In its issue 12, the summer 2023 edition, ActuIA magazine devoted an interview to Karim Beguir, co-founder and CEO of InstaDeep, on multi-agent artificial intelligence. On 31 July 2023, a few weeks after publication, BioNTech finalised the takeover of his company. Three years later, InstaDeep has kept its brand, its leader and its industrial clients. It has also gained a supercomputer in the Paris region. Its headcount exceeded 400 people at the end of 2024, compared with around 240 when the takeover was announced.
The role played during the pandemic
The relationship with BioNTech dates back to 2019. On 25 November 2020, while the Mainz-based laboratory was developing its vaccine with Pfizer, the two companies created a joint AI innovation lab in London and Mainz. The teams worked on selecting neoantigens for cancer vaccines and on optimising RNA sequences for BioNTech's RiboMab and RiboCytokine platforms, according to the statement announcing the takeover.
The best-known project remains the early warning system for SARS-CoV-2 variants, presented on 11 January 2022. It combines modelling of the Spike protein and a language model applied to virus sequences. Tested weekly from 16 September 2020 to 23 November 2021, it flagged 12 of the 13 variants designated by the World Health Organization, on average 58 days before their designation, according to the two companies. Omicron was classified as high risk on the very day its sequence was posted online, among more than 70,000 variants discovered in October and November 2021. In December 2021, 12,000 new variants were being sequenced each week. Only an AI system could follow this flow in real time, Karim Beguir explained to Rest of World in September 2024. ActuIA had detailed how it worked at the time.
A record takeover for an African start-up
Founded in Tunis in 2014 by Karim Beguir and Zohra Slim, based in London since 2015, InstaDeep had raised $7 million in 2019 from AfricInvest, then $100 million in January 2022, with BioNTech and Google among the investors. A year later, on 10 January 2023, BioNTech announced the takeover of the 94.7% of the capital it did not yet own: £362 million in cash and shares, plus up to £200 million tied to performance targets. The total could reach £562 million, or $682 million, according to Reuters. It was then BioNTech's largest acquisition.
It was also, according to TechCrunch, the largest takeover of an African start-up, ahead of Paystack, DPO Group and Sendwave. Google had paid around £400 million for DeepMind in 2014. The deal was finalised on 31 July 2023 for around €500 million, including contingent payments. It brought BioNTech some 290 employees and a network of research partners, including Google DeepMind.
BioNTech's AI engine
InstaDeep has operated since as a British subsidiary of BioNTech, and Karim Beguir, honoured in February 2024 with a TIME100 Impact Award, still leads it. Its image segmentation tools annotate tumour tissues five times faster than BioNTech's experts did by hand, he told TechCrunch in October 2024. There he compared AI applied to biology to the state of language processing in 2020, with GPT-3: systems that were starting to work, with a wide margin for progress.
InstaDeep has built its own infrastructure. On 1 October 2024, at BioNTech's first AI Day, it presented Kyber, a cluster of 224 Nvidia H100 GPUs and 86,000 CPU cores, for around 0.5 exaflop in FP16. It is installed in the Equinix data centre in Saint-Denis. The Paris office had more than 130 engineers and researchers in January 2025. AIchor, the platform that distributes computing among teams, processed 15,000 experiments per month on average in 2025, according to the report of the second AI Day, held in London on 1 October 2025.
The models followed. The Nucleotide Transformer, open-source DNA language models, exceed one million downloads and 500 citations. Their third version, NTv3, pre-trained on 9 trillion base pairs from around 128,000 species, reads sequences of one million nucleotides, according to its preprint of December 2025. According to InstaDeep, its AbBFN2 model humanises an antibody, that is, makes it tolerable to the human immune system, in under 20 minutes, when the step takes months in the laboratory. InstaNovo, which identifies peptides from mass spectra, is used at BioNTech to search for new targets and new biomarkers. The teams also claim T-cell receptors whose affinity has been multiplied up to 50,000 times in three optimisation cycles. In July 2026, they presented at the ICML conference MOJO, a model that classifies tumours of 33 cancer types from gene expression and DNA methylation.
Karim Beguir summed up the method at that AI Day. "AI is not a single exponential, but a triple exponential: data, compute and models," he said (translated from English). BioNTech and InstaDeep are building, according to him, "the entire AI stack" to make the group a world reference in immunotherapy.
Revenue more than tripled in two years
The 2024 accounts of InstaDeep Ltd, filed with the UK companies register in August 2025, give the measure of this ramp-up. The British company's revenue reached £61 million, against £34.4 million in 2023, a rise of 77% after 92% the previous year. The operating loss is almost eliminated. It represented 38% of revenue in 2023, against 3% in 2024. Research and development spending doubled, to £30.5 million, and equipment investments reached £12.5 million. These accounts do not consolidate the subsidiaries in Paris, Tunis and the United States, all three profitable in 2024.
Agents for trains, printed circuit boards and water networks
The subject of the 2023 interview, multi-agent AI, remains a research focus. Presented at the NeurIPS conference in December 2025, Oryx learns to coordinate dozens of agents from recorded data, without interaction with the real system, for uses such as warehouse logistics or railway networks. On the ground, InstaDeep is running a five-year project to automate train planning for Deutsche Bahn and is working with Fraport on airport operations, according to TechCrunch. BioNTech has allowed it to keep these non-health clients, because its AI skills are nourished by several sectors, Karim Beguir explained to the same media outlet.
New partners arrived after the takeover. In June 2024, Syngenta Seeds began a collaboration on maize and soya around AgroNT, InstaDeep's genomic model trained on crop species.
DeepPCB, launched in November 2019, applies reinforcement learning to printed circuit board routing, then to component placement. On a board, these two steps remain an engineering job of intuition and manual iterations: each moved component changes the electrical, mechanical and manufacturing constraints of the others. It is a combinatorial optimisation problem, of the same family as the logistics or railway planning on which InstaDeep built its first systems. Google showed in 2021, in Nature, that a reinforcement-trained agent places the blocks of a chip in under six hours, with a result comparable to that of experienced engineers. In September 2024, InstaDeep launched with Google Cloud a usage-billed version of DeepPCB, DeepPCB Pro. Its placement module handles boards up to 1,000 components. According to use cases published by InstaDeep, an optimised placement in 30 minutes is followed by routing in under an hour, compared with over ten hours after a quick placement, with fewer vias, those connections between the layers of the board.
On 4 March 2026, the German pump manufacturer Wilo announced a long-term partnership, signed by the two CEOs in Dortmund, to build an AI platform applied to water networks. A first pilot project is under way. On 24 June 2026 in Berlin, the KI Park association awarded its AI Innovation Award to Wilo for Speaking Waters, an anomaly detection solution for water networks co-designed with InstaDeep.
A new boss at BioNTech in 2027
The takeover had a purpose, stated by Ugur Sahin on 10 January 2023: to make BioNTech "a technology company where AI is integrated into all aspects of our work". Three years later, the computing, the models and the teams are there. InstaDeep's tools have been applied to selecting neoantigens for BioNTech's personalised cancer vaccines, whose most advanced candidate, autogene cevumeran, is in phase 2 against pancreatic cancer with Genentech. The molecules designed with its more recent models, humanised antibodies, T-cell receptors or nanoparticles, remained at the preclinical stage at the AI Day of 1 October 2025.
The next chapter will unfold under different leadership. On 10 March 2026, BioNTech announced that its co-founders, Ugur Sahin and Özlem Türeci, would step down from their roles by the end of 2026 to create a company dedicated to the next generation of messenger RNA, in which BioNTech will take a minority stake. On 3 August, the supervisory board appointed Guido Oelkers, until then CEO of Sweden's Sobi, to succeed Ugur Sahin by 1 February 2027 at the latest. The next day, the group published quarterly revenue of €105.6 million, down 59% year on year. It also lowered its R&D spending forecast for 2026 to €2 to €2.3 billion, from €2.2 to €2.5 billion previously. That statement does not mention artificial intelligence.
As of 9 October, BioNTech had not announced a third AI Day, after those of 1 October 2024 and 1 October 2025. Guido Oelkers comes from a group specialising in rare diseases. InstaDeep approaches this change of leadership with its non-health clients and its own computing resources.
Tunis, Kigali, Cape Town: what InstaDeep is building in Africa
InstaDeep has offices in Tunis, in Kigali, opened in October 2023, and in Cape Town. The Kigali team leads its geospatial intelligence work: it predicts, 30 days ahead, from satellite images, the breeding areas of locusts, with an open software framework, InstaGeo, that other organisations can reuse. The company has supported since its beginnings the Deep Learning Indaba, the gathering of African machine learning researchers, with travel grants for students from the continent. Its Tunisian subsidiary posted £282,000 in net profit in 2024.
Karim Beguir grew up in Tataouine, in southern Tunisia. A graduate of École Polytechnique, he went through the Courant Institute at New York University, then financial engineering at JPMorgan and Bank of America, before leaving a finance job in London in 2014 to found InstaDeep at home. Asked by Rest of World in which country he would start over, he replied: "If it weren't Tunisia, I don't think I would ever have done it." He has since helped the French government prepare the AI Action Summit of February 2025, according to Sifted, and spoke there in plenary session on 11 February 2025.
On 19 May 2026, at the AI Infrastructure Forum in Tunis, he argued for Tunisia to enter the global AI value chain by converting its solar potential into computing power and data centres, rather than exporting electricity, according to Managers. Around 4% of the Tunisian desert could, according to him, cover up to 10% of Europe's energy needs. The window is short, he warned: the next 24 months will be decisive. "The most important refined product in history is intelligence," he said. Eleven years after selling, from Tunisia, his first AI system to a London client, the founder of InstaDeep is proposing that his country become a refinery of it.
Our articles will then appear first in Google Top Stories.