Why China’s Answer to Surge AI Got a $1 Billion Valuation On Just $30 Million in Orders

Unless you’ve been under a rock over the weekend, you’d know that the debate about AI safety took a big leap forward, as Anthropic CEO Dario Amodei on Saturday called for leading AI companies to slow the development of advanced AI, drawing supportive responses from OpenAI CEO Sam Altman, Google DeepMind chair Demis Hassabis and SpaceX CEO Elon Musk.

On Monday, Microsoft joined in, while Chinese authorities appeared skeptical. This is likely to dominate discussion in the AI sector and on Wall Street in the coming days. Stay tuned…

Onto today’s news…

When representatives of Silicon Valley venture capital firm Dimension Capital visited China on a research trip recently, one trend that grabbed their attention was the rise of local AI data and model evaluation startups akin to Surge, Mercor and Scale in the U.S.

The Chinese startups represent a new phase in the country’s AI advancement. Access to high-end data is increasingly critical for developers of cutting-edge models capable of handling complex tasks that require expert-level knowledge. The startups’ sophisticated benchmarks also help evaluate and refine these models’ capabilities for real-world tasks in different fields.

“On the ground we saw a nascent cohort of Chinese upstarts building evaluation and verification infrastructure” instead of data annotation services for more rudimentary forms of AI, Dimension said in a letter to its limited partners about the trip. Dimension singled out UniPat, a Beijing-based startup whose specialized benchmarks have been used by developers and researchers to evaluate American and Chinese AI models.

UniPat, founded in December, recently closed a funding round at a $1 billion pre-money valuation, according to a person with knowledge of the matter. UniPat’s major competitor, Humanlaya, has just closed a new round at a post-money valuation of nearly $1 billion, according to two other people with knowledge of the round.

As of the end of May, UniPat had $30 million worth of orders from customers, according to a person with knowledge of the matter, making the $1 billion valuation it recently fetched seem extremely high as a multiple of forward revenue. By comparison, Mercor crossed over $2 billion in gross revenue run rate in June while sporting a roughly $10 billion valuation.

Venture capital investors backing the Chinese startups say the premium is justified because, compared to the U.S., China’s domestic market for high-end data and post-training services for advanced AI models is still relatively new.

While many Chinese AI developers are improving their models by distilling American models, AI researchers in the country say firms like UniPat and Humanlaya are boosting model quality by providing human experts to help shape models’ intelligence. This happens during the reinforcement learning process, also known as the post-training phase of model development.

Most major Chinese model makers, including Alibaba, ByteDance, Moonshot and DeepSeek, have bought data or other services from UniPat and Humanlaya, according to people with knowledge of the startups’ businesses.

Some Chinese model developers also buy data and services from American firms including Surge. (UniPat and Humanlaya would find it difficult to sell data to American AI labs due to the Chinese government’s strict regulations on overseas data transfers.)

UniPat and Humanlaya, like their American counterparts, pay human experts from industries such as law, finance, healthcare and science to produce specialized data to refine models and strengthen them in those domains. The startups also produce AI-generated, or synthetic, data using their own algorithms, which AI developer customers can use to train new models. The startups also develop model evaluation benchmarks in specific areas such as coding, biomedical research and visual intelligence.

UniPat, which stands for “universe of patterns,” was co-founded by Li Kuan, who previously worked at Alibaba on infrastructure for reinforcement learning. UniPat started out as a research lab focusing on synthetic data, but soon expanded its effort in human expert data. UniPat has more than 100 employees in Beijing, and roughly half of them are researchers who previously worked for AI labs in China.

Humanlaya, whose name combines “human” with the Himalayas, was co-founded by Jasper Jiao, a former ByteDance employee who helped build that company’s own expert data platform, Xpert. (Xpert is a quasi-competitor to Humanlaya, but ByteDance also is a customer of Humanlaya.) Roughly half of the startup’s 80 employees are experts in specific areas; they include a Harvard Law School graduate and a medical research expert with a doctorate from Peking University. Both UniPat and Humanlaya have received funding from HSG, formerly known as Sequoia Capital China.

In multiple internal meetings with staffers earlier this year, Jiao talked about how the success of the likes of Surge and Mercor in the U.S. points to a promising future for Chinese companies like Humanlaya, according to two employees who attended the meetings.

Humanlaya said more than 100,000 human experts from 300 industries and academic disciplines have signed up to join the startup’s expert data task platform called TalentsAI, which is similar to Surge’s DataAnnotation platform. (After signing up, they still need to pass an exam to actually qualify for AI work.) UniPat, which has been working with experts by reaching out to them individually, is planning to launch a competing platform called Bencher later this month.

Due to lower labor costs in China, Humanlaya and UniPat pay their experts less than what U.S. data firms do. For example, a medical doctor in the U.S. gets paid $150 per hour to work with a data firm, but Humanlaya pays about $45 per hour for a similar doctor in China.

But the Chinese data startups also generate much smaller revenues from their customers than in the U.S., where OpenAI and Anthropic pay billions of dollars a year for posttraining data. By contrast, China’s AI sector is still warming up to the idea of spending big on such data. Chinese model makers can hire research interns for less than $50 a day and get them to work on generating synthetic data too, giving the companies less incentive to acquire data externally.

Chinese data startups internally estimate that the country’s current market size by revenue for such AI training data is roughly a tenth the size of the market for their American counterparts, but they are betting that the Chinese market will grow rapidly.—Jing Yang also contributed to this piece.

Here’s what else is going on…

Big Number

Microsoft is planning to more than triple the size of its Azure cloud unit’s data center capacity to over 38 gigawatts by 2032, up from 12 gigawatts today, Bloomberg reported Thursday. A gigawatt of capacity would be able to power a city the size of San Francisco. In a statement Friday, a Microsoft spokesperson said that the reported figures “do not accurately reflect Microsoft’s results and projections” and said Microsoft’s SEC filings “should serve as the sole source of record for reported financial results.”

Overheard

China on Monday rejected calls from top U.S. tech executives on pacing frontier AI development while maintaining a lead over China to ensure America’s national security, calling it “fearmongering.”

A swarm of OpenAI agents conducted a cyberattack on software service RubyGems in May, months before the company’s agents hacked model platform Hugging Face, researchers at AI safety organizations Nightingale Collective and AI Futures Project found. RubyGems allows software engineers to download ready-made pieces of code in the Ruby coding language.

Salesforce, ahead of its annual Dreamforce customer conference next week, is removing the word “Agentforce” from the names of some of its top products, according to two people with direct knowledge of the matter. The reversal comes after Salesforce made Agentforce a centerpiece of its product sales and marketing strategy since launching it in late 2024, with CEO Marc Benioff previously suggesting that Salesforce might rebrand its official name as Agentforce.

Anthropic has disrupted several attempts to use its Claude models for potentially malicious activity, including research into how to adapt bird flu to a human-transmittable strain with “pandemic potential,” the company said in a report Thursday. Amid increasing worry about AI misuse and rogue models, the Anthropic report says “biological misuse is one of the most serious risks of frontier AI models.” In one case Anthropic identified, a scientist used Claude to write a grant for research—apparently to be done at a military research institute—into engineering more infectious strains of the debilitating mosquito-borne virus chikungunya.

OpenAI said Thursday it will pause new subscriptions to its $200-a-month plan for Astra, its new flagship model, another sign of the compute crunch facing AI companies. Thibault Sottiaux, who leads OpenAI’s core products, cited strains on the company’s systems and a desire to preserve service for existing users. Astra, released this month, has met “unprecedented demand,” Sottiaux said Wednesday on X. “We’re pulling all the levers possible to sustain the demand,” he wrote, adding that he had never seen growth like it.

People on the Move

Microsoft chief communications officer Frank Shaw will leave the company at the end of this year, he announced in a LinkedIn post Friday. Shaw has spent more than 28 years at the company, including 17 years overseeing its communications. During that time, he steered the company’s public messaging through its $26 billion LinkedIn acquisition, its $69 billion deal to buy Activision, and its often-tumultuous relationship with OpenAI.

Court Watch

The Department of Justice is looking into whether Nvidia tried to avoid antitrust scrutiny in its $20 billion deal to license the technology of chip startup Groq and hire most of its employees in December, the New York Times reported Wednesday.

The U.S. Senate formally launched an investigation into the July hack of Hugging Face, when a swarm of OpenAI agents broke out of a testing environment and gained access to the open-source repository.

Earnings Report

Adobe Systems shares dipped 2% late Thursday after its third-quarter earnings report, as some investors appear to unsure about how much its AI products will drive future revenue growth. Adobe, which began breaking out revenue for its AI products in March, is continuing to highlight this area of its business. On an earnings call, Adobe chair and CEO Shantanu Narayen said annual recurring revenue from AI products—the value of its contracts over the next 12 months—surpassed $650 million during its August quarter, up more than 150% from last year and $150 million from last quarter.

Oracle reported 30% growth in revenue of $19.3 billion for the quarter ending Aug. 31, a slightly better growth rate than the company projected in June, even as it burned $5 billion to expand the data centers that are driving its accelerating topline growth.

Deals and Debuts

See The Information’s Generative AI Database for an exclusive list of private companies and their investors.

Nvidia has discussed investing in Anthropic’s upcoming initial public offering that would raise as much as $100 billion at a valuation of around $2 trillion, Reuters reported Friday. Nvidia could invest as much as $10 billion in Anthropic at the IPO price, according to the report. Anthropic last raised capital at a $965 billion valuation in May.

SpaceX has signed another cloud computing rental deal to an identified customer that will pay it about $1.11 billion per month starting on Dec. 1, the company’s chief financial officer, Bret Johnsen, said in an interview at Goldman Sachs’ Communacopia + Technology conference on Thursday.

Shanghai Enflame Technology surged on its first trading day Friday after the Chinese AI chip developer raised about 6.12 billion yuan ($911 million) in an initial public offering, underscoring strong investor appetite for domestic alternatives to Nvidia chips.

Google appears to have completed its talent deal with Mechanize, an AI coding startup, with the team going to Google DeepMind, according to Business Insider. The price has been reported at $1.5 billion or more.

Italian digital conglomerate Bending Spoons is buying whiteboarding software firm Miro at a valuation of $1.355 billion, the latest in a series of purchases the Italian firm has done at knock-down prices.

Positron AI, a company designing chips and servers for running AI models rather than training them, raised $875 million in a Series C funding round led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, Dylan Patel's SemiAnalysis Capital and Jim Clark.

Kepler Computing, a company that develops energy-efficient AI memory and logic chips, came out of stealth with $468 million in late-stage venture funding from GlobalFoundries, Intel Capital, AMD Ventures, Baillie Gifford and Bill Gates's Gates Frontier fund, according to Wired.

Mach Industries, a defense manufacturer that develops unmanned aircraft, jet engines, rocket motors and the explosives that go in them, raised $600 million in a continuation of its Series C funding round, with backers including Ribbit Capital, Infinite Capital, Bedrock Capital and Sequoia.

Ayar Labs, a company that replaces the copper wiring between AI chips with light, so that data moves between processors optically instead of electrically, raised $150 million in additional funding from Wiwynn, Alchip, AMD, Intel, MediaTek and Nvidia.

TAR, a company that develops off-grid power plants — solar and battery systems that connect directly to an AI data center rather than to the grid — raised $120 million in a Series A funding round led by Spark Capital.

Maven Robotics, a company whose robots stack and move boxes in warehouses and factories, came out of stealth with $100 million in a Series A funding round from RoboStrategy, LocalGlobe, Vine Ventures and XTX Ventures.

Heron Power, a company developing power electronics that let renewable projects, battery storage and data centers connect straight to medium-voltage transmission lines without a transformer, raised a $60 million credit facility from J.P. Morgan and TriplePoint Capital.

Xapien, which runs background checks and due-diligence research on companies and people using AI, raised $56 million led by Spectrum Equity.

Archy, a San Jose, Calif.-based startup that develops AI technology for dental practice management, raised $50 million in Series C funding led by JMI Equity.

Vheda Health, a company whose software identifies the members of a health plan most likely to end up in hospital and pushes them toward care first, using connected devices and human coaching, received a $47 million investment from Agora.

Arlequin AI, a company developing a different kind of AI model, raised €28 million in a Series A funding round led by redalpine and OTB Ventures.

Epsilon Health, a company that operates as a radiology practice built around AI, raised $27.6 million led by AlleyCorp.

GenHealth.ai, a company whose AI agents do the paperwork side of medicine, raised $16.5 million in a Series A funding round led by Flare Capital Partners.

Buildcheck AI, a company whose software reads construction blueprints and finds the errors, omissions and clashes between disciplines before crews build them, raised $12 million in a Series A funding round led by Telescope Partners.

Graph AI, a company whose software automates drug-safety reporting for pharmaceutical companies, raised $13.3 million in a Series A funding round led by Insight Partners.

Intermezzo, a company that automates the plumbing of international payroll, using AI agents, raised $10 million from Crosslink Capital, UKG Ventures, CloudPay, Character Capital, Behind Genius Ventures and angel investors.

Enigmata, a company whose technology encrypts a company's records and documents in a way that AI and analytics tools can still work with directly, without ever decrypting them, raised $6.5 million in a seed funding round led by Blockchange Ventures.

Verily, the Alphabet health company, received an investment from Nvidia, with participation from the CU Healthcare Innovation Fund II. The amount was not disclosed.

Cognition is acquiring Dioxus Labs, the team behind Dioxus, an open-source toolkit for offering apps in the Rust programming language. Terms were not disclosed.

SemiAnalysis, the semiconductor and AI-infrastructure research firm run by Dylan Patel, acquired Citrini Research, a markets research shop, for an undisclosed sum, according to Bloomberg.

OpenAI launched ChatGPT for Financial Services, a version of its work product aimed at investment bankers and equity researchers, built with Morgan Stanley and Evercore as design partners.

Universal Music Group and ElevenLabs signed a multi-year licensing agreement and said they will develop a new AI music platform that lets fans make remixes, mashups and new interpretations of tracks by participating artists. No financial terms were disclosed.

Ant International, Visa and Mastercard began work on a shared standard for identifying and verifying AI agents that make purchases for people, which they are calling a Know-Your-Agent interoperability framework, according to Reuters.

Amazon is letting advertisers test buying ads on ChatGPT via Amazon’s advertising technology platform, Amazon said on Sept. 10.

Chip startup d-Matrix said Thursday it plans to use Nvidia’s networking hardware to connect its chips for running AI models to each other and to Nvidia’s Vera central processing units, which are used in data centers.

Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [email protected].

If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.

Juro Osawa is a reporter covering tech in Asia, from Alibaba and Tencent to startups. He previously worked for The Wall Street Journal. He is based in Hong Kong and can be found on Twitter at @JuroOsawa.