools.
Snapshot
- Employees
- ~16,000 full-time
- Dec 31, 2025; about 1,000 roles cut in Oct 2026
- Technology budget
- $3.39B (2025)
- Technology and development expense; $3.0B in the first nine months of 2026
- AI/ML headcount
- Not disclosed
- AI Platform, personalization and research teams sit in product and technology
- Who owns AI
- Elizabeth Stone, CPTO
- Leads product, engineering and data since Feb 2026; CTO from Oct 2023
- Platform
- AI Platform, built in-house
- Post-training framework, vLLM serving and a model gateway (2026)
- Models
- In-house foundation models, open-weight LLMs, OpenAI
- OpenAI behind generative search (May 2025); most generative AI tools bought (Oct 2025)
- Reach
- 325M+ paid memberships
- Served by its recommendation models (Jan 2026)
- Use cases
- ~300 titles used generative AI (2026)
- Plus recommendations, artwork, localization, search tests and ads
Key findings
Netflix said on Jul 16, 2026 that generative AI workflows had been used in roughly 300 of its titles in 2026, mostly in post-production, and Ted Sarandos said the 17 minutes of AI-enhanced footage in The American Experiment were made twice as fast and at half the cost of earlier options.
Netflix runs its AI stack itself: a recommendation foundation model pre-trained monthly and fine-tuned daily feeds production personalization (Oct 2026), and the AI Platform team post-trains open-weight LLMs and serves them on vLLM on GPUs provisioned on AWS (Feb and Jul 2026).
Netflix acquired InterPositive, Ben Affleck's filmmaking AI company, on Mar 5, 2026 and brought its whole team in-house; Netflix's filings record a March 2026 acquisition for about $587 million in cash, which TheWrap identified as InterPositive (Jul 2026).
Elizabeth Stone, Netflix's CTO since Oct 2023, became Chief Product and Technology Officer in Feb 2026, leading product, engineering and data after Chief Product Officer Eunice Kim left in Sep 2025.
Netflix has published no business return credited to AI: it cites engagement gains from newer recommendation models (Apr 2026), and in Oct 2026 it cut about 1,000 jobs without tying the cuts to AI and lowered its 2026 ad revenue forecast to about $2.7 billion.
Where Netflix sits
- Stage
- Experimenting
Netflix has reported no business return it credits to AI: it cites engagement gains from newer recommendation models (Apr 2026) and single-production examples, such as 17 minutes of AI footage in The American Experiment made at half the cost of earlier options (Jul 2026), and it does not separate AI's share of its ad revenue growth. Netflix has also reported no count of employees using AI tools; Chief Product and Technology Officer Elizabeth Stone, who leads product, engineering and data, says AI fluency is expected of everyone (Jul 2026). Netflix is at Experimenting despite two decades of machine learning in its product and generative AI workflows on about 300 titles in 2026; one return it credits to AI in a use case, or a reported count of employees using AI tools, would move it to Piloting.
How the stages are defined - Ownership
- Elizabeth Stone, Chief Product and Technology Officer (since Feb 2026)
- Platform
- Hybrid: AI Platform and foundation models built in-house; most generative AI tools bought
- Models
- In-house recommendation and media models; open-weight LLMs post-trained in-house; OpenAI for generative search
- Hosting
- AWS for the vast majority of computing, including GPUs and self-served LLMs on vLLM; OpenAI behind generative search (2025)
- Reach
- Recommendations for 325M+ paid memberships (Jan 2026)
How they deploy
Buy most tools, build where data and scale count
Greg Peters said in Oct 2025 that the vast majority of Netflix's generative AI uses involve buying solutions and integrating them into existing tools and products, with targeted investment in product experiences, content production and advertising, often built on foundation models Netflix gets open source or commercially. Peters said in Apr 2026 that Netflix builds where its data is unique and large and where it has products or business processes at scale to attach the technology to.
One foundation model for personalization
Netflix built a foundation model for recommendations so that member preference learning sits in one large model instead of many specialized ones, trained on hundreds of billions of interactions from more than 300 million users (Mar 2025). Netflix now pre-trains the model from scratch each month, fine-tunes it daily and feeds it to production applications through embeddings, as a subgraph inside other models, or by fine-tuning (Oct 2026). Greg Peters said in Apr 2026 that the newer model architectures let Netflix add content types more quickly and had already raised engagement.
Tools for creators, not replacements
Ted Sarandos said in Apr 2025 that making films 10% better is a bigger opportunity than making them 50% cheaper, and has said since that AI gives filmmakers better tools without replacing the artists. Netflix released generative AI production guidance for creative partners by Oct 2025, acquired InterPositive in Mar 2026 for tools built by and for filmmakers, and Sarandos said in Jul 2026 that savings from generative AI will likely be reinvested in more content.
Self-hosted LLMs
Netflix runs the full LLM stack itself, from deployment through inference, inside its existing production environment instead of relying only on hosted APIs, according to its AI Platform team (Jul 2026). Netflix exposes an OpenAI-compatible API on its own serving system so that a team can move from a hosted model to a fine-tuned self-hosted one for quality, latency, cost or data privacy with minimal code changes.
AI across advertising
Netflix said in Oct 2025 that it would use AI to test new ad formats, generate the most relevant ad creative and placement, and develop media plans faster, and in Jan 2026 that it had begun testing AI tools that help advertisers create custom ads from Netflix's intellectual property. Netflix said at its May 2026 upfront that it is testing AI agents that manage, optimize and buy ads, and in Jul 2026 that it had expanded AI-powered tools across the full advertising lifecycle. Greg Peters said in Oct 2025 that Netflix planned to pivot in 2027 to investments in data capabilities such as ML-based optimization, advanced measurement and targeting.
- Deployment model
- Netflix's product and technology organization, led by Chief Product and Technology Officer Elizabeth Stone, builds shared AI infrastructure: an AI Platform org with model runtime, inference and post-training teams, the Mako ML compute platform that provisions GPUs on AWS, the Model Scoring Service on NVIDIA Triton Inference Server, and a model gateway. Personalization, search, data platform and advertising teams build applications on that infrastructure. In content production, Netflix's creative partners use generative AI tools, supported by Eyeline, Netflix's production and innovation studio, which absorbed Scanline VFX in Oct 2025 and handles generative AI in its Studios division, and by the InterPositive team that joined in Mar 2026.
- Model providers
- Netflix trains its own models for its core product: a transformer-based recommendation foundation model, and MediaFM, a tri-modal audio, video and text model pretrained on portions of the Netflix catalog (Feb 2026). For language tasks, Netflix's post-training framework starts from open-weight models loaded from Hugging Face, supports architectures such as Qwen3, Gemma3 and GPT-OSS, and adapts models to the Netflix catalog and member histories (Feb 2026), and GenRec post-trains a Netflix-adapted open-source LLM for ranking (Jul 2026). CNBC reported from Netflix's May 2025 product event that OpenAI powers Netflix's generative search features, with the models trained by a Netflix team, and Eyeline's head of studio told TheWrap in Oct 2025 that Eyeline uses commercially available video diffusion models under enterprise licenses rather than building its own. Greg Peters said in Oct 2025 that Netflix builds on foundation models it gets open source or commercially; Netflix's filings and engineering posts name no other commercial model provider.
- Data and infrastructure
- Netflix runs the vast majority of its computing on Amazon Web Services and says it cannot easily switch those operations to another cloud provider (2025 Form 10-K). Mako, Netflix's internal ML compute platform, provisions GPUs on AWS for training and post-training with PyTorch, Ray and vLLM (Feb 2026), and Netflix says personalization runs on GPUs that train its models and increasingly serve them live, with training jobs on a Kubernetes-native compute platform (Oct 2026). Netflix's Model Scoring Service serves XGBoost, TensorFlow, PyTorch and LLM models behind one interface on NVIDIA Triton Inference Server, with vLLM as the standard LLM engine since summer 2025 (Jul 2026). Netflix streams to members through Open Connect, its own content delivery network.
Platforms and internal tools
Recommendation foundation modelDescribed Mar 2025
Netflix's recommendation foundation model learns member preferences from full interaction histories and content data at large scale and passes what it learns to other models. Netflix pre-trains the model monthly, fine-tunes it daily to add newly launched titles, publishes profile and title embeddings to an Embedding Store, and uses it in production through embeddings, subgraph integration and fine-tuning (Oct 2026).
- Users
- More than 325 million paid memberships (Jan 2026)
MediaFMFeb 2026
MediaFM is Netflix's in-house multimodal content embedding model, pretrained on tens of millions of shots from the Netflix catalog using audio, video and subtitle text. Netflix built MediaFM for cold start of new titles in recommendations, promotional assets, ads relevancy, clip tagging and internal content analysis, and MediaFM became Netflix's default video preview embedding on all platforms after beating the alternatives in online tests (Aug 2026).
- Models
- In-house transformer encoder
AI Platform: post-training and LLM serving2025 to 2026
Netflix's AI Platform team runs a post-training framework on PyTorch, Ray and vLLM over GPUs provisioned on AWS through Mako (Feb 2026), and serves LLMs in-house through the Model Scoring Service with vLLM on NVIDIA Triton, an OpenAI-compatible API alongside gRPC, and constrained decoding for structured output (Jul 2026). The AI Platform team also runs a model gateway and evaluation infrastructure used by Netflix's data platform (Oct 2026).
- Models
- Open-weight LLMs post-trained in-house
Pensive AIBy Oct 2026
Pensive AI adds LLM error extraction and on-demand root-cause debugging to Pensive, the auto-diagnosis system in Netflix's data platform, for data and ML workloads that fail at a rate of roughly 100,000 on peak days. Pensive AI extracts errors for 100% of failures, including those that rules left unclassified, and is available through APIs, MCP and the Data Portal interface (Oct 2026).
Eyeline
Eyeline is Netflix's production and innovation studio, which Ted Sarandos described in Jul 2025 as the production innovation group inside Scanline VFX; in Oct 2025 Netflix merged Scanline into the Eyeline brand, with a Studios division for virtual production, volumetric capture and generative AI and a Labs research division. Eyeline worked with the creators of El Eternauta on an AI-powered building collapse (Jul 2025), and its volumetric capture technology was combined with generative AI to de-age characters in Happy Gilmore 2 (Oct 2025).
InterPositiveAcquired Mar 5, 2026
InterPositive, founded by Ben Affleck, develops AI tools for filmmakers trained on a proprietary dataset filmed on a controlled soundstage, with models focused on filmmaking techniques such as missing shots, background replacement and lighting rather than performances. Netflix brought the entire InterPositive team in-house, and Ted Sarandos said in Jul 2026 that it was early days for the tools.
Netflix Ads SuiteBy Oct 2025
Netflix Ads Suite is Netflix's own first-party ad tech stack. Netflix said in Jul 2026 that it had expanded AI-powered tools across the full advertising lifecycle, from planning and creative production to campaign management, optimization and reporting.
Vendors and partners
Amazon Web Services
Cloud for the vast majority of Netflix's computing, including the GPUs that train and serve its models · Netflix's 2025 Form 10-K says it cannot easily switch its AWS operations to another cloud provider, and its Mako ML compute platform provisions GPUs on AWS (Feb 2026).
NVIDIA
Triton Inference Server under Netflix's Model Scoring Service · Netflix's Model Scoring Service runs NVIDIA Triton Inference Server for model loading, batching and GPU scheduling, with vLLM as the LLM engine (Jul 2026).
Hugging Face
Source of open-weight models and tokenizers for Netflix's post-training framework · Netflix's AI Platform team made the Hugging Face AutoTokenizer its single source of truth to keep training and vLLM serving consistent (Feb 2026).
OpenAI2025
Model provider behind Netflix's generative search features · Bloomberg reported in Apr 2025 that Netflix was testing OpenAI-powered search on iOS in Australia and New Zealand, and CNBC reported from Netflix's May 2025 product event that OpenAI powers the search features, with models trained by a Netflix team.
Braintrust
Tracing backend for Pensive AI · Netflix's Pensive AI team built its tracing backend on Braintrust, which shows prompts, completions, tool calls and latency in one view (Oct 2026).
Leadership
AI, data and platform leaders
Executives
Board
AI org map
Central product and technology organization builds the AI platform and models; content, studio and advertising teams apply them
Netflix is run by co-CEOs Ted Sarandos and Greg Peters, who together act as its chief operating decision maker (2025 Form 10-K). Netflix promoted CTO Elizabeth Stone to Chief Product and Technology Officer in Feb 2026, combining the CTO and chief product officer roles over product, engineering and data, and her organization holds the AI Platform, personalization and data platform teams. Generative AI in production runs through Eyeline, which absorbed Scanline VFX in Oct 2025, and the InterPositive team acquired in Mar 2026. Netflix named Peter Rice Chief Content Officer on Oct 15, 2026, effective Oct 26, succeeding Bela Bajaria.
Netflix does not disclose how many people work on AI. Netflix's AI work sits mainly in the product and technology organization under Elizabeth Stone, including the AI Platform org with its model runtime, inference and post-training teams, the personalization and data science teams behind the recommendation foundation model, MediaFM and GenRec, and the data platform team behind Pensive AI. Generative AI production work runs through Eyeline, which absorbed Scanline VFX in Oct 2025, and the InterPositive team, and Netflix had about 16,000 full-time employees at Dec 31, 2025 before cutting about 1,000 roles in Oct 2026, mostly in product, technology, legal and human resources.
Reporting lines
Co-CEOs
Ted Sarandos and Greg Peters
Product and Technology
Elizabeth Stone, CPTO
Product, engineering and data since Feb 2026; Netflix has not published Stone's reporting line
AI Platform
Model runtime, inference and post-training teams, the model gateway and evaluation infrastructure; leader not named; placement inferred from engineering posts
Personalization and recommendation research
Recommendation foundation model, MediaFM, GenRec and GenPage; placement inferred
Data Platform
Pensive AI and the Data Portal; placement inferred
Content
Peter Rice, Chief Content Officer
Peter Rice succeeds Bela Bajaria, named Oct 15, 2026 and effective Oct 26; reporting line inferred
Advertising
Amy Reinhard, President of Advertising
Netflix Ads Suite and AI ad tools; reporting line inferred
Eyeline
Jeff Shapiro, CEO
Netflix's production and innovation studio, which absorbed Scanline VFX in Oct 2025; generative AI sits in its Studios division; reporting line not published
InterPositive team
Ben Affleck, Senior Advisor
Joined Netflix in Mar 2026; reporting line not published
Netflix's leadership page, checked Nov 1, 2026, lists Elizabeth Stone as Chief Product and Technology Officer and no separate chief product officer, AI or data executive, and does not yet list Peter Rice. Netflix has published no headcount for its AI Platform, personalization or research teams, so the split of its AI workforce is not shown.
Named people
Use cases
Netflix runs AI in production in recommendations and the homepage, artwork and preview personalization, subtitle and promotional localization, the Clips vertical video feed, generative AI tools in film and series production (about 300 titles in 2026), advertising workflows and data platform debugging, with conversational search, LLM-written recommendation explanations, LLM-based ranking and custom AI ads still in testing. Netflix has published no business-level return credited to AI for any of these use cases or for AI as a whole; it reports engagement gains and single-production examples such as The American Experiment, whose AI footage Ted Sarandos said cost half as much as earlier options.
| Use case | Area | Status | ROI |
|---|---|---|---|
| Recommendations and homepage personalization Netflix's homepage rows, titles and artwork are chosen by machine learning models trained on how members watch, and a recommendation foundation model now feeds those applications through embeddings, subgraph integration and fine-tuning. | Member experience | In production | |
| Artwork and video preview personalization (MAPS) Netflix uses multimodal content embeddings in artwork personalization, query-aware artwork ranking in search and video preview personalization, so personalization starts closer to a title's launch; MediaFM became the default video preview embedding on all platforms after winning Netflix's online tests. | Member experience | In production | |
| Conversational, natural language and voice search Netflix unveiled generative AI search powered by OpenAI at a May 2025 product event, beta tested a conversational search experience that lets members describe what they want to watch in natural language (Oct 2025), and said in Jul 2026 that it is adding voice search and AI-powered natural language search and using LLMs to improve title discovery. | Member experience | Pilot | |
| LLM-written recommendation explanations Netflix tested short LLM-written explanations of why a title was recommended on title detail pages, generating hundreds of thousands a week in online experiments including a month-long test on the mobile app, with a second LLM gating quality. | Member experience | Pilot | |
| LLM-based ranking and generative homepage (GenRec, GenPage) GenRec post-trains an internal foundation LLM as a recommendation ranker that Netflix says can match or exceed its production system, and GenPage, a single generative model that builds the whole homepage, beat the production baseline on Netflix's core engagement metric in online tests. | Member experience | Pilot | |
| Clips feed and multi-ratio video cropping Netflix built an automated cropping system using neural shot detection and active speaker detection to produce vertical and other aspect-ratio versions of footage for the Clips feed in its mobile app and for marketing. | Member experience | In production | |
| Generative AI in film and series production Netflix's creative partners use generative AI for set references, pre-visualization, VFX, sequence prep and shot planning, with the largest concentration in post-production; examples include El Eternauta's building collapse, de-aging in Happy Gilmore 2 and crowd and battle sequences in Glory, Brasil 70 and The American Experiment. | Content production | In production | |
| InterPositive filmmaking tools InterPositive's AI tools address filmmaking problems such as missing shots, background replacement and incorrect lighting; Ted Sarandos said in Apr 2026 that adoption was building among creators who tried them, and in Jul 2026 that it was early days. | Content production | Pilot | |
| Subtitle and promotional asset localization Netflix uses generative AI to localize promotional assets in many languages (Oct 2025) and AI to improve subtitle localization (Jan 2026), so titles reach viewers in more countries. | Marketing and localization | In production | |
| AI tools across the advertising lifecycle Netflix uses AI in ad planning, creative production, campaign management, optimization and reporting, including automated workflows for ad concepts, AI models for campaign planning that Netflix says sped those processes up, AI-driven media plan tooling and AI that adapts advertiser assets to different canvases; Netflix said in May 2026 that it is testing AI agents that manage, optimize and buy ads. | Advertising | In production | |
| Custom AI-generated ads from Netflix titles Netflix began testing AI tools in 2025 that match advertiser creative with the worlds of Netflix titles, tested them with DoorDash, Target and TurboTax, and said in May 2026 that it would bring the capability to every ad-supported region by the end of 2026. | Advertising | Pilot | |
| Pensive AI failure diagnosis Pensive AI extracts errors from the logs of every failed data and ML workload and offers on-demand LLM root-cause analysis through the Data Portal, APIs and MCP. | Engineering and data | In production | |
| Natural language Graph Search Netflix added LLM-based natural language search to Graph Search, its platform for searching federated enterprise data, and planned to evaluate it in a live production environment (Jan 2026). | Engineering and data | Pilot |
Adoption
Netflix spread AI first through its product: machine learning has picked titles and artwork for members for two decades, and since 2025 Netflix has moved personalization onto a recommendation foundation model and multimodal embeddings while testing LLMs for search, explanations and ranking. AI in production went from AI-powered de-aging in Pedro Páramo (Apr 2025) and El Eternauta, the first generative AI final footage in a Netflix original (Jul 2025), to generative AI workflows in about 300 titles in 2026. Netflix has published no figures on employee use of generative AI tools; Elizabeth Stone said in Jul 2026 that Netflix treats AI fluency as an expectation for everyone rather than a skill for certain levels.
By the numbers
Rollout, first pilot to today
Mar 2025
Netflix described a foundation model for personalized recommendation trained on hundreds of billions of interactions from more than 300 million users.
Apr 2025
Ted Sarandos said AI-powered tools delivered de-aging in Pedro Páramo for a fraction of what it cost on The Irishman, and Greg Peters said Netflix was building interactive search based on generative technologies.
May 2025
Netflix unveiled a redesigned homepage and app with a vertical video feed and a generative AI search tool that CNBC reported is powered by OpenAI, with models trained by a Netflix team.
Jul 2025
Netflix said a building collapse in El Eternauta, made with Eyeline using AI tools, was the first generative AI final footage in a Netflix original, and that it had piloted conversational search.
Oct 2025
Netflix reported generative AI de-aging in Happy Gilmore 2, a beta of conversational search, generative AI localization of promotional assets and new production guidance for creators, and said it would use AI to test ad formats in Q4.
Oct 2025
Netflix merged Scanline VFX into Eyeline, whose Studios division runs generative AI work on commercially licensed video diffusion models alongside virtual production and volumetric capture.
Jan 2026
Netflix said it was using AI for subtitle localization, merchandising and ad concept workflows, and testing AI tools for custom ads built on Netflix titles.
Feb 2026
Netflix promoted CTO Elizabeth Stone to Chief Product and Technology Officer, a new role combining the CTO and chief product officer functions over product, engineering and data.
Mar 2026
Netflix acquired InterPositive, Ben Affleck's filmmaking AI company, and brought its whole team in-house.
Apr 2026
Greg Peters said recommendation systems built on newer model architectures had already raised engagement on Netflix in Q1 2026.
May 2026
Netflix said at its upfront that it offers AI-driven media plan tooling, is testing AI agents that manage, optimize and buy ads, and would bring AI matching of advertiser creative to Netflix titles to every ad-supported region by the end of 2026.
Jul 2026
Netflix said generative AI workflows had been used in about 300 titles in 2026 and that its AI-powered tools now span the full advertising lifecycle.
Oct 2026
Netflix described its recommendation foundation model in production across personalization applications through embeddings, subgraph integration and fine-tuning.
Economics & controls
Spend and ROI
Technology budget
Netflix spent $3.39 billion on technology and development in 2025, up 16% from $2.93 billion, mostly because of a $438 million rise in personnel costs, and $3.0 billion in the first nine months of 2026, up from $2.5 billion, with technology and development growing faster than revenue in Q3 2026. Netflix does not break out AI spending within technology and development or its cloud costs on AWS.
What AI itself costs
Netflix has not disclosed what it spends on AI models, GPUs or AI tools. Netflix's Form 10-Q filings record a March 2026 acquisition for about $587 million in cash, which TheWrap identified as InterPositive (Jul 2026). Netflix's engineers call GPU capacity among the fastest-growing and most expensive parts of its infrastructure, with inference a growing share as more AI features reach members (Oct 2026), and treat cost as a design constraint: the AI Platform team self-hosts LLMs partly for cost and data privacy (Jul 2026), and the Pensive AI team caches each debugging answer so a repeat request costs no further inference (Oct 2026).
Reported ROI
Netflix has reported no business-level return credited to AI. Netflix cites engagement instead: Greg Peters said newer recommendation models raised engagement in Q1 2026, and Netflix researchers estimated in Oct 2026 that random recommendations would cut engagement by 16% compared with the current system. Ted Sarandos said in Jul 2026 that savings from generative AI in production will likely be reinvested in more content.
Productivity markers
Netflix has shared single-production examples rather than portfolio figures: the El Eternauta VFX sequence was finished 10 times faster than with traditional tools (Jul 2025), the de-aging in Pedro Páramo cost a fraction of the de-aging on The Irishman (Apr 2025), and the 17 minutes of AI-enhanced footage in The American Experiment were made twice as fast at half the cost of earlier options (Jul 2026). Netflix also said in Jan 2026 that AI models had significantly sped up ad campaign planning.
Advertising revenue
Netflix's ad revenue rose more than 2.5 times to over $1.5 billion in 2025, and Netflix said in Jul 2026 that ads growth was fueled by its slate together with AI-powered tools and workflows, the Netflix Ads Suite and broader programmatic capabilities, without separating AI's share. Netflix lowered its 2026 ad revenue forecast to about $2.7 billion from about $3 billion on Oct 20, 2026, citing a slightly slower rollout of some ad product capabilities.
Headcount
Netflix had about 16,000 full-time employees at Dec 31, 2025. Netflix told staff on Oct 14, 2026 that it would cut nearly 1,000 jobs, about 6% of its workforce and its biggest round of cuts in years, mostly in product, technology, legal and human resources, and the co-CEOs' memo framed the cuts as discipline about where to put people and resources, without citing AI. Netflix's Q4 2026 guidance includes about 120 basis points of operating margin cost for the cuts, and Spencer Neumann declined to quantify the savings.
Governance and risk
Guidance for creative partners
Netflix released generative AI production guidance for its creative partners in 2025 so they use the tools responsibly (Oct 2025 shareholder letter). Netflix says creators decide how their films and shows are made, and Bela Bajaria said in Mar 2026 that new tools should expand creative freedom, not replace the work of writers, directors, actors and crews.
Training data and talent
InterPositive trained its first model on a proprietary dataset it filmed on a controlled soundstage and focused its smaller models on filmmaking techniques rather than performances, with built-in restraints to protect creative intent, according to Ben Affleck (Mar 2026). Netflix pretrains MediaFM on portions of its own catalog (Feb 2026).
Humans over LLM output
Netflix's engineers gate LLM-written recommendation explanations with a second LLM judge that is benchmarked against trained human raters, retrained when it drifts, and changed only with human approval (Oct 2026). Pensive AI reconstructs error text from the original logs instead of letting the model rewrite it, and evaluation thresholds gate every code change before it merges (Oct 2026).
Board oversight
Netflix's board oversees enterprise, strategic and operational risk, and its audit committee oversees cybersecurity risk. Netflix's 2026 proxy statement says stockholders asked about its use and governance of generative AI in engagement meetings, and cites director Brad Smith's experience leading Microsoft's work on artificial intelligence.
Disclosed risks
Netflix's 2025 Form 10-K warns that competitors could gain an advantage by using generative AI more effectively, that partners' multi-service discovery interfaces, including those powered by generative AI, could come between Netflix and its members, and that generative AI could intensify cybersecurity risks.
Recent signals
Oct 26, 2026
Netflix described Pensive AI, which uses LLMs to extract errors from every failed data and ML workload and offers root-cause debugging through its Data Portal, APIs and MCP.
Oct 16, 2026
Netflix described AutoRCA, an internal agentic AI system whose specialist agents reason over an ontology-based knowledge graph to rank likely root causes of incidents, with agents proposing fixes for human review.
Oct 14, 2026
Netflix said its recommendation foundation model, pre-trained from scratch each month and fine-tuned daily, is now used in production personalization through embeddings, subgraph integration and fine-tuning.
Oct 14, 2026
Netflix described how it gates LLM-written recommendation explanations, generated by the hundreds of thousands each week in tests, with a human-anchored LLM judge.
Aug 28, 2026
Netflix described how multimodal content embeddings personalize artwork and video previews soon after a title launches, and said MediaFM became its default video preview embedding on all platforms after winning online tests.
Jul 30, 2026
Netflix described GenRec, a recommendation ranker built by post-training an internal foundation LLM, which it says can match or exceed its production system.
Jul 19, 2026
Elizabeth Stone told Lenny's Podcast that Netflix treats AI fluency as an expectation for everyone and discussed managing the volume of AI-generated output.
Jul 17, 2026
Netflix's AI Platform team described serving LLMs in-house on vLLM and NVIDIA Triton with an OpenAI-compatible API.
Jul 16, 2026
Ted Sarandos said on the Q2 2026 call that generative AI workflows had been used in roughly 300 Netflix titles, that The American Experiment's 17 minutes of AI-enhanced footage cost half as much as earlier options, and that savings will likely be reinvested in content.
May 13, 2026
Netflix said at its upfront that it is testing AI agents to manage, optimize and buy ads, and that AI matching of advertiser creative to Netflix titles, tested with DoorDash, Target and TurboTax, would reach every ad-supported region by year end.
Apr 16, 2026
Greg Peters said newer recommendation model architectures had raised engagement in Q1 2026 and set out where Netflix invests in generative AI: content production, the member experience and advertising.
Mar 5, 2026
Netflix acquired InterPositive, the filmmaking AI company founded by Ben Affleck, who joined Netflix as Senior Advisor.
Feb 23, 2026
Netflix introduced MediaFM, an in-house tri-modal model pretrained on its catalog for recommendations, promotional assets and ads relevancy.
Feb 13, 2026
Netflix described its LLM post-training framework, which adapts open-weight models on GPUs provisioned on AWS through its Mako compute platform.
Feb 3, 2026
Netflix promoted CTO Elizabeth Stone to Chief Product and Technology Officer, combining the CTO and chief product officer roles.
Quick answers
- How does Netflix deploy AI?
- Netflix deploys AI mainly inside its own product and studio: a recommendation foundation model and open-weight LLMs it post-trains and serves itself on AWS GPUs, generative AI tools that creative partners used on about 300 titles in 2026, and AI tools across its advertising workflow. Netflix has run machine learning in recommendations for two decades, reports no dollar return from AI, and has put Chief Product and Technology Officer Elizabeth Stone over product, engineering and data since Feb 2026.
- What does Netflix use AI for?
- Netflix runs AI in production in recommendations and the homepage, artwork and preview personalization, subtitle and promotional localization, the Clips vertical video feed, generative AI tools in film and series production (about 300 titles in 2026), advertising workflows and data platform debugging, with conversational search, LLM-written recommendation explanations, LLM-based ranking and custom AI ads still in testing. Netflix has published no business-level return credited to AI for any of these use cases or for AI as a whole; it reports engagement gains and single-production examples such as The American Experiment, whose AI footage Ted Sarandos said cost half as much as earlier options.
- Who leads AI at Netflix?
- AI at Netflix is led by Elizabeth Stone (Chief Product and Technology Officer); Ben Affleck (Senior Advisor; founder of InterPositive). Ownership of AI on the org chart: Elizabeth Stone, Chief Product and Technology Officer (since Feb 2026).
- What AI platform does Netflix use?
- Netflix builds its own AI stack: a foundation model that centralizes member preference learning, the MediaFM content model, and an AI Platform that post-trains open-weight LLMs and serves them on vLLM inside Netflix's production environment (Jul 2026).
- Which AI models does Netflix use?
- Netflix trains its own recommendation and media models and post-trains open-weight LLMs, CNBC reported in May 2025 that OpenAI powers its generative search, Eyeline licenses commercial video diffusion models (Oct 2025), and Greg Peters said in Oct 2025 that most generative AI uses at Netflix are bought tools.
- How many Netflix employees use AI?
- Netflix's recommendation models shape the homepage for more than 325 million paid memberships (Jan 2026), and creative partners used generative AI workflows on about 300 Netflix titles in 2026 (Jul 2026).
- How far has AI spread at Netflix?
- Netflix has reported no business return it credits to AI: it cites engagement gains from newer recommendation models (Apr 2026) and single-production examples, such as 17 minutes of AI footage in The American Experiment made at half the cost of earlier options (Jul 2026), and it does not separate AI's share of its ad revenue growth. Netflix has also reported no count of employees using AI tools; Chief Product and Technology Officer Elizabeth Stone, who leads product, engineering and data, says AI fluency is expected of everyone (Jul 2026). Netflix is at Experimenting despite two decades of machine learning in its product and generative AI workflows on about 300 titles in 2026; one return it credits to AI in a use case, or a reported count of employees using AI tools, would move it to Piloting.
Open questions
- Beyond OpenAI for generative search and the video diffusion models Eyeline licenses, which commercial models does Netflix use behind its model gateway and its ad tools?
- How many Netflix employees use generative AI assistants or coding tools, and which tools has Netflix sanctioned for staff?
- How much of Netflix's $3.39 billion 2025 technology and development expense, and of its AWS spend, goes to AI compute, GPUs and models?
- What business return have generative AI tools delivered across the roughly 300 titles that used them in 2026, beyond single-production examples?
- How much of Netflix's ad revenue do its AI ad tools drive, and which ad product capabilities rolled out more slowly than planned in 2026?
- Did the Oct 2026 job cuts in product and technology have any link to AI, and what annual savings does Netflix expect from them?
- How many productions use InterPositive's tools since Netflix's roughly $587 million acquisition, and has Netflix confirmed that price itself?
- When will GenRec, GenPage and LLM-written recommendation explanations move from tests into production, and what did the tests show beyond engagement?
- Who leads Netflix's AI Platform org, how many people work on AI and machine learning, and to whom does Elizabeth Stone report?
Sources
Netflix
- Netflix leadership
- Innovation for filmmaking, by filmmakers: why InterPositive is joining Netflix (Mar 5, 2026)
- Netflix opens Eyeline Studios Hyderabad (Mar 12, 2026)
- Netflix Upfront 2026: get closer (May 13, 2026)
- Peter Rice named Netflix Chief Content Officer (Oct 15, 2026)
- Netflix Tech Blog: Ontology-driven observability, an end-to-end knowledge graph at Netflix scale (Oct 2026)
- Netflix Tech Blog: Foundation model for personalized recommendation (Mar 2025)
- Netflix Tech Blog: Integrating Netflix's foundation model into personalization applications (Oct 2026)
- Netflix Tech Blog: MediaFM, the multimodal AI foundation for media understanding (Feb 2026)
- Netflix Tech Blog: MAPS, multimodal asset personalization at scale (Aug 2026)
- Netflix Tech Blog: Scaling LLM post-training at Netflix (Feb 2026)
- Netflix Tech Blog: In-house LLM serving at Netflix (Jul 2026)
- Netflix Tech Blog: Reliable GPU compute for large training and inference jobs (Oct 2026)
- Netflix Tech Blog: GenRec, towards LLM-native recommendation (Jul 2026)
- Netflix Tech Blog: GenPage, towards end-to-end generative homepage construction (Jun 2026)
- Netflix Tech Blog: The lifecycle of LLM-as-a-judge (Oct 2026)
- Netflix Tech Blog: Making failure diagnosis reliable with Pensive AI (Oct 2026)
- Netflix Tech Blog: The AI evolution of Graph Search at Netflix (Jan 2026)
- Netflix Tech Blog: Multi-ratio video resizing for mobile feed and marketing (Oct 2026)
- Netflix Tech Blog: Measuring the impact of personalized recommendations (Oct 2026)
Filings, letters and earnings calls
- Q3 2026 letter to shareholders (Oct 20, 2026)
- Q2 2026 letter to shareholders (Jul 16, 2026)
- Q1 2026 letter to shareholders (Apr 16, 2026)
- Q4 2025 letter to shareholders (Jan 20, 2026)
- Q3 2025 letter to shareholders (Oct 21, 2025)
- Q3 2026 earnings interview transcript (Oct 20, 2026)
- Q2 2026 earnings interview transcript (Jul 16, 2026)
- Q1 2026 earnings interview transcript (Apr 16, 2026)
- Q3 2025 earnings interview transcript (Oct 21, 2025)
- Q2 2025 earnings interview transcript (Jul 17, 2025)
- Q1 2025 earnings interview transcript (Apr 17, 2025)
- Netflix 2025 Form 10-K (filed Jan 23, 2026)
- Netflix Form 10-Q for Q3 2026 (filed Oct 21, 2026)
- Netflix 2026 proxy statement (filed Apr 16, 2026)
News and interviews
- Associated Press via ABC30: Netflix is laying off 6% of its workforce (Oct 14, 2026)
- Variety: Netflix promotes CTO Elizabeth Stone to Chief Product and Technology Officer (Feb 3, 2026)
- Lenny's Newsletter: Netflix CPTO Elizabeth Stone on AI and the future (Jul 19, 2026)
- CNBC: Netflix unveils revamped homepage and app with OpenAI-powered search tool (May 7, 2025)
- Engadget, citing Bloomberg: Netflix is reportedly testing a search function powered by OpenAI (Apr 12, 2025)
- TheWrap: Inside Netflix's merger of Scanline VFX and Eyeline Studios (Oct 15, 2025)
- TheWrap: Netflix paid $587 million in cash for Ben Affleck's AI startup InterPositive (Jul 17, 2026)
Data confidence
- Confirmed
- Stated in a primary or reputable public source.
- Estimated
- Derived or approximated from public reporting.
- Unconfirmed
- Reported, but not yet corroborated.
Netflix profile · Enterprise AI diffusion report by Imagine · facts checked against their sources on Nov 1, 2026
