This past Monday, Rocket Lab announced a deal to acquire established satellite direct-to-device (D2D) operator Iridium for about $8B, in a half cash, half stock transaction expected to close mid-2027. The move turns the rocket launch company into a global mobile satellite network expected to compete head-to-head with SpaceX’s Starlink. Rocket Lab has been on an acquisition spree lately, leveraging its high-flying stock – up 182% over the past year, reaching a $63B market cap as of this writing – as currency. It comes at a time when SpaceX is signaling rising ambitions in mobile, aiming for a Starlink Mobile service that would be “on par with terrestrial mobile networks.”

With Iridium, Rocket Lab gets a low-Earth orbit (LEO) network of 66 operating satellites (plus 14 in-orbit spares and 1 on-ground spare). Most of the satellites were launched between 2017-2019 as part of the Iridium NEXT generation, with a remaining operating lifespan of 8-11 years. Iridium’s constellation is fewer in number than SpaceX’s 10,700+ satellites, ranking #6 among operators. Notably, the Iridium NEXT constellation was launched using SpaceX’s Falcon 9 rockets. Iridium’s constellation is considered by many to be the only communications operator that can provide practically useful, continuous coverage of 100% of the planet, including the poles. (Starlink and OneWeb/Eutelsat are near-global but have constraints.)

Iridium NEXT satellites were designed for direct-to-device, carrying more powerful transmitters to pick up relatively weak cellphone signals. Because of this, they are larger than SpaceX’s 1st-gen satellites (860 kg vs. 260 kg), which were designed to reach panel-style antennas. Iridium NEXT satellites are more comparable in size to SpaceX’s newer 2nd-gen satellites (V1 Mobile) – its first direct-to-device satellites – which weigh 800 kg. Notably, SpaceX’s 3rd-gen (V2 Mobile) satellites – which it hopes to launch in mid-2027 on the long-delayed Starship rockets – will be much larger at about 1900 kg and will be able to support gigabit speeds. Rocket Lab reportedly has plans to expand Iridium’s next-gen direct-to-device satellite network.

Just as important as the constellation, Rocket Lab gets Iridium’s spectrum licenses – especially its highly valuable, relatively unique licenses to D2D-friendly, mobile satellite services (MSS) spectrum in the L-band. Iridium operates in the “Big LEO” (low-Earth Orbit) bands designed for mobile handsets, which were named for the “big” satellites of the 1990s like Iridium’s and Globalstar’s (vs. the “little” LEO systems used for pagers and tracking). The Big LEO bands have advantages in being able to use a smartphone’s internal antenna (i.e. they don’t require a large dish like Starlink’s); being relatively immune to rain and tree cover; and globally harmonized (vs. country-by-country) so devices can work across regions. The Big LEO bands include a slice of the 1.6 GHz L-band for uplink/bidirectional (1610-1626.5 MHz) and a slice of the 2.4 GHz S-band for downlink (2483.5-2500 MHz). Iridium specifically operates in the 1617.775-1626.0 MHz slice of the L-band. Globalstar – which is being acquired by Amazon ($11.6B) – has the remaining L-band in the 1610-1618.725 MHz slice, and exclusivity on the Big LEO S-band. The FCC recently affirmed the “Big LEO” status quo and license holders’ rights, rejecting petitions from SpaceX and others to share the Big LEO spectrum.

Iridium’s L-band is relatively slow but highly weather-resistant and reliable, which is useful for safety-critical services and military applications. Iridium satellites operate slightly higher at 780 km to cover more of the Earth (vs. SpaceX’s 340-550 km), but that means its service has higher latency. Iridium’s fastest Certus 700 service tops out at 704 Kbps down / 352 Kbps up, which it has described as broadband in the past but is much slower than terrestrial broadband. By comparison, Starlink broadband offers 30 to 400+ Mbps and is aiming for 1+ Gbps speeds. Starlink primarily operates on the Ku-band (shared with FCC approval) and terrestrial wireless spectrum (partnering with carriers like T-Mobile), and will also operate on the S-band (recently acquired from EchoStar for $17B).

Rocket Lab would get Iridium’s profitable (albeit slow-growing) business with $872M in revenue (2025), 27% operating income, and 13% net income. The combination offers Rocket Lab “firm recurring revenue” and would edge it closer to profitability. (In Q1 2026, Rocket Lab saw a $56M operating loss and $45M net loss vs. Iridium’s $51M operating income and $22M net income.) Iridium has a ready commercial business with 2.6M+ active subscribers globally and a 500+ partner ecosystem, as well as established government contracts (29% of revenue).

The acquisition is likely to accelerate Rocket Lab’s ambition to become a “fully vertically integrated space powerhouse” with launch services, satellite production, and connectivity applications. While Rocket Lab still trails far behind SpaceX in scale, it has emerged as the primary US private-sector rival to SpaceX in launch services, where it has demonstrated reliability. To date, Rocket Lab has conducted 91 launches and deployed 262 satellites, with its small-lift Electron the 2nd-most launched US rocket. (The Electron specializes in dedicated, custom missions, but Rocket Lab’s medium-lift Neutron set to debut this year will allow for larger missions.) Rocket Lab is now looking to compete in the lucrative services layer, turning its rockets into more recurring revenue.

Direct-to-device is a major driver for Rocket Lab’s acquisition. The industry has been moving in that direction for the past few years. The 3GPP released its standards for non-terrestrial networks (NTN) in 2022, enabling regular smartphones and IoT (Internet of Things) devices to connect directly to satellites. Since then, Apple has been working with Globalstar on satellite messaging and is now planning for advanced satellite features, reportedly engaging in talks with SpaceX and T-Mobile last year.

While the headlines have focused on smartphones, the more attractive near-term opportunity for Rocket Lab with the acquisition might be satellite-based IoT (Internet of Things) connectivity. About 2M of Iridium’s 2.55M existing subscribers are already IoT devices (e.g. shipping containers, ocean buoys, oil/gas rigs, farm equipment, Garmin consumer devices). In May 2026, Iridium acquired the remaining stake in its Aireon aircraft-tracking data JV, which tracks 190K flights per day and helps private-sector and government customers manage operations, route traffic, and maintain safety.

Iridium has been working since early 2024 to build out a standards-based Iridium NTN Direct service that would allow for D2D/IoT roaming. NTN Direct represents a shift from Iridium’s traditional proprietary handsets, although the service will still require NTN-capable devices/chipsets. Iridium has been upgrading its constellation via over-the-air software updates in anticipation of a 2026 commercial launch focused on IoT. It is also signing roaming agreements with wireless operators like Deutsche Telekom and Vodafone for IoT – a less threatening partnership arena for operators than direct-to-cell.

Iridium also offers an alternative PNT (Positioning, Navigation, and Timing) system to GNSS (Global Navigation Satellite System) constellations like the US-owned GPS, the EU’s Galileo, and China’s BeiDou. Iridium satellites, leveraging its $115M acquisition of GPS backup provider Satelles from 2024, can deliver authenticated time and location signals up to 1000x stronger than GPS, and can even work indoors. Its system can protect devices relying on navigation satellites – such as drones and other autonomous vehicles – from jamming/spoofing. Iridium’s partners have been testing an ultra-small PNT ASIC chip that can be embedded in devices, to be rolled out this year.

Rocket Lab is buying some of the hardest pieces of a satellite telecom operator – spectrum, customers of a certain profile, operating history, and brand trust. Iridium already sells reliability-sensitive services to a wide array of customers, including government/military, commercial, industrial, aviation, and maritime sectors. That, combined with the differences in spectrum and speed/latency, suggests a potentially different lane in the market from Starlink.

SpaceX and its Starlink business have certainly been putting pressure on the rest of the industry. Starlink has reached 12M+ active users for its broadband service, and is already commercializing its Starlink Mobile D2D service with T-Mobile ($10/month add-on for limited applications). The “T-Satellite” service allows texting, voice chat via WhatsApp, and select apps, used outdoors under a clear sky. The limited service would automatically work with most smartphones from the last 4 years, connecting to satellites that act like “cellphone towers in space.” Backpackers are reporting issues with interference and reliability, and speeds of just 3-4 Mbps (vs. Starlink’s much faster non-D2D service). Despite having 10M monthly users, the service is “not meaningfully competitive with existing terrestrial cellular networks,” and is generally being sold as an add-on for use in rural and low-connectivity areas (e.g. national parks). T-Mobile is reporting that T-Satellite represents just 0.0002% of its total network usage.

Satellite connectivity is becoming a spectrum-and-distribution war, not just a satellite war. Most of the satellite players are looking to access key spectrum and partner with mobile operators, country by country. It’s driving a wave of consolidation, including SpaceX’s acquisition of Echostar’s spectrum, Amazon’s acquisition of Globalstar, Rocket Lab’s acquisition of Iridium, and even SpaceX CEO Elon Musk’s toying with the idea of acquiring a terrestrial mobile carrier.

Mobile operators are recognizing their rising importance and negotiating power, and signaling plans to work with multiple satellite partners. In May 2026, Verizon, AT&T, and T-Mobile agreed to pool their spectrum in concept and standardize their approach around D2D. The joint venture will reportedly be an D2D intermediary “buying capacity on a wholesale basis to resell to the individual carriers.” All 3 of the big carriers – including T-Mobile – have declined SpaceX’s requests to become an MVNO (mobile virtual network operator). The big carriers view Starlink as evolving from potential partner to competitor – a shift unsettling for the incumbents. SpaceX has pushed back on the joint venture, saying it raises “real collusion concerns.”

In addition to toying with the idea of a carrier acquisition, SpaceX says it could launch a Starlink retail offering selling mobile contracts to consumers in competition with the big carriers, and even build its own US terrestrial mobile network. Although some industry players believe this could be an attempt to gain negotiating leverage with operators, SpaceX has already reportedly begun talks with home-internet provider Charter Communications on a consumer mobile offering that could use the same infrastructure as Charter’s own MVNO service Spectrum Mobile. SpaceX also recently revealed a prototype of a handset-like AI device, which could be a foray into gaining adoption for 5G NR-NTN-compliant hardware (a barrier to SpaceX maximizing the value of its spectrum).

Where we’re heading is a world where coverage will no longer be a carrier map. Instead, “coverage” will be a blended terrestrial-satellite service layer negotiated among big carriers, satellite companies, and device makers. If the big carriers have their way, satellite connectivity will mean resilient low-throughput fallback – not a replacement for terrestrial 5G – and eventually, lucrative revenue-sharing deals on direct-to-cell broadband. (AST SpaceMobile has signed a 50/50 deal with Vodafone, for instance.) The strategic wedge will start with messaging, emergency access, location sharing, and low-data apps, and climb toward richer broadband alongside progress in rockets, satellites, handsets, standards implementation, and ecosystem partnerships.

As Rocket Lab puts it, this is a “defining moment for the space industry” but not necessarily in the way that Rocket Lab means it. Satellite is on its way to becoming a normal carrier add-on, an operating-system feature, and an element of hyperscaler infrastructure. But given that no single satellite operator offers the best price-performance ratio for every use case, the most likely future is not one winner. SpaceX (with T-Mobile) and Amazon-Globalstar (with Apple) will dominate early satellite access for consumer messaging and low-data apps, helping make satellite connectivity feel invisible inside devices and cloud services. Rocket Lab-Iridium will become a premium resilience layer for government, enterprise IoT, PNT, aviation, maritime, and safety-critical applications. AST could become a significant player when true direct-to-cell satellite broadband emerges. And over time, this could evolve into a connectivity fabric with many satellite players and mobile operators involved – with the consumers being the main winners.

Apr 24 2026 (3 Shifts): Amazon Leo, Blue Origin and the space race

Sep 12 2025 (3 Shifts): SpaceX buys $17B in spectrum for Starlink

A week and a half ago, OpenAI unveiled Jalapeño – its first custom AI chip, a specialized accelerator designed for inference (i.e. generating outputs rather than training models). Co-developed with custom-silicon leader Broadcom, OpenAI’s new ASIC (application-specific integrated circuits) processor went from initial design to tape-out – when the locked design is sent to the foundry for fabrication – in just 9 months. This is much faster than the normal 18-24 month design cycle for frontier-class AI chips. The chip is expected to go into limited production late in 2026, with ramp-up in 2027. Broadcom says it “represents what may be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors.” The chip will handle the workloads behind OpenAI’s ChatGPT, Codex, API, and agentic products.

OpenAI began exploring making its own chips in 2023, building out its chip team led by ex-Google senior TPU engineers. It started working with Broadcom in early 2024, although the collaboration wasn’t reported until Oct 2024. By Oct 2024, the chip team had grown to 20 people, which quickly doubled to 40 people by Feb 2025. Around late 2024, OpenAI secured manufacturing capacity through Broadcom with chipmaking giant TSMC (Taiwan Semiconductor Manufacturing Company) to make its first chip in 2026. (Broadcom is fabless – i.e. the AI company is like the homebuyer/architect, Broadcom is the homebuilder, and TSMC is the subcontractor.) In Oct 2025, OpenAI formally announced a multi-year deal with Broadcom to produce a massive 10 GW of custom AI chips from H2 2026 thru 2029.

The project, code-named Nexus, is estimated to cost $180B in chip production alone, not including other datacenter costs. Phase I will cost $18B for 1.3 GW of chips, which reportedly could be financed by Broadcom under the condition that Microsoft buys 40% of the chips for its datacenters and rents them back to OpenAI. (Some of the remaining chips will be installed in Oracle datacenters.) Last month, Broadcom revealed a new fund backed by Apollo and Blackstone to finance custom chips for Anthropic, OpenAI, and other AI players. (The initial $35B financing package will fund 1+ GW of chips for Anthropic.)

The 9-month tape-out was achieved through a narrower inference-focused chip mandate, Broadcom as a deep silicon-implementation partner, Celestica for board, rack, and system integration, and AI-assisted design and optimization using OpenAI’s own models. According to OpenAI president Greg Brockman, “The degree to which our models have been able to accelerate it was very surprising to us.” The Jalapeño ASIC uses a commonly used systolic array architecture with high-bandwidth memory (HBM) (also used by Nvidia), with extensive networking capabilities leveraging Broadcom’s Tomahawk chips. Celestica will make the custom server systems.

This may be OpenAI’s TPU-like moment. According to Broadcom CEO Hock Tan, Jalapeño’s performance is on par with Nvidia’s Blackwell chips and Google’s TPUs (tensor processing ​units). Unlike Nvidia’s general-purpose GPUs (graphics processing units), custom ASICs like Jalapeño can be tailored for AI workloads, and designed to perform tasks faster, using less power, and with lower cost. (The tradeoff is that ASICs are less flexible than GPUs.) Jalapeño’s architecture is specifically optimized for inference, around the “kernels, memory movement, networking, scheduling, serving patterns, and latency requirements” that matter for frontier AI models. The aim was to get closer to the hardware’s theoretical limits – across power, throughput, and latency – by reducing data movement and balancing compute, memory, and networking constraints. As a result, “early testing shows that Jalapeño will deliver performance per watt substantially better than current state-of-the-art.”

Keep in mind that reaching tape-out is not the same as proven deployment. The company has not yet released its technical report, which might disclose detailed benchmarks, memory configuration, HBM capacity, latency curves, throughput-per-dollar, and cost-per-token. While OpenAI’s accelerated path to tape-out lends it strategic credibility, it won’t be a gamechanger until the chips are in production and yield/performance are conclusively established.

If Jalapeño lives up to OpenAI’s claims, it would be a testament to full-stack chip design. In the words of OpenAI’s hardware head, “When the model, software, kernels, serving stack, networking, and silicon are designed together, you make different decisions than you would if each layer were optimized separately.” Modern inference is not a straightforward throughput problem. Reasoning models, long-context sessions, tool-use agents, and code agents create bursty, often latency-sensitive workloads where utilization, memory traffic, KV cache handling, and cross-chip communication matter as much as raw floating-point operations. The bet is that OpenAI knows its own inference bottlenecks better than any chip vendor can, and that this knowledge is worth embedding in silicon and turning into custom-hardware economics. According to OpenAI, engineering samples of Jalapeño running GPT-5.3-Codex-Spark in its labs have been operating at the production target power and performance.

OpenAI has many reasons for urgency behind this project. First, it needs chips to support 2027 demand (OpenAI has 1B+ monthly active users). Broadcom’s TSMC allocation for OpenAI – slated for 2026 – would also have been at risk if the team didn’t move fast. Furthermore, it can’t be ignored that proven chip production will be fodder for OpenAI’s upcoming IPO, which is currently looking like it might be delayed until 2027 to protect its $1T+ valuation.

Perhaps the most important reason for urgency is that custom AI chips could help drive down OpenAI’s chip costs by 20-30%. This is on top of the recent revelation that OpenAI has discovered optimizations that can “more than halve” the cost of inference. Inference is becoming the economic center of AI in terms of value and cost – and it’s already the larger part of the AI market (vs. training). Training creates the model but inference is where each query and response, API call, and agentic action consumes power, compute, memory bandwidth, and dollars. Driving down the cost of inference is critical for OpenAI’s economics. OpenAI will spend $50B on compute this year – far more than its projected $30B+ in annualized revenue – and is projected to spend as much as $600B on compute from 2026-2030. Given the scale of its expected losses, custom silicon is deeply strategic for OpenAI – and the sooner, the better. For a major AI player like OpenAI, custom chips are ultimately about controlling its own destiny.

AI hardware is starting to iterate more like software. In large part, this is because AI models are helping design the chips that will run future AI models. The flywheel goes from better models to better chips to lower inference costs to more usage (and more ambitious products) to better hardware telemetry, and back around to better models and better chips. AI players are looking ahead towards faster cycles across multiple chip generations. OpenAI has a multi-generation “Intelligence Processor” chip plan, with design work on the next-gen Serrano chip already underway. Meta is aiming to develop and deploy 4 generations of chips in the next 2 years – a much faster chip cycle than normal. Tesla, which took nearly 2 years to reach tape-out with its AI5, is aiming for a 9-month design cycle for AI7, AI8, AI9, and later generations.

Like the rest of the AI chip market, custom silicon is experiencing a surge in growth. On the co-design side, Broadcom and Marvell together represent about 95% of the market. The market is led by Broadcom (70-90% of the market), which will see about $25B in custom AI chip revenue (based on $42B in AI semiconductor revenue, 60/40 split between custom silicon and AI networking) for the 4 quarters from Q4 FY25 to Q3 FY26 (projected). In FY27, Broadcom could see as much as $60B+ in custom AI chip revenue (based on the same split on a projected $100B+ in AI semiconductor revenue). Smaller player Marvell is projecting $1.8B in custom silicon revenue for this year (FY27), and aiming for $10B+ by FY29. There are also other custom-silicon players like MediaTek, Alchip, Nvidia, Intel, Qualcomm, and Samsung (which is in talks with Anthropic), among others.

Given its dominance, it shouldn’t be too surprising that most of the major AI players work with Broadcom. Broadcom has 6 core customers for its custom AI chips (what it calls “XPUs”) – Google, Anthropic, OpenAI, Meta, and 2 unnamed others (possibly ByteDance and Apple). Marvell has 2 major customers, Microsoft and Amazon.

Not all custom-silicon customers have found success with their programs. The notable successes have been Google (which was early to the game with its TPUs, now also used by Meta, Anthropic, and Apple) and Amazon (which gained steam through its 2015 Annapurna Labs acquisition and now has $225B in revenue commitments for Trainium). In contrast, Meta and Microsoft have struggled – in part, due to a shifting market altering the requirements and the need to pivot from training to inference. (Inference chips for known workloads are easier to make than training chips that can compete with Nvidia.)

AI players are keen to diversify their supplier base, to reduce cost and overreliance on a single vendor. (Broadcom is known to be a tough negotiator.) Google has been making a diversification push and is now working with Marvell as a secondary partner, in addition to bringing more in-house. Microsoft has reportedly been in talks with Broadcom to take over Microsoft’s ASIC program from Marvell. It plans to leverage OpenAI’s work with Broadcom, since Microsoft – under its Sep 2025 renegotiated partnership agreement with OpenAI during the for-profit conversion – gets IP rights/access to OpenAI’s chip designs. Meta, although its inference chips are being built with Broadcom, also works with Marvell on networking ASICs/components. (Marvell’s networking and connectivity chips were the basis for Jensen Huang’s opinion – which skyrocketed the stock – that Marvell will be the “next trillion-dollar company.”)

If OpenAI’s numbers prove out, every frontier AI lab will feel pressure to own at least part of its silicon roadmap. Across the industry, this could eventually relieve dependency on constrained AI chip supply, improve bargaining power on the buyers’ side, and drive down the cost for users – the case that OpenAI is trying to make. Custom AI chips won’t solve all the industry’s constraints – they won’t eliminate the need for GPU companies like Nvidia and AMD; or for memory-chip makers like Micron, SK Hynix, and Samsung; or for foundries like TSMC or advanced packaging; or for power, land, and datacenters. But it could shift the bases of competition towards higher-level hardware architecture, rack-scale systems, networking, connectivity, and software.

Jan 9 2026 (3 Shifts): TPUs & custom AI chips will be a big category

Oct 17 2025 (3 Shifts): OpenAI’s crafty deal-making in chips

This past Tuesday, music publisher Primary Wave revealed that it was taking its playbook for IP rights acquisition into the book realm, launching Atticus Works with a $100M+ commitment. Atticus Works plans to acquire and revive literary and theatrical catalogs, in the same vein as music catalogs, which have seen deals worth in the billions of dollars. In 2024, Queen’s music catalog went for £1B and Michael Jackson’s catalog was valued at $1.2B, and Garth Brooks is lately shopping his catalog aiming for $2B. Books are starting to look more like music and film/TV libraries, where proven IP can be re-marketed, re-packaged, licensed into new formats – and valued as a portfolio. AI, in particular, is making it even easier to squeeze new value out of old content.

Atticus is targeting major authors and playwrights with works of literary fiction (e.g. Ernest Hemingway), sci-fi and mystery (e.g. Agatha Christie, Stephen King), narrative nonfiction (e.g. Robert Caro), children’s books (e.g. Dr. Seuss), and plays (e.g. Arthur Miller). The focus is on “timeless stories” that are market-tested and ready to be reimagined and redeployed into new form factors.

Atticus isn’t the only investor eyeing this arena. Earlier this year, International Literary Properties (ILP), which invests in and manages literary and theatrical IP, secured a 5-year $100M+ credit facility to support acquisitions, after doubling the size of its portfolio in 2025. It describes its model as acquiring all or partial rights to books, plays, and musicals. Its portfolio includes works tied to Georgette Heyer, Langston Hughes, W. Somerset Maugham, Tennessee Williams, and others, including The Westing Game (and an unpublished sequel) and Clue: The Musical.

Publishing is already a backlist business. In the US, backlist books account for around 70% of print sales. For bookstore owners, backlist books might represent 25-50% of sales. For authors, 50% (median) of their income is earned from backlist and back-in-print books. For publishers, backlist books as a percentage of total sales ranges from 25-80%.

BookTok – a passionate subcommunity of readers on TikTok (370B+ views and 82M+ posts) – has shown that old books can be rediscovered and re-enter the demand curve without a traditional launch cycle. About 59M print book sales were attributed to BookTok in 2025 – up from 46M copies 2 years prior – suggesting that BookTok accounts for about 8% of US print book sales. BookTok has revived books like It Ends With Us, which was published in 2016, experienced a BookTok-driven surge of popularity in 2021, and was adapted into a movie starring Blake Lively in 2024. In another example, the book Stone Maidens became the #1 book on Amazon 11 years after publication, after a BookTok post by the author’s daughter. At one point, novels by 7 of the top 10 BookTok authors had film/TV adaptations under development. The backlist is no longer just “long tail” – it is reactivatable inventory.

The core investor thesis is that the old-book asset base is large, under-utilized, and under-managed. It is producing cash flow but it has not been optimized for modern discovery, sequels, audiobooks and podcasts, other social/digital formats, screen adaptations, foreign and foreign-language rights, AI licensing, merchandise, and digital offshoots (e.g. games, education). In some cases, the original authors have passed away and the IP owners are the authors’ estates, which, in some cases, may be readily open to a deal on the right financial terms. If a major literary/talent agency has interests in multiple estates, this opens the door further to larger, multi-estate deals.

AI licensing is a particularly untapped and potentially fruitful avenue for books. Books are not just stories – they are structured, edited, rights-bearing knowledge. In fiscal 2026, book publisher Wiley generated $49M of AI revenue – up 23% – with lifetime AI revenue surpassing $110M. HarperCollins has tested a model in partnership with an AI company where nonfiction authors can opt into a 3-year AI training license, with a $5000 fee per book split 50/50 between author and publisher. While this was small money at the individual-title level, it established that backlist books could become a rights library of licensable AI inputs.

For AI players, the legal overhang is real. Anthropic agreed to pay $1.5B plus interest to settle claims that it used pirated books in its training. (This was only about $3,100 per work, however.) Meta is currently being sued by a group of major publishers over alleged use of pirated books and journal articles to train its AI. This rising wave of litigation is changing how AI players are sourcing data for training, making cleanly licensed catalogs more valuable.

AI licensing requires portfolios of catalogs at a certain scale. This suggests either a financed roll-up strategy (e.g. Atticus, ILP) or a standards/platform approach. The investors who win won’t just be buying books – they’ll need to build operating platforms around discovery, managing rights and reuse. Last year, startup Created by Humans ($5M in seed funding) launched an AI licensing platform that lets authors set their own terms (e.g. training rights) while allowing AI players to browse and license content. Other AI licensing startups have also arisen to tackle this problem, although Created by Humans stands out for its focus on authors/books.

This is likely to push book catalogs towards a tiered market. At the high end, iconic estates and genre franchises will be priced as multi-format IP (e.g. print, digital, audio, translation, stage, screen, merchandise, AI). In the middle, backlists will be bundled into thematic portfolios (e.g. children’s, sci-fi/fantasy, romance, biographies), where buyers are not underwriting individual titles but rather a diversified pool of recurring royalties and options. At the low end, neglected titles may become rights-cleanup projects, where value is created by recovering rights, refreshing metadata/covers, relaunching formats, and making the catalog discoverable to the ecosystem.

The economics of book publishing have always been different than, say, filmmaking. While both are hits-driven businesses, the cost to write a book is roughly the cost to sustain 1-2 authors, plus a chunk of editor time. By contrast, the cost to make a movie – which typically involves a lot more people and significantly greater hard costs – can easily run into the millions. This disjoint is the reason why there are always many more great books that can be made into movies than movies that end up being made – it’s a funnel process. AI has the potential to change this by streamlining the conversion from a piece of “pillar content” – a rich story on the page – to a screen adaptation or an audiobook or a game or something else altogether. (AI’s ability to make “weirdly fun” games at scale are currently driving a booming proliferation of games.) AI could reshape the funnel structure, unlocking the value inherent in the deep well of underutilized books. It’s no wonder authors and books are gaining respect.

We’re still at the “bottom of the first inning” and we’re unlikely to see a wholesale financialization of literature overnight. The market is too fragmented, authors and estates care too much about stewardship, and the adjacent arenas (e.g. AI licensing) are too politically charged. What we are likely to see, however, is a shift in how publishers and estates think about their IP and manage their backlists, especially at the top tier.

Sep 12 2025 (3 Shifts): The open RSL Standard for licensing content

Oct 15 2021 (3 Shifts): Investors are on the hunt to put billions of dollars directly behind recurring revenue