Michael Gordon

Reviewed 20+ data supplier pitches today. 1. Stop password-protecting sample. 2. Tell me the target price and licensing terms upfront. 3. Put the volume, update cadence, and eval results up front and center. Few did that - I cannot forward your pitch along if it's half-baked. Help me Help You! If you really want to sell data, make a sample public, show the eval results and how it performed against benchmarks, and sell a story! We developed, alongside a handful of data buyers, a vendor checklist here - complete it, download the zip, and share the PDF. brickroad.network/supplier-o… If you want us to host and display it in front of our clients agents, there is a fee 😱

One of the challenges with tools like "Invent a Dataset" or "Build Your Dream Dataset" (which we built last year at @TryBrickroad) is psychological. It's also why OTC data marketplaces haven't worked - data doesn't sell itself. There is a reason why the $40B data brokerage market remains primarily human-driven.

Everyone Wants to Sell Data. Here's Why You Won't - and How You Might. Let me explain. At some point over the last few years, almost every company has considered data licensing as a revenue stream. AI opened up a new buy-side archetype for unstructured data in a way that made most data suddenly valuable - in theory. We've talked to major league sports orgs, hospital networks, every major credit card and payments company, and even certain governments about data licensing. Most decide not to go down the path for a handful of important reasons: 1. It's worth less than you think - much less. in finance, data suppliers typically aim to earn a tenth of a tenth (1%) of the value their dataset returns to the firm. When your dataset is one of ten combined to generate $1M for a firm, you're looking at around $1k a year from that firm. It gets worse when selling to a lab. A lab has to spend expensive compute to process your data, and it isn't generating a monetary return from your data that you can take a cut of. There's no denominator. Lab pricing is anchored to what your data replaces - compute, annotation, the next-best alternative - not to returns. That's why doing evals before selling is so important. 2. It costs money - upfront - no matter what anyone tells you. Between your database and a signed licensing deal sits a wall of technical and back-office work: data normalization, evals, backtesting, IP stripping, marketing, and legal. Most companies see this list, reason that it's going to take 6-12 months before they get their first receipt, and shelve the project. 3. There is a $40B market of institutionalized frictions. If data were easy to sell, everyone would sell it, and broker margins would evaporate. For incumbent data vendors who have been doing this for decades, the frictions around relationship building, normalization, validation, and access are their moat. It's a $40B moat that you're up against. Now: here's how you might sell data. The companies that make the most money selling their data typically share three core traits. 1. Their data is actually scarce - and it's recurring, ideally with low latency. If someone, somewhere, cannot do their job without it and cannot get it anywhere else - then you're golden. but importantly, it can't be a one-time dump: buyers pay subscriptions for data that refreshes. A scarce asset that regenerates every day is a business. A scarce asset you can only sell once is a transaction. 2. They invest to validate demand and value. There is a reason @mercor, @handshake, @micro1_ai and others are investing so heavily in RL gyms. Successful data deals are done on the basis of known results and savings. 3. They treat data licensing as a secondary revenue stream, but not a side hustle. There's nuance here. First, the best data is a byproduct - exhaust from great operations. Focus on keeping your core operations sustainable so they keep producing more data. Second, you do need to invest up front in a dedicated owner, a real budget, and a multi-year horizon. Data buyers can smell a tourist, and they don't do second meetings.

We bootstrapped @TryBrickroad in the early days by brokering data deals to AI labs. From that work, we learned that there is a massive missing infrastructure layer needed to power efficient access to data. The access necessary to square with the worlds growing computational needs. Most people stop at data production and brokerage. Its neither sustainable, not civilizationally impactful. Build shit that matters.

The Inevitable Consolidation of Data Markets - Mercor, Data Production, and the Missing Technology Layer TLDR: The data industry doesn't need more brokers or production companies. It needs better technical rails. I've been seeing a lot of narratives here about the state of data and RL environments. Few of these tales are grounded in actual data market dynamics - likely because few have actually sold data, very few have done it at any significant scale (vc scale), and even fewer have the right incentives to solve the most critical issues in the data markets with technology. A couple of things to keep in mind when exploring this space. 1. Data procurement is a $400B industry run almost entirely on human-to-human rails - it thrives off of artificial frictions and economic inefficiencies (for an in-depth analysis, see nitter.cf/Freeman_Lewin/status/2…). Outside of a few commodities markets - plastics, scrap metal - I know of no more inefficient market. 2. It is inevitable that the human-centricity of data markets will, and must, change in the next 24 months. Buyer archetypes are shifting, the information frontier is expanding, and access is becoming the bottleneck. The work Mercor, Surge, Scale, Micro1, Handshake, and the RL labs have done in this space is not trivial. Mercor's rumored $20B valuation is cheap in the grand scheme of the production layer. But production is only a small portion of what makes up the market. The access layer — massive by comparison — simply doesn't have a technical winner yet. The dominance of production companies is the market pricing in what doesn't exist: the technical rails underneath it. Mercor and its peers are winning because the access layer hasn't been built. That won't stay true for long. You see, labs rely on large production houses because accessing what's actually valuable - closed, private corporate data repositories - is hard. It's simply easier and cheaper to go to a producer for marginally good data than to negotiate with the companies that hold alpha as a byproduct of their operations. To get deals done in the current environment, every buyer negotiates bilaterally with every seller. That is an n×m problem - and it hampers the progress of every lab, production house, financial firm, government, and corporation that relies on external data to drive strategy, operations, sales, and alpha generation. Bilateral deal-making carries significant fixed costs - discovery, diligence, contracting. It's why almost every established data producer has a minimum-value entry point. Below a certain deal size, nobody can serve you profitably. The massive long tail of data buyer TAM is not being serviced as a consequence of these transaction costs(cf. details at brickroad.network/thesis.pdf). The human-centricity of the data market is also meeting a demand problem. We are scaling computation — LLMs, agents — into every corner of the economy at an accelerating rate. But computations cannot scale their ROI without inputs. What good is all that horsepower if you can't get it on the road? The rate at which we're deploying compute and the rate at which we can access frontier data (cf. brickroad.network/blog/the-i…) have diverged. If we can't close that gap, we'll be hard-pressed to earn a return on our computations. And finally, the buyer archetype is changing. Data acquisition was once the primary domain of regulated industries like finance and banking. Now the archetype is entirely different: agents buying data via APIs for context, AI labs doing pre- and post-training on unstructured multi-modal data, corporations buying data not just for strategy but to train their own models, and governments purchasing real-time feeds for intelligence and national security. More buyers, buying more types of data, in smaller increments, at faster rates. All these forces point in one direction. The data layer needs a technology that can scale at the rate of compute. That's what we're building at @TryBrickroad