*

Code-gen startups are disrupting the software industry,

but face

mounting losses

*

Big tech firms like Google and Microsoft are entering the

AI

coding market

*

AI coding tools are allowing tech giants to shed expensive

human

software engineers

(Updates headline)

By Anna Tong and Krystal Hu

SAN FRANCISCO, June 3 (Reuters) - Two years after the

launch of ChatGPT, return on investment in generative AI has

been elusive, but one area stands out: software development.

So-called code generation or "code-gen" startups are

commanding sky-high valuations as corporate boardrooms look to

use AI to aid, and sometimes to replace, expensive human

software engineers.

Cursor, a code generation startup based in San Francisco

that can suggest and complete lines of code and write whole

sections of code autonomously, raised $900 million at a $10

billion valuation in May from a who's who list of tech

investors, including Thrive Capital, Andreessen Horowitz and

Accel.

Windsurf, a Mountain View-based startup behind the popular

AI coding tool Codeium, attracted the attention of ChatGPT maker

OpenAI, which is now in talks to acquire the company for $3

billion, sources familiar with the matter told Reuters.

Its tool is known for translating plain English commands

into code, sometimes called "vibe coding," which allows people

with no knowledge of computer languages to write software.

OpenAI and Windsurf declined to comment on the acquisition.

"AI has automated all the repetitive, tedious work," said

Scott Wu, CEO of code gen startup Cognition. "The software

engineer's role has already changed dramatically. It's not about

memorizing esoteric syntax anymore."

Founders of code-gen startups and their investors believe

they are in a land grab situation, with a shrinking window to

gain a critical mass of users and establish their AI coding tool

as the industry standard.

But because most are built on AI foundation models developed

elsewhere, such as OpenAI, Anthropic, or DeepSeek, their costs

per query are also growing, and none are yet profitable.

They're also at risk of being disrupted by Google, Microsoft

and OpenAI, which all announced new code-gen products in May,

and Anthropic is also working on one as well, two sources

familiar with the matter told Reuters.

The rapid growth of these startups is coming despite

competing on big tech's home turf. Microsoft's GitHub Copilot,

launched in 2021 and considered code-gen's dominant player, grew

to over $500 million in revenue last year, according to a source

familiar with the matter.

Microsoft declined to comment on GitHub Copilot's revenue.

On Microsoft's earnings call in April, the company said the

product has over 15 million users.

LEARN TO CODE?

As AI revolutionizes the industry, many jobs - particularly

entry-level coding positions that are more basic and involve

repetition - may be eliminated. Signalfire, a VC firm that

tracks tech hiring, found that new hires with less than a year

of experience fell 24% in 2024, a drop it attributes to tasks

once assigned to entry-level software engineers are now being

fulfilled in part with AI.

Google's CEO also said in April that "well over 30%" of Google's

code is now AI-generated, and Amazon CEO Andy Jassy said last

year the company had saved "the equivalent of 4,500

developer-years" by using AI. Google and Amazon declined to

comment.

In May, Microsoft CEO Satya Nadella said at a conference

that approximately 20 to 30% of their code is now AI-generated.

The same month, the company announced layoffs of 6,000 workers

globally, with over 40% of those being software developers in

Microsoft's home state, Washington.

"We're focused on creating AI that empowers developers to be

more productive, creative, and save time," a Microsoft

spokesperson said. "This means some roles will change with the

revolution of AI, but human intelligence remains at the center

of the software development life cycle."

MOUNTING LOSSES

Some "vibe-coding" platforms already boast substantial

annualized revenues.

Cursor, with just 60 employees, went from zero to $100

million in recurring revenue by January 2025, less than two

years since its launch. Windsurf, founded in 2021, launched its

code generation product in November 2024 and is already bringing

in $50 million in annualized revenue, according to a source

familiar with the company.

But both startups operate with negative gross margins,

meaning they spend more than they make, according to four

investor sources familiar with their operations.

"The prices people are paying for coding assistants are

going to get more expensive," Quinn Slack, CEO at coding startup

Sourcegraph, told Reuters.

Both Cursor and Windsurf are led by recent MIT graduates in

their twenties, and exemplify the gold rush era of the AI

startup scene. "I haven't seen people working this hard since

the first Internet boom," said Martin Casado, a general partner

at Andreessen Horowitz, an investor in Anysphere, the company

behind Cursor.

What's less clear is whether the dozen or so code-gen

companies will be able to hang on to their customers as big tech

moves in.

"In many cases, it's less about who's got the best

technology -- it's about who is going to make the best use of

that technology, and who's going to be able to sell their

products better than others," said Scott Raney, managing

director at Redpoint Ventures, whose firm invested in

Sourcegraph and Poolside, a software development startup that's

building its own AI foundation model.

CUSTOM AI MODELS

Most of the AI coding startups currently rely on the Claude

AI model from Anthropic, which crossed $3 billion in annualized

revenue in May in part due to fees paid by code-gen companies.

But some startups are attempting to build their own models.

In May, Windsurf announced its first in-house AI models that are

optimized for software engineering in a bid to control the user

experience. Cursor has also hired a team of researchers to

pre-train its own large frontier-level models, which could

enable the company to not have to pay foundation model companies

so much money, according to two sources familiar with the

matter.

Startups looking to train their own AI coding models face an

uphill battle as it could easily cost millions to buy or rent

the computing capacity needed to train a large language model.

Replit earlier dropped plans to train its own model.

Poolside, which has raised more than $600 million to make a

coding-specific model, has announced a partnership with Amazon

Web Services and is testing with customers, but hasn't made any

product generally available yet.

Another code gen startup Magic Dev, which raised nearly $500

million since 2023, told investors a frontier-level coding model

was coming in summer 2024 but hasn't yet launched a product.

Poolside declined to comment. Magic Dev did not respond to a

request for comment.