Callosum Raises $100m With the British State on the Cap Table. Here’s Why Construction Should Care.
Callosum, founded last year by two Cambridge PhDs, has raised a $100m seed led by Atomico, with Plural, DCVC and the UK’s £500m Sovereign AI Fund on the cap table. It does not make chips. It makes software that chops an AI task into pieces and sends each piece to whichever model and whichever chip handles it most cheaply. That matters to construction for one unglamorous reason: inference cost is the hidden input price under every AI tool your vendors are trying to sell you, and it sits under the data center demand keeping large parts of the market busy.
Six months ago Callosum came out of stealth with $10.25m. On Thursday it announced ten times that, and the British government is a shareholder. The valuation has not been disclosed, which is the first of several gaps worth holding onto.
Most readers of this newsletter will never buy a GPU or care which accelerator is racked where. The reason this deal belongs in a construction publication is that it targets the running cost of AI rather than the training cost, and running cost is what gets passed to you. Every estimating tool, drawing-review platform, and site-camera vendor pitching your business right now is paying an inference bill on your behalf. Their pricing and their margin sit on top of that bill.
There is a second thread. State money is now buying equity in AI compute companies rather than just funding the sheds they sit in. If you build, power, or bid on data centers, the direction of public capital is a leading indicator you can actually use.
A $100m seed, one of Europe’s larger ones, for software that routes AI work across mixed hardware instead of assuming a uniform wall of Nvidia GPUs. The UK state is a disclosed but unquantified investor. The performance claims are the company’s own and vary depending on which announcement you read.
What Callosum actually sells
Orchestration, not silicon
The product is a cloud service called Tailored Inference. It takes an AI task, which in practice is rarely one clean instruction but a chain of steps, and breaks those steps into standalone units the company calls blocks. Each block gets routed to the model best suited to it. Something trivial goes to a cheap, small model. Something that needs real reasoning goes to a frontier model. Then the service picks the chip that runs each of those models most efficiently, according to SiliconANGLE’s write-up of the launch.
Think of it the way you would think about crewing a job. You do not put your most expensive specialist on every task on the programme. You match the trade to the work. Callosum’s argument is that the AI industry has been doing the equivalent of sending a chartered engineer to sweep the site, over and over, at scale, and paying for it.
The service supports accelerators from more than half a dozen companies at launch, and Cerebras announced a partnership the same day. The founders, Danyal Akarca and Jascha Achterberg, met while doing PhDs at Cambridge and have worked at Intel and Google DeepMind between them.
The money, and who is on the cap table
Atomico led it, and then the state showed up
Atomico led the round with participation from Plural, DCVC, and the UK Sovereign AI Fund, per Bloomberg. Plural, the fund set up by former Wise executive Taavet Hinrikus, had led the earlier $10.25m alongside the government research agency ARIA and angels including Charlie Songhurst, Stan Boland and John Lazar.
The public money is the part that should make you sit up. Back in April, Callosum became the first equity investment made by the UK’s Sovereign AI Unit, a £500m vehicle chaired by James Wise and launched with backing from technology secretary Liz Kendall and chancellor Rachel Reeves. Six other companies named in the same announcement, including Cosine, Prima Mente, Cursive, Doubleword, Twig Bio and Odyssey, got up to a million GPU hours each on national supercomputing capacity instead of cash.
Neither the size of the state’s cheque nor the equity stake has been disclosed, then or since, a gap that has already attracted parliamentary attention according to The Next Web. Companies House filings should settle it next quarter. Slower than a press release, considerably more reliable.
Why inference economics reach the jobsite
Your vendor’s compute bill is buried in your subscription
Training a frontier model is a problem for about ten companies on earth. Running one is a problem for everybody who builds on top of it, and that includes the entire ConTech stack. When a drawing-review platform reads a 4,000-page tender pack, someone pays for the tokens. That cost lands in the vendor’s gross margin first and your renewal quote second.
This is why the adoption picture in construction has been so uneven. We have written before about the quiet brake on ConTech AI, where the constraint is rarely whether a tool can be built but whether a conservative buyer can absorb it. Price is a large part of absorption. A tool that costs a vendor 70% less to run is a tool that can be priced for a mid-market subcontractor rather than only for a tier-one with an innovation budget.
There is a supply-side version of this too. A meaningful share of US non-residential activity is currently propped up by hyperscaler capex, something we broke down in our look at how data centers are keeping US construction afloat. Anything that changes the cost of running AI eventually changes how much compute the market wants and where it wants it built. That is not an immediate effect. It is a two-to-five-year one, and it belongs in your scenario planning rather than your next bid.
The deals that move construction, one email a week
Funding, M&A, and the context underneath them, cut down to one sharp read every Tuesday.
Join 3000+ Readers
The numbers to hold loosely
Every performance figure here comes from the company or its lead investor
Callosum has quoted several different sets of results and they do not line up neatly. At the February launch the claim was roughly twice the accuracy, seven times the speed and a quarter of the cost on complex agentic work. Atomico’s note on this round cites four times the speed, 70% lower compute cost and a 10% lift in task success on agentic financial services workloads run with Cerebras. SiliconANGLE reports a separate claim of some inference tasks completing 3.7 times faster than a named frontier model.
Different benchmarks, different workloads, different comparison points. Normal for a company at this stage, and none of it independently verified. If a ConTech vendor tells you their tool got cheaper because of clever routing underneath, ask which benchmark, against what baseline, on whose hardware. Same question you would ask about a productivity claim on a piece of plant.
Two other unknowns sit on the deal. The valuation was not disclosed. Neither was how much of the $100m came from the Sovereign AI Fund as against private capital, which is the figure that would show how heavily the British state is now underwriting its own compute layer.
The read, depending on where you sit
Three positions, three different things to watch
If you run projects, this changes nothing this quarter. What it does is give you a fair question for your next software renewal: has your vendor’s underlying cost fallen, and did any of that reach your price? Most will not have an answer. The ones that do are worth keeping.
If you are building a ConTech product, the routing layer is becoming a real lever rather than a research curiosity. Whether you use Callosum or one of the alternatives, the cost of a token is no longer a fixed input you have to accept. That reshapes what you can afford to give away in a pilot, which is often the whole battle in this industry.
If you allocate capital, note the pattern rather than the company. The UK is buying equity in the software layer of AI infrastructure while the physical layer keeps drawing private money, visible across recent ConTech boards including this month’s funding roundups. Britain is not trying to out-build the American labs on frontier models. It is buying positions in the plumbing. Whether that is strategic patience or smaller ambition dressed up, the next few disclosures will tell you.
| Detail | February 2026 | August 2026 |
|---|---|---|
| Amount | $10.25m | $100m |
| Lead investor | Plural | Atomico |
| Other backers named | ARIA, plus angels including Charlie Songhurst, Stan Boland, John Lazar | Plural, DCVC, UK Sovereign AI Fund |
| State involvement | ARIA (research agency) | Sovereign AI Fund equity, amount undisclosed |
| Valuation | Not disclosed | Not disclosed |
| Product status | Stealth exit | Tailored Inference APIs generally available |
It sells software, not hardware. Its Tailored Inference service splits an AI task into smaller units, sends each one to the AI model best suited to that step, then deploys each model on whichever chip runs it most efficiently. It supports accelerators from more than half a dozen manufacturers, with Cerebras joining as a partner alongside the funding announcement. (Source)
Atomico led, with participation from Plural, DCVC, and the UK Sovereign AI Fund. Bloomberg described the public fund’s contribution as significant but did not put a figure on it. Callosum did not disclose a valuation. (Source)
A £500m government vehicle chaired by James Wise, launched with backing from technology secretary Liz Kendall and chancellor Rachel Reeves. Callosum was its first equity investment, announced in April. Six other companies in the same announcement received up to a million GPU hours each on national supercomputing capacity instead of cash. (Source)
No. Every figure in circulation comes from Callosum or its lead investor, and the numbers differ across announcements. The February launch claim was around twice the accuracy, seven times the speed and a quarter of the cost on agentic work. Atomico’s note on this round cites four times the speed, 70% lower compute cost and a 10% lift in task success on financial services workloads run with Cerebras. Treat all of them as vendor claims until a third party benchmarks the platform. (Source)
Because you pay them indirectly. Every ConTech product built on large models carries a per-use compute cost that shapes the vendor’s pricing and margin. When that cost falls, tools that were only viable for tier-one contractors with innovation budgets start becoming viable for mid-market firms. Adoption in construction has consistently been constrained by what buyers can absorb rather than by what engineers can build. (Source)
Three things. The valuation was not disclosed. The size of the Sovereign AI Fund’s cheque and its equity stake have not been published, a gap that has drawn parliamentary attention. And the full investor list has not been confirmed by the company. Companies House filings should resolve the ownership question in the next quarter. (Source)
No. Velaura raised $110m for low-power inference chips, and UK AI startups collectively are valued at around $256bn according to a recent industry count. The distinction is that Callosum sells orchestration software rather than silicon, which is a cheaper thesis to fund and a faster one to ship. (Source)
Related Articles
“You Can Only Go As Fast As Your Customers”: The Quiet Brake on ConTech AI
Data Centers Are Keeping US Construction Afloat
SoftBank’s $100B Robotics Bet: Why Roze AI Wants to Build the Data Centers AI Runs On
Latest Construction Technology Funding Rounds, 17th Aug 2026
The Next Web, Callosum raises $100m with the British state on the cap table
Bloomberg, AI Startup Callosum Raises $100 Million to Make AI Tasks Cheaper
SiliconANGLE, AI workload optimization startup Callosum raises $100M
GOV.UK, first backing through the UK’s Sovereign AI Unit
Tech Startups, Callosum raises $100M backed by Sovereign AI Fund
