At the opening panel of Rise of AI 2025 in Berlin, moderator Larissa Holzki, Chief Correspondent Tech at Handelsblatt, asks Peter Sarlin, CEO of Silo AI, and Jonas Andrulis, founder and CEO of Aleph Alpha, to find good news for Europe against a week of headlines dominated by the US Stargate infrastructure project, China's DeepSeek model, and multi-billion-dollar funding rounds for OpenAI, Anthropic and xAI. Andrulis argues infrastructure spending is a distraction unless it follows genuine value creation, while Sarlin says Europe's path runs through products, not the biggest data centers, pointing to a fragmented single market as the real barrier to scale. The panel covers Aleph Alpha's tokenizer-free transformer architecture, developed with Silo AI, claimed to give large efficiency and fine-tuning gains for low-resource languages; Silo AI's $665 million sale to AMD, framed by Sarlin as a European success rather than a loss; and Andrulis's account of public criticism after Aleph Alpha avoided the race for general-purpose frontier models. Both argue Europe should put AI at the core of transforming industries like biopharma, automotive and manufacturing rather than chasing efficiency gains alone, and call for stronger, more concentrated European AI ecosystems and a genuine single market.
Key takeaways
- Jonas Andrulis argues infrastructure spending alone is a distraction; Europe needs value creation and sovereignty first, with infrastructure following.
- Peter Sarlin says Europe's core barrier is a fragmented single market that makes scaling hard, not a lack of capital or talent.
- Aleph Alpha and Silo AI built a tokenizer-free transformer architecture, claiming a 400% efficiency gain for Finnish over Llama.
- Silo AI's $665 million sale to AMD is framed by Sarlin as strengthening, not diminishing, Europe's position, since the company stays headquartered in Helsinki.
- Andrulis says Aleph Alpha deliberately avoided the race for general-purpose frontier models, calling it "the fastest depreciating asset in tech history."
Detailed notes
Reacting to a week of US and Chinese AI headlines
Holzki opens by listing recent headlines: the US Stargate project with up to $500 billion in planned AI infrastructure investment, Chinese startup DeepSeek releasing a model said to outperform OpenAI's latest GPT-4, OpenAI closing a $40 billion round at a $300 billion valuation, and large raises for Anthropic and xAI. Sarlin calls the overall progress good for humanity but acknowledges Europe is not matching that scale of investment, while arguing Europe can still compete since AI is not just about building general-purpose models, and that capacity from emerging AI "gigafactories" can build competitive edge if spent well. Andrulis agrees but calls infrastructure investment on its own "a distraction": Europe needs value creation and value capture, meaning innovation and sovereignty, first, since infrastructure should follow value creation rather than precede it. Both reject settling for small wins, with Andrulis saying Europe has missed past technology waves but still has talent, leadership and capital, and "we have to pick up the pace."
Why Europe lacks tech platforms, and where it can still lead
Sarlin argues success is not about the biggest AI lab or data center but about products that deliver end-to-end value to users, and that Europe's core problem, referencing Mario Draghi's competitiveness report from the previous autumn, is that its "single market" is in practice extremely fragmented, making it hard for a digital platform to scale; entrepreneurs repeatedly cite market access, not capital or talent, as the reason they leave for other regions. He points to biopharma, automotive and robotics as domains where Europe already has major companies and is less dependent on owning a tech platform outright, framing AI assistants as only the first chapter of a wider transformation spreading into life sciences, robotics and entertainment.
A tokenizer-free architecture and the AMD deal
Andrulis describes an architecture Aleph Alpha built with Silo AI's team that removes the tokenizer from the transformer, so the model no longer relies on a fixed vocabulary optimized mainly for English, general-purpose text; he says this makes it far better suited to fine-tuning on new languages or company data, citing a 400% efficiency gain for Finnish over Llama, and calls it "a gamechanger for sovereign AI strategies." Sarlin adds that Silo AI has worked on low-resource languages for 15 years, using cross-lingual signal transfer between languages like Finnish and English, and that the new approach adapts a powerful model's reasoning capability to a low-resource language rather than training from scratch. On Silo AI's $665 million sale to AMD, one of Europe's largest AI exits, Sarlin says the company remains headquartered in Helsinki with offices across Europe and gained resources and a global agenda rather than losing its European identity; Andrulis calls it a success for Europe and praises AMD as a collaborative partner. Both note AMD's hardware is strong but its software stack still needs to catch up to Nvidia, an area Silo AI helps with after being among the first to prove large-scale LLM training on AMD GPUs.
Handling hype, criticism and a different competitive bet
Andrulis describes being highly visible and facing criticism over valuation and model quality, saying some claims were simply untrue but that scrutiny is expected once a company is prominent; his advice to founders is to avoid burning money on "the fastest depreciating asset in tech history," the race for the best general-purpose model, since benchmarks like MMLU reward memorized trivia that is irrelevant to transforming a company like Deutsche Bank or Bosch. He also cites his own path, selling a previous company and working in Apple's AI R&D leadership before founding Aleph Alpha in 2019, as what made the company possible. Sarlin separately funded 13 professorships at Finland's ELLIS Institute, following Germany's Tübingen institute, through his own foundation to support European AI research, while stressing that research alone is not enough without commercial follow-through.
Putting AI at the core, not just buying efficiency
Sarlin distinguishes buying AI tools for efficiency and productivity gains, which he compares to adopting PCs decades ago and says every company should do, from putting AI at the core of strategy to redesign processes such as drug discovery or autonomous driving, which requires larger budgets and is no longer "plug and play." Andrulis says Aleph Alpha built a platform with a company called Faria, shown recently at Hannover Messe, to rethink industrial and production processes with customers, and frames the European goal not as beating the US in a zero-sum race but building a prosperous Europe that reflects its values while addressing demographic and climate change through innovation, collaboration and technology.
Q&A highlights: building European ecosystems
Asked about joining forces, Andrulis argues Germany and Europe need to overcome internal rivalry, including between German states, to use their shared market and talent pool, and cites SAP as proof a European tech platform can succeed, since most major global business processes run through SAP systems; he notes several startups have already spun out of former Aleph Alpha employees. Sarlin argues Europe is too fragmented across many small, parallel AI initiatives and should instead concentrate resources into a few strong ecosystems, citing Paris as a model that built a strong academic cluster, attracted labs from Meta, OpenAI and Google, and produced companies such as Mistral, alongside German examples like Heilbronn and Tübingen. Asked for the good AI news each hopes for this year, Andrulis wants European partners such as SAP, Bosch, Schwarz Group and Deutsche Bank to lead a sovereign AI transformation that spreads value more broadly rather than concentrating it in a handful of companies, while Sarlin hopes the EU Inc. initiative reduces market fragmentation enough for one company to scale across all European countries, which he says is otherwise essential for scaling any consumer-facing AI company in Europe.
Quotes
Infrastructure investment is not the game. This is a distraction.
Jonas Andrulis
Eventually, Europe has to have products.
Peter Sarlin
The goal is not to be bigger than the US tech giants or to win against the US. The goal is to build a great future for Europe.
Jonas Andrulis
Questions answered in this talk
Why does Jonas Andrulis think building AI infrastructure alone is not enough for Europe?
At Rise of AI 2025, Jonas Andrulis argues that infrastructure investment is a distraction unless it follows genuine value creation and sovereignty; he says infrastructure should be funded by value already created, not built first in hope that value creation follows automatically.
What is Aleph Alpha's tokenizer-free transformer architecture?
According to Jonas Andrulis at Rise of AI 2025, Aleph Alpha and Silo AI built a transformer architecture without a fixed-vocabulary tokenizer, so it represents all languages equally well and fine-tunes better on new languages or company data, claiming a 400% efficiency gain for Finnish over Llama.
Why does Europe struggle to produce large AI tech platforms?
Peter Sarlin explains at Rise of AI 2025 that Europe's main obstacle is its fragmented single market, which makes it hard for digital platforms to scale, rather than a lack of capital or talent, and points to SAP as a rare example of a European platform that did scale globally.