From HRT Interns to Rivals: How Chinese-American Founders Came to Control AI

On October 19, Alexandr Wang posted a photo on X of three former colleagues gathered for the first time in years. The caption — "got the HRT gang back together!" — referred to Hudson River Trading, the New York quantitative firm where Wang, Cognition CEO Scott Wu, and Hyperliquid founder Jeff Yan had all worked as interns about a decade ago. The image looked like a reunion. The timing suggested something else.

Six days earlier, Wang's team at Meta Superintelligence Lab had released Muse Wonder Full, a frontier-class reasoning model built for software engineering tasks. Meta has declined to make it available through its public API. Instead, the model ships exclusively inside Muse Code, Meta's terminal-based AI coding agent — whose primary commercial target is Devin, the autonomous software engineer built by Cognition, the company Scott Wu now leads.

The photo is a small human detail inside a structural story about the AI industry's new power map. Follow any major segment of the AI stack — foundation models, coding agents, customer service automation, AI search, data infrastructure, semiconductors — and the founder or CEO of a leading company is either Chinese-American or the child of Chinese immigrants. More striking, in many cases they know each other personally, having come through the same recruiting channels, the same quantitative trading firms, and often the same early employers.

Meta's Frontier Model Strategy Locks Out Competitors — Including One Its Own Chief Helped Build

Wang joined Meta in mid-2025, when the company acquired a 49% stake in his data-labeling company Scale AI for $14.3 billion and recruited him to run a new internal unit called Meta Superintelligence Lab. Scale was the company where Scott Wu had worked as a founding engineer after leaving Harvard, and where Cognition's CTO Steven Hao spent four years as an early engineer. Hao left Scale in May 2022; by the end of that year, he and Wu had founded Cognition.

On September 8, Cognition announced a Series E funding round of more than $2 billion at a post-money valuation of $48 billion. Andreessen Horowitz and Accel led the round; Founders Fund, General Catalyst, and Avenir participated. Annualized revenue was running close to $900 million — up from $492 million at the spring funding round. Cognition's named enterprise customers include Citigroup, GE Aerospace, Mercedes-Benz, NVIDIA, and Modal.

That was the same day Meta's consumer AI assistant, Muse, launched to U.S. users with an agent capable of sending email, booking travel, and making online purchases.

The coding-agent confrontation hardened further on October 12–13, when Meta released Muse Wonder and Muse Wonder Full. On the Coding Agents Index published by AI benchmarking service Artificial Analysis, Muse Wonder Full scores 68 — matching Claude Sonnet 5.5 running inside Claude Code. Unlike Muse Wonder, which is available through Meta's public API at $4 per million input tokens and $20 per million output tokens, Muse Wonder Full carries no per-token price. It ships only to subscribers of Muse Code's Power ($50 per month) and Maximum ($200 per month) plans. Developers building products outside Meta's ecosystem cannot license the more capable model.

This restriction matters because Devin's competitive position has always rested partly on the assumption that Cognition could buy access to whatever frontier model best suited a given task. In a June 2026 podcast, Scott Wu described Cognition as aiming to be the "Switzerland" of the coding-agent market — a neutral orchestrator that routes tasks across OpenAI, Anthropic, Google, and open-source models, optimizing cost and quality for each job. That strategy depends on neutrality upstream. When the most capable tool is locked inside a direct competitor's product, neutrality stops working.

Read more: Meta releases Muse Wonder, its first frontier AI model built for long-running agents

Cognition's SWE-1.7 model, which powers Devin's most demanding tasks, is built on Moonshot AI's Kimi K2 foundation model, with additional reinforcement learning training applied by Cognition. The dependency on a third party's base highlights the exposure. Devin also uses Claude Fable 5.1 and OpenAI's GPT-6 Astra for specific subtasks, but the tightest coupling in the stack — the model that handles the most complex software engineering — runs on a Chinese foundation model that Cognition itself does not control.

The Talent Pipeline That Built Both Sides of This Competition

Cursor, acquired by SpaceX for $60 billion in August 2026, built its most recent self-trained model, Composer 2, on top of Moonshot AI's Kimi K2.5 — licensed and hosted through Fireworks AI, an inference platform founded by Jiao Lin. Lin previously led PyTorch infrastructure at Meta and managed more than 300 engineers. Fireworks AI has reached a valuation above $15 billion, with customers that include Cursor and Perplexity.

These overlapping connections — shared former employers, shared model suppliers, shared investors — are not coincidental. They are the output of a specific talent pipeline.

The pipeline's first stage is competitive mathematics and computer science. Johnny Ho, Perplexity's co-founder and chief strategy officer, scored a perfect 40 on the 2012 International Olympiad in Informatics. Scott Wu won multiple national math olympiad medals. Jensen Huang, Nvidia's CEO, placed in mathematics competitions as a student in Taiwan. Industry observers have noted that Chinese educational institutions have disproportionately produced the olympiad-trained talent now defining the AI industry's leadership.

The pipeline's second stage is quantitative trading. Jane Street, Citadel, and Hudson River Trading have spent years recruiting directly from olympiad competitions. When Wang, Wu, and Yan interned at HRT around 2016, the firm was running one of the first formalized programs for competition math talent; Citadel has said it has had between 20 and 30 interns per year who competed in olympiads. The trading desks became the primary bridge between elite mathematics education and technology leadership, offering salaries and intellectual challenges that alternative career paths could not match.

The third stage is the AI lab. A 2024 MacroPolo analysis found that more than a third of NeurIPS paper authors held undergraduate degrees from Chinese universities, and that roughly three-quarters of them worked at U.S. institutions. Reporting on this cohort has documented the pressures that come with their visibility: 16-hour workdays, extreme compensation volatility, and some founders declining interviews out of concern that public identification as Chinese could affect investor or customer relationships.

The Data Layer Reshaped by Wang's Departure

Scale AI's position in the data-labeling market was also restructured by the Meta transaction. Before the deal closed, Scale's customer list included Google and OpenAI, two of the three largest model developers in the world. Neither company wanted its training-data requirements visible to a platform now co-owned by Meta. Google and OpenAI both shifted significant data orders to outside suppliers.

The primary beneficiary was Surge AI, founded in 2020 by Edwin Chen. Chen had previously held engineering roles at Google, Facebook, and Twitter. He built Surge without venture capital, on the explicit reasoning that outside funding distorts incentive structures and inflates headcount. According to Forbes, Surge generated approximately $1.2 billion in revenue in 2024 with a staff of roughly 110 employees, while Scale — with more than 1,000 employees — reported $870 million in revenue for the same year. Chen's roughly 75% stake in Surge placed him on the Forbes 400 in 2025 with an estimated net worth of $18 billion.

The Chip Tier: Four Chinese-American CEOs Supply Everyone Else's Compute

While the application and data layers feature direct competition between Chinese-American founders, the infrastructure layer presents a different structure: four Chinese-American executives lead the companies that supply AI compute to the entire industry, including to each other's competitors.

Jensen Huang co-founded Nvidia in 1993, when it was one of dozens of hardware startups attempting to stake a claim in the emerging consumer graphics market. The company's data-center revenue reached $89 billion in its most recent fiscal quarter. Lisa Su has led AMD since 2014; the company's market capitalization crossed $1 trillion on September 21, 2026 — a 187% gain in the year — while still trailing Nvidia's roughly $5.45 trillion. Hock Tan, who leads Broadcom and was born in Penang, Malaysia, generated $16.7 billion in AI semiconductor revenue in fiscal Q3 2026, a 221% year-over-year increase, driven by custom accelerators built for Google's TPU program, Meta's MTIA chip, and a co-development agreement with OpenAI. Custom accelerators represented 73% of Broadcom's AI business. Tan has set a target of $115 billion in AI revenue for fiscal 2027.

The fourth member of the group is Lip-Bu Tan, appointed Intel's CEO in March 2025 after Pat Gelsinger's departure. Tan's tenure was briefly destabilized in August 2025, when President Trump demanded his resignation over prior investments in Chinese semiconductor companies. Tan went to the White House. The outcome: the U.S. government announced an $8.9 billion equity investment in Intel representing a 9.9% stake, and Nvidia separately committed $5 billion.

The chip-layer dynamic creates an unusual incentive structure. Both Meta and OpenAI have signed large procurement agreements with Nvidia while also contracting for custom accelerators. OpenAI signed a 10-gigawatt capacity agreement with Nvidia in September 2025 and a 6-gigawatt custom chip agreement with Broadcom in October 2025. Meta replicated both approaches in February 2026, agreeing to purchase millions of Nvidia Blackwell and Rubin GPUs and signing a matching 6-gigawatt agreement with AMD with terms nearly identical to OpenAI's AMD arrangement, including the same maximum of 160 million AMD share warrants. Each major buyer is effectively hedging the same group of Chinese-American CEOs against each other.

What the PayPal Mafia Analogy Gets Wrong

The analogy to the PayPal Mafia — the group of founders who built Musk's eBay-acquired payment company and later spread across social media, aerospace, venture capital, and hedge funds — comes up regularly in profiles of this cohort. It is instructive mainly for where it breaks down.

The PayPal generation benefited from what one might call the open-range period of the early internet: the relevant product categories (social networking, video hosting, consumer reviews, professional networking, payments) were each large enough to sustain independent billion-dollar companies, and they did not structurally conflict with each other. Reid Hoffman introducing Peter Thiel to Mark Zuckerberg cost Thiel nothing; an investment in Facebook did not compete with Founders Fund's position in SpaceX. When the PayPal group experienced internal conflict — most visibly when the board removed Musk as CEO while he was mid-flight on a honeymoon — the resolution came through reinvention. Thiel went to venture; Musk built rockets; others built social platforms and professional networks. Each person found open territory.

The AI industry as it has developed offers narrower territory. Foundation models, coding agents, enterprise customer-service automation, AI search, data labeling, and specialized inference infrastructure are the six categories that have generated valuations above $10 billion in the current cycle. The capital is heavily concentrated in those six segments, and the segments are structurally linked: whoever controls the best foundation model has leverage over every downstream product that depends on API access. When Meta's most capable coding model becomes available only inside Meta's coding agent, the coding-agent category shrinks from six viable players to five.

Scott Wu's "Switzerland" framing, offered just four months ago, assumed that no model supplier would foreclose access to a competitor. The Muse Wonder Full release answered that assumption directly. Cognition's response has been accelerated self-reliance — training SWE-1.7 on top of Kimi K2 rather than relying on API access to a competitor's frontier model. The $2 billion Series E provides the runway to continue that investment.

The One Exit Ramp

In the reunion photo posted by Wang, the third figure is Jeff Yan, who did not join the AI market's main battlegrounds. Yan has built Hyperliquid, a decentralized perpetuals exchange, in a market largely distinct from the AI industry's capital concentration.

Yan once described his relationship with competitive failure plainly: "Losing is common. Most people lose. Usually only one person wins." The AI product stack may be structured to prove him right. With hundreds of billions of dollars committed to building the infrastructure and applications of a single technology cycle, and with fewer than ten categories capable of sustaining the companies that absorbed that investment, the probability that every founder from the HRT cohort navigates to non-overlapping territory is close to zero.

The closest historical parallel for what happens next may not be the PayPal Mafia at all — a group that dispersed because the internet was large enough to absorb them all — but the competitive hardware industry of the 1990s, where Jensen Huang, Lisa Su, Hock Tan, and their peers learned an earlier version of the same lesson: that in a market defined by compounding technical advantages and concentrated capital, the number of long-run winners is small, and the people who trained together eventually compete for the same customers.