On April 16, 2026, Samasource Impact Sourcing Inc issued formal redundancy notices to 1,108 workers at its Nairobi delivery center after Meta Platforms terminated a major data-annotation and content-processing contract. The engagement had been central to Sama’s Kenyan operation for years. Thirty-four days later, on May 20, Meta began notifying roughly 8,000 of its own global employees that they were being cut, part of a restructuring the company says is necessary to fund an AI infrastructure program running at up to $145 billion in capital expenditure this year.

Both decisions belong to the same logic. Workers in Nairobi spent years labeling images, moderating harmful video, and processing data streams that trained Meta’s AI systems. Those systems are now the primary reason Meta says it needs fewer humans. That circuit, from outsourced labor in Kenya to in-house automation in California, has run to its conclusion at both ends simultaneously.

Nairobi’s Workforce and the Contract That Powered It

Kenya’s emergence as a digital-outsourcing hub was not accidental. The government invested in fiber-optic connectivity, English-language education, and policies designed to attract global technology companies seeking lower labor costs. By 2025, Kenya’s business-process outsourcing sector (BPO, meaning contracted digital services delivered remotely for international clients) was valued at roughly $270 million annually, employing around 36,000 workers in roles spanning customer support, financial processing, and AI data services, according to industry figures compiled by Kenyan investment media.

Sama occupied a prominent position in that ecosystem. The San Francisco-headquartered firm promoted itself as an “impact sourcing” operation, deliberately placing workers from underserved communities into digital roles connected to global tech clients. Meta contracted Sama in 2019 to moderate Facebook content across sub-Saharan Africa, a task requiring local language knowledge and cultural judgment that automated systems of the era could not handle. Over years of mounting legal disputes and labor complaints, Sama exited direct content moderation and pivoted its Kenyan workforce into AI data annotation, the systematic labeling of images, video, and audio that large-scale machine-learning systems require at every stage of development. That AI-labeling workstream is what Meta has just ended.

The formal redundancy process is being conducted under Section 40 of Kenya’s Employment Act, which requires written notification to employees, the labor office, and relevant trade unions, with a minimum severance of 15 days per year of service. Sama said it had engaged Meta in an attempt to preserve the contract but acknowledged those discussions were unsuccessful. A Swedish media investigation by Svenska Dagbladet and Göteborgs-Posten, conducted alongside Kenyan journalists and published in the weeks before the notices arrived, found that footage captured by users of Meta’s AI glasses was being reviewed and labeled by contracted workers in Nairobi to train Meta’s underlying models. The investigation raised questions about user consent and the accountability distance between the platform and the people doing its most sensitive processing work that went publicly unanswered as the workforce handling that review was being dissolved.

For many of the workers now without contracts, the relationship had paid meaningfully above Kenya’s formal minimum wage but placed every strategic decision, including the decision to end it, exclusively with the foreign client.

What Sama Workers Were Building for Meta

Sama’s Nairobi center ran two related streams of work. Content moderation required workers to review graphic violence, hate speech, and extremist material flagged on Facebook across African markets. AI data annotation involved labeling images, video clips, and audio samples so Meta’s machine-learning systems could be trained to perform those judgments automatically. Both types of work required sustained human attention. Both fed the same AI pipeline that Meta is now funding at a scale that renders the human version economically redundant.

The financial terms of that arrangement have been disputed since the relationship began. Several independent assessments establish the structure of the compensation.

• 2.5× the formal minimum wage: average compensation found by a J-PAL evaluation of Sama’s Nairobi delivery center, with benefits including healthcare, pension, and meal subsidies
• $414/month: what some former moderators told The Associated Press they were paid while reviewing traumatic content for eight-hour daily shifts
• $2.20/hour: the hourly rate described by former moderator Daniel Motaung, who reviewed beheadings and child abuse material under a Sama contract for Meta

The gap between those wages and the scale of the infrastructure they supported is not incidental. A 2026 peer-reviewed study presented at the ACM Conference on Human Factors in Computing Systems, based on interviews with Kenyan data workers, described their position as a regime of entrapment in Kenya’s AI supply chain, held in place by short-term contracts, non-disclosure agreements, and a structural barrier to moving into higher-value roles. The workers who labeled Meta’s training data were, in aggregate, subsidizing an AI buildout in which they held no equity and over which they had no say.

Sama, not Meta, was the legal employer of record throughout. That structure meant Meta could close the contract without the Nairobi job losses appearing anywhere on its own headcount figures. The 1,108 workers simply ceased to be a line item.

A Legal Fight That Has Outlasted Two Contracts

Kenya’s courts have been working through the consequences of Meta’s outsourcing model since 2022. In March 2023, 43 former content moderators, backed by Foxglove, a U.K.-based technology rights nonprofit, filed suit alleging Meta had directed its replacement contractor, Luxembourg firm Majorel, not to hire former Sama employees, treating the contract cancellation as retaliation for worker organizing rather than genuine redundancy. A separate case involving 185 moderators from across African markets sought $1.6 billion in compensation for psychological harm and the absence of adequate safeguards against the cumulative toll of years spent reviewing violent imagery. Settlement talks opened by court order collapsed in October 2023, with Foxglove describing Meta’s conduct in those negotiations as insincere. Meta argued throughout that Kenyan courts had no jurisdiction over a company not incorporated in Kenya and that the moderators were employees of Sama, not of Meta itself.

As is standard in our industry, client programmes evolve, and we work closely with our partners to manage these transitions responsibly.

Annepeace Alwala, Sama’s country lead and vice-president for global delivery, offered that statement when the April 2026 redundancy notices were filed. It mirrors language the company used in earlier contract transitions. The Kenyan judiciary has been less willing to accept the underlying framing. In 2024, Kenya’s Court of Appeal ruled that Meta could be sued within Kenya, rejecting years of jurisdictional arguments. The 185-moderator case is now proceeding on that basis, with liability claims against a company whose newest round of contract terminations has just added several hundred more potential plaintiffs to the broader conversation about working conditions in its outsourced ecosystem.

Meta’s Financial Equation

The Numbers Behind the Restructuring

Meta’s decision to cut roughly 8,000 direct employees globally while accelerating AI infrastructure spending reflects a specific corporate arithmetic. Analysts at Evercore estimate the May round will generate approximately $3 billion in annualized savings, against a capital expenditure guidance range for 2026 that runs between $115 billion and $145 billion, nearly double the $72.2 billion the company spent in 2025. The savings, in other words, cover roughly two cents of every dollar the company plans to spend on AI infrastructure this year.

The May round is the largest companywide reduction since Mark Zuckerberg’s “Year of Efficiency” campaign in 2022 and 2023, which removed roughly 21,000 positions. Meta’s full-year revenue crossed $200 billion for the first time in 2025. These cuts are not a response to financial pressure; they are a reallocation of cost from people to machines, and the savings they generate are a rounding error against the infrastructure bill they are meant to partially offset.

The AI Infrastructure Replacing Those Roles

In June 2025, Meta hired Alexandr Wang, the former chief executive of Scale AI, as its first chief AI officer through a transaction that included a $14.3 billion investment in Scale AI. Wang now leads Meta Superintelligence Labs, the division that debuted its first major model in early May. Alongside that organizational shift, the company is building Prometheus, a one-gigawatt AI supercluster in Ohio, and Hyperion, a 2,250-acre facility in Louisiana rated for five gigawatts of AI computing capacity.

Scale AI, before Wang’s departure, was itself a significant buyer of outsourced data-labeling work from contractors in Kenya and comparable lower-cost markets. Meta’s move toward absorbing that capability in-house describes precisely what happened to Sama at the bottom of the same supply chain, only one level up. The company that organized the labelers has been acquired by the platform. The workers who did the actual labeling were not part of that transaction, and there is no equivalent mechanism for them to follow the work upward.

The Automation Floor Under Kenya’s Outsourcing Sector

The Sama layoffs are the most visible single event in a structural problem that has been accumulating for years. Kenya’s government has publicly set a target of 1 million BPO jobs within five years and secured a $10 million investment pledge from ADEC Innovations at the Kenya International Investment Conference held weeks before the redundancy notices arrived. The aggregate numbers for the sector look optimistic. The contract-level reality is more fragile.

• Concentration risk: A small number of U.S. tech companies account for a disproportionate share of Kenya’s BPO volume; the Sama-Meta contract showed how a single termination can remove over a thousand positions overnight with no domestic absorber capable of replacing the demand.
• Automation exposure: Research by Caribou Digital and Genesis Analytics, conducted with support from the Mastercard Foundation, found 40% of tasks in Africa’s tech outsourcing sector could be automated by 2030, with customer-experience roles (44% of current sector employment) carrying the highest near-term risk.
• Gender disparity: The same research found tasks performed by women are on average 10% more vulnerable to automation than tasks performed by men in the same sector, concentrating displacement in the workforce segment with the fewest alternative pathways into higher-value digital roles.
• Value-chain position: Kenya ranks 11th globally as an outsourcing destination by the 2026 Ataraxis Global Outsourcing Talent Index, a genuine competitive strength; it does not address the fundamental issue that every contract renewal or cancellation is decided by a handful of foreign clients whose own technology strategies, not Kenya’s workforce policy, determine how many contracts exist.

A 2026 CHI Conference paper drawn from interviews with Kenyan data workers described the sector’s core tension in terms the workers themselves had identified: the AI models that required human training data to become functional are, once functional, being deployed to remove the humans who generated that data from the next version of the pipeline. The workers being let go helped build the reason they are being let go.

Two outcomes will determine whether the Sama layoffs become a turning point or a template. Kenya’s appellate courts now have jurisdiction over Meta, and the 185-moderator case is proceeding. If those plaintiffs win meaningful compensation, every global platform that has outsourced sensitive AI labor to African contractors will need to recalculate how it structures liability and what counts as an employment relationship. If the case resolves without consequence, the model that built Kenya’s digital-labor economy, and that is now eroding it, carries on under exactly the same terms as before.

EBSCO Information Services, one of the world’s largest academic database providers, launched EBSCOhost AI Exchange on May 20, 2026, positioning the company as a licensed intermediary between peer-reviewed scholarly content and the AI tools now generating first-pass research answers for millions of students, clinicians, and knowledge workers. The platform connects AI systems directly to EBSCO’s holdings of journals, reference databases, and conference papers under a framework the company says enforces licensed access, source attribution, and content delivery aligned with each institution’s existing subscription rights.

The Ipswich, Massachusetts company timed the launch one day after announcing a partnership with Perplexity, the AI-powered answer engine widely used in higher education and professional research settings. The pairing makes a single argument: that AI platforms will need a credentialed, licensed channel into academic publishing, and EBSCO means to own that channel.

How EBSCOhost AI Exchange Works

The Exchange places itself between three groups: academic publishers holding content rights, AI systems that want to cite verified sources, and the libraries or institutions already paying for EBSCOhost subscriptions. According to EBSCO’s official EBSCOhost AI Exchange launch announcement, the platform provides a governed framework supporting "licensed access, proper attribution and content delivery aligned with existing subscriptions and permissions."

For an AI tool connected to the Exchange, the process runs through an API (application programming interface, the technical layer that lets AI software query an external database directly without a human browser session). The AI sends a query, retrieves relevant journal content, and returns a citable response that links users back to the original article, gated by their institutional login. Organizations already paying for EBSCOhost databases gain a machine-readable connection that AI vendors can integrate without negotiating separately with each individual journal publisher.

EBSCO says the platform supports four categories of AI deployment: commercial tools such as consumer-facing answer engines, institutional models built by universities or hospitals, enterprise platforms deployed by corporations and government agencies, and RAG applications. RAG, or retrieval-augmented generation, describes the technique where an AI system queries an external database before drafting a response, rather than relying solely on what was absorbed during training. Most enterprise AI vendors building factual-answer products are working with RAG pipelines right now, and EBSCO is positioning the Exchange as a licensed, subscription-aligned data source for exactly those systems, one that requires no separate publisher negotiation on the AI company’s side.

Perplexity and the First Live Integration

The EBSCO-Perplexity partnership announced May 19 brings EBSCOhost databases into Perplexity’s Premium Sources tier and extends to Perplexity Computer, the platform’s multi-step research task tool. What a user sees depends on their institutional affiliation: general users receive a reference citation and a link to the source article, while users with valid institutional EBSCOhost logins can click through to the full journal text, tracing any AI-generated answer back to the underlying research paper.

The deal’s significance runs beyond the two companies involved. Perplexity has grown into one of the most active AI search tools in research and education settings, meaning real-world usage of the integration will accumulate quickly and produce observable data on whether licensed scholarly sourcing changes how users interact with AI answers. Sam Brooks, Executive Vice President at EBSCO Information Services, framed the platform’s purpose plainly: "AI is the front door to research for many users, and that makes quality sources more important than ever." Financial terms of the arrangement have not been publicly disclosed.

Libraries and Publishers Find Structural Allies

The 70,000 libraries, academic institutions, hospitals, and research organizations worldwide that already hold EBSCOhost subscriptions are the structural anchors of the Exchange’s value proposition. Those institutions have paid for database access for decades while AI tools simultaneously trained on or summarized academic content through channels that returned neither readers nor revenue to the library systems those institutions fund. For students who now reach for a general-purpose AI tool before opening a library database, the library’s holdings have become effectively invisible, a problem that has placed database contract renewals under scrutiny at institutions of every size.

EBSCOhost AI Exchange proposes a corrective flow: AI research answers that route back through institutional subscription infrastructure, giving library contracts a visible role in the AI-powered research experience. Budget holders fielding internal questions about the ongoing value of database spending now have a specific counterargument: the library’s licensed content is what makes the AI answer citable and independently verifiable, two attributes a general-purpose AI tool without licensed sourcing cannot reliably provide.

For publishers, the platform addresses attribution rather than access alone. Melissa D’Amato, Senior Vice President of Publisher Services at EBSCO, said publishers need "a responsible path into AI-supported discovery, one that keeps their content visible, valued, and properly attributed." That sentence names a fear that has deepened throughout the AI boom: journal content appearing in AI-generated summaries without source credit, without driving reader traffic to publisher platforms, without producing licensing revenue.

The stakeholders the Exchange is structured to serve, by primary concern:

• Libraries: AI query traffic routed back through institutional subscriptions, reinforcing the case for existing database contracts at a moment when their value is under scrutiny
• Academic publishers: structured attribution and content visibility inside AI answer workflows, replacing the current pattern of invisibility
• AI platforms: a governed, licensed source of peer-reviewed content with defined attribution rules, reducing copyright exposure
• Researchers and students: traceable citations in AI-generated answers rather than outputs that cannot be independently verified

Academic Content in the AI Licensing Market

A Benchmark Set by a Few Large Deals

EBSCO is entering a licensing market for academic content that has precedent but not yet meaningful volume. Several major publishers have completed direct AI arrangements in recent years. Springer Nature concluded a deal with Google worth a reported $23 million. Wiley’s chief executive described executing a "$23 million content rights project with a large tech company." Taylor and Francis secured $10 million upfront plus recurring payments extending through 2027, combining immediate payment with ongoing usage-based revenue in a structure more complex than pure one-time licensing.

Those were direct publisher-to-AI-company transactions, each requiring separate bilateral negotiations. EBSCOhost AI Exchange takes a different architecture: EBSCO aggregates content from its publishing network and offers AI companies a single integration point, cutting the transaction cost of repeated one-on-one talks. For smaller academic journals that would never appear on OpenAI’s or Google’s direct negotiation list, that aggregation opens an AI-distribution path otherwise unavailable to them.

The broader AI content licensing market has been shifting from ad-hoc scraping toward structured marketplace infrastructure, with Microsoft and Amazon moving in 2026 toward formal content deal frameworks under defined commercial terms. Academic content is the segment of that shift where EBSCO’s existing publisher relationships translate directly into negotiating leverage. No AI company without decades of academic publishing relationships could replicate that starting position quickly, which is the point EBSCO is making by launching the Exchange now rather than waiting for a larger competitor to build a substitute.

The Author Attribution Gap

A question the market has not resolved, and which the Exchange sidesteps rather than answers, involves the researchers whose work gets licensed. Cambridge University Press sent published authors addendum requests to allow their content to be licensed to undisclosed AI partners. Some publishers signed blanket licensing agreements without informing the authors whose work was covered, a practice documented in Boston College Law School’s generative AI and scholarly publishing copyright guide. Others offered opt-in arrangements, presenting each author with a separate agreement before licensing their content to AI systems.

EBSCOhost AI Exchange focuses on institutional access and publisher attribution rather than individual researcher notification or compensation. For a discovery platform built on top of existing publisher contracts, that boundary is defensible. But as scholarly content becomes a primary grounding layer for AI-generated research answers, the question of whether any of the value flowing through content licensing reaches the scholars who created the underlying work remains open, and no major platform launch in this cycle has attempted to answer it.

The Open Question Behind the Launch

Paywalled Content Cannot Be Scraped

The platform rests on an assumption that AI companies want licensed, attributed academic content enough to route queries through a single governed gateway rather than sourcing content through other means. That assumption is stronger today than it was two years ago, given copyright litigation that has reshaped the AI training data market and the reputational damage AI tools sustain when they produce fabricated journal citations. But it remains an assumption, and the commercial AI market has a long history of finding lower-cost alternatives to licensed content when they exist.

On the open web, blocking AI training crawlers has increasingly come to mean sacrificing search visibility, because Google’s indexing operations and its AI data collection have merged into one system. Publishers face a binary choice between total access and total exclusion. Most cannot afford the latter. Academic databases sit entirely outside that dynamic. EBSCOhost’s content lives behind institutional paywalls, was never part of the public web index, and cannot be obtained by scraping. Content that requires a valid institutional subscription must be licensed to reach AI systems at all. That structural fact gives EBSCO negotiating leverage that most open-web content owners surrendered when they chose discoverability over restriction.

Three major EBSCO partnership announcements landed in a single month: the Exchange launch on May 20, the integration with a top-tier AI search platform on May 19, and a May 11 deal with GetFTR to route researchers from open-web sources directly to full-text articles on EBSCOhost. The concentration reads less like coincidence than coordinated market positioning under a new chief executive, designed to signal momentum to AI companies and library budget holders simultaneously.

The Variables EBSCO Cannot Control

Whether AI companies move toward licensed academic content at scale is a separate question from whether they should. Open-access preprint repositories, synthetic grounding data, and training-only licensing deals could each reduce demand for a live-query API like the Exchange. The platform is most valuable to AI companies building retrieval-based products where freshness, provenance, and citability matter to end users. Its value is lower for companies building general LLMs (large language models, AI systems trained on large text datasets to generate responses rather than retrieve specific documents in real time) that need historical training archives rather than live query access to current literature.

• 70,000 libraries, hospitals, and academic institutions worldwide hold active EBSCOhost subscriptions, forming the institutional base the Exchange is built to serve
• $23 million one-time payment: the benchmark set by both the Springer Nature-Google deal and Wiley’s undisclosed partner agreement for academic journal archives
• Three major EBSCO partnership announcements arrived in May 2026 alone, covering AI answer access, full-text routing via GetFTR, and the AI Exchange marketplace itself
• One day separated the first live AI integration from the Exchange launch, a gap consistent with coordinated rather than sequential market positioning

Allen Powell, appointed Chief Executive Officer of EBSCO Information Services on May 4, 2026 after more than three decades at the company, steps into a concentrated institutional bet on one thesis: that licensed, attributed scholarly content becomes a non-negotiable input for any AI research tool that needs to be trusted by libraries, hospitals, and academic institutions. Powell previously served as interim CEO, executive vice president, and chief financial officer, giving him unusual continuity with the strategy he now leads.

If the first live AI integration demonstrates measurable citation quality gains and other AI platforms follow toward the Exchange, EBSCO occupies a chokepoint with no obvious substitute and a library customer base already paying for the content. If AI companies conclude that open-access preprints, synthetic grounding, and training-only archives are sufficient for their users, the Exchange serves a specialized market where citation provenance carries legal or professional weight, including healthcare, government research, and university settings with accreditation requirements. Those are substantial markets. They are not, however, the full size of the bet being placed.

OpenAI shuttered the Sora standalone app on April 26, 2026, less than six months after its public debut. On May 19, Google answered with Gemini Omni Flash, announced at I/O 2026 and already live across the Gemini app, Google Flow, and YouTube Shorts before the press cycle had finished. The model does not just generate video from text prompts; it edits through conversation, treating each new instruction as a continuation of the prior one rather than a fresh start.

Runway Gen-4.5 already leads professional editors on character consistency; Kling 3.0 delivers competitive quality at a fraction of the cost. Google’s bet is on workflow: feed the model a reference image or an existing clip, and talk it into whatever you actually wanted, with each exchange preserving what came before.

The Conversational Layer Arrives

Prior text-to-video tools operated as one-shot generators. Prompt once, receive a clip, evaluate it, restart if the output drifted from the brief. Google’s official Gemini Omni introduction positions the model as a break from that cycle: every instruction builds on the last, with character identity, scene continuity, and physical logic carried forward across multiple exchanges. Google describes it as Nano Banana for video, a reference to the image-generation model that preceded Omni in the multimodal product line and helped millions restore and redesign photos before the approach was extended to moving images.

Every Omni output carries a non-optional SynthID digital watermark, combined with C2PA Content Credentials, the industry standard for provenance metadata that documents how media was created and modified. The watermark is designed to survive compression, cropping, and common file transforms. Verification is available via the Gemini app and Chrome, with Google Search verification announced as forthcoming.

The three capabilities Google emphasizes on its product pages sit in a deliberate sequence: an improved intuitive understanding of physical forces including gravity, kinetic energy, and fluid dynamics; world knowledge drawn from Gemini’s training in history, science, and cultural context; and character consistency across multi-turn revisions, where prior video models have tended to drift on identity between edits. That third capability is also Runway’s central selling proposition, which is where the competitive signal lives.

What Omni Flash Generates

Multi-Turn Editing in Practice

The conversational workflow lets users revise without regenerating from scratch. A starting scene can be transported to a new environment, a specific object removed, a camera angle shifted to over-the-shoulder, all through separate instructions that each preserve what the prior step established. Google DeepMind’s Gemini Omni product page frames the editing premise with a direct statement from the launch:

Your video becomes the starting point for something you never could have filmed yourself.

The line appeared in Google’s official Omni product blog, describing what conversational editing changes about video production: existing footage becomes raw material for visual transformations no camera operator could produce on location.

According to Google’s official documentation, editing tasks Omni Flash supports through plain-language conversation include:

• Background environment swaps that preserve the foreground subject
• Wardrobe, style, and artistic treatment changes across a clip
• Specific object substitution mid-shot
• Lighting intensity and mood adjustments via single instructions
• Camera angle and composition changes without restarting generation

At launch, Omni Flash generates clips capped at 10 seconds. Google’s product blog describes this as a deployment decision rather than a model constraint, suggesting the limit is expected to extend as supporting infrastructure scales. Audio input accepts only voice references at launch; other audio input types are announced as coming later.

Physics and World Knowledge

Generative video has long struggled with physical coherence: a marble that defies gravity, water flowing upward, hands multiplying between frames. Omni Flash claims an improved intuitive understanding of forces like gravity, kinetic energy, and fluid dynamics for more realistic scene generation. Independent benchmarks comparing Omni Flash directly to Veo 3.1 or Runway Gen-4.5 on physics accuracy had not been published as of this writing.

The world-knowledge angle is the more architecturally distinctive claim. Omni draws on Gemini’s training to connect language, imagery, and meaning in ways that go beyond pattern matching, a distinction Google draws by contrasting photorealism with meaningful storytelling. A prompt for a claymation explainer of protein folding produces a stop-motion clip with scientifically coherent folding sequences, according to Google’s launch demos. An alphabet video requiring 26 unusual objects stresses whether the model understands concepts or mimics visual patterns: the model has to reason about what unusual means across 26 letters simultaneously.

YouTube as the Distribution Weapon

No AI video tool has walked into a user base comparable to YouTube’s. Google is rolling Omni Flash out to YouTube Shorts and the YouTube Create App at no additional cost to users, inside the creation tools hundreds of millions of people open daily, with no separate subscription or API key required. Paid access runs through Google AI Plus, Pro, and Ultra subscriptions globally, with subscribers receiving Omni Flash in the Gemini app and Google Flow alongside generation credits. Developer and enterprise API access is positioned as arriving in the coming weeks, with no specific date confirmed at I/O.

The go-to-market structure separates Omni from every standalone AI video tool competing for the same creator audience. Runway, Kling, and Seedance all charge per-video, per-second, or through monthly credit bundles. Google is using YouTube’s distribution to put the model in front of users at no marginal cost, then monetizing through the subscription tier above. That pricing architecture is difficult to replicate for a company that does not own the world’s largest video platform.

Google Flow received simultaneous upgrades. Google Flow’s creative studio gained Omni Flash for conversational iterative editing with improved character consistency across scenes. Flow Agent, announced alongside Omni, acts as a creative partner for brainstorming, planning, and scene reasoning under the user’s direction. Flow Tools lets any subscriber build custom video-processing presets using plain language, no coding required, and an early-access partner built a lo-fi and glitch aesthetic post-processing tool that other creators can remix. Flow Music, rebranded from ProducerAI in April 2026, gains Omni Flash for conversational music video direction. Both Google Flow and Flow Music are launching dedicated mobile apps alongside these upgrades: Android beta first for Flow, iOS first for Flow Music, with each platform’s reverse release to follow.

Rivals Caught in a Restructured Market

The Sora Exit and the Vacuum It Left

OpenAI announced in late March 2026 that Sora’s standalone web app and mobile experience would shut down April 26. The Sora application programming interface (API, the gateway that lets developers integrate video generation into their own software) continues through September 24, 2026, giving production pipelines time to migrate. Multiple analysts have framed the exit as a compute-economics failure: Sora produced compelling clips it could not commercialize at a price that recovered generation costs, and by Q1 2026 four competitors had matched or exceeded its quality benchmarks.

The market Sora helped validate has not contracted. Venture capital investment in AI video reached $4.7 billion in 2025, and the overall revenue market is growing at a compound annual growth rate of 34.2%. Sora’s exit clarified the tier structure rather than creating a vacuum.

• $4.7 billion in VC investment into AI video in 2025, a 189% increase from 2023
• $2.4 billion in current AI video market revenue
• 34.2% compound annual growth rate for AI video generation
• April 26, 2026: Sora app shut down; the API runs through September 24

Where the Incumbents Stand

Runway’s position is the most interesting pressure test. Gen-4.5 leads on character consistency and reference image control, making it the default for professional advertising and post-production workflows where brand continuity across dozens of clip variations is the actual deliverable. Those workflows are not migrating to a 10-second chat-based generator overnight. But Omni’s free YouTube tier competes directly for the social creator segment that Runway has been building toward with its faster Turbo-tier variants, and that is the segment where the next wave of paying subscribers forms.

The Safety Brake Google Won’t Remove Yet

The single most conspicuous absence from Omni Flash at launch is audio and speech editing of generated videos. Users can supply a voice reference to shape new audio in a generated scene, but they cannot take an existing clip and alter what the people in it are saying. Google’s official launch language: the company is “still working to test this and better understand how we can bring this capability to users responsibly.”

The logic is not hard to follow. Omni Flash launched into a political calendar that had just produced one of the most contentious election cycles in recent US memory. A tool that lets any subscriber revoice a real video through a chat interface, with output carrying Google’s own watermark as a provenance signal, creates liability the company has not yet resolved. The withholding is a deliberate deployment decision on a capability the model apparently already has, not an architectural gap waiting to be filled.

The avatar feature is the bounded exception. Users can build a digital likeness of themselves using their own voice and appearance, with a structured onboarding process that requires recording oneself reading a series of numbers aloud. That stored likeness can be reused across future sessions without re-uploading. Google has constrained the feature to the user’s own likeness; it is not a general-purpose face-replacement or revoicing tool.

Google’s content transparency expansion announced alongside Omni situates the watermarking inside a broader cross-industry standard. SynthID has been used to mark more than 100 billion images and videos across Google’s services, and on the same day Omni launched, OpenAI separately announced it was adopting SynthID for ChatGPT-generated images. The infrastructure Google is building around Omni’s provenance is being assembled at an industry level, not locked to a single product.

Developer API access arrives in the coming weeks according to Google, at which point independent benchmark comparisons with Veo 3.1, Runway Gen-4.5, and Seedance 2.0 become possible across standardized test sets. That benchmark profile matters less than the audio question. If Google clears its internal safety review and ships conversational voice editing before the end of this year, Omni Flash becomes the first broadly available tool that can credibly alter what any video subject appears to say. If the holdback extends, Runway’s professional-tier differentiation survives longer than its current market position would suggest it should.