Mistral AI on Monday launched Forge, an enterprise model training platform that allows organizations to build, customize, and continuously improve AI models using their own proprietary data — a move that positions the French AI lab squarely against the hyperscale cloud providers in one of the most consequential and least understood markets in enterprise technology.

The announcement caps a remarkably aggressive week for Mistral, which also released its Mistral Small 4 model, unveiled Leanstral — an open-source code agent for formal verification — and joined the newly formed Nvidia Nemotron Coalition as a co-developer of the coalition’s first open frontier base model. Together, these moves paint the picture of a company that is no longer content to compete on model benchmarks alone and is instead racing to become the infrastructure backbone for organizations that want to own their AI rather than rent it.

Forge goes significantly beyond the fine-tuning APIs that Mistral and its competitors have offered for the past year. The platform supports the full model training lifecycle: pre-training on large internal datasets, post-training through supervised fine-tuning, DPO, and ODPO, and — critically — reinforcement learning pipelines designed to align models with internal policies, evaluation criteria, and operational objectives over time.

“Forge is Mistral’s model training platform,” said Elisa Salamanca, head of product at Mistral AI, in an exclusive interview with VentureBeat ahead of the launch. “We’ve been building this out behind the scenes with our AI scientists. What Forge actually brings to the table is that it lets enterprises and governments customize AI models for their specific needs.”

Why Mistral says fine-tuning APIs are no longer enough for serious enterprise AI

The distinction Mistral is drawing — between lightweight fine-tuning and full-cycle model training — is central to understanding why Forge exists and whom it serves.

For the past two years, most enterprise AI adoption has followed a familiar pattern: companies select a general-purpose model from OpenAI, Anthropic, Google, or an open-source provider, then apply fine-tuning through a cloud API to adjust the model’s behavior for a narrow set of tasks. This approach works well for proof-of-concept deployments and many production use cases. But Salamanca argues that it fundamentally plateaus when organizations try to solve their hardest problems.

“We had a fine-tuning API relying on supervised fine-tuning. I think it was kind of what was the standard a couple of months ago,” Salamanca told VentureBeat. “It gets you to a proof-of-concept state. Whenever you actually want to have the performance that you’re targeting, you need to go beyond. AI scientists today are not using fine-tuning APIs. They’re using much more advanced tools, and that’s what Forge is bringing to the table.”

What Forge packages, in Salamanca’s telling, is the training methodology that Mistral’s own AI scientists use internally to build the company’s flagship models — including data mixing strategies, data generation pipelines, distributed computing optimizations, and battle-tested training recipes. She drew a sharp line between Forge and the open-source tools and community tutorials that are freely available today.

“There’s no platform out there that provides you real-world training recipes that work,” Salamanca said. “Other open-source repositories or other tools can give you generic configurations or community tutorials, but they don’t give you the recipe that’s been validated — that we’ve been doing for all of our flagship models today.”

From ancient manuscripts to hedge fund quant languages, early customers reveal what off-the-shelf AI can’t do

The obvious question facing any product like Forge is demand. In a market where GPT-5, Claude, Gemini, and a growing fleet of open-source models can handle an enormous range of tasks, why would an enterprise invest the time, compute, and expertise required to train its own model from scratch?

Salamanca acknowledged the question head-on but argued that the need emerges quickly once companies move beyond generic use cases. “A lot of the existing models can get you very far,” she said. “But when you’re looking at what’s going to make you competitive compared to your competition — everyone can adopt and use the models that are out there. When you want to go a step beyond that, you actually need to create your own models. You need to leverage your proprietary information.”

The real-world examples she cited illustrate the edges of the current model ecosystem. In one case, Mistral worked with a public institution that had ancient manuscripts with missing text from damaged sections. “The models that were available were not able to do this because they’ve never seen the data,” Salamanca explained. “Digitization was not very good. There were some unique patterns and characters, and so we actually created a model for them to fill in the spans. This is now used by their researchers, and it’s accelerating their publication and understanding of these documents.”

In another engagement, Mistral partnered with Ericsson to customize its Codestral model for legacy-to-modern code translation. Ericsson, Salamanca said, has built up half a decade of proprietary knowledge around an internal calling language — a codebase so specialized that no off-the-shelf model has ever encountered it. “The concrete impact is like turning a year-long manual migration process, where each engineer needs six months of onboarding, to something that’s really more scalable and faster,” she said.

Perhaps the most telling example involves hedge funds. Salamanca described working with financial firms to customize models for proprietary quantitative languages — the kind of deeply guarded intellectual property that these firms keep on-premises and never expose to cloud-hosted AI services. Using Forge’s reinforcement learning capabilities, Mistral helped one hedge fund develop custom benchmarks and then trained the model to outperform on them, producing what Salamanca called “a unique model that was able to give them the competitive edge that was needed.”

How Forge makes money: license fees, data pipelines, and embedded AI scientists

Forge’s business model reflects the complexity of enterprise model training. According to Salamanca, it operates across several revenue streams. For customers who run training jobs on their own GPU clusters — a common requirement in highly regulated or IP-sensitive industries — Mistral does not charge for compute. Instead, the company charges a license fee for the Forge platform itself, along with optional fees for data pipeline services and what Mistral calls “forward-deployed scientists” — embedded AI researchers who work alongside the customer’s team.

“No competitor out there today is kind of selling this embedded scientist as part of their training platform offering,” Salamanca said.

This model has clear echoes of Palantir’s early playbook, where forward-deployed engineers served as the critical bridge between powerful software and the messy reality of enterprise data. It also suggests that Mistral recognizes a fundamental truth about the current state of enterprise AI: the technology alone is not enough. Most organizations lack the internal expertise to design effective training recipes, curate data at scale, or navigate the treacherous optimization landscape of distributed GPU training.

The infrastructure itself is flexible. Training can happen on Mistral’s own clusters, on Mistral Compute (the company’s dedicated infrastructure offering), or entirely on-premises within the customer’s own data centers. “We have all these different cases, and we support everything,” Salamanca said.

Keeping proprietary data off the cloud is Forge’s sharpest selling point

One of the sharpest points of differentiation Mistral is pressing with Forge is data privacy. When customers train on their own infrastructure, Salamanca emphasized that Mistral never sees the data at all.

“It’s on their clusters, it’s with their data — we don’t see anything of it, and so it’s completely under their control,” she said. “I think this is something that sets us apart from the competition, where you actually need to upload your data, and you have a black box effect.”

This matters enormously in sectors like defense, intelligence, financial services, and healthcare, where the legal and reputational risks of exposing proprietary data to a third-party cloud service can be deal-breakers. Mistral has already partnered with organizations including ASML, DSO National Laboratories Singapore, the European Space Agency, Home Team Science and Technology Agency Singapore, and Reply — a roster that suggests the company is deliberately targeting the most data-sensitive corners of the enterprise market.

Forge also includes data pipeline capabilities that Mistral has developed through its own model training: data acquisition, curation, and synthetic data generation. “Data is a critical piece of any training job today,” Salamanca said. “You need to have good data. You need to have a good amount of data to make sure that the model is going to be good performing. We’ve acquired, as a company, really great knowledge building out these data pipelines.”

In the age of AI agents, Mistral argues that custom models still matter more than MCP servers

The timing of Forge’s launch raises an important strategic question. The AI industry in 2026 has been consumed by agents — autonomous AI systems that can use tools, navigate multi-step workflows, and take actions on behalf of users. If the future belongs to agents, why does the underlying model matter? Can’t companies simply plug into the best available frontier model through an MCP server or API and focus their energy on orchestration?

Salamanca pushed back on this framing with conviction. “The customers that we’ve been working on — some of these specific problems are things that no MCP server would ever solve,” she said. “You actually need that intelligence. You actually need to create that model that will help you solve your most critical business problem.”

She also argued that model customization is essential even in purely agentic architectures. “There are some agentic behaviors that you need to bring to the model,” Salamanca said. “It can be about reasoning patterns, specific types of documentation, making sure that you have the right reasoning traces. Even in these cases where people are going completely agentic, you still need model customization — like reinforcement learning techniques — to actually get the right level of performance.”

Mistral’s press release makes this connection explicit, arguing that custom models make enterprise agents more reliable by providing deeper understanding of internal environments: more precise tool selection, more dependable multi-step workflows, and decisions that reflect internal policies rather than generic assumptions.

The platform also supports an “agent-first” design philosophy. Forge exposes interfaces that allow autonomous agents — including Mistral’s own Vibe coding agent — to launch training experiments, find optimal hyperparameters, schedule jobs, and generate synthetic data. “We’ve actually been building Forge in an AI-native way,” Salamanca said. “We’re already testing out how autonomous agents can actually launch training experiments.”

Mistral Small 4, Leanstral, and the Nvidia coalition: the week that redefined the company’s ambitions

To fully appreciate Forge’s significance, it helps to view it alongside the other announcements Mistral made in the same week — a barrage of releases that together represent the most ambitious expansion in the company’s short history.

Just yesterday, Mistral released Leanstral, the first open-source code agent for Lean 4, the proof assistant used in formal mathematics and software verification. Leanstral operates with just 6 billion active parameters and is designed for realistic formal repositories — not isolated math competition problems. On the same day, Mistral launched Mistral Small 4, a mixture-of-experts model with 119 billion total parameters but only 6 billion active per query, running 40 percent faster than its predecessor while handling three times more queries per second. Both models ship under the Apache 2.0 license — the most permissive open-source license in wide use.

And then there is the Nvidia Nemotron Coalition. Announced at Nvidia’s GTC conference, the coalition is a first-of-its-kind collaboration between Nvidia and a group of AI labs — including Mistral, Perplexity, LangChain, Cursor, Black Forest Labs, Reflection AI, Sarvam, and Thinking Machines Lab — to co-develop open frontier models. The coalition’s first project is a base model co-developed specifically by Mistral AI and Nvidia, trained on Nvidia DGX Cloud, which will underpin the upcoming Nvidia Nemotron 4 family of open models.

“Open frontier models are how AI becomes a true platform,” said Arthur Mensch, cofounder and CEO of Mistral AI, in Nvidia’s announcement. “Together with Nvidia, we will take a leading role in training and advancing frontier models at scale.”

This coalition role is strategically significant. It positions Mistral not merely as a consumer of Nvidia’s compute infrastructure but as a co-creator of the foundational models that the broader ecosystem will build upon. For a company that is still a fraction of the size of its American competitors, this is an outsized seat at the table.

Forge takes aim at Amazon, Microsoft, and Google — and says they can’t go deep enough

Forge enters a market that is already crowded — at least on the surface. Amazon Bedrock, Microsoft Azure AI Foundry, and Google Cloud Vertex AI all offer model training and customization capabilities. But Salamanca argued that these offerings are fundamentally limited in two respects.

First, they are cloud-only. “In one set of cases, it’s very easy to answer — they want to run this on their premises, and so all these tools that are available on the cloud are just not available for them,” Salamanca said. Second, she argued that the hyperscalers’ training tools largely offer simplified API interfaces that don’t provide the depth of control that serious model training requires.

There is also the dependency question. Salamanca described digital-native companies that had built products on top of closed-source models, only to have a new model release — more verbose than its predecessor — crash their production pipelines. “When you’re relying on closed-source models, you are also super dependent on the updates of the model that have side effects,” she warned.

This argument resonates with the broader sovereignty narrative that has powered Mistral’s rise in Europe and beyond. The company has positioned itself as the alternative for organizations that want to own their AI stack rather than lease it from American hyperscalers. Forge extends that argument from inference to training: not just running models you own, but building them in the first place.

The open-source foundation matters here, too. Mistral has been releasing models under permissive licenses since its founding, and Salamanca emphasized that the company is building Forge as an open platform. While it currently works with Mistral’s own models, she confirmed that support for other open-source architectures is planned. “We’re deeply rooted into open source. This has been part of our DNA since the beginning, and we have been building Forge to be an open platform — it’s just a question of a matter of time that we’ll be opening this to other open-source models.”

A co-founder’s departure to xAI underscores why Mistral is turning expertise into a product

The timing of Forge’s launch also arrives against a backdrop of fierce talent competition. As FinTech Weekly reported on March 14, Devendra Singh Chaplot — a co-founder of Mistral AI who headed the company’s multimodal group and contributed to training Mistral 7B, Mixtral 8x7B, and Mistral Large — left to join Elon Musk’s xAI, where he will work on Grok model training. Chaplot had previously also been a founding member of Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati.

The loss of a co-founder is never insignificant, but Mistral appears to be compensating with institutional capability rather than individual brilliance. Forge is, in essence, a productization of the company’s collective training expertise — the recipes, the pipelines, the distributed computing optimizations — in a form that can scale beyond any single researcher. By packaging this knowledge into a platform and pairing it with forward-deployed scientists, Mistral is attempting to build a durable competitive asset that doesn’t walk out the door when a key hire departs.

Mistral’s big bet: the companies that own their AI models will be the ones that win

Forge is a bet on a specific theory of the enterprise AI future: that the most valuable AI systems will be those trained on proprietary knowledge, governed by internal policies, and operated under the organization’s direct control. This stands in contrast to the prevailing paradigm of the past two years, in which enterprises have largely consumed AI as a cloud service — powerful but generic, convenient but uncontrolled.

The question is whether enough enterprises will be willing to make the investment. Model training is expensive, technically demanding, and requires sustained organizational commitment. Forge lowers the barriers — through its infrastructure automation, its battle-tested recipes, and its embedded scientists — but it does not eliminate them.

What Mistral appears to be banking on is that the organizations with the most to gain from AI — the ones sitting on decades of proprietary knowledge in highly specialized domains — are precisely the ones for whom generic models are least sufficient. These are the companies where the gap between what a general-purpose model can do and what the business actually needs is widest, and where the competitive advantage of closing that gap is greatest.

Forge supports both dense and mixture-of-experts architectures, accommodating different trade-offs between performance, cost, and operational constraints. It handles multimodal inputs. It is designed for continuous adaptation rather than one-time training, with built-in evaluation frameworks that let enterprises test models against internal benchmarks before production deployment.

For the past two years, the enterprise AI playbook has been straightforward: pick a model, call an API, ship a feature. Mistral is now asking a harder question — whether the organizations willing to do the difficult, expensive, unglamorous work of training their own models will end up with something the API-callers never get.

An unfair advantage.

Open-ear wireless earbuds are having a moment, and it is not just about comfort. More listeners are realizing that complete isolation is not always ideal when you are walking at night, jogging before sunrise, or navigating chaotic commuter hubs like Penn Station where situational awareness can matter as much as sound quality.

Against that backdrop, Cleer Audio has introduced the ARC 5, the latest addition to its ARC series of open fit true wireless earbuds. Building on the momentum of the ARC 4+ unveiled at CES 2026, the new model retains THX Headphone Certification, Dolby Atmos optimization, and Qualcomm Snapdragon Sound with aptX Lossless and head tracking, while continuing Cleer’s push into spatial audio for listeners who want immersion without cutting themselves off from the world around them.

What’s New in the Cleer ARC 5

The ARC 5 introduces several meaningful upgrades over previous models, including THX Spatial Audio, a new AMOLED HD touchscreen smart case, thinner ear hooks for improved comfort, lighter overall weight, and extended battery life. The goal is simple: deliver a more immersive listening experience without sacrificing the situational awareness that makes open fit earbuds appealing in the first place.

THX Spatial Audio: With THX Headphone Certification, the ARC 5 is engineered for balanced frequency response, clear vocals, accurate stereo imaging, and controlled bass performance. The addition of THX Spatial Audio builds on that foundation with support for Dolby Atmos and head tracking, creating a more immersive soundstage while maintaining the open ear design that keeps listeners aware of their surroundings.

AMOLED Touchscreen Smart Case: Cleer also introduces a new AMOLED Smart Case that goes beyond basic charging duties. The touchscreen allows users to control playback, adjust EQ settings, monitor battery levels, and access features like UV C sterilizationwithout reaching for a smartphone. It essentially turns the charging case into a small command center for the earbuds.

Improved Comfort and Fit: The ARC 5 features thinner ear hooks designed to reduce pressure while maintaining stability during movement. A redesigned acoustic structure also positions the drivers more effectively, improving clarity, spatial imaging, and sound projection. The result is a sleeker and more comfortable design suited for workouts, commuting, or extended listening sessions.

Lightweight Design: Each earbud weighs 11.5 grams, while the total weight with the smart charging case is 117 grams, keeping the system portable whether you are using the earbuds or carrying them in a bag or pocket.

Extended Battery Life: Battery life has also been extended, with the ARC 5 delivering up to 60 hours of total playback with the charging case via USB C charging. A 65 minute quick charge can provide up to two hours of listening time, making it easier to top up before heading out the door.

Although the spotlight is on the ARC 5’s new additions, Cleer has carried forward many of the core features that made the previous ARC 4 series appealing, including the following:

Open Ear Design: The ARC 5 retains Cleer’s flexible ear hook and hinge design, which helps improve stability and comfort during extended use. The open ear architecture allows listeners to enjoy music while remaining aware of their surroundings, making the earbuds well suited for commuting, office use, and workouts where situational awareness still matters.

THX Headphone Certification: THX Headphone Certification remains a central part of the ARC platform. It ensures open ear sound accuracy with balanced frequency response, clear vocals, detailed highs, and controlled bass performance. The certification process also requires precise channel matching for accurate imaging and low distortion even at higher listening levels, while meeting THX’s specialized standards for optimized low frequency performance in open ear designs.

Bass Enhancement: Because open ear earbuds naturally struggle with low frequency reinforcement, Cleer relies on its DBE 4.0 (Dynamic Bass Enhancement) processing to add depth and weight to the bass. The system is designed to reinforce the low end without muddying the midrange or creating an overly bloated presentation.

Bluetooth Connectivity: The ARC 5 continues to support Bluetooth 5.4, Qualcomm Snapdragon Sound, and Bluetooth multipoint connectivity. Support for aptX Lossless provides the bandwidth needed for higher quality wireless playback with lower latency, while integrated voice control enables hands free operation. The overall connectivity package is clearly aimed at everyday usability rather than simply stacking features on a spec sheet.

Call Quality: Call performance is supported by Qualcomm cVc (Clear Voice Capture) microphones that reduce background noise and suppress echo during conversations. Support for aptX Voice further improves speech intelligibility, helping voices sound clearer on both ends of the call. That focus matters because open ear designs can struggle with call clarity if microphone processing is not properly implemented.

Cleer+ App: The ARC 5 also integrates with the Cleer+ app for iOS and Android. The app provides access to EQ adjustments, ambient and noise control settings, firmware updates, battery monitoring, and the user manual. In other words, the practical tools users actually end up relying on during daily use.

With THX Certification, THX Spatial Audio, Dolby Atmos optimization, improved ergonomics, and a feature rich control system, Cleer Audio has positioned the ARC 5 as a high performance open ear wireless earbud option aimed at listeners who want strong audio quality without completely shutting out the world around them.

Performance & Award-Winning Design

“ARC 5 represents a major leap forward in open ear audio,” said Hideaki Yamaguchi (Yama), CEO of Cleer Audio Americas, “and by combining THX Certification and THX Spatial Audio with improved ergonomic design, a thinner hook, lighter weight, an HD touchscreen smart case, and a wide range of advanced features, we have created earbuds that deliver immersive cinematic audio, exceptional comfort, and strong value while still keeping users aware of their surroundings.”

The ARC 5 has already earned two major international design honors, receiving both the Good Design Award 2025 and the Red Dot Design Award 2025, recognition that highlights the product’s blend of technological innovation and thoughtful industrial design.
                                                                

Comparison 

The Bottom Line

With each generation of Cleer’s open ear lineup from the original ARC through ARC II, ARC 3, ARC 3 Gaming Edition, ARC 3 Sport Pro, ARC 4 and ARC 4+, the company has steadily refined the formula that made the series popular in the first place. The ARC 5 continues that pattern. It keeps the core elements that defined earlier models while improving ergonomics with a thinner ear hook, lighter earpieces, and longer overall battery life.

One of the more notable additions is THX Spatial Audio, which expands the listening experience beyond the ARC series’ existing Dolby Atmos optimization and head tracking support. The ARC 5 also introduces a touchscreen smart charging case that allows users to control playback, monitor battery status, and adjust settings without reaching for a smartphone. While the concept of a smart case is not entirely new, Cleer brings the feature to its open ear lineup at a lower price than similar implementations such as the case included with the JBL Tour Pro 3 wireless earbuds.

What makes the ARC 5 stand out is its focus on open ear listening with certified sound quality. The combination of THX Headphone Certification, spatial audio support, Snapdragon Sound connectivity, and a comfortable ear hook design targets listeners who want immersive audio without blocking out the outside world.

At the same time, the ARC 5 deliberately avoids features common in traditional in-ear earbuds. There is no active noise cancellation, no sealed ear tip design, and therefore none of the deep isolation or heavy low end reinforcement that those designs deliver. That tradeoff is intentional.

The ARC 5 is aimed at listeners who value situational awareness, comfort, and credible sound quality over total isolation. Commuters navigating crowded transit hubs, runners and cyclists who need to hear traffic, and office listeners who do not want to be completely cut off from their surroundings are the most likely audience.

If you have followed the ARC series from the beginning, the ARC 5 represents a logical next step in Cleer’s evolution of open ear wireless earbuds and is well worth serious consideration.

Price & Availability

The Cleer Audio ARC 5 Open-Ear Wireless Earbuds are now available (as of March 17, 2026) in black or white for $219.99 USD.

Related Reading:

The FBI has resumed purchasing reams of Americans’ data and location histories to aid federal investigations, the agency’s director, Kash Patel, testified to lawmakers on Wednesday.

This is the first time since 2023 that the FBI has confirmed it was buying access to people’s data collected from data brokers, who source much of their information — including location data — from ordinary consumer phone apps and games, per Politico. At the time, then-FBI director Christopher Wray told senators that the agency had bought access to people’s location data in the past but that it was not actively purchasing it.

When asked by U.S. Senator Ron Wyden, Democrat of Oregon, if the FBI would commit to not buying Americans’ location data, Patel said that the agency “uses all tools … to do our mission.”

“We do purchase commercially available information that is consistent with the Constitution and the laws under the Electronic Communications Privacy Act — and it has led to some valuable intelligence for us,” Patel testified Wednesday.

Wyden said buying information on Americans without obtaining a warrant was an “outrageous end-run around the Fourth Amendment,” referring to the constitutional law that protects people in America from device searches and data seizures.

The RIAA’s 2025 year end revenue report gives us a pretty clear look at where the music business stands right now. The recorded music industry in the United States generated $11.535 billion in wholesale revenue, a 3.1% increase over 2024. Streaming continues to dominate the market, vinyl remains the strongest part of the physical media business, and CDs keep drifting further into legacy territory.

Strip away the nostalgia and the usual industry hype and the story becomes pretty straightforward. Streaming pays most of the bills, vinyl still attracts buyers who want something tangible, and the rest of the formats are slowly losing ground. The numbers don’t argue. They just show how people are choosing to listen to music in 2025.

Before anyone starts comparing every figure to older headlines, one detail matters: RIAA says its reports are now based on wholesale data to align with global reporting standards. That means accuracy requires using the 2024 and 2025 figures shown inside the new report itself, not mixing them with older retail based totals from prior annual summaries.

Streaming Still Owns the Block

Streaming generated $9.4745 billion in 2025, up 3.1%, and accounted for roughly 82% of total U.S. recorded music revenue. Paid subscriptions were the engine inside the engine: 106.5 million premium accounts produced $5.881 billion, up 6.8% in revenue, while the broader paid subscription category represented 55.3% of all U.S. music revenue. That is the part of the business paying the rent, keeping the lights on, and making the rest of the machine possible. Spotify and Apple Music still dominate the landscape, but streaming’s share of total U.S. recorded music revenue has not pushed past the mid-80% range, which suggests something important: physical media’s double-digit share of the market has stabilized rather than disappearing.

There is a second story tucked inside those streaming numbers. Not every digital segment is moving in lockstep. Free streaming slipped 0.6% to $1.789 billion, and other streaming fell 3.8% to $1.309 billion. So yes, streaming is still the dominant format by a mile, but the healthiest part of it is clearly the paid side.

That tells us consumers are still willing to pay for access when the experience is frictionless and the catalog feels essentially infinite; though it does raise a fair question about how much of that catalog is genuine music versus the growing volume of AI-generated content now flooding streaming platforms.

Vinyl Is Not a Gimmick Anymore

Vinyl was the headline physical format again, and this time the number has real swagger: $1.0429 billion in U.S. revenue on 46.8 million units, up 9.3% in revenue and 7.9% in units year over year. According to the RIAA, 2025 marked vinyl’s 19th consecutive year of growth, and U.S. vinyl accounted for nearly 50% of global vinyl revenue.

This is no longer a boutique side hustle for crate diggers and weekend collectors. Vinyl has grown well beyond the small group of audiophiles who kept the format alive after CDs, downloads, and eventually streaming pushed it to the margins. Today it is a serious commercial format with real momentum.

Record Store Day alone tells part of the story. What started as a niche celebration has evolved into two major annual events, with collectors and music fans lining up before sunrise for a shot at limited pressings. Independent record stores are also seeing renewed life in college towns and major cities across North America. Many have adapted by turning themselves into community spaces selling coffee, food, turntables, record accessories, and entry level hi-fi gear alongside vinyl. The format is no longer surviving on nostalgia. It has rebuilt an ecosystem around it.

The more revealing point is not merely that vinyl grew. It is how decisively it beat the rest of physical media. Vinyl generated more than three times the revenue of CDs in 2025 and outsold them by a wide margin in units as well. The format has moved beyond novelty because it offers something streaming cannot: ownership, ritual, display value, collectibility, and a direct emotional transaction between listener and music. Streaming gives you access to everything. Vinyl gives you a reason to care about something. Different drug. Different high. Same customer wallet, sometimes.

CDs Are Not Dead, But They Are Bleeding Out Slowly

CDs brought in $312.4 million on 29.5 million units in 2025. That represented a 7.8% decline in revenue and an 11.6% drop in units compared with the prior year. The compact disc has not disappeared. It still sells in meaningful numbers, and there are niches where it remains relevant, including collectors, catalog buyers, K-pop fans, box set purchasers, and listeners who want a physical copy without paying vinyl prices. But the long-term direction is clear. CDs are no longer the center of the physical media market.

The biggest remaining advantage for CDs is price. New releases on CD are usually far cheaper than new vinyl, and the used market is enormous. Walk into almost any record store and you will still find large CD bins with thousands of titles selling for $1 to $5. For listeners building a music library, it is still one of the most affordable ways to own albums.

There is also continued support from the hardware side of the industry. A number of hi-fi manufacturers still produce portable CD players, traditional CD players, transports, and SACD machines, aimed at listeners who value physical playback and the potential for higher quality sound through a dedicated system.

But it would be a mistake to frame this as a revival similar to vinyl. The data does not support that narrative. CD sales continue to decline year after year, even if the format maintains a loyal audience. On a good player, a well-mastered CD can absolutely outperform most streaming playback, but that alone is not enough to drive a broad market comeback. Vinyl has turned collecting into an event. CDs have not. Nobody is lining up outside record stores before sunrise waiting for Compact Disc Day. Not in 2025, and not anytime soon.

Downloads Continue Their Slow Walk to the Exit

Permanent digital downloads kept shrinking. Total download revenue fell 0.8% to $272.6 million. Download singles dropped 2.2% in value, while download albums fell 9.2%. The category still exists, but mostly as a residue of older buying habits and edge cases where ownership of files still matters. In the larger picture, downloads now look like a transitional format that has already served its purpose. They were once the escape car from the CD era. Now they are just parked outside with a flat tire.

What the Data Actually Tells Us

The 2025 RIAA numbers show a music business that is healthier than the doom merchants want to admit and more complicated than the vinyl evangelists sometimes pretend. Streaming is the mass market utility. Vinyl is the premium physical object. Everything else is fighting for leftovers. Paid subscriptions are where the financial strength sits, vinyl is where the emotional and collectible upside sits, and CDs are losing relevance even though they still matter in absolute dollars.

That also means the industry has stopped being a simple format war. Streaming and vinyl are not really enemies. They represent different behaviors. One is about access, portability, and habit. The other is about ownership, fandom, and intent. Consumers are using both because they solve different problems.

The 2025 numbers make that clear: digital revenue climbed to about $9.7 billion while physical formats generated roughly $1.38 billion, and both categories grew at the same time. In other words, the market can support subscriptions and collecting when consumers see value in both. 

There is also a broader context that is easy to overlook. Music consumption in the United States now generates billions more in revenue than the theatrical movie business, with recorded music alone clearing $11.5 billion in 2025. People simply spend more time listening to music than watching films. It happens everywhere: at home, in the car, on the train, at the gym, and on airplanes. That constant presence makes music one of the most durable forms of media consumption, and it is likely that the gap between music and movies will continue to widen as streaming subscriptions and mobile listening keep expanding. 

When you compare it with other physical media categories, the contrast is interesting. Print books remain remarkably resilient, with roughly 762 million copies sold in the United States in 2025, a small increase over the previous year. But the growth is modest and largely flat compared with the pandemic peak earlier in the decade. 

Taken together, the numbers show something simple. Streaming dominates music access. Vinyl anchors the physical collector market. CDs remain a niche but affordable ownership format. And across all of it, music continues to be one of the most consistent and pervasive forms of entertainment people consume every day.

Related Reading:

The MacBook Neo, along with the Mac mini, is Apple‘s most affordable computing option at just $599. And yet it is a formidably capable device, possibly, a once-in-a-generation product. Why? One unknown I wanted to clear up was how well it ran Windows 11 in a virtual machine.

So I asked the team at Parallels to benchmark a Windows-equipped Neo, and what they delivered shocked me.

Yesterday’s discount on the Sonos Ace over-ear headphones wasn’t the only sale you can find on new additions to your Sonos setup. You can also nab the Sonos Roam 2 for just $139 as part of the Amazon Spring Sale. This Bluetooth speaker has excellent sound despite its relatively compact size, and of course it plays nicely with your other Sonos speakers.

Unlike older Sonos products, the Roam 2 now has Bluetooth in addition to Wi-Fi. When you’re home, the speaker joins your network and acts just like any other Sonos speaker in your setup. Take it on the go, and you can easily connect your phone and keep the tunes rolling. The Sonos app isn’t always the best at finding new speakers, but in this case it fired right up and connected to the Roam 2, good news for the easily frustrated. It has a fun sound profile that’s great for picnics or backyard hangs, with solid bass and balanced mid and upper ranges. Some other Bluetooth speakers might get louder, but the Roam 2 makes up for it by joining a chorus of other speakers around your home.

While the first-generation Roam suffered from some long-term battery health issues, Sonos has assured us that the Roam 2 more than fixes the problem, and at least in the short time our reviewer Parker Hall spent with it, it wasn’t an issue anymore. The outside is also slightly prone to smudges and scuffs, something to keep in mind if you prefer your equipment looking pristine. It’s waterproof, though, and quite sturdy, so just know that any marks you see on the housing are just surface level.

I spotted the discount in both black and white, but they’re both marked as “Limited Time Deals” so you’ll want to move fast if you’re interested. For anyone wondering what else is out there, or for non-Sonos users, make sure to check out our guide to the best Bluetooth speakers, or swing by our roundup of the best deals on hand-tested and approved products in the Amazon Spring Sale to see what else sparks your fancy.

The New York digital comics platform is combining its 300,000-title library with INKR’s AI localisation engine, and bringing in new leadership to execute the expansion.

The problem with getting manga into the hands of readers outside Japan is not demand. Manga is the fastest-growing category in American book publishing; global interest has been building for years, accelerated by streaming adaptations of franchises like Demon Slayer, Attack on Titan, and Jujutsu Kaisen.

The problem is infrastructure. Translating, reformatting, and distributing a comics series across languages and screen sizes is still a largely manual process, and the industry’s publishing toolchain has never been built to handle it at speed or at scale.

GlobalComix, the New York-based digital comics platform, is betting it can fix that. On Wednesday the company announced three moves at once: a $13 million funding round, the appointment of Henrik Rydberg as chief executive, and the acquisition of INKR, a Singapore-founded AI localisation platform for comics.

Together, the announcements describe a company that wants to be not just a reading destination but the infrastructure layer beneath the global comics publishing industry.

The $13 million round was co-led by SBI US Gateway Fund, the US arm of SBI, one of Japan’s most active venture capital firms with more than 1,200 portfolio companies, and Point72 Ventures, the venture arm of Steve Cohen’s Point72 Asset Management.

Point72 Ventures previously led GlobalComix’s $6.5 million Series A in July 2023 and returns here as co-lead. Additional participants include Scrum Ventures, Wise Ventures, Wicklow Capital, and Upside VC.

The Japan-US investor pairing is deliberate. SBI’s network spans Japanese media and publishing, the market that produces manga, while Point72 brings continuity and US market perspective.

Shohei Yamada, Managing Partner of SBI US Gateway Fund, described GlobalComix as “building the infrastructure that connects creators, publishers, and readers worldwide,” adding that he believed it had the potential to make manga and comics “accessible to anyone, anywhere.”

Ishan Sinha, now a Partner at Point72 Ventures who led the 2023 Series A, said the addition of INKR’s AI team and localisation technology “meaningfully expands what the platform can support for creators and publishers.”

The INKR acquisition brings the most technically substantive element of the announcement.

INKR was founded in 2019 by Ken Luong, Khoa Nguyen, and Hieu Tran, a team based in Singapore and Ho Chi Minh City, and launched its app in October 2020. 

The platform’s core product is an AI localisation engine that automates the most labour-intensive steps in preparing a comic for a new language market: text and object detection, image cleaning, translation, and typesetting.

The company says the technology reduces localisation time from days to hours and has been used to localise more than 15,000 comics, though that figure comes from GlobalComix’s press materials and has not been independently verified.

GlobalComix’s platform currently hosts more than 300,000 titles from publishers including Marvel, DC, Kodansha, Image Comics, and Tokyopop, alongside more than 25,000 independent creators.

The company’s ambition is to combine INKR’s localisation pipeline with its existing distribution and monetisation infrastructure, effectively creating a vertically integrated system: a publisher brings a Japanese title in, the AI engine prepares it for English, French, or Brazilian Portuguese markets, and GlobalComix handles distribution and revenue.

The global manga market is estimated to exceed $20 billion annually, with demand for translated content growing across the West.

Whether GlobalComix can capture meaningful share of that workflow, against established players including Viz Media, Yen Press, and digital platforms like WEBTOON, depends on whether the AI localisation quality is good enough for professional publishing standards and whether publishers will trust a startup with their most valuable IP.

The acquisition of INKR, whose technology is described as already trusted by publishers in Japan and Korea, is the clearest attempt to answer that second question before it is asked.