By Sagoh 
By [Your Name/Editorial Desk]
In the modern venture capital ecosystem, there is an obsession with "traction." For most SaaS founders, this means proving a repeatable revenue model, customer acquisition costs, and a clear path to product-market fit. But for Alexander Kardos-Nyheim, a former law student who successfully navigated the high-stakes world of AI research, these metrics were not just insufficient—they were a distraction from the only thing that truly mattered: scientific breakthroughs.
Kardos-Nyheim, founder of the AI research firm Safe Sign Technologies, recently shared a provocative perspective on the current state of artificial intelligence. Having sold his company to Thomson Reuters in 2024—the first pre-revenue acquisition in the giant’s 170-year history—he argues that the industry’s current fixation on "application-layer" products is a recipe for long-term obsolescence.
The Genesis of Safe Sign: A Study in Capital Efficiency
The story of Safe Sign Technologies is one of defiance against conventional startup wisdom. While most founders are pushed to prioritize minimum viable products (MVPs) and rapid monetization, Kardos-Nyheim took a different approach. As a 21-year-old law student in the U.K., he assembled a formidable team of researchers from academic and industry powerhouses, including the University of Cambridge, DeepMind, Harvard, and MIT.
Their goal was not to build a chatbot for legal assistants, but to solve the fundamental problem of legal reasoning in large language models. The result was a suite of published papers that vaulted their model into the top tier of global performance, all while operating on a budget that was a fraction of what major labs were spending.
"We were a quieter version of the ‘DeepSeek story,’" Kardos-Nyheim notes. By focusing on novel algorithms and extreme capital efficiency, the team proved that high-quality, foundational science does not require the billions currently being funneled into Silicon Valley’s largest incumbents.
The Funding Chasm: Why Science Lost the Beauty Contest
Despite the technical superiority of their work, the path to funding was anything but straightforward. Kardos-Nyheim recounts a recurring theme during his early fundraising rounds: investors repeatedly prioritized immediate revenue and "bolted-on" products over the raw intellectual property of his research.
In the U.K., his pitch often fell on deaf ears. The local investment community, seemingly risk-averse or perhaps blinded by traditional software-as-a-service (SaaS) metrics, passed on the opportunity. It was only when he looked toward the United States that he found the appetite for long-horizon, research-heavy investment.
This experience highlights a systemic flaw in the current investment landscape: a failure to value the "foundational layer." When investors demand a finished product before they will fund the science, they effectively force founders to cut corners on the very innovations that could create long-term value.
Chronology of a Paradigm Shift
- The Inception: Alexander Kardos-Nyheim identifies a gap in legal reasoning capabilities within existing AI models.
- The Assembly: A team of elite researchers is recruited, united by the scientific challenge rather than the promise of quick commercial gain.
- The Research Phase: Safe Sign Technologies publishes breakthroughs in model architecture, outperforming established models while operating with extreme capital efficiency.
- The Funding Struggle: Multiple rejections in the U.K. highlight a disconnect between investors seeking short-term product-market fit and the long-term needs of deep-tech research.
- The Acquisition: Thomson Reuters acquires Safe Sign in 2024. For the first time in its 170-year history, the firm buys a pre-revenue company, betting entirely on the scientific potential of the technology.
- The Present: Kardos-Nyheim shifts his focus to angel investing, advocating for a return to "foundational" building in an industry crowded with "wrappers."
Supporting Data: The Concentration of Capital
The current state of the AI market is defined by extreme concentration. In the first quarter of 2026, foundational AI startups raised a staggering $178 billion. However, this figure masks a stark reality: approximately 97% of that capital was absorbed by a handful of established giants—OpenAI, Anthropic, and xAI.
For a new founder, the data suggests that the "race is over." The temptation is to become an application-layer company, building "wrappers" atop the APIs provided by these incumbents. Yet, Kardos-Nyheim warns that this is a dangerous path.
"Most application-layer companies, built on a model they do not own, adapt to the pricing and access decided for them by the firms upstream," he explains. These companies are effectively competing for shelf space in a market where the landlord—the foundation model provider—can decide to absorb their entire business model with a single update or pricing adjustment.
Implications: The Death of the Wrapper
The implications for the next generation of AI startups are clear. If you build at the application layer, you are building on sand. If you build at the foundational layer, you are building the bedrock of the future.
Why Deep Tech is the Only Durable Bet
The genuine challenges in AI remain unsolved:
- Interpretability: Understanding how models reach their conclusions.
- Safety: Ensuring that models behave predictably in high-stakes environments.
- Inference Cost: Driving the computational expense down to make advanced reasoning accessible at scale.
- Training Efficiency: Doing more with less, as the era of "infinite compute" budgets inevitably cools.
Founders who focus on these problems are building "moats" that are impossible to ignore. While the "wrappers" will be priced out of the market by the very companies that own the underlying models, the scientific breakthroughs in training and architecture will remain the primary drivers of value for years to come.
Official Responses and Industry Outlook
Industry analysts have noted that the "Safe Sign model"—prioritizing science over immediate commercialization—is gaining traction among a new breed of deep-tech venture firms. These investors are increasingly wary of the "wrapper" trend, acknowledging that the value in the AI stack is shifting downward, back toward the models and the researchers who build them.
Thomson Reuters, in its decision to acquire Safe Sign, signaled that large, legacy enterprises are becoming more comfortable with acquiring "R&D-as-a-service." They recognized that the most efficient way to maintain a competitive advantage in a fast-moving market is to bring the foundational thinkers in-house, rather than waiting for them to launch a commercial product that might eventually compete with their own core services.
Conclusion: Build for the Future, Not the Quarter
For the aspiring entrepreneur, the message from Kardos-Nyheim is both a challenge and a warning: Stop trying to look "fundable" for this quarter.
The venture capital market is famously slow to change and often chases the familiar, yet the history of the industry—from the early days of DeepMind to the explosive growth of the foundational giants—proves that the most important companies always begin as "unfundable" research projects.
The future of technology will not be defined by the superficial applications built to chase current trends. It will be defined by the hard, foundational work that is currently being ignored by the mainstream investment crowd. The founders who have the courage to tackle these deep scientific problems, even when it is unfashionable, are the ones who will ultimately force the market to come to them.
As we look toward the latter half of the decade, the question for any AI founder remains: Are you building a temporary product, or are you building a component of the infrastructure upon which the future will depend? The answer to that question will determine whether you are a footnote in the AI revolution or the force that shapes it.
Azzam Bilal Chamdy 
Lina Hope 
Asep Darmawan 