Integration in AI

Welcome back to Consortium.

I recently came across an interesting post by Gokul Rajaram that covered the Diffusion Prime Thesis (coined by Nihar Bobba) on how vertical AI companies will not remain vertical for very long. The key to expansion will be going horizontal, but not in the traditional sense that most software companies have in the past.

Thesis

The core thesis follows from the observation that unlike traditional SaaS companies where COGS would diminish and gross profit margins would rise as the company would scale, AI companies today face mounting costs as they grow in size (customer size and number, revenue, contracts, etc.) due to one thing - inference.

This tension wasn't new to me as I'd already been circling it while looking at a Series B AI-native company, where inference costs scaling with usage volume were quietly eating into COGS in a way traditional SaaS never had to contend with.

Today, inference costs have led to considerable gross margin compression (10-17%) for listed SaaS companies that are adding generative AI capabilities. Additionally, the average AI product company is experiencing gross margins of ~50% relative to the large 80% margins that pre-AI SaaS companies enjoyed (ICONIQ, 2026).

Although there are meaningful solutions such as intelligent model routing, semantic caching, batch processing, and RAG to mitigate this risk (and cut inference costs down 50-70%), companies have been thinking of more aggressive approaches towards improving long term unit economics health.

An example of this is how companies have now started "owning the stack". This translates to the idea that a company acquires different layers in its architecture that allows significant cost control.

Some examples of this are:

1. Legora acquiring Qura for its legal research infrastructure and partnering with Wolters Kluwer and InfoTrack to pull proprietary statutory and regulatory data directly into its workflows.

2. Sierra acquiring Opera Tech, Receptive AI, and Fragment for buying local market presence in Japan, voice AI capabilities, and strong talent plus geographic expansion into France.

The difference between this form of horizontal expansion and what we have traditionally seen in SaaS (buying companies solely for geographic expansion and market consolidation) is that these forms of acquisitions now allow for full stack technological vertical integration.

Currently, the data layer, the routing layer (orchestration APIs vs. in-house agentic capabilities), distribution layer (FDE installment for local setup), and talent layer are all being actively pursued.

However, the real gap currently exists in the compute area while even the fine-tuning area has experienced companies like Wonderful that are making strides in localised fine tuning for language, cultural norms, and regulatory environments.

HR Tech Experience

Aside from Legora and Sierra, we are seeing a similar trend take place in the HR tech space.

Phenom acquired Be Applied for cognitive assessment capabilities and Included AI for its agentic people analytics capabilities.

Smaller acquisitions include Docebo buying 365Talents to develop an agentic career mobility platform and Perceptyx buying Lyceum to create an orchestrating layer for learning outcomes.

However, not all meaningful acquisitions qualify under this thesis.

Remote's acquisition of Atlas and Payoneer's acquisition of Boundless are more a counter pattern to this as they take the traditional form of acquisitions.

The primary incentives here were creating a more consolidated workforce management system (Remote <> Atlas) and integrating a payments solution into a compliant global workforce ecosystem (Payoneer <> Boundless).

For Founders & Funders

What are the implications of this for both tech firms as well as venture funds?

For Tech Firms & Founders

1. There will be a two tier market - smaller firms that don't look to scale and only serve a select set of customers and won't incur strong inference costs or, if they do, can mitigate this with caching, routing, batching, and other processes that shave off COGS.

   The incentive to not expand is two tier - one, not having to endure the intensity that comes with having to take on large capital and thus risk valuations and returns.

   Two, building for acquisition is a real strategy as bigger firms that want to expand will start to eat up firms in the architecture stack and more diffusions will take place.

   However, founders that are incentivised simply by the exit path will risk not creating enduring and more valuable companies, thus inversely creating a company that is not an attractive acquisition.

2. As the stack matures, founders that are looking to build in the AI space have options within the architecture.

   The data layer is growing in maturity, however there is ample room in the orchestration and routing space.

   Although horizontal agentic routing capabilities are being built by the larger companies (Legora and Sierra's OSs), smaller companies have the opportunity to build vertical specific orchestration points that create real value for incumbents, and strong defensibility against other horizontal competitors.

For Venture Funds

1. Capital infusion will be even larger.

   Currently, we have seen Sierra raise $950M at a ~$16B valuation and Legora has tripled in valuation to $5.6B in five months.

   With firms growing larger and looking to grow through this inorganic route of stack consolidation, venture firms need to be ready to fund the few winners that can afford to do this - making returns more predictable.

2. As more founders are looking to start companies across the technical architecture, venture funds have more options to underwrite layer specific bets through explicit theses as the path to exit is predefined.

   This allows supporting the current incumbents through funnelling smaller layer-based companies towards them.

   In essence, venture funds become arbiters between large companies that will return the fund and smaller companies that will accelerate this process.

However, there is a risk factor that is involved.

As concentration occurs amongst the large firms, smaller layer specific bets will get acquired for cheap as the aim for some founders might end up being:

"Build the next product that Sierra or Legora or Phenom wants to buy."

---

If you have reached this far, thank you very much for reading this issue from Consortium.

I have loved every bit of research, writing, reading, scrapping, and re-writing. I hope you have enjoyed this as much as I have, and have hopefully learnt something new along the way.