Arena CEO: There Will be a $100BN US Open-Source Model & Data is a Trillion Dollar Market

Why does it matter? Because the AI threat is already inside your building.

The reassuring story about Chinese AI — they're just distilling us, they're hardware-constrained, self-hosting neutralizes the risk — is collapsing on all three fronts simultaneously. Anastasios Angelopoulos, founder and CEO of Arena, dismantles each assumption in sequence, with data from the evaluation platform that sits at the center of the global model ecosystem.

• Kimi K3 beat every American frontier model on a meaningful subset of tasks, proving China has moved beyond distillation to genuine original innovation
• Self-hosting a Chinese open-source model doesn't neutralize the backdoor threat — the attack vector is baked into the training, not the API call
• AI-generated ghost candidates are already passing live technical interviews at top AI companies and vanishing when hired
• Data only becomes obsolete at AGI — making it structurally more defensible than GPUs or algorithms, and worth at least $100B by 2030

AI-generated ghost candidates are passing live technical interviews at top AI companies — and senior engineers can't tell the difference

The fake candidate sat through live technical interviews with world-class Arena engineers, impressed everyone in the room, and then ceased to exist the moment a hiring decision landed. This is not a story about resume spam. "Some guy looks perfectly normal. They're passing all of our technical interviews. They're like such an amazing blah blah blah. And then what happens at the end of it? You try to hire them and it's vaporware. Person doesn't exist."

Anastasios is categorical: this is happening across American businesses, not just Arena. The motive varies — data access, corporate espionage, nation-state intelligence gathering, double-pay schemes — but the attack pattern is consistent. Get through the interview process before anyone verifies you're human, then either extract what you can during the process or, if somehow onboarded, gain inside access.

Arena's response: requiring all onboarding to happen in person. If you want a company laptop, you come to the office. You shake hands. You get verified as real. Figma has done the same.

Any company granting system access without physical verification has an open door to an attack vector that already works, at scale, against some of the most technically sophisticated hiring teams in Silicon Valley.

Kimi K3 didn't just beat American models — it destroyed the story that kept the US complacent

The American reassurance narrative on Chinese AI rested on a single load-bearing assumption: they're only keeping up by distilling our models. Kimi K3 collapsed it.

"It violates a narrative that has been persistent in the United States, which is that the Chinese are just distilling American models. And that's the only way that they're able to keep up." Kimmy didn't just keep up — "Kimmy actually beat all American models, including Fable, in some subset of tasks." Front-end coding, specifically — the domain of a huge fraction of working developers worldwide.

This doesn't prove China has abandoned distillation. It proves distillation is only part of the story. Something else is happening inside those labs, above and beyond copying, that's lifting performance past what American labs are currently producing. On Arena's leaderboards today, the top American open-source model sits at position ten. Nine Chinese models rank above it.

The hardware constraint is real but not permanent: export controls might be hindering them now while simultaneously incentivizing them to build their own chip ecosystem. If that succeeds, the US loses its most structural advantage.

Treating Chinese labs as second-class competitors who can't innovate is no longer defensible. The lead is real; the structural advantage is not.

Hosting a Chinese model on your own servers doesn't neutralize the backdoor — the attack is in the training

Security teams that signed off on Chinese open-source model deployment because it's "self-hosted" built their strategy on a misconception.

The threat isn't in the API handshake. It's upstream, in the training run. Walk through the attack: a chatbot with access to all your internal company data, running cleanly on your own infrastructure, trained somewhere you don't control. "What if the other side that's interacting with the chatbot can build in a certain code word or a certain character sequence that then jailbreaks that model and gets it to reveal all the data to me?" One trigger phrase — and the model dumps everything it knows. "That is totally something that you can build into a model and have companies host it on their own infrastructure. It's an attack vector."

The only real defense is AI watching AI. Guardian models — running in parallel with every agentic deployment, monitoring traces in real time, smart enough not to be outsmarted by the agents they're watching. "We're going to need AI to be guarding AI because humans are going to be too slow to do that."

A self-hosting policy is not a security architecture. It's a liability.

Data becomes obsolete only at AGI — it's a more durable asset than GPUs, and the market will hit at least $100B by 2030

A trillion-dollar market hiding in plain sight, consistently mispriced as a commodity.

"In order for data to become irrelevant, humans need to become irrelevant — and that means we've achieved AGI." That framing is why data is structurally unlike every other layer of the AI stack. GPUs commoditize as fabs scale. Algorithms commoditize — "people know how to use the transformer," which is why former frontier-lab founders all look the same walking in the door. Data doesn't commoditize, because sourcing it is the hardest, dirtiest, most labor-intensive part of model training. Nobody wants to do it, which is exactly why it's defensible.

Frontier labs currently spend roughly 10 to 20% of their GPU budget on data. Believe in GPU market growth and you have to believe in data market growth as a mathematical consequence. Anastasios puts the floor at "at least a hundred billion dollars by 2030, if not a trillion."

Revenue concentration objections — the knock that these companies only sell to OpenAI, Anthropic, and Meta — he dismisses flatly: "TSMC has revenue concentration. There's businesses that are many hundreds of billion dollar public market businesses that have revenue concentration." The real story is what comes next: enterprise expansion. Every business that trains its own model needs its own data. That market hasn't started.

Software moats are dead, so enterprises will fine-tune their own models on proprietary data — and someone will build a trillion-dollar American open-source company from that demand

The logic is tight, even if the company that captures it doesn't fully exist yet.

In an AI world, software can be produced instantaneously. "Software is no longer really a moat because it can be produced instantaneously." What remains: network effects and data. A Coca-Cola or a Cisco — massive proprietary data, no claim to frontier AI — will take an open-source model, fine-tune it on their corpus, and own their stack end to end. "The business incentives make this inevitable." They won't hand that data to an external third party that might compete with them one day.

The business model for whoever captures this is taking shape: revenue sharing with inference providers like Fireworks or Together when they cross a revenue threshold, plus the AI modernization wave — helping every enterprise retool, restructure its data pipelines, and integrate models into workflows. "That is going to be massive massive massive," he says, "and that is another way for them to become multiundred billion or trillion companies."

Anastasios has been vocal about Thinking Machines as a candidate. Mistral is in the conversation. The trillion-dollar American open-source company doesn't exist yet. The structural demand for it already does.

75+ NeolLabs exist and two-thirds will be worth nothing — the first round is protected, the next round is where it ends

"There's at least 75 Neilabs for sure. Like 2/3 of those are going to be worth nothing or like they're going to be bought out for parts."

Early-round investors aren't panicking, and the math explains why. Write $200M at seed into a team from a top frontier lab, the floor is an acqui-hire in the $500M to $1B range. It's nearly a free option. The risk is acceptable, which is why multi-billion-dollar valuations with zero revenue keep closing.

The second round is a different calculation. A $10B NeolLab valuation needs roughly $4 billion in revenue within two to three years to justify a 25-30x multiple to a hundred-billion-dollar outcome. Most of these companies have no revenue and no credible path to it. "Not enough just to like create a model and then have a party about it. Hey, we created an AI. That is like old news. Today it's about not just can I create a model, but do I have a sustainable business model around that?"

When evaluating a NeolLab at round two, ignore the benchmark score. Ask what specific, funded path to $4B in revenue exists within three years. Most of them don't have an answer.

Frontier labs are moving up the application stack — but stealing Harvey's legal GTM would take Anthropic a decade

Claude Design is already eating into Figma. The direction is unambiguous: as inference commoditizes, model providers migrate up toward the application layer to own more of the value chain and escape margin compression. Harvey and Lagora are watching their most resourced potential competitor build above them.

But the threat is not symmetric across verticals. Design tools are pick-up-and-go — a designer grabs the tool, the value is immediate. Legal is a different organism. "You've got to go into Clifford Chance or any of the big firms, build relationships with 50-year-old white male partners who want to play golf and be told that they're great... do deployment to junior lawyers who don't want to use you because they think you're going to take their jobs. The deployment in the GTM is the heavy lifting and that's real work."

Anthropics' appetite is real, but focus is finite. "Priority number 12 for Anthropic is probably not high enough for Harvey and Lagora to be too scared." The moat for vertical AI companies isn't the model — it's the GTM complexity that a frontier lab optimizing for a thousand other priorities won't replicate on any near-term timeline.

Every defensible position in AI that seemed structural is proving not to be — except the hard, unglamorous ones

Closed-source dominance: contested. "China only distills": gone. Self-hosted safety: a misconception. Software moats: dead on arrival. What's left standing is the stuff nobody wanted to build — dirty data pipelines, in-person hiring verification, decade-long legal GTM relationships. The durable assets in this cycle are the ones that require real operational investment, not the ones that felt like elegant abstractions.

The companies and investors that get this right will be the ones who stopped waiting for a clean, software-like margin structure and started building the unglamorous infrastructure underneath the models. Everything else is renting someone else's moat.

AI sovereignty isn't a buzzword. It's what you're left with when every other option disappears.

---

Topics: AI models, open source AI, Chinese AI, Kimi, model evaluation, Arena, data market, AI security, enterprise AI sovereignty, NeolLabs, AI backdoors, frontier labs, model commoditization, AI hiring, deepfake candidates