Build vs Buy: What Ramp's In-House AI Model Router and Revolut's Nvidia-Backed AI Research Unit Signal for Fintech AI Strategy
Within days of each other, two of the most well-capitalised fintechs in the market made the same underlying bet on AI infrastructure. Ramp launched "Router," an in-house AI model router that dynamically selects between large language models for its finance-automation product. Revolut launched its own in-house AI research unit to build foundational models, developed in partnership with Nvidia. Two different companies, two different products, the same strategic instinct: build the AI infrastructure layer themselves rather than buy it as a managed service. For a scaling fintech watching this from the outside, the obvious question is whether that instinct should be theirs too.
What Ramp and Revolut actually did
Ramp's "Router" is a model-routing layer: rather than committing its finance-automation product to a single large language model provider, Ramp built infrastructure that dynamically selects which model handles a given task. That's a meaningfully different bet than simply calling an external AI API — it's an investment in the orchestration layer that sits between the product and whichever underlying models it uses.
Revolut's move is further upstream still. Its in-house AI research unit, developed in partnership with Nvidia, is building foundational models rather than only routing between existing ones. That's a deeper and more capital-intensive commitment: foundational model research sits closer to the frontier of AI infrastructure than orchestration tooling does, and a partnership with a hardware and compute provider like Nvidia signals a multi-year resourcing commitment, not a short-term experiment.
Both moves are evidence of the same underlying pattern, even though they sit at different points on the infrastructure stack: two well-capitalised fintechs independently concluding that AI infrastructure is strategic enough to own rather than rent, within a short window of each other.
It's worth being precise about what isn't being claimed here. Neither company's decision has been independently audited by this article, and no financial or resourcing figures beyond the fact of the investments themselves are being asserted. What's being analysed is the strategic signal the two moves send when read together, not a detailed assessment of either company's internal technical architecture.
Why this is a market-structure signal, not two isolated decisions
A single company building in-house AI infrastructure is a data point about that company. Two independent, well-capitalised fintechs making structurally similar bets within days of each other is closer to a market-structure signal — evidence that a segment of the category has reached a scale and maturity where in-house AI infrastructure investment has become a genuinely live strategic option, not just something attempted by a handful of outliers.
That matters for how a scaling fintech should read this news. The instinct might be to treat Ramp and Revolut's moves as a roadmap to copy. The more useful read is structural: these are two companies with the balance sheet, the engineering headcount, and the AI-native product surface to make in-house infrastructure investment pay off at their current scale. Whether that same calculus holds for a different fintech depends entirely on factors specific to that business, not on what Ramp or Revolut did.
What "build vs buy" means for AI infrastructure specifically
AI infrastructure build-vs-buy decisions differ from a typical software build-vs-buy call in one important way: the build side isn't a single decision, it's a spectrum with genuinely different commitment levels. A team can build a thin orchestration layer on top of third-party models — closer to what Ramp's Router represents — without taking on the far larger commitment of foundational model research, which is closer to what Revolut's unit represents. Treating "build AI infrastructure" as one binary choice obscures a decision that actually has several distinct tiers, each with a different cost, timeline, and payoff:
- Consume. Call third-party model APIs directly, with no proprietary orchestration layer. Fastest to ship, lowest upfront cost, but limited differentiation and no insulation from a provider's pricing or model changes.
- Orchestrate. Build a routing or orchestration layer on top of multiple third-party models — Ramp's Router sits here. This is a meaningful but bounded engineering investment: it buys flexibility, cost optimisation across providers, and some insulation from any single model provider's roadmap, without the cost of training models from scratch.
- Fine-tune. Adapt existing foundation models on proprietary data for a specific use case. A deeper commitment than orchestration, requiring real data infrastructure and ML expertise, but still building on top of existing frontier models rather than replacing them.
- Research. Build foundational models in-house, as Revolut's unit is doing with Nvidia. This is the deepest and most capital-intensive tier, requiring sustained research investment, specialised talent, and typically a compute partnership at the scale Revolut has secured — realistic for only a small number of companies at any given time.
Most of the public conversation about "fintechs building their own AI" compresses these four tiers into one story. The useful version of that story keeps them separate, because the right tier for a given company is a different question at each level.
What actually justifies moving up a tier
The case for moving further up this stack — from consuming models to orchestrating them, or from orchestrating to fine-tuning or deeper research — rests on a small number of concrete conditions, not on following what a larger, better-capitalised competitor has done. A scaling fintech has a reasonable case for investing in an orchestration layer like Ramp's when it is running AI-dependent product surfaces at meaningful volume across more than one use case, such that cost and latency differences between model providers start to matter at scale, or when no single provider's model reliably outperforms the others across the company's full range of tasks. The case for going further — fine-tuning or foundational research — depends on having proprietary data that genuinely differentiates model performance for the company's specific use case, and the sustained capital and specialised talent to support a multi-year research investment, which is exactly the combination Revolut's Nvidia partnership reflects and which very few companies at any given scale actually have.
Absent those conditions, moving up the stack is usually a cost and distraction rather than an advantage. Engineering effort spent building orchestration or model infrastructure is effort not spent on the product itself, and for most scaling fintechs, the product is still the more defensible differentiator than the underlying AI infrastructure.
The case for a delivery-partner model instead
For most fintechs below the scale and capital position of a Ramp or a Revolut, the more capital-efficient path is a delivery-partner model: working with a technical partner that brings AI integration and orchestration expertise, without the company itself carrying the full build, hire, and maintain burden of in-house infrastructure. This isn't a lesser version of what Ramp or Revolut did — it's a different, often more appropriate answer to the same underlying question, suited to a company whose differentiation lives in its product and customer relationships rather than in owning the AI infrastructure layer itself.
A delivery-partner engagement can also be the right way to build toward the orchestration tier specifically: a partner can help design and implement a model-routing layer suited to the company's actual usage patterns, without requiring the company to build that research and engineering capability from scratch in-house. That gets a scaling fintech most of the practical benefit Ramp's Router demonstrates — flexibility across providers, cost optimisation — without requiring Ramp's scale of engineering investment to get there.
What this looks like in practice, tier by tier
A finance-automation product consuming a single external model API directly is making a reasonable, low-commitment choice if its AI features are still a small part of the product and speed to market matters more than cost optimisation across providers. The moment that changes — multiple AI-dependent features, meaningful volume, real sensitivity to latency or per-call cost — is usually the moment an orchestration layer starts paying for itself, which is the tier Ramp's Router sits at. Fine-tuning becomes worth considering only once a company has accumulated genuinely differentiated proprietary data that a generic foundation model can't match on the company's specific task. Foundational research, Revolut's tier, becomes rational only for a small number of companies with both the data scale and the capital to sustain a multi-year research programme — which is precisely why a Nvidia-backed research unit is notable rather than ordinary.
A framework for deciding where a fintech sits on this spectrum
Rather than asking "should we build our own AI infrastructure like Ramp or Revolut," a more useful question for a scaling fintech is: which of the four tiers above actually matches our current AI-dependent product surface, our data advantage, and our capital position — and what would have to change for the next tier up to make sense. For most companies today, that answer is somewhere between consuming and orchestrating, with fine-tuning and foundational research remaining genuinely out of reach until the business reaches a different scale. Revisiting that answer periodically, rather than defaulting to whatever the most-covered infrastructure story in the market happens to be this quarter, is what keeps the decision tied to the company's actual position rather than to industry narrative.
Where Innovify fits
Helping scaling fintechs work out exactly where they sit on this build-vs-buy spectrum — and building the right tier of AI infrastructure for their actual scale and data position, rather than over- or under-investing relative to it — is core to the work Innovify's AI Labs does with clients. That often means building an orchestration layer sized to a company's real usage patterns, without the company needing Ramp's or Revolut's scale of in-house investment to get there.
FAQ
What did Ramp and Revolut actually launch?
Ramp launched "Router," an in-house AI model router that dynamically selects between large language models for its finance-automation product. Revolut launched an in-house AI research unit, developed in partnership with Nvidia, to build foundational models.
Does this mean every fintech should build its own AI infrastructure?
No. Ramp and Revolut's moves reflect their specific scale, capital position, and AI-dependent product surface. Most scaling fintechs are better served by a lighter-weight approach — consuming or orchestrating existing models, often via a delivery partner — rather than replicating a foundational-research investment like Revolut's.
What's the difference between an AI model router and foundational AI research?
A model router, like Ramp's, selects dynamically between existing third-party models to optimise cost, latency, or performance per task. Foundational AI research, like Revolut's unit, builds new models from scratch — a deeper, more capital-intensive commitment typically requiring a dedicated compute partnership.
When does it make sense for a fintech to build an in-house AI orchestration layer?
Generally when the company is running AI-dependent product surfaces at meaningful volume across multiple use cases, such that cost and performance differences between model providers start to matter at scale — a bar many scaling fintechs haven't yet reached.
Is a delivery-partner model a worse option than building in-house?
Not inherently — it's a different answer suited to a different scale and capital position. A delivery partner can help a fintech reach meaningful orchestration-layer benefits without the full engineering and hiring investment that an in-house build like Ramp's required.
Conclusion
Ramp's model router and Revolut's foundational AI research unit are evidence that AI infrastructure has become strategic enough, for a segment of well-capitalised fintechs, to own rather than rent. That's a genuine market-structure signal — but it's a signal about those companies' specific scale and capital position, not a template for every scaling fintech to copy. The more useful exercise is working out which tier of the build-vs-buy spectrum actually matches a company's current position, and treating that as a decision to revisit as the business changes, rather than a race to match what the best-capitalised players in the category are doing right now.
