The AI product developed by a Shanghai-based young woman has gone viral across all online platforms, with a valuation of 50 billion yuan.

A new $7.5 billion (approximately 500 billion RMB) unicorn has emerged in Silicon Valley.

On October 9, AI startup TypeSafe AI announced the completion of an $870 million (approximately 5.8 billion RMB) financing round, led by a16z and followed by HSG. Just 24 days ago, its valuation was only about $200 million. In a very short period of time, its valuation has skyrocketed 37.5 times.

Silicon Valley investors are scrambling to place bets, simply because TypeSafe AI's business is unique: making enterprise software intelligent.

Valuation skyrockets 37 times in 24 days

Diogo Almeida, founder of TypeSafe AI, was once involved in the creation of ChatGPT.

In his early years, he worked as a researcher at OpenAI and participated in relevant research on InstructGPT and Reinforcement Learning from Human Feedback (RLHF). These technologies later became the important foundation of ChatGPT.

After leaving OpenAI in 2024, Almeida co-founded TypeSafe AI with former Meta researcher Sasha Sheng and engineer Erik Gafni. Sasha Sheng is a Shanghai native who graduated from Shanghai Jiao Tong University with a bachelor's degree in mechanical engineering.

The three of them spent two years on secret R&D, and the company made its first public debut only on September 15 this year.

On that day, TypeSafe AI launched its first model Jev, and announced the completion of a $40 million seed round of financing led by US venture capital firm DCVC. The company's valuation reported by the media at that time was about $200 million.

Jev is a highly vertical model: it helps enterprise software make decisions. For example, an insurance company can use it to determine which claims applications require manual review; a recruitment website can use it to determine which candidates meet the standards.

One word sums up the feature of Jev: fast.

In the following weeks, the fate of this company changed.

Three days after Jev was released, cloud platform Vercel released a set of data: within only 24 hours after Jev went online, nearly 13% of the platform's paid teams tried to use it. This proportion is twice that of the GPT-5.6 series in the same period, setting a new record for the adoption speed of new models after going online on Vercel.

By October 15, Vercel disclosed that in the first two weeks after Jev went online, 35.3% of paid teams had used it. Even though it was not released until mid-September, Jev still became the model with the largest number of requests on the platform in September.

TypeSafe AI also claimed that shortly after the product was launched, about one-third of the Fortune 500 companies in the United States had used Jev. However, the company has not yet made the full list of these companies public.

Capital moved quickly in response to the news.

On October 9, TypeSafe AI announced a new round of financing: $870 million, with a valuation of $7.5 billion.

Calculated from the approximately $200 million valuation on September 15, in just 24 days, the company's valuation has increased 37.5 times.

The lead investor a16z sent heavyweight figure Martin Casado. He co-founded software-defined networking company Nicira and sold it to VMware for $1.26 billion in 2012. After that, he joined a16z and was long responsible for software infrastructure investment.

After this financing, Casado also joined the board of directors of TypeSafe AI.

Moreover, according to HSG partner Pat Grady cited by investment research institution Sacra, within 7 days after Jev went online, its annualized revenue run rate had reached $100 million.

What exactly did TypeSafe AI do right?

In an interview with TechCrunch, he once explained that in the past few years, large language models have been very good at understanding and generating human language, but these capabilities are not necessarily suitable for computer software automation.

For example, ChatGPT can write an article to help people analyze complex problems, but if an enterprise only needs to judge whether a transaction is risky, letting it generate a long analysis will instead be time-consuming and costly.

Almeida believes that the AI needed by humans and computers is not the same. So he decided to do the opposite: abandon text generation, and specifically train an AI that helps software make judgments — this is Jev.

Li Mingshun, Chairman of Xingxing AI and Founder of Shunfu Capital, told Pencil News that TypeSafe AI is sought after not only for its technology, but also for the background of its team. The founders come from OpenAI, and such a team not only has the potential to develop independently, but may also become the acquisition target of tech giants.

Help customers save $500,000 per year

TypeSafe AI helps customers use low-cost AI to replace expensive AI.

Jack & Jill is an AI recruitment platform that has served more than 5,000 enterprises and 400,000 professionals. It uses AI to help job seekers find suitable positions, and also helps enterprises screen candidates.

For every 1,000 candidates screened, Gemini costs $0.755, while Jev only costs $0.092, reducing the cost by 88%. The median time to complete the initial screening of a position is also shortened from 20.3 seconds to 10.3 seconds.

Less than 10 days after the first test, Jack & Jill replaced all the initial screening processes with Jev.

Calculated based on the current business scale, it can save about $265,000 per year; considering business growth, it is expected to save about $500,000 in the next 12 months.

This is exactly the opportunity for TypeSafe AI.

Why is Jev so cheap? The reason is that Jev abandons text generation. After receiving a task, traditional large models usually need to generate responses step by step. Jev directly outputs a score or a probability, eliminating the text generation process.

Jev is suitable for judgment tasks with clear boundaries, and cannot replace general large models to complete work such as complex analysis and long text generation.

An e-commerce enterprise can use it to assign risk scores to orders; an insurance company can use it to determine which claims applications require manual review; a customer service company can use it to classify work orders.

He Jun (pseudonym), head of agent at a leading large tech company, told Pencil News that enterprises are increasingly inclined to let different models perform their respective duties. For example, let Jev first judge which category the task belongs to, call GPT or Claude when complex reasoning is needed, and use a more cost-effective model when completing summary sorting.

In this way, enterprises do not need to pay for high-performance models for simple tasks.

AI investment, targeting the unseen areas

Silicon Valley investors are re-evaluating AI business opportunities.

In the past few years, the most high-profile companies in the AI industry have mostly competed for two markets: chat and content generation for ordinary users, and AI programming for programmers.

But no matter how powerful AI programming is, the software written may not be smarter. This is exactly the point a16z put forward when investing in TypeSafe AI: AI enables developers to write software faster, but it does not make the software itself generally intelligent.

What investors want to bet on now is to let AI directly enter the interior of the software and become part of the automatic business execution. And this is the area invisible to users.

In its investment article, HSG compares this change to databases: in the future, developers may call AI just like calling databases, allowing the software to continuously make judgments during operation.

This is why TypeSafe AI's financing has attracted institutions such as a16z, HSG and DCVC.

There is also an interesting economic problem here.

The name Jev comes from British economist William Stanley Jevons.

There is a famous "Jevons Paradox" in economics: when the use efficiency of a resource increases and its cost decreases, the demand may grow rapidly, and the final consumption will instead increase.

TypeSafe AI hopes that the same will happen to AI.

In the past, because model invocation was expensive, an enterprise might only let AI check part of the orders and spot check part of the customer service records. Now that each judgment is cheap enough, enterprises may expand the inspection scope, and even let AI evaluate every business operation.

AI demands that did not exist before are thus created.

However, there are many other companies that want to make money from this business.

On October 7, AI evaluation company Arize released a report comparing Jev with professional decision-making models from companies such as OpenAI, Liquid AI and Cloudflare.

In a test to judge whether AI responses contain fictional content, Jev achieved a balanced accuracy of 85.5%, the relevant OpenAI model reached 81.5%, and Liquid AI reached 84.8%. Although Jev is leading, the gap between some models is not statistically significant.

From the release of Jev to the release of this evaluation, only three weeks have passed, and many similar products have appeared on the market.

Some developers have even launched the open source model Kev. Arize's test shows that in an AI Agent task routing test, the performance of Kev is very close to that of Jev.

TypeSafe AI's financing also reminds entrepreneurs: making a low-cost model is just the beginning of business. Once competitors can provide similar accuracy at a lower price, customers may leave at any time. The real competitiveness depends on whether the company can deeply integrate into customers' businesses, accumulate stable invocation volume, and provide irreplaceable services.

This article does not constitute any investment advice.

This article is from WeChat Official Account "Pencil News" (ID: pencilnews), author: Pencil News, published with authorization from 36Kr.