Callosum raises $100m for mixed-silicon scheduling, and the 2x accuracy claim is still its own

Callosum has raised $100m, six months after leaving stealth with $10.25m, and Bloomberg reported that the UK's public AI fund is among the backers ^[1][2][3]. The money is notable for where it is not going: not into chips, not into a frontier model, but into software that decides which chip runs which part of a job ^[4].

The product is orchestration. Callosum's software spreads a workload across whatever mixed hardware is in the rack, Nvidia, AMD, Google and others, routing each part of a task to the chip and model best suited to it instead of assuming a homogeneous GPU grid ^[4]. That is a bet on the shape of the next buildout. Most inference capacity has been sold as uniform fleets; Callosum is selling the assumption that fleets will stay heterogeneous and that the scarce skill is scheduling across them.

The performance case is where care is needed. The company claims that on complex agentic work such as autonomous computer use, its approach delivers twice the accuracy, seven times the speed and a quarter of the cost of a conventional GPU setup ^[5]. Those are three separate step changes from a routing layer, all asserted by the vendor, on a workload category with no standard benchmark named in the announcement, and no customer has published figures of its own ^[5][15]. Callosum had not published a statement of its own at the time of writing, which makes the numbers harder still to interrogate ^[15].

The founders, Danyal Akarca and Jascha Achterberg, met during PhDs at Cambridge, have published in Nature journals, and have stints at Intel and Google DeepMind between them ^[6]. Their thesis is explicitly contrarian: "Big labs are currently betting that one model will rule them all," one of them said at launch, according to the report. "We think that's wrong." ^[7] The company name comes from the corpus callosum, the fibre bundle connecting the brain's hemispheres, and the argument is that intelligence emerges from separate systems coordinating rather than one system scaling ^[8]. Whether that holds as engineering is still open in the field ^[9].

The state's role is the murkier half. In April, Callosum became the first equity investment 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 ^[10]. Neither the cheque size nor the equity stake was disclosed then or since, a gap that has already drawn parliamentary attention ^[11]. Six other companies 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 rather than cash ^[12]. Bloomberg's report did not disclose a valuation, the full investor list has not been confirmed by the company, and how much of the $100m is public money is unclear ^[13].

The February seed was led by Plural, the fund founded by former Wise executive Taavet Hinrikus, with the research agency ARIA and angels including Charlie Songhurst, Stan Boland and John Lazar ^[16]. Talks about a round of up to $100m, roughly £75m, were reported in May ^[17]. The step up is close to ten times the seed in six months ^[18].

Watch Companies House filings next quarter, which should settle the ownership question ^[14]. Watch also for the first customer willing to attach its name to the 7x figure.