Perspectives
Tags
Owning Intelligence, Not Renting It: Adaption’s Sara Hooker and Sudip Roy
November 12, 2026
All content
Portfolio
AI Infrastructure
Co-founders Sara Hooker and Sudip Roy started Adaption on a simple bet: that the next era of AI belongs to companies that can keep learning after they ship, not just the ones with the biggest training runs. Sara spent years studying why good research ideas win or lose based on the hardware underneath them, and Sudip built the systems that make models fast enough to matter in production. That combination, research instinct paired with production discipline, is why we backed them, and why the team has moved quickly from a research thesis to a working product in Adaptive Data. Partner Yaz El-Baba sat down with Sara and Sudip at a moment when the industry's assumptions about scale were already starting to wobble, to talk through where they think intelligence is actually headed, and why they built Adaption to get there. Yaz: For most of the last decade, the story in AI was simple: bigger model, better results, whoever scales fastest wins. That story's been changing lately. Where are we right now in the cycle? Still scale-at-all-costs, or already shifting toward efficiency? Sara: We are in an era of change, where innovation matters again. For the last ten years, scaling pre-training model size was enough. That created a concentration of power in a handful of providers. Returns have flattened, and all bets are off now. At a trillion parameters, the economics break in both directions: models become too expensive to train and serve. This reckoning is a good thing. It means new ways of training models matter again. We are building models that learn continuously, become smarter than they were when first deployed, and are owned by companies around the world. Yaz: Sara, your paper, "On the Slow Death of Scaling," argues scaling laws are running out of road. What are the earliest signs you're seeing that it's already playing out? Sara: No frontier lab will 5x the size of the training model this year. The returns from scale are in test-time compute, adaptive data, and real-world interactions, not in model size alone. Sudip: We're seeing the same thing from the customer side. As AI adoption scales into production, users are increasingly cost and latency-conscious. They're realizing they don't need the full general capability of a massive model; they need specialized capability at lower cost and lower latency. That's exactly what Adaption is built to accomplish. Yaz: This summer, a wave of companies, including Nvidia and Microsoft, signed an open letter on open weights arguing there needs to be a balanced ecosystem between open and closed models. Adaption signed on too. Why? Sudip: There's a narrative floating around the industry that open source models are somehow less secure. That's a narrow way of looking at things. Adaption's mission is to transfer ownership of intelligence across the full stack to the enterprises and communities we serve, and an open ecosystem is an important cornerstone of that mission. Yaz: You've said before that open weights are safer because they can be inspected by anyone, similar to Linux and public encryption standards. Taking the flip side, when does that stop being true? Sudip: I do think open weights are safer, because they can be probed by a much larger audience. But there's a key difference between how traditional open source software like Linux has been developed and how open-weight models are published today. With Linux, the entire process, including all the code, is public. With an open-weight model, what's delivered is almost always the end outcome, not the process it came from or the original source data used to create it. A handful of open-weight models also publish their source data, but for the most part that process is still a black box. Still, it's a step in the right direction, and hopefully more players in the space move toward openness over time. Yaz: Your work around the concept of “Hardware Lottery” argued that available hardware and software can determine which ideas succeed, independent of their merit. Open weights broaden access to models, but they don't equalize access to compute. How does the hardware lottery shape what you're building? Sara: The hardware lottery has only narrowed in recent years. Nearly every chip built is optimized for one operation: the matrix multiply at the heart of deep learning. Almost nobody notices, because the winning ticket has been paying out for a decade. Compute access doesn't overcome the hardware lottery, but making intelligence as fast as possible accelerates innovation and improves overall access. Yaz: You've spent a lot of energy rethinking how models are hosted for end users, especially given the compute constraints in the market today. Where do you see serving infrastructure heading? Sudip: A few trends stand out. One is bespoke, custom hardware built specifically for inference, purpose-built ASICs rather than general-purpose chips. Another is decentralized inference that does not run everything in a data center. We're already seeing smaller models run on laptops, and we’ll soon see low-power devices like watches make local decisions that check in with a centralized system periodically. The one I find most interesting from a systems perspective is that training and inference have always been treated as separate paradigms, optimized independently. That gap will erode as continuous learning systems mature and models update regularly, even during inference. Yaz: The highest-value internet data is largely tapped out at this point. It's not a coincidence that Adaption's first product launch was Adaptive Data. Why start there? Sudip: Enterprises are sitting on massive volumes of data accumulated over decades in business; data that's never been seen by open or closed models. Enterprises are disincentivized to hand it over to frontier labs, because it's their moat. We built Adaptive Data to make it seamless to connect that historical data to downstream AI systems, whether at training time or inference time. It's a key piece of transferring ownership of intelligence back to enterprises. Yaz: As more ownership of intelligence shifts to the enterprise and the end user within it, where does adaptability go in the next year? What changes for the end user compared to the current experience with agents and LLMs? Sudip: At one extreme, you could imagine a personal AI that knows your tastes and the context relevant to you well enough to make decisions almost as if you were making them yourself. We're far away from that. In the interim, adapting to a particular organization is a much more tractable step, and it's within reach over the next year: a system, not just a model, where the harness on top and the data feeding it are built specifically for that enterprise, instead of shipping something generic out of the box.
Verda on Why Europe Needs to Build Its Own Cloud, Not Rent Someone Else's
November 11, 2026
All content
Portfolio
AI Infrastructure
Most conversations about AI infrastructure start with chips and end with hyperscalers. Ruben Bryon starts somewhere else: with the fact that almost none of that infrastructure is European, and almost none of the companies building it are trying to stick around long enough for that to matter. Verda, the company Bryon founded in 2020 as DataCrunch and rebranded last November, builds AI cloud infrastructure in the Nordics, vertically integrated from the data center up through the software layer developers actually touch. We backed the company's Series B, a round Bryon is using to do two things at once: accelerate the product, and accelerate the physical build-out of infrastructure that, in his view, Europe doesn't have enough of and needs to own outright. We sat down with him right after the round closed to talk about sovereignty, why he turned down other term sheets, and why he's building Verda to still exist in thirty years. Joe Floyd: Congrats on the Series B. What are your ambitions now that you've raised a big round? Ruben Bryon: We're ramping up R&D to improve the product and deliver more value to our customers. But a big piece of this raise is also accelerating the build-out of the physical infrastructure we're doing in Europe. Joe: Why does Europe need more of its own sovereign physical infrastructure? Why does it need a Nordic cloud? Ruben: There's still a lot of opportunity in Europe, a lot of pockets of power for building out this kind of infrastructure. There's the obvious latency argument for locality of compute, but sovereignty matters too. As open source models get more useful and more relevant to actually run in production for real workloads, it becomes important that the compute running them exists on the European continent and is operated by European companies. Joe: You're also moving beyond physical infrastructure into the software layer. Why does that matter, for Verda and for Europe broadly? Ruben: The value that cloud providers bring is the same as road infrastructure or electrical infrastructure. We want this compute to exist so our customers don't have to think about what they need in place to get an environment that can scale quickly to meet demand. That shows up in different shapes: instant clusters, containerized environments, sandbox environments for agents. We want it to be easily accessible, and we want a European company to be able to offer it in a way that's mindful of the basic unit economics, covering the full vertical from the data center layer up. Joe: What does success look like for you in twenty or thirty years? Ruben: One of the frustrations that led me to start the company is that Europe doesn't have enough companies with a long-term enough vision to build past zero to one. A lot of companies come into existence, build something genuinely interesting, become a gravity well for people who want to build something too, and then end up selling to a larger company, or worse, to private equity. The people who built it are left disillusioned, because they were hoping to build something generational. What we want to build is something that stands the test of time instead of chasing a quick exit, a real tech company in Europe that becomes that gravity well for local talent, so they don't have to move to the Bay Area to be part of something ambitious. Joe: Verda, the name, refers to green. Tell us about your ambitions for AI and the planet. Ruben: We build the majority of our infrastructure in the Nordics and Europe, where there's access to large pools of electrical power that's straightforward to source from green power. The grid itself is already green on a global scale compared to other countries, which lets us make power purchase agreements for 100% renewable power. More companies are giving up on that, which is a shame. Anyone who can build a gas turbine or reciprocating engine to turn fossil fuel into electricity to generate tokens is running those plants flat out because demand is so high. I think that's a mistake. There's a growing backlash to it, and I think it has merit. As a species, we should think hard about how to build out this AI infrastructure without doing it at any cost. We want to set an example on having a low carbon footprint in the data centers we operate.

Robots That Go Brrr
November 4, 2026
All content
Physical World AI
Robots have gotten smarter. Now we have to make them fast. Robotics latency, System One decision models, Jev, TypeSafe AI, Physical Intelligence, action chunking, frontier VLA models, reusable skills, local controllers, edge compute, Jetson Thor, physical AI silicon, batch size one, tail latency, power limits.

Who Owns the Intelligence Economy?
November 2, 2026
All content
Market POV
Future of Work
The rules of the AI economy are being set right now, and no one is setting them on purpose. We’ve seen this happen before. Between 1920 and 1927, American broadcasters staked out radio frequencies faster than the law could follow. Herbert Hoover, then Commerce Secretary, held four national radio conferences to broker voluntary rules of cooperation among broadcasters. At the same time, he told Congress that those voluntary rules wouldn't be enough. The public interest principle that he believed should govern the airwaves needed a real law behind it, or it had no teeth. By the time Congress acted with the Radio Act of 1927, the ownership structure had already hardened. The Communications Act of 1934 mostly ratified it. The window between a new technology's arrival and the hardening of its power was much shorter than anyone expected. We are in a similar window with AI, and far more is at stake. Radio was a fight over a few hundred megahertz of spectrum and one industry's advertising revenue. The resource this time is human intelligence: how people reason, decide, and create. If we treat that as something to extract rather than cultivate, we drain the source that every future breakthrough depends on. The defaults are already being set, mostly without anyone deciding to set them. Every organization building an AI system on top of its employees' work is making an implicit claim about who owns what that system learns. Every platform that captures the judgment behind a decision, the correction that improved a result, the reasoning that closed a deal, is accumulating something that used to walk out the door with an individual. None of this is malicious. But if the rules aren't shaped intentionally, the companies controlling the most powerful models will set the terms, and everyone else will pay rent. Before these norms harden, we're proposing a starting place: a social framework for consideration and conversation. In the coming months, we'll bring founders, researchers, legal scholars, and labor voices together to stress-test it. What follows is a first attempt, not the final answer. Why a social framework There is precedent for this working. The Internet Engineering Task Force built much of the architecture of the modern Internet through open standards adopted without government mandate. There is also a cautionary precedent. The W3C's "Do Not Track" privacy standard won broad browser adoption and regulatory support, and collapsed anyway, because companies faced no real cost for ignoring it. Princeton researcher Arvind Narayanan, who participated in the process, later argued that once the threat of legislation faded, the effort mostly gave the appearance of self-regulation without substance. The three layers The intelligence economy needs a social contract with three layers: platform, organization, and individual. 1. At the platform level: choice No single model should become the unavoidable infrastructure of intelligence. The durable advantage in the intelligence economy sits above the model, not inside it, and that becomes more true as foundation model capabilities converge. What matters is the learning loop an organization builds on top of whichever model it runs: the domain-specific training, the institutional memory, the orchestration that a competitor cannot replicate by switching providers. That advantage only exists if the market rewards cultivating better judgment over time. The status quo is the opposite. The provider's incentive is to make switching expensive, which means organizations are often locked into whichever model they adopted first. Enterprise AI agreements grant the provider broad rights over what happens inside a session by default; opting out is theoretically possible and functionally rare. But access to AI capability should not require surrendering intelligence as a condition of service. Organizations should be able to choose which models they build on, port their workflows when they leave, and decline participation in model training without losing access to what they're paying for. 2. At the organization level: sovereignty The learning loops, workflows, and domain expertise organizations build on top of AI belong to them. The fine-tuning that turns a generic model into something that reflects how a particular organization operates is an asset, and it deserves deliberate protection. But protecting something requires being able to name it. Expert-network firms like GLG, AlphaSights, Third Bridge, and Guidepoint have spent decades converging on a definition: proprietary information is information that belongs to an organization rather than to the individual who happens to know it, that isn't publicly available, and that has competitive or economic value specifically because it's kept secret. The AI economy needs this same distinction, especially… When employees use public AI tools to do their work, they teach those tools how the organization operates. A long session with an external agent encodes decision logic, process sequencing, and priorities, some of the most valuable things a model can learn. Enterprise contracts usually prohibit training on internal data, but the protection around behavioral traces and process knowledge is far less settled. Not every workflow deserves the same protection. Routine, standard operations are one thing. But the outliers are another: the rep who ignores the standard script and closes the deal anyway, or the engineer's fix that isn't in any runbook. Those moments are the actual source of improvement. It’s the generative part of the loop. Protection should follow along the lines of that distinction: default transparency for the routine work, real sovereignty claims reserved for the judgment that changed the outcome. 3. At the individual level: appreciation Organizations should be able to capture and learn from the judgment employees apply at work. But in an era when jobs are on the line and models get smarter by learning from human judgment, how will humans grow smarter using AI? Does using these systems make someone sharper or duller? What does each side keep when the relationship ends? This is the human layer, the least settled of the three. It is the one place the framework reaches past ownership into design. Whether AI appreciates human intelligence, beyond harvesting it, depends on how it’s designed in the workplace, with growing human intelligence as the goal. And that growth in contribution is what compounds, both for the organization and travels with people if they leave. We have long warned about "Crutch Software": systems that do the work but don't help improve the people using them. Most enterprise AI today is crutch software at scale. It hands over an answer without asking the user to judge whether that answer is right, or how it could be better. A 2025 randomized controlled trial shows the cost. Undergraduates who used ChatGPT as a study aid retained significantly less 45 days later than students who studied the traditional way. The tool gave correct answers but removed the recall that makes learning sticky. The author’s term for it: “a cognitive crutch.” The crutch is a design choice, not a property of the technology. A Harvard randomized controlled trial tested an AI tutor built the opposite way: it asked questions and withheld answers until students reasoned their way to them. Students using it learned more than peers in a well-run active-learning classroom, in less time. Same underlying models as the crutch. Different design, opposite result. We believe the technology people use should make them better at their work. This is the next layer of our Coaching Networks thesis. The first generation watched the best performers and spread their moves to everyone else. The next generation builds human capacity to invent new moves, which means designing friction back into the system. Without the friction of understanding and internalizing new concepts, workers will atrophy as AI ascends. Next, the improving judgements people make in the intelligence economy leave the relationship worth more than when they came in and that they get to take it with them. Work product belongs to the employer. When a doctor leaves a hospital, she takes her clinical judgment; the hospital keeps the patient records and the protocols. Trade secret law draws the same line between an employer's protectable secrets and an employee's general skill. AI blurs it by capturing both a person's output and the approach behind it. A cognitive fingerprint could restore the line. Think of it as a passport: a portable, anonymized record of how a person approaches problems, with the deals, documents, and confidential processes excluded. The organization keeps what it learned. The person keeps the ability to apply their judgment elsewhere. That portability gives people a reason to keep growing and contributing, and companies that support it will recruit against companies that don't. Why Emergence Humans are the only mutation engine in the age of AI. AI can master any game with fixed rules. Markets keep changing the rules, and people are the first to see the new game. The job of a well-designed AI system is to find those breakthroughs and spread them. The temptation now is to use AI to reduce human work. We think that's the wrong target. The goal is to make human thinking more powerful, more durable, and more portable. People will keep working. The question is whether the intelligence they generate compounds in their favor or flows to a platform they don't control. It is early enough that the answer is still a choice. McKinsey found nearly two-thirds of organizations have not yet begun scaling AI across the enterprise, even with 88% using it somewhere. Most companies are still making single-vendor architecture decisions by default, under speed pressure, before they've had a reason to think about switching costs. That is exactly the window where a genuine-choice principle does the most good. We are not a policy shop, a law firm, a standards body, or a regulator. We are a venture firm that has backed enterprise software companies since 2003, and we are arguably the wrong institution to write this framework alone. We are also not a disinterested party: we make money when the companies we back get more valuable, and we won't pretend the practical and the principled are perfectly separable here. We believe in a free market for intelligence. Open and closed models can coexist, and the strongest ecosystem probably needs both. But it cannot afford captivity. More than twenty years in this industry have shown us that technology creates the most durable value when more people participate. Open standards made the internet market bigger for everyone. The intelligence economy will create more value if the people whose thinking feeds it have a reason to trust the system they're building. That is a good bet, and one we'd like to make. A future where cognitive labor is extracted without acknowledgment, where the people who trained the systems are treated as raw material rather than co-authors, is a glum one. An opening, not a conclusion Nobody knows yet where the lines should sit. Not the companies building these systems, not the regulators, not us. But the moment to decide is not after the architecture is built. It is now, while it is still being designed. The convenings start soon, and we will share what we hear. We expect to be wrong about parts of this framework. Do Not Track died because ignoring it was free. What does ignoring this framework cost, and who imposes it? If you have a reaction, we want it.
The Hardest Job at the Bank
October 28, 2026
All content
Portfolio
Vertical AI
Casap earned banks’ trust by tackling dispute resolution. Lotti Siniscalco shares why Emergence is investing again and what comes next for AI agents in banking.

Built on Bedrock: Doubling Down on the Autonomy Platform for Heavy Construction
October 26, 2026
All content
Portfolio
Physical World AI
This August, on a water treatment site in Nevada, an excavator dug to grade and loaded dump trucks for a full shift with nobody in the cab. Not a demo lot. Not a fenced-off test. It was the first time an excavator had done real commercial work entirely on its own. Since then, Bedrock has scaled rapidly, providing autonomous excavation services building data centers, manufacturing facilities, and more across the country. Today we're doubling down. Emergence is co-leading Bedrock's $600 million Series C alongside CapitalG. When we first backed Bedrock in 2025, we bet that autonomy could change the economics of heavy construction. In February we wrote that once autonomy delivers throughput, safety, and schedule certainty at the same time, it stops being experimental. This summer it stopped being experimental. This round is about what comes next. Why construction can't wait Industries representing more than 20% of global GDP face critical skilled labor shortages, and construction is exhibit A. Demand for housing, factories, data centers, and infrastructure keeps climbing while the workforce to build them shrinks, and productivity has been flat for decades. The operator seat is where it bites hardest. The U.S. needs to fill roughly 46,000 construction equipment operator jobs every year through 2034, most of them replacing people who retire, and it takes about five years to train a competent replacement. Nearly nine in ten contractors have open craft positions right now, and worker shortages are the single biggest reason projects run late. The industry doesn't have five years. In February we argued that construction needs a new operating model, not better software. Bedrock's answer is autonomy that reasons about the work, on the machines contractors already own. Proven, not promised Bedrock started with the hardest machine on the site. An excavator is an articulated hydraulic beast that has to read the soil, the trucks, and the people around it and decide what to do next. Solve that first and the dozer, the loader, and the haul truck arrive later as easier versions of a problem already beaten. The team's lineage shows: Boris Sofman, Kevin Peterson, and much of the engineering bench came from Waymo, where they helped build the systems that put driverless cars on public roads, and from Caterpillar. While the founding team had unparalleled domain expertise in autonomy and construction, they, nor anyone else, had ever priced an autonomous excavator. This is the definition of an emerging business model, exactly where Emergence loves to partner. Across a series of working sessions, we pressure-tested pricing and packaging against what we'd watched contractors actually pay for: utilization, throughput, safety, and schedule certainty. The result is a model built around the economics of the job site rather than the robot. And because Bedrock retrofits the fleet a contractor already runs, adoption doesn't start with a capital purchase. This model is now translating into rapid commercial traction. "Construction has spent decades trying to squeeze more out of the same two constraints: people and machines. Bedrock changes the math on both. Autonomy that runs on the fleet contractors already own doesn't just make a job site more productive. It changes what the industry can afford to build." Jake Saper, General Partner Why we're doubling down In 2015 we led DroneDeploy's Series A and spent the better part of a decade alongside that team, learning what it takes to get robotics into contractor workflows and build a commercial model around how contractors actually operate. This July, Procore acquired DroneDeploy for $845 million. Those lessons shaped how we underwrote Bedrock, and they're why we moved to double down to co-lead the Series C. "We’ve worked closely with the leaders in construction robotics since 2015, and the pattern is consistent. The hardware was never the hard part. The company that wins is the one that fits the machine into the customer's existing workflow and owns the deployment data. Bedrock chose to be a neutral autonomy layer across fleets contractors already run, instead of asking them to buy new iron." Kevin Spain, General Partner One autonomous excavator is a productivity story. A coordinated fleet of excavators and haul trucks working alongside a human crew is a different way to plan a project, staff it, and decide what gets built at all. Dump trucks, dozers, and loaders are next on Bedrock's roadmap, and the Series C is the capital to get there. We said in February that Bedrock could become the default autonomy platform for heavy construction. Eight months of deployments later, we’re millions of cubic tons of dirt moved closer to that goal. Congratulations, Boris, Kevin, Laurent and the whole Bedrock team. We're thrilled to deepen our partnership. Read why we first backed Bedrock here, and Kevin’s piece on why assemblers win in robotics.

Selling the Outcome: A Conversation with Gainsight CEO Chuck Ganapathi
October 16, 2026
All content
AI-Native Services
Arcee AI on Why Open-Weights Are a National Bet, Not a Business Model
October 9, 2026
All content
Portfolio
conversation Mark McQuade CEO Lucas Atkins CTO Arcee AI Emergence General Partner Joe Floyd companies solely focused building open-weight models U.S culturally different labs call blue-collar industry tends run elitist think every-person company Look OpenAI Anthropic others actually ship rarely UI top model custom variant trained tuned alongside harness built specifically product rest access capability end three holding disproportionate control shortage opinions right now whether AI's future runs open closed backs choice market builders enterprises locked single architecture handful deciding capabilities exist gets meanwhile made specific bet matters moved post-training people's pre-training foundation scratch capital-intensive pivot people around thought mistake sat Series anyway leading revolution investment glad start driving commitment background early Hugging Face traces back mission democratizing putting hands making customizable ways typically allow gives ownership asset instead renting lab year ago customizing little decided move training bad idea assumption couple big players keep releasing competitive indefinitely relying point failure sit later talked anyone enterprise government consumer venture level thinks wrong mentioned customizability core advantage umbrella motivates broader ecosystem decisions behind API matter optimizing every eventually starts running backbone Markets tolerate long profit political economic safety genuinely transformational say kind concentration small number real risk builder's side practical advantages working weights lets serve customers lower price larger Google raise enormous amounts capital goes compute serving traffic massive scale put toward work directly signatories letter Microsoft Nvidia enough Arcee's name Signing giants vouching support mattered entire business catch surpass China Chinese started arriving months West ahead places sent shock especially fear government's response restrict generally statement startups restricting answer American Western vibe shift since leadership lead overall need xAI gap taken approach state-supported mutually reinforcing pushing deliberate strategy project needs treat something lucrative adoption hobby Shifting makes distinct space staffed team alumni traditional pedigreed resumes skilled scrappy thirty total grounded outcomes credentials able capitalize wave hungry comes nontraditional backgrounds mostly engineering rather research development discipline well-defined steps execute structured dark art world understand complexity architectures pipelines process require original isolate researchers narrow slices pipeline often sixteen across lifecycle data curation mid-training context extension SFT RL Everyone understands full fast develop phase nuance siloed organization ambitions post-Series next eight want DeepSeek Moonshot frontier achievable tiptoeing anymore
Rethinking Communication for Human and AI Teams
September 24, 2026
All content
Portfolio
Building something iconic?
We'd love to meet you
Want to stay in the loop?
Subscribe to our newsletter