A $100 Million Seed Bets Against Vendor Lock-In

The $100 Million Round, In Full

Callosum, a London-based AI infrastructure company, announced on August 20, 2026, that it had raised a $100 million seed round, an unusually large check at the seed stage for any startup, let alone one still built around a single core idea. The round was led by Atomico, with Plural, DCVC, and the UK's Sovereign AI Fund also participating, according to Callosum's own announcement on its company blog. Bloomberg and the European startup outlet Sifted both corroborated the raise and the investor list independently on the same day.

Callosum names two chip-vendor partnerships as part of its routing approach: Cerebras and Rebellions, both suppliers of AI accelerators built as alternatives to Nvidia's GPUs. Naming those two specifically, rather than describing the approach only in the abstract, is itself part of what a seed round of this size is buying: enough capital to stand up working integrations with more than one non-Nvidia vendor before the company has to prove out a broader commercial thesis.

The Bet: No Model, No Chip, Gets Locked In

Callosum's entire product thesis is that an enterprise should not have to commit to one AI model or one chip vendor to run inference at scale. The company describes its approach as heterogeneous compute: a routing layer that sits between a business's applications and whichever combination of models and chips actually serves a given request most cheaply or reliably, rather than hard-wiring that business to a single provider's stack. Cerebras and Rebellions are not incidental logos on a slide; they are the concrete evidence that the routing promise is real enough for two non-Nvidia chip makers to build integration work into their own roadmaps.

A $100 million seed round is also large enough to be an anomaly against nearly any other seed-stage company; most seed rounds fund a founding team and a first product, not routing infrastructure across multiple foundation models and multiple chip architectures at once. That is not a claim about whether Callosum will succeed, it is a description of what the round size is actually paying for: an unusually wide starting footprint for a company whose product only works if it stays neutral between vendors who would each rather it did not.

A Sovereign Fund's First Disclosed Bet Skips the National-Champion Play

The UK Sovereign AI Fund's participation in this round is also, on the reporting so far, notable for what it is not: it is not a bet on a single national foundation model. Coverage of the raise suggests this is the fund's first disclosed investment check, and if that holds, the first disclosed move by a UK-government-linked AI fund went to a routing layer designed to work across other people's models and chips, not to a lab trying to build one model to rival OpenAI or Anthropic.

That choice is worth reading as a policy signal in its own right. The more familiar sovereign-AI playbook, visible in several government-linked initiatives elsewhere, is to back a national champion model and hope it can compete directly with the largest US labs. Putting a first disclosed check into multi-vendor routing infrastructure instead suggests the UK fund is placing its early industrial-AI risk on structurally avoiding lock-in for the country's AI users, rather than on picking a single model to win.

The Contract Question This Round Puts on Every Vendor's Desk

For any EU or UK business currently negotiating an AI inference contract, whether with OpenAI, Anthropic, or a single cloud provider's managed inference service, this round is a concrete prompt to ask a question before signing, not a tip about which company to invest in. A $100 million seed check for multi-vendor routing infrastructure means well-funded, credible alternatives to single-vendor lock-in now exist as a live commercial category, not just a theoretical one.

The practical move for a procurement or engineering team is to ask its own current or prospective AI vendor directly about portability: what does it cost, in time and engineering effort, to move workloads off this vendor's specific model or chip stack if pricing or performance changes later, and does the contract itself make that move harder. Multi-year AI inference contracts are being signed across Europe right now, and the existence of funded routing infrastructure is a reason to get that answer in writing before committing to one, not a reason to switch to any particular alternative.

The UK-EU Split in Where Sovereign AI Money Goes

Where the UK put its early industrial-AI money is also a useful contrast to draw against the rest of Europe's own sovereign-AI efforts. Several EU member states and the European Commission have stood up their own sovereign-AI-oriented funds and cloud initiatives over the past two years, most of them framed publicly around building or hosting European models and European compute capacity. A UK fund choosing, in its first disclosed check, to back vendor-neutral routing rather than a model or a data center is a different bet on where the risk in AI dependency actually sits.

Whether that bet pays off for the UK is a separate question from what it means for a buyer today. What the reporting supports is narrower and more useful: a serious investor group, including a government-linked fund, has now put real money behind the idea that no enterprise should have to be captive to one AI vendor, and that idea is worth testing against your own contract before you sign a multi-year one.

Read next: Stripe Is Buying the AI Router Built to Avoid Lock-In  |  CodeRabbit's $1.5B Round Lets AI Approve Your Code