If Software Is the New Marketing, What's the New Product?
Software has no moat, but intelligence may. Four forms of intelligence — in the model, the data, the people, and the network — with very different defensibility.
Recall Labs CEO Andrew recently pointed out that “software is the new marketing.” As software’s cost to produce rapidly falls, moats for pure software are disappearing. This is increasingly true for even sophisticated software. At the same time, the cost to produce content is falling even faster, so marketing channels are absolutely overloaded with AI slop.
Software is becoming marketing because it’s cheap to produce but not free, and with a far easier install path for even technical and niche tools (“Claude, install this”) shipping OS tools or skills or full products is becoming the best way to cut through the noise of AI content slop and get attention for a person or company’s capabilities.
This begs the question: if software is the new marketing, what’s the new product? For most companies in the knowledge economy, the product is shifting to intelligence. Software is a layer to access or leverage intelligence for these companies. (For others that touch financial transactions, physical goods, transport, etc. the dynamics are different). Software has no moat, but intelligence may.
The key question is: where will intelligence live, and who will control it? We’re seeing the emergence of four forms of intelligence, with major questions still about how distributed or concentrated it will be.
Intelligence in the model
Foundation labs ship consumer products — ChatGPT, Claude Code, Gemini — to create leverage for what the underlying models can do. Given the capex and scale involved, this will likely be concentrated among a handful of players and a long tail of open source alternatives that keep the big labs honest on pricing power.
We may yet see a proliferation of foundation models in specialized domains (the ‘medicine model’), but each new generation suggests that generalized models may eat more intelligence. It’s clear either way that models are eating a lot of intelligence. It’s less clear what could be held out and keep value as a complement.
Intelligence in data
This is where a lot of software companies want to plant their flag: proprietary data, insights and intelligence.
The pitch is that models are trained on the open internet, and certain additional data, but curated, proprietary, hard to recreate data can augment that in highly valuable ways. It’s clear there’s some value here, but less clear how much pricing power and defensibility there will be. There is evidence on both sides of this.
Tales from the “models ate my business” side (and there are many many more):
- When Anthropic launched a Meta ads connector, Ryze AI’s deal close rate collapsed from 70% to 20% in weeks.
- Jasper AI’s revenue roughly halved after its marketing-copy moat turned out to be one context window away from commoditized.
- Chegg’s converted visits dropped 89% year-over-year once Google’s AI Overviews absorbed the queries it monetized.
Meanwhile, the big labs are sourcing domain data directly — Anthropic and OpenAI both launched healthcare-specific products in January, integrating medical coding databases and clinical records. At the scale these labs are operating at, they can find training data extremely widely. For inference data & context, data portability regulation is accelerating: the EU Data Act took effect in September 2025 (US CFPB’s personal financial data rules hit the largest institutions in April 2026), giving customers the right to port data from connected devices, and the labs are also racing to connect to enough of this that it becomes table stakes for any piece of software — people simply won’t keep using software that refuses to share because the bar has been raised for context awareness.
But models won’t eat all data. The most clear example is that companies whose data intelligence is generated through physical processes (or other hard to re-create processes), seems defensible for now. OpenAI and Anthropic are not going to immediately compete on self-driving with Tesla and Waymo, who have been building their training sets for years. Fero Labs predicts product quality from individual manufacturing line telemetry — each steel mill’s blast furnace has unique dynamics no foundation model has access to. CAPE Analytics layers years of labeled insurance loss data over aerial imagery to produce property-level actuarial risk scores that a general vision model can’t replicate. Recursion Pharmaceuticals runs 2.2 million samples per week through automated wet labs, generating 50+ petabytes of biological data that emerges from physical experimentation, not text corpora.
Data that’s fundamentally textual, public, or derivable from public information is being largely absorbed because it can be relatively easily replicated in an age of AI agents, abundant data, and synthetic data. Data generated through physical-world interaction — sensors, labs, regulatory processes, verified relationships — resists absorption (at least until intelligent robotics scales).
Intelligence in the people (services)
a16z recently framed the shift as “services-led growth” — selling outcomes to clients rather than tools to workers. Forward-deployed engineer roles are up 800-1000% year-over-year. Decagon prices per conversation handled, not per agent seat.
The inversion from Software as a Service to Services by Software goes deeper than pricing, though. Software is becoming the marketing of choice for highly capable people and teams to advertise their hands-on services.
Trail of Bits, the security firm, built and open-sourced Slither and Echidna. The tools prove they understand the attack surface better than anyone — so they get hired for the audits. thoughtbot, the Rails consultancy, open-sourced Factory Bot and dozens of other tools — collectively over a billion downloads. They don’t cold-call. Their GitHub is their sales deck. The audits fund better tools.
This arc is accelerating and reaching wider. Patrick McKenzie built small SaaS products for years — but the real business was always consulting, where he went from $100/hour to $30,000/week. The software was proof of work. Simon Willison co-created Django, built Datasette, and now consults on AI tooling — with a readership in the millions that came from building in public. The tools and experience with software are the pipeline.
When anyone can build the tool, what matters is the judgment behind it. Knowing what to build for a specific client, how to weave it into their messy organization, and where it creates leverage they couldn’t see themselves. That’s the intelligence. The software proves you have it.
In the age of AI, everyone has more leverage, and the best people and teams are differentiated even more than before from the rest. They’ll command higher prices, and while software as a service margins will fall, services enabled by software margins will rise. But how do you know who is the best, in a world filled with slop? Working, effective software is the proof of work.
Intelligence in the network
The last form is the trickiest. It’s about keeping multiple parties intelligent with respect to each other — making sense of information scattered across organizations, systems, and people who don’t naturally talk.
Jamin Ball’s recent writing on systems of record makes the case well: as automation increases, the importance of a canonical source of truth increases, not decreases. “The more we automate,” Ball writes, “the more important it becomes that someone has done the unglamorous work of deciding what the correct answer is and where it lives.”
But a source of truth is, at the end of the day, a database. Within a single enterprise, coordination software is often a well-designed application on top of shared data — a pretty thin moat. Enterprises providing systems of record are going to have a much tougher time charging exceptional margins when CTOs and CIOs believe they can replicate much of the functionality and interoperability built atop the core database easily, and just do it on their fully owned and controlled DBs.
The moat gets thicker when coordination crosses organizational boundaries, where you need governance rules, dispute resolution, and trust protocols between competing interests. That’s intelligence embedded in how parties stay coherent — not just what they store. Visa operates a rule book across billions of parties — dispute resolution, fraud scoring, settlement logic built over decades. Surescripts handles 95% of US e-prescribing by coordinating trust between EHRs, pharmacies, and benefit managers — coordination so effective the FTC sued them for monopolization. The joint decision making to migrate here is a bigger moat.
But even strong moats have limits: Ball’s own Amadeus example is telling. Amadeus still owns the system of record for travel, but Booking and Expedia captured seven times the value by building the front door on top. Top companies are racing to introduce stablecoin systems to compete with Visa. And regulations + technology are making interoperability more of a standard, reducing the strength of switching costs for cross-boundary coordination layers. There’s still very real value in anchoring multi-party trust, but the moat is probably far more in the trust than the technology.
What intelligence do you offer?
The knowledge economy world is being repriced. For the vast part of the economy that we used to call the software industry, the product is shifting from the code to the intelligence underneath it. The major question for companies and investors navigating this transition is what intelligence do generalized models eat, and what intelligence can effectively complement it.
Companies that figure this out will create mind-blowing software at mind-blowing speeds to market just how powerful their particular intelligence is as a complement to SOTA models. Companies that don’t figure this out may find themselves waiting in fear of each new Anthropic release.