If you’re a vertical SaaS operator, frontier labs may never replace your business entirely. But they still may be able to pull margin from you. On the surface, your relationship with frontier labs might seem straightforward: they sell tokens; you use them to power AI in your products. But frontier labs have already made excursions into domains like coding, design, and legal. Could this kind of pressure affect margins in your vertical too?
There are at least three opportunities on the table for frontier labs to pull value from vertical SaaS platforms: token economics, carve-outs, and front-door strategy. In this essay, I’m going to explore what those are and map out some opportunities for vertical SaaS platforms to defend themselves.
It’s a common strategy for upstream suppliers to capture value downstream. Oil exploration and production companies, for example, have moved downstream into refinement, transportation and storage, and finished goods.1
So the question isn’t whether frontier labs have the opportunity to capture value from vertical SaaS platforms; it’s whether the pressure of commoditization will make such a strategy attractive.2 And high-tech products do become commoditized. Memory chips, displays, and hard disks are just a few examples.3 When a product becomes commoditized, commodity producers look downstream to capture value.
So if we’re going to see if frontier labs pose a threat to vertical SaaS, one way to exmaine that is to see if commoditization pressures are likely to happen to them. The signs point to yes. Model performance, to date, hasn’t provided anyone with a sustainable advantage. One frontier lab will set a new benchmark, and another will match it soon after.4 Switching costs are still relatively low, both for consumers and enterprise. And pricing pressure is present. Palantir CEO Alex Karp has publicly griped that enterprise CEOs aren’t seeing value from the tokens they’re spending billions on; cheaper, open-weight models are gaining traction (like this one from China); and Meta has publicly said it will compete on price.
Frontier labs don’t need to sell pure commodities to look for value elsewhere; the move toward commoditization (and the resulting margin pressure) would be sufficient. Remember, frontier labs make lots of revenue today5, but profits are a question mark. My thesis is that enough margin pressure already exists; the next question is whether vertical SaaS would be an attractive target.
One could make a strong case against why that wouldn’t be the case:
Ripping out and replacing a system of record is painful and expensive, but that’s the position many vertical SaaS platforms already occupy.
Partner integrations may be difficult to replicate, or even exclusive.
Industry practices can be difficult to understand and bake into a product. Epic Systems, for example, has spent decades building software to work within the byzantine healthcare system.
Customers don’t always just want software; they want the service and support to help them use it well. Frontier labs may not want to be in that business.
Vertical SaaS businesses are often enmeshed in the financial underpinnings of their customers; again, hard to rip and replace.
Software revenue isn’t always even the main business model. Toast, for example, generated 80% of its revenue from payment processing. That’s an orthogonal business model to how frontier labs work.
Vertical SaaS platforms also have reams of proprietary customer data that unlock value like benchmarking or pattern recognition.
To borrow from Helmer’s 7 Powers6, vertical SaaS platforms enjoy high Switching Costs from serving as the system of record and being enmeshed in their customers’ businesses; Cornered Resources in the form of proprietary data and partnerships; and, to a lesser degree, Process Power from understanding their industry deeply.
There’s a structural barrier, too. Vertical SaaS businesses often operate in industries with strict regulatory and compliance requirements, and they often absorb the risk of non-compliance on behalf of their customers [examples]. Frontier labs may simply see such a role as unattractive.
Yet that doesn’t mean vertical SaaS platforms are out of the woods.
Frontier labs don’t need to displace vertical SaaS platforms to capture value for themselves. Between the system of record and a pure LLM is a whole middle ground where value can be contested. I see that contest taking place on at least three fronts: token economics, carve-outs, and front-door strategy.
Here’s a breakdown of each.
In short, frontier labs may have an incentive to steer token consumption toward end users and away from wholesale buyers like vertical SaaS platforms.
There are two primary models for token consumption in a vertical SaaS context. The first is the wholesale route: an end user initiates an AI prompt within a vertical SaaS platform, the action is passed off to the LLM, and the result is passed back to the vertical SaaS platform. The other is the retail route: the user initiates an action in an LLM directly, which then queries the vertical SaaS platform via an API or MCP. The result is passed back to the LLM.
The compute may be similar, but the economics may be different.
In the wholesale model, the vertical SaaS platform buys the tokens. Since it’s buying tokens “in bulk” and may choose to allocate its spend across different LLMs, this gives the vertical SaaS platform a degree of negotiating power. It’s not unlike a factory sourcing power from different utility providers.
In the retail model, it’s the end user who buys the tokens. Since they’re consuming tokens in smaller quantities, they have less buying power. It’s more akin to buying a mobile phone plan from a major carrier, where the carrier sets the terms of engagement.
Yes, wholesale token sales do provide some benefits to a frontier lab. Usage is more even and predictable, so the cost of that compute may be easier to keep in check as well. But ultimately, wholesale buyers have more power over suppliers. If I were a frontier lab, then, I may want to steer token usage to where I had more leverage: the retail channel, especially if the margins were more favorable.
For a vertical SaaS platform that generates consumption-based revenue (like Clay does with selling enrichment tokens), this would pull consumption out of their platform and redirect it to the LLM. Ouch.
Point solutions have always been a thorn in the side of vertical SaaS businesses. It’s harder to charge more for a product add-on when that same functionality can be provided by a point solution, with better results or for a lower cost. Could frontier labs absorb point solutions themselves to capture some of this value?
For example, I recently spoke with a vendor who offers AI voice products (think of an AI receptionist). It’s a bit more advanced than what vertical SaaS platforms offer themselves, so it’s an integration partner that works across multiple platforms, not a direct competitor. But what’s to keep a frontier lab from acquiring a point solution themselves? Doing so could create further stickiness for end users (over other LLMs) and potentially prevent vertical SaaS platforms from profitably acquiring that functionality for their own products.
Or instead of acquiring point solutions directly, perhaps frontier labs could serve as a distribution channel for them. They could provide a marketplace where those point solutions get woven into the LLM product itself, using whatever system of record product happens to sit underneath.
Whether through direct acquisition or as a distribution partner, it’s not hard to imagine a frontier lab offering a “suite” of vertical-specific functions. These tools wouldn’t compete with system of record providers, but they would pull value away from them.
While vertical SaaS platforms serve as the core system of record, many of their customers rely on other tools, too. Some may be vertical-specific, others general purpose (Google Docs, Notion, QuickBooks, Slack). And wherever data is fragmented, that’s an opportunity for frontier labs. Why keep switching back and forth between different software vendors and different sets of agents, when an LLM can serve as a single “front door” for access and coordination?
The move to make SaaS accessible via MCPs is a perfect illustration of this playing out right now.
Vertical SaaS companies recognize that customers would rather just access the system of record through an LLM instead of in their software directly. MCP access is a concession to that. But where value gets redirected with this move is still unclear. The legal space is one good example.
Earlier this year, Anthropic released “Claude for Legal.” It’s not a standalone product the way Claude Code is; it’s a plug-in that gives Claude enhanced capabilities for operating in a legal context. Adjacent to this is Harvey, a kind of “AI layer” that sits between Claude and the underlying legal management software. Claude for Legal is both a complement and a competitor to Harvey, depending on the context.
Some legal firms may use Harvey because it lets them fine-tune Claude for Legal to match the nuances of their specific practice. But other firms may forgo Harvey entirely, choosing instead to use Claude for Legal through Cowork or its Microsoft 365 plugins. In both situations, though, the underlying vertical SaaS platform (like Clio, for example) is relegated to the background. As I mentioned before, the more time users spend with tools outside a vertical SaaS platform, the less they may be willing to pay for it.
The relationship between frontier labs, vertical SaaS platforms, point solutions, and AI layers could play out in many different ways. But for a frontier lab, any situation that relegates vertical SaaS platforms to mere systems of record (or better yet, headless systems of record without a front end) works to their advantage. In the oil industry, some extractors stayed in that lane; others moved downstream; and businesses competed and partnered with each other at the same time. The endgame got complicated; trying to make similar predictions within vertical SaaS is impossible.
What really matters is whether you’re prepared. Whenever there’s an opportunity to capture value, someone will try to capture it. If you’re operating a vertical SaaS company now, here are a few questions to consider.
On token economics: Are you structuring your platform so that users have a compelling reason to pay for token usage through you, or are you incentivizing customers to purchase tokens directly from LLMs themselves?
On point solution distribution: Are you investing in defensible features that would be difficult or impossible for frontier labs to acquire or distribute? Or are you falling prey to the same vulnerability that point solutions are subject to, by creating “features” that will later show up in someone else’s product?
On front-door strategy: How is the front end of your platform creating value in ways that an LLM cannot? What’s your plan for continuing to be the value-creation layer, instead of a commodity, headless system of record? Are there orthogonal business models you can create, or invest further in, that offer value-capture opportunities frontier labs couldn’t access?
Vertical SaaS platforms can no longer afford to worry only about their direct competitors. Even if frontier labs never try to replace them outright as the system of record, they have a legitimate shot at capturing value that could have accrued to vertical SaaS instead. What’s your plan for keeping that from happening?
John Rougeux is the founder of Flag & Frontier, a strategy consultancy that helps executive teams align around their strategy and narrative when the future is up for grabs.
Standard Oil (under Rockefeller) moved into pipelines and tanker cars, and struck agreements with railroads in the late 1800s. Today, companies like Shell and ExxonMobil have moved to a more integrated model, with businesses ranging from oil exploration and drilling, refinement, gas stations, motor oil, and even adjacent products like plastics and rubber.
I don’t want to make the claim that the threat of commoditizing is the only factor that could push upstream producers to move downstream, or even a necessary one. But it’s worth noting that from the 1940s to the 1970s, oil prices were set by suppliers; a group of oil companies known as the “Seven Sisters.” That became undone in later years as spot pricing and futures trading came into play. As pricing shifted from being set by suppliers to being set by buyers, that only commoditized oil further. There’s a great book called “The World For Sale” that covers this in more depth; I just finished it this summer and highly recommend it.
In the 1990s, commodity DRAM prices fell by 51% in 1996 and another 65% in 1997. A 65-inch LCD panel cost roughly $1,000 to manufacture in 2015, but costs about $200–300 today. Price per gigabyte of hard disks has fallen from tens of thousands of dollars per megabyte to pennies today.
I’m slightly oversimplifying. Not all models excel at the same things. My point isn’t to see that they are purely interchangeable, but that model performance alone, thus far, has only provided a short-term advantage.
Anthropic’s last reported annual run rate was $47B; OpenAI’s was around $25B.
From the book, 7 Powers, The Foundations of Business Strategy, by Hamilton Helmer.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.