Anthropic's prospectus contains a fascinating detail.

In this document, which runs to hundreds of pages, an investment firm called Spark Capital is mentioned more than 50 times throughout.

That frequency even exceeds that of quite a few tech giants that have poured billions of dollars into Anthropic.

The reason is simple: Spark Capital is one of Anthropic's earliest institutional investors, and its partner remains the only representative from a traditional venture capital firm on the company's board to this day.

Many people may not know that when Anthropic was first founded, almost no VC was willing to invest. Even before its Series C round, the company largely had to rely on personal connections within the tech circle to secure funding.

This was especially true after ChatGPT burst onto the scene, when investors had even more reason to hesitate: OpenAI had already seized the initiative and had Microsoft's backing—so how could Anthropic possibly win?

But Spark chose to bet firmly.

Starting from Series C, Spark participated in every round of Anthropic's financing, investing $777 million in total. Based on a valuation of $1.5 trillion, it made a full $10 billion in three years.

Why was it Spark, of all firms, that dared to bet on OpenAI's competitor while OpenAI was at its peak?

Looking through this company's investment record over the past decade or so, as well as public interviews with several core partners, I believe this investment was far more than just good luck.

Spark realized very early on that what AI is changing is not only the competitive landscape of the tech industry, but also the rules of the venture capital game itself:

Those relatively static competitive advantages in traditional investment logic are becoming increasingly unreliable in the AI era. Compared with a temporary technological lead, technical judgment and the ability to respond to change matter more.

/ 01 / A $75 Million Investment That Changed Its Fate

When it comes to Spark Capital, one has to mention its partner Yasmin Razavi.

This woman, who grew up in Tehran, Iran, went to Canada in 2006 to visit her older brother who was studying there—and decided on the spot to stay. Later, she got into the University of Toronto, joined McKinsey, and then attended Harvard Business School. She joined Spark in 2017 and was promoted to partner in 2020.

And the most successful investment of her career was Anthropic.

For all of Anthropic's current glory, almost no VC was willing to invest in it when it was first founded.

In 2021, Anthropic had no product and no revenue, yet it asked for $500 million in financing right out of the gate.

At the time, Anjney Midha, who came from a scientific background, helped Anthropic look for investors and met 22 VCs in one go. Twenty-one rejected it on the spot, and the remaining one ultimately did not work out either.

One investor even asked in response: What is GPT-3?

With no other option, Anthropic could only turn to tech-circle tycoons who had already achieved financial freedom.

In May 2021, Anthropic completed a $124 million Series A financing led by Skype co-founder Jaan Tallinn, with other investors including Facebook co-founder Dustin Moskovitz and others.

In April 2022, it completed a $580 million Series B financing, one of whose biggest backers was Sam Bankman-Fried, who later became notorious because of FTX's bankruptcy.

The real turning point came in 2023.

ChatGPT burst onto the scene, and Microsoft poured enormous amounts of money into OpenAI. The commercial prospects of large models were finally validated, but a new question also arose:

Since OpenAI had already seized the initiative and had Microsoft's backing, why still invest in Anthropic?

Most VCs still chose to wait and see. But Yasmin Razavi did not think so.

In an interview with Forbes in May this year, she recalled that in order to push this investment through, she specifically wrote two detailed investment memos to persuade Spark's other partners.

Her logic mainly had three points:

First, large models may represent a true technological paradigm revolution.

Borrowing scientific philosopher Thomas Kuhn's theory of paradigm shifts, she argued that the development of deep learning, unsupervised learning, and Transformers over the past decade had already created the conditions for LLMs to bring about a new technological paradigm.

Second, the scope of AI applications will far exceed that of ordinary software tools.

Razavi listed fields such as law, healthcare, accounting, software development, mining, and agriculture, arguing that AI is redefining the way humans and computers work together.

Third, safety is not only a matter of technical ethics, but also the foundation for developing reliable AI products.

She specifically mentioned the Anthropic team's accumulated research in Scaling Laws, RLHF, and Constitutional AI, arguing that this team has both top-tier model capabilities and long-term commitment to safety and alignment issues.

More importantly, she did not believe that the large-model industry had been completely taken by OpenAI. On the contrary, as AI becomes a new technological paradigm, companies with original research capabilities and the ability to continuously iterate on models still have the chance to become the next generation of tech giants.

In May 2023, Spark led Anthropic's $450 million Series C financing and put in $75 million of its own money. This was the largest check in Spark's history, equivalent to more than 5% of its $1.4 billion growth fund.

Facts proved that this investment was only the beginning.

The prospectus shows that Spark subsequently participated in five more rounds of Anthropic financing, with cumulative investment reaching $777 million. Based on a $1.5 trillion valuation, the book value of Spark's shares could exceed $10 billion, for a cumulative investment return of about 14x.

Among them, the shares corresponding to the initial $75 million investment could potentially return as much as 80x.

But what Spark gained was not only financial returns. Razavi began serving as an Anthropic director in March 2023, even earlier than the official announcement of the Series C financing.

After joining the board, Razavi began helping the company identify which specific applications could create the greatest commercial value.

Anthropic co-founder Daniela Amodei once told Forbes that before Claude Code matured, Razavi repeatedly advised the company to treat AI programming as a separate and important market.

Later, we all know the story: Claude Code became one of Anthropic's fastest-growing products.

/ 02 / Spark Only Favors Two Types of AI Application Companies

If we look only at Anthropic, it is easy to understand Spark's success as a one-off investment. But if we lengthen the timeline, we will find that this company's layout in AI is actually quite systematic.

Judging from its public portfolio, Spark has bet on almost every important link in the AI industry chain.

In the foundation model field, in addition to Anthropic, it has also invested in AI protein design company Profluent and robotics foundation model company Generalist.

At the computing power layer, Spark has invested in Lightmatter, MatX, and Enfabrica, laying out optical interconnect, AI-specific chips, and high-performance networking interconnect, respectively.

At the infrastructure layer, it has invested in data annotation company Scale AI, vector database Qdrant, model inference platform Baseten, and AI data center company Crusoe.

It has not neglected the application layer either, including AI healthcare company Abridge, audio and video editing platform Descript, AI video generation company Pika, and AI meeting notes tool Granola.

These investments look scattered, but the logic behind them is highly consistent: seeking companies in each link that truly have the ability to build long-term competitive moats.

Especially for the AI application layer, Spark divides AI startups into three categories:

The first category is Adaptation—that is, adding AI features to existing software, such as AI search and AI assistants.

Although such products can improve the experience, it is difficult to form a long-term moat based on AI features alone.

The second category is Evolution—that is, using AI to redesign workflows.

Representative examples are AI healthcare company Abridge and AI meeting notes tool Granola.

The former uses AI to automatically organize doctor-patient conversations and generate clinical documents. The latter combines users' own records with meeting content to generate notes that better fit individual needs.

The third category is Revolution—that is, AI creates products and markets that simply did not exist before.

Hyatt admitted frankly that Spark's most successful investments are mainly concentrated in the third category, as well as part of the second. For companies in the first category that simply add AI features, it is basically not interested.

Because in their view, truly valuable technological innovation is not just making the original things cheaper and faster, but enabling users to accomplish things they could not accomplish before.

Hyatt also specifically warned that there is a trap in AI Agent startups.

Many companies use large models to automate traditional manual services and quickly gain revenue by reducing costs. But if competitors can all call the same models, then the cost advantage established today may disappear tomorrow.

A true competitive moat must come from the product itself, as well as continuously accumulated data, user feedback, and workflow capabilities.

This may also be the key to understanding Spark's AI investment map: it invests both in companies that create technological paradigms and in companies that use new technologies to redefine products.

/ 03 / In the AI Era, the Rules of VC Investment Are Also Changing

Spark is not only rejudging what kinds of AI companies are worth investing in, but also rethinking traditional VC investment methods.

In 2025, Spark partner Nabeel Hyatt put forward a very interesting viewpoint in an interview.

In a 2025 interview, Hyatt proposed that SaaS investing over the past decade or so had increasingly come to resemble a math problem with a standard solution.

Investors could evaluate a software company's growth potential through metrics such as ARR, customer retention rate, sales efficiency, and market size.

But investing in AI is different. What capabilities models will have in a few months, what product forms will emerge, and which jobs will be redefined—none of these questions have answers at the moment of decision-making.

So he made a distinction: in the past, VC was good at solving Puzzles—difficult problems under known conditions; AI investment is more like a Mystery—an exploration whose answer has not yet been written.

This distinction seems abstract, but it is very intuitive when applied to Anthropic.

In 2023, any investor trying to estimate Anthropic's value with traditional financial models would have found it hard to find a convincing investment reason for a company with limited revenue, high computing costs, and OpenAI ahead of it.

But what Spark valued was precisely the things that financial models find hard to measure: a top research team, and the entirely new markets that large models might create.

Of course, betting on technological change does not mean investment risk can be ignored. Especially in the AI industry, the technology that looks leading today may be caught up with or even surpassed by competitors within a few months.

So the question is: if a technological lead can be broken at any time, what exactly is the long-term competitive moat of an AI company?

Hyatt was also asked this question in an earlier interview.

At the time, the emergence of DeepSeek led many investors to begin questioning how much competitive advantage large-model companies could still build by investing tens of billions of dollars in training models.

But Hyatt does not believe that huge capital investment itself is the only moat for model companies. Compared with leadership in a certain generation of models, Spark values more the technical judgment of top teams and their ability to continuously execute and iterate.

From this perspective, those relatively static competitive advantages in traditional investment logic are becoming increasingly unreliable in the AI era. Compared with a temporary technological lead, whether a company can continuously adapt to change and even actively lead the next round of technological change is the key to determining its long-term value.

Text/Aqi