Axiom: AI Mathematician Scores Perfect on Putnam as Series A Hits $1.6B Valuation
24-year-old founder Carina Hong closes $200M round, builds team from Meta AI veterans and renowned mathematicians including Ken Ono
Seven months after founding Axiom, Carina Hong has closed a $200 million Series A at a $1.6 billion valuation, cementing her position as the youngest founder in the current wave of Neo Labs sweeping Silicon Valley. The company's AxiomProver system achieved the sixth perfect score in the 98-year history of the Putnam Mathematics Competition in December 2025, marking a watershed moment for AI-driven mathematical reasoning. What makes this particularly striking is the nature of the achievement: while human solutions often rely on geometric intuition, AxiomProver generated thousands of lines of Lean code using brute-force enumeration—a fundamentally different approach that nonetheless arrived at correct proofs.
The financing represents a dramatic escalation from Axiom's seed round, which closed at $64 million on a $300 million valuation just months earlier. Hong describes the seed fundraising process as exhausting rather than difficult: "You say the same thing again and again. You get the same questions again and again. I can record it and send it to all of you, right?" The round saw competitive bidding that tripled the initial offer, with B Capital's Howard Morgan—co-founder of Renaissance Technologies alongside Jim Simons and also a co-founder of First Round Capital—ultimately leading. Hong recalls that Morgan displayed more optimism about the business model than she herself held, telling her directly about potential applications she hadn't fully articulated.
From Verve Coffee to Formal Proofs
The origin story is decidedly unconventional for a Silicon Valley unicorn. Hong met her co-founder and CTO Shubho Sengupta at Verve Coffee in Palo Alto, where both were regular patrons. Sengupta, a veteran of Meta AI and one of the early developers of CUDA at Baidu's Silicon Valley AI lab in 2015-2016, worked alongside figures like Dario Amodei (now founder of Anthropic). Their relationship began not with a pitch but with eighteen months of conversations about the history of science. Hong was a Stanford law and mathematics PhD student reading legal cases and watching dogs in the courtyard; Sengupta was a seasoned researcher looking for intellectual engagement outside Meta's walls.
The turning point came during a constitutional law class when Hong realized that if AI could parse the precise language of constitutional interpretation—a textualist approach requiring strict definitional rigor—it should certainly be able to handle mathematical proofs. She had been following her friend Kenny Luo's work formalizing undergraduate mathematics in Lean, the proof verification language, since 2020. "If AI can do law, it's to see something more structured," Hong explains. "Mathematics is a more structural presentation." By September 2024, she had decided to start a business, though she spent two months trying to convince herself not to. "I actually think if I were an undergraduate student I would never drop out of school and go to Y Combinator. I feel like I'm the least likely entrepreneur to start a business."
The Brute Force Player Builds a System
Hong's self-assessment as "not a talented player" but rather someone who achieves through sheer persistence permeates her approach to both mathematics and company-building. At MIT, she struggled with Euclidean geometry problems that were "must-get" questions in math competitions, ultimately solving them through complex number methods that took two to three times longer than geometric intuition would require. "I usually just add each point, then use this complex number method for each line. Just work hard to achieve miracles. In the end I don't need to understand the geometric meaning behind it." This experience shaped her conviction that AI could benefit brute-force mathematicians most: "Maybe it really is how many possibilities do I want to enumerate? How many days do I have to go through to verify a very standard argument."
The technical architecture reflects this philosophy. Axiom's system combines multiple types of AI: reinforcement learning agents for applied AI work, code generation and compiler teams (including core members from Meta's CodeWorld 32 billion parameter model), and pure mathematicians working on Lean's Mathlib. The company deliberately avoided hiring mathematicians until reaching fifteen employees, believing early focus should be on engineering challenges rather than mathematical intuition. When they did bring on mathematicians, it was with specific intent: Professor Ken Ono joined to create adversarial benchmarks, constantly testing where the system breaks rather than celebrating successes.
Hong draws a distinction between Axiom's approach and competitors who rely heavily on Monte Carlo Tree Search, the method used in DeepMind's AlphaProof. "We think this is too expensive. Let's see how much money we have. We don't think we can do this Monte Carlo tree." Instead, the team explored sub-agents and skills-based approaches, finding unexpected transfer learning between mathematical proof and code verification. On the Verina Benchmark for code verification, AxiomProver achieved results that "surprised us all," demonstrating that strong theorem-proving ability translates directly to code validation capabilities—a critical insight for future business applications.
Ken Ono's Unexpected Email
The recruitment of Ken Ono, a 57-year-old tenured professor at the University of Virginia and prominent number theorist, generated significant attention when announced. The reality was more organic than dramatic. Ono had supervised Hong during an REU (Research Experience for Undergraduates) program when she was a waitlisted undergraduate, eventually collaborating on multiple papers. In November 2024, he sent an email saying he was coming to the Bay Area and that if he joined OpenAI or DeepMind, he hoped she wouldn't be upset. Hong's response was immediate: "I said you can come to my place."
What drew Ono to choose Axiom over OpenAI came down to focus and culture. "He thinks our team is more mathematical," Hong explains. "If you come to us, maybe there is more of an atmosphere in mathematics. The DNA and focus of this company is mathematics, rather than a general AGI where math is a part of it, maybe more of a marketing component." Ono's role extends beyond research; Hong describes him as having "a high school basketball coach" personality—relentlessly optimistic, team-oriented, and skilled at making others believe difficult goals are achievable. His presence also attracts other mathematicians who might otherwise be skeptical of AI's role in their field.
The question of whether AI threatens mathematicians' livelihoods gets a nuanced answer from the Axiom team. Ono and others increasingly believe that as AI handles more proof work, human mathematicians will "learn at different levels of abstraction," focusing on higher-order questions about what problems are worth solving and how to structure theories. Hong points to an analogy: programming evolved from punch cards to low-level assembly to Python to natural language coding. Mathematics may follow a similar trajectory, with mathematicians working at increasingly abstract levels while AI handles mechanical verification and even some proof generation.
The Math Benchmark Crisis and Research Challenges
Axiom's rapid progress created an unexpected problem: benchmark saturation. The team took mini-f2f (high school level) to saturation in roughly two months, then did the same with Putnam. "We are relatively short of benchmark sets at the research level, to tell the truth, because there aren't many in research-level mathematics," Hong admits. The famous FrontierMath challenge poses different issues—many questions can't even be formalized in Lean because requisite definitions don't exist in Mathlib, the core mathematics library. Hong tried to participate in that challenge but couldn't convert the problems to Lean: "The question involves many definitions. These definitions are not at all in this library."
This points to a deeper technical challenge that Hong identifies as potentially more difficult than proof itself: auto-formalization, the conversion of natural-language mathematics into formal Lean code. "If I take this human mathematics that's already been discovered and written, and convert it into formalized data, you will get less compliment than if you prove something. But this technology is actually harder than proof. At least the same level of difficulty. I actually think it's more difficult." The process requires creating detailed blueprints—taking a 20-page paper and expanding it to 200-500 pages of granular proof steps that can then be formalized. Terence Tao and others have done this manually for large collaborative projects, but automating it remains an open problem.
The company's current technical priorities involve scaling inference through sub-agents rather than relying on computationally expensive tree search. Hong describes an "aha moment" when the team scaled proofs from 40-node trees to 4,000-node trees, demonstrating that "learning from experience"—treating each problem-solving trajectory as training data for skill acquisition—could work at scale. Another critical insight involved the relationship between mathematics and code: "AI's past work good at math sounds very silly to talk to people like this, but math to code generation, from math to chip verification for these codes, these are obviously transformed." The implication is that Axiom's core technology isn't just about mathematics—it's about verified reasoning applicable to any domain with formal specifications.
Business Model and the Validation-First Strategy
Hong is disarmingly frank about business model uncertainty. During the seed round, she told investors: "Our business model is not very certain. I have been a quantitative trader myself. I don't know if this thing is quantitative trading. I think maybe not. I said at that time that as a business model it must not be a final business model. This is just an example of an AI that is better at mathematics being useful." The candor reflects her background—she nearly joined Bridgewater Associates before the pandemic redirected her to Professor Ono's REU program. Quantitative finance represented a clear application, but Hong believes it's not the end state.
The company has settled on validation as the first market: chip verification, circuit verification, and code verification where formal methods currently rely on expensive, labor-intensive SMT solvers and tools like Cadence Jasper. Amazon's automated reasoning team spent three to five years writing 260,000 lines of code to verify a single hypervisor memory isolation component. "AI does not improve these engineers' lives," Hong observes. "But in these fields, you will have better pricing power." The vision extends to a future where any programmer can request a function with a guarantee that it works correctly, no test cases required—though Hong acknowledges this is "5-10 years" away and requires solving the specification problem, which remains extremely difficult.
Despite the validation focus, Hong maintains that AI for Math is fundamentally a moonshot with binary outcomes. "Either the moon landing is successful or the moon landing fails. Yes, the moon landing was successful, moon landing failed," she says, explicitly comparing Axiom to SpaceX. "There is no possibility in the middle." This creates tension between long-term research ambitions and near-term execution pressure. The Series A's high expectations could push the company toward incremental commercialization, but Hong believes the team's DNA demands otherwise: "If we just pursue profit-seeking and commercialization too much, we will actually be out of balance. We can take a look at our competitor companies. In the early days they talked more about scientific ambition. We think we are ideal realists and realistic idealists. Both are good."
Competition, Resources, and the Neo Lab Phenomenon
Axiom exists in a peculiar competitive position. The primary competitor, founded two years earlier with five times the initial funding, reached IMO level (solving 5 of 6 problems) after two years of work. Axiom achieved Putnam perfection in four months. Yet the company operates with just 30 people compared to the competitor's 50-75, and raised what Hong considers insufficient capital for their ambitions: "Everyone's common thought is that we underraised. We didn't raise enough for our seed round." The resource constraints forced strategic choices—no pre-training, focus on post-training and system design, and aggressive recruiting of talent willing to accept startup compensation over the million-dollar packages offered by established AI labs.
The recruiting challenge is existential. In early 2025, most AI talent had six competing offers, and Hong lacked leverage: "I have 6 offers on hand. I became that person who has to dance. Maybe you can't be Peter Thiel." She describes recruiting decisions as the most painful aspect of entrepreneurship, situations where "you may think there is a conflict between people and things, but actually you are not good enough." The company's bottoms-up culture—inherited from the Facebook veterans who form much of the early team—helps attract talent who want research freedom rather than top-down direction. Hong deliberately cultivates this: "If you want to know what this means, you can go get one LLM, go give him some data. I think our National Foundation needs to give funding to pure mathematics because mathematics is an ecosystem."
The broader Neo Lab wave provides context for Axiom's emergence. Hong views it as a reaction to structural inefficiencies in large organizations: "Competition leads to mediocrity will result in not having some long-term ideas. He has to come out and do it himself, do Neo lab. The birth of Neo lab is because of curiosity and creativity—it is our basic human need." She believes large frontier labs face a fundamental tension: they'll continue pushing general capabilities while specialized startups can move faster on vertical problems. OpenAI's approach to AI for Math focuses on informal natural language proofs; DeepMind has multiple competing teams working on both formal and informal approaches; Anthropic uses math reasoning to improve general reasoning scores rather than making it a core focus.
The Technical Frontier: From Proof to Conjecture
Hong articulates a taxonomy of AI for Math capabilities that extends beyond proof. "You need to have a very good prover. You need to have a very good conjecture system that can ask mathematical questions. The reward of this AI is if I can take this prover to make the reward signal for this conjecture." The conjecture problem remains largely unsolved because unlike proofs—which have binary verification (proven or not)—conjectures lack clear reward signals. How does an AI know if its conjecture is interesting, important, or meaningful versus trivial? The current state-of-the-art uses crude heuristics like proof length relative to statement length as an "elegance filter," but this breaks down beyond high school mathematics.
A second major challenge is library learning: having AI not just prove theorems within existing definitions but create new definitions and build theory. "Let this machine build a library from the definition, including theorems above, including proving these theorems, then derive new theorems and add more definitions. This process is a process of architectural theory. This is quite difficult." The difficulty lies in verification—how can you ensure a machine-generated definition doesn't create logical inconsistencies that manifest as paradoxes hundreds of steps later? In a hypothetical future where all human mathematics has been formalized in Lean, this becomes critical: "If that definition is one a human being does not have, a definition that has never been written, how can I guarantee these definitions of this system do not conflict with each other?"
The relationship between formal proof in Lean and natural language reasoning creates fascinating technical tensions. Hong frames it as "math is code and code is math" through the Curry-Howard correspondence, where every mathematical proof can be converted to a computer program. But this translation isn't symmetric—converting English to Lean is far harder than converting Lean to English, because the model has seen vastly more English tokens. Axiom uses cycle consistency (translate to Lean, translate back, check differences) to validate conversions. Hong believes language models can generate good conjectures through their ability to break boundaries, but verification must happen in the formal space. "Language can make a good conjecture. Then to a certain extent, generation is generation, then verify—these two go back and forth together."
The Mathematician's Identity Crisis
Hong's personal journey reflects broader questions about what it means to be a mathematician in an age of AI-assisted proof. At MIT, she felt like "the stupidest one in the whole math department" surrounded by IMO gold medalists whose news she had followed since childhood. Her response wasn't despair but acceptance: "My expectation is that I am the stupidest one, so I actually adapt very well. I feel lucky every day I can talk to these people, take class together, do homework together." This framing—grateful apprentice rather than competing genius—shaped her approach to both mathematics and company-building.
The question of whether she wants to be remembered as a mathematician draws a surprisingly philosophical answer. "I used to think it was important. I used to cry sometimes. I am still a very emotional person. Then later I felt, actually, I want to be an apprentice. I just want to go to as many intellectual fields as possible and learn as much as possible." She describes jumping from math to physics (to remedy being "very bad at physics in high school") to neuroscience to AI to quantitative trading to law, each transition driven by curiosity about unfamiliar domains. "This might be a good thing—how can I let one apprentice, an entrepreneur's apprentice, not die? 99% of startups will die. Why am I now just don't die?"
The dinner-with-a-mathematician question elicits a telling response: Erdős, because "the descriptions he told you, after all, it is combinatorics, you can understand what you have for dinner on the table." Erdős represents mathematics as social practice, the itinerant collaborator with hundreds of co-authors and peculiar personality. Hong contrasts this with Grothendieck, whose dinner would be "of great significance" but whose abstract algebraic geometry might be wasted on someone without the requisite background. The implication is that mathematical value exists not just in theorems but in the community and culture surrounding them—something AI might augment but cannot replace.
The Addiction to Suffering and the Pendulum of History
When asked about her self-reward mechanism, Hong describes a shift from deriving energy from small teams and camaraderie to drawing it from the work itself. "I think most founders are addicted to suffering. Yeah, that's right. Addicted to Misery." She frames this as necessary rather than noble: "Usually these conversations are what you end up choosing—selected investors. They have some very exciting conversations. But from these large numbers of rather boring process, there were some very exciting conversations." The venture capitalists who sought "chip on the shoulder" founders—people with past scars converted into relentless drive—found in Hong someone whose entire academic career was a series of perceived failures transformed into motivation.
This manifests in how Axiom handles challenges. When the AxiomProver initially couldn't solve one Putnam problem that required geometric insight, Professor Evan Chen (coach of the U.S. IMO team) drew a picture, and everyone in the room immediately saw the solution. The AI eventually found a completely different proof through exhaustive enumeration. "I think comparing humans and AI, the math behind the solution we got, it's a very interesting process," Hong reflects. The incident crystallized a core tension: intuitive mathematicians provide elegance and insight; AI provides exhaustive reliability and verification. Both are necessary.
Looking forward, Hong frames AI for Math as enabling a transition from "math poor to math rich society." Supply will explode. But she rejects the framing that this makes mathematics less special: "I think mathematicians to a large extent, to a large extent, the significance of our mathematicians lies in solving problems. They will provide 0.01% intuition about what math problems are more worthwhile for us to concentrate our computing power to solve." In a world of infinite computing power, this resource allocation role becomes less critical—"we're going to do this, we want to do everything that human beings can think of, all curious math problems solved." But in practice, computing power remains finite, and human judgment about what's important versus trivial remains irreplaceable.
The final philosophical question—what if AI makes mathematics boring by solving everything—prompts Hong to return to human nature. "You give me one with 1 million lines of Lean code, I must see how it is made inside. You won't let me see it, I will go and see it. This is the embodiment of human curiosity." She points to examples like Stefano Ermon making diffusion models faster, or Periodic Labs bringing top AI talent to messy real-world materials science problems, or Recursive working on hardware-AI co-optimization. "Behind this is a basic force. This basic force is human curiosity. Some people say it is a bubble now, or should we say it's the moon landing now? Some companies, once he becomes him, he will go to the moon. Then there are companies that go to the moon—they really work hard, everyone is really 996 every day working towards that dream together. It didn't work out. That's the bubble."
Axiom's bet is that formal verification, recursive self-improvement, and a systems approach to reasoning will create a defensible position in a market where DeepSeek can commoditize base models overnight. The company positions itself as "more like SpaceX" than a typical AI lab—tackling engineering challenges with huge technical barriers that extend beyond training bigger models. Whether this bet succeeds determines not just the company's fate but potentially the timeline for when machines can contribute meaningfully to humanity's frontier of mathematical knowledge. Hong seems comfortable with the binary nature of this outcome, perhaps because her entire trajectory has been a series of improbable successes that looked like failures in progress. As she puts it: "There is always a way. You have to believe it."