There's something almost quaint about a seven-person startup in New York handing $10,000 to each of the world's most advanced AI models and watching them try—mostly unsuccessfully—to make money trading cryptocurrency.
But that experiment, conducted over several weeks last fall, became the foundation for a $15 million bet announced May 15, 2026. Nof1, the research lab behind what it calls Alpha Arena, has secured funding co-led by SUI Group Holdings and Karatage Opportunities, convinced that the very public struggles of OpenAI's, Anthropic's, and Google's flagship models reveal something crucial: even the smartest general-purpose AI isn't built for markets.
And if you're wondering how badly things went—across 32 autonomous trading runs conducted between October and November 2025, only six finished in the black.
The Unforgiving Math of Perpetual Contracts
Alpha Arena gave each model a real stake: $10,000 and access to cryptocurrency perpetual contracts on Hyperliquid, a decentralized exchange. The focus was Bitcoin, Dogecoin, and Solana. No simulations, no paper trading—actual capital at risk.
Some models showed flickers of competence. Grok, xAI's chatbot, demonstrated early promise during Season 1, which ran from October 17 through November 3. So did DeepSeek, the Chinese upstart that has rattled the AI establishment with surprisingly capable performance at a fraction of the training cost. But promising isn't the same as profitable, and the broader pattern was bleak. Most models, for all their linguistic sophistication and reasoning prowess, couldn't consistently navigate the volatility and execution demands of real-time trading.
The results seem to have crystallized a thesis for Jay Azhang, Nof1's founder. Azhang previously managed a small fund that grew from $3 million to $20 million in assets—a respectable run in an industry littered with flameouts—and he's now positioning Nof1 as the first AI research lab dedicated exclusively to financial markets. The pitch, essentially: large language models are generalists. Markets demand specialists.
Building From the Ground Up
Rather than continue tweaking prompts and hoping frontier models improve, Nof1 plans to train its own. The new capital will fund Season 2 of Alpha Arena, this time featuring custom-built agents equipped with web search, extended reasoning time, and what the company describes as multi-step execution capabilities. It's a departure from the plug-and-play approach of Season 1—an acknowledgment, perhaps, that off-the-shelf intelligence isn't enough.
Beyond Season 2, the roadmap gets more ambitious. Nof1 intends to launch a consumer platform offering "the world's first coding agents for markets," according to the company's announcement. The phrasing is bold, though details remain vague. What exactly that looks like—whether it's automating strategy development, backtesting infrastructure, or something else entirely—isn't yet clear. But the bet is straightforward: specialized models, trained from scratch with financial data and market structure embedded in their architecture, will outperform general-purpose systems in trading contexts.
It's worth noting that this isn't an entirely new idea. Quantitative hedge funds have been using machine learning in various forms for years, though typically not the transformer-based large language models that have dominated recent AI progress. Whether Nof1's approach can bridge that gap, or whether financial markets simply present a different kind of challenge than the ones LLMs have mastered, remains an open question.
The Money Behind the Bet
SUI Group Holdings, a publicly traded firm on the Nasdaq under the ticker SUIG, put in $3 million of the $15 million round. The investment fits within what the company describes as an "agentic finance" strategy linked to the Sui blockchain ecosystem—a Layer 1 platform that's been positioning itself as a faster, more developer-friendly alternative to Ethereum. SUI Group has indicated it plans to deploy Nof1's models to generate yield for its treasury once they're operational, effectively making itself an early customer.
Karatage Opportunities, which co-led the round alongside SUI Group, shares some notable personnel overlap. Marius Barnett and Stephen Mackintosh, Karatage's co-founders, also serve as SUI Group's board chair and chief investment officer, respectively. The press release notes that independent directors at SUI Group reviewed and unanimously approved the investment—a disclosure that acknowledges the potential for conflicts but doesn't entirely dispel them.
AlleyWatch estimates Nof1 has now raised $25 million in total equity funding, though details about earlier rounds remain sparse. The company was founded in 2023, which means it's been operating for roughly two years with what appears to have been modest initial backing before this larger infusion.
What Comes Next
There's a certain irony to the setup. The AI boom of the past few years has been driven largely by the success of models that can do a little bit of everything—write essays, debug code, answer medical questions, generate images. Nof1's bet is that markets are different. That the volatility, the adversarial nature of trading, and the sheer complexity of execution require something more targeted than a system trained to be helpful, harmless, and polite.
Whether that thesis holds depends on what Season 2 reveals. If custom-trained models can consistently outperform the frontier systems that stumbled through Season 1, Nof1 will have demonstrated something valuable. If they don't—well, the lesson might be that markets remain elusive regardless of the architecture.
For now, the lab has capital, a clear target, and a public benchmark that showed just how hard the problem is. And perhaps that's the point. Sometimes the most useful thing a startup can do is prove what doesn't work.
