INFORMATION

Anthropic’s IPO and the B2D Strategy

2026-11-22
Anthropic just began trading under the ticker ANT, in the largest IPO on record. Their revenue numbers are impressive: $11.6 billion in the second quarter of 2026, against $6.7 billion reported for OpenAI for the same period.


The story investors are hearing is a simple one: Anthropic passed OpenAI because it sold to businesses, while OpenAI chased consumers. Tech investor Dan Niles, for example, wrote on the day of the IPO that Anthropic was “focused on day one on enterprises who are willing to pay,” while OpenAI focused on consumers used to getting internet services for free.


广告


For the most part, this is correct. In the twelve months to September 2026, customers paying Anthropic more than $100,000 a year accounted for 74% of its revenue, with the remaining 26% coming from roughly 40 million smaller customers. OpenAI, by contrast, entered 2026 with about 60% of its revenue coming from consumers.


Anthropic IPO_Source Nasdaq
(Source: Nasdaq)
B2C vs B2B…and B2D
Why is Anthropic winning enterprise customers over OpenAI? IMHO, model quality alone doesn’t explain the gap. On broad benchmarks, the two labs’ top models trade places within a point or two, and Anthropic’s edge in coding is real but narrow.


Partner Content
View All
Geehy Launches G32R502 1kW Bidirectional Inverter Solution
Geehy Launches G32R502 1kW Bidirectional Inverter Solution
By Geehy Semiconductor 11.20.2026
Qnity at One: Randy King on Shrink, Stack and the Materials Behind AI
Qnity at One: Randy King on Shrink, Stack and the Materials Behind AI
By Randal King, Senior Vice President and Chief Technology and Sustainability Officer, Qnity 11.18.2026
Silicon Motherboard Enables Double-Sided Integration of ICs and Passive Components
Silicon Motherboard Enables Double-Sided Integration of ICs and Passive Components
By Nisshinbo Micro Devices 11.16.2026
The real difference might be in their go-to-market strategies: It’s not simply that Anthropic chose a business-to-business (B2B) strategy while OpenAI preferred a business-to-consumer (B2C) option. If you dig deeper, Anthropic won developers first, and the enterprise followed them, in what is called a business-to-developer (B2D) strategy. Anthropic’s own prospectus describes software developers as “powerful advocates for Claude” who “often serve as internal champions for adopting new technologies.”




It turns out this is not the first time this approach was used, with different outcomes. This is the subject of this piece.


The Claude Code journey
Yes, Anthropic is currently stronger than OpenAI among enterprise customers. But this statement deserves three observations:


The first is that, as we mentioned, the growth was developer-led rather than enterprise sales-led. The engine was Claude Code, which reached $1 billion in run-rate revenue within six months of its public launch. By February 2026, business subscriptions to Claude Code had quadrupled since the start of the year, and enterprise use made up over half of its revenue. It spread through engineers who adopted a tool, well before procurement teams signed enterprise deals.


The second is that OpenAI is moving toward the enterprise as fast as it can. In March, Fidji Simo, then OpenAI’s CEO of Applications, told staff the company was “acting as if it’s a code red” and turned it toward coding and business users. By August, CFO Sarah Friar told investors that enterprise revenue had overtaken consumer, months ahead of the company’s own forecast.


The third is that revenue and market penetration are different measures. OpenAI says its products reach more than one billion weekly active users and more than two million businesses.


Lead with developers
We are used to hearing about two possible go-to-market strategies: sell to the enterprise (B2B) or win the consumer (B2C). But, as we mentioned, some players took a third path: They won developers first and moved up from there.


AWS is a strong example: S3 launched on March 14, 2006, followed by EC2 a few months later, and anyone with a credit card could sign up. In those early years, Andy Jassy focused AWS’s sales efforts on smaller firms and startups, companies for which building data centers was financially impossible, and fast-growing customers such as Netflix went all-in on AWS. When enterprises later came knocking, Jassy recalled, the team realized it had to staff up. Today, AWS has reached a $184 billion annualized run rate, selling almost entirely to other businesses, and it remains the largest cloud provider.


GitHub and MongoDB took the same route in software tools. GitHub became the place where developers host and share code. When Microsoft agreed to buy it for $7.5 billion in 2018, it said explicitly that the deal would accelerate GitHub’s use by enterprises. Today, GitHub has more than 225 million developers. MongoDB grew from an open-source database that individual developers downloaded and liked into an enterprise data platform with about $2.5 billion in annual revenue.


In the meantime, Eliot Horowitz, co-founder and former CTO of MongoDB, founded Viam, a software platform designed to apply the exact same B2D playbook to robotics and smart hardware.


Finally, Zoom initially charted a similar journey: It spread bottom-up through individual users (though not necessarily developers), and as a result won enterprise sales. This started well before Covid: Its 2019 IPO prospectus noted that 55% of its 344 customers paying more than $100,000 a year had started with at least one free host. Covid then accelerated the pattern, and Zoom kept winning until Microsoft bundled Teams free with Office. IMHO, only a bundle that strong could have broken the pattern.


B2D in the semiconductors industry
One may think that B2D would work only in the software context, as the entry barrier is lower. But in fact, we can find many examples in the semiconductor and hardware context.


Nvidia is the strongest example: CUDA launched in 2007 on the GeForce 8800 GTX, a $599 gaming card released the previous November, which put a programmable parallel processor within reach of students and researchers at consumer prices. In 2012, AlexNet was trained on two GTX 580 gaming GPUs.


Nvidia never had to sell deep learning to a CIO. By the time enterprises came looking for AI solutions, their engineers already wrote CUDA, and the data center became Nvidia’s main business.


In fact, the semiconductor industry has practiced B2D for decades: Microcontroller vendors such as STMicroelectronics and Texas Instruments sell development boards for a few dollars, such as ST’s STM32 Nucleo and TI’s LaunchPad. By 2017, ST had delivered one million development kits for its STM32 microcontrollers, and TI today lists more than 4,000 development kits and boards. The logic is simple: The engineer who prototypes with a part is likely to specify it for production long before procurement places a volume order. ST and TI own that relationship directly, and they sell into thousands of products across many markets.


More recently, Tenstorrent, a U.S. startup developing AI silicon and systems, made this choice explicitly. It initially set up a sales force to sell to enterprise customers. However, in December 2025, CEO Jim Keller announced he’d be cutting 7.5% of the company and moving its sales focus from enterprise customers to individual developers and small businesses.


“The number one rule of enterprise sales is never sell to little people, because it’s a waste of time, but we’re building a new product, a new technology.” (Jim Keller, Tenstorrent)


In the long run, he said, the big customers would make the most money, “but that’s not where we are in our journey.” So far, though, the evidence that the developer route is paying off is thin. Tenstorrent’s biggest customer remains ai&, a Japanese AI company led by former Tenstorrent executive David Bennett, which has deployed more than 120 Tenstorrent Galaxy systems. IMHO, it is still too early to tell whether developers will indeed lead the big customers in, or whether the big customers simply came first.


From developers to volume
A developer community still has to be converted into revenue, and in hardware that conversion is often hard. Three newer companies from the maker world show the range of outcomes:


Raspberry Pi made the full journey. It began as a low-cost computer for education and hobbyists, and over the years, its developer community carried it into industrial products. Today, 75% of its board and module unit sales go to industrial and embedded customers. Raspberry Pi owned its developer community, and it gave those developers a path into production.


Arduino started in a similar place, with an even larger community, but never made the jump to enterprise on its own. It tried: In 2020, it launched Arduino Pro, a line of industrial-grade boards. Still, at least as many Arduino projects focus on education and experimentation as on future products. In October 2025, Qualcomm agreed to acquire Arduino, citing its more than 33 million active users and promising them a clear path to commercialization through Qualcomm’s technology and partners. IMHO, that promise says a lot both about Arduino’s potential and about what it lacked.


Finally, Hailo, an Israeli edge AI startup, has the most dramatic story of the three. On paper, it did everything right. It built a developer community of roughly 10,000 monthly active users, about 80% of them coming through Raspberry Pi. Unlike Arduino, it also reached real business customers: It sold more than half a million units to more than 100 customers, including ASUS, HP, and Fujitsu. Investors bought the story, and in April 2024, Hailo raised $120 million at a valuation of $1.2 billion.


Yet the community and the customer list never turned into enough revenue. In June 2026, Hailo laid off half of its workforce, its planned SPAC merger collapsed, and in July, Microchip agreed to acquire it for what Globes described as a small amount, after Hailo struggled to produce significant revenue.


IMHO, three things worked against Hailo: Most of its developers arrived through Raspberry Pi, so the relationship belonged more to Raspberry Pi than to Hailo. The volume decisions in cameras, cars, and robots were made by OEM procurement over long design-win cycles, rather than by the hobbyists who tried the kit. And it competed with Nvidia’s Jetson, whose robotics stack has attracted more than 3 million developers.


Is Claude Code the new CUDA?
Nvidia’s CUDA is widely considered a moat that few competitors can cross, precisely because of developer preference. Nvidia says that CUDA has over 4 million developers and 3,000+ optimized applications, integrated into every major AI framework. Moving away from it means rewriting code, retraining teams, and revalidating entire pipelines.


It seems that a similar dynamic is taking place with Claude Code. A recent article in Business Insider, titled “Anthropic IPO Rides on Claude Code As Startups Splurge on AI,” names some of the startups spending the most. Most of them lean on Claude, though not exclusively.


At Grafana Labs, an observability startup, CTO Tom Wilkie said that token bills have reached tens of millions of dollars a year, with about two-thirds going to Claude and most of the rest to OpenAI’s Codex. Chris Yang, cofounder of the startup Menos AI, said Claude has been improving so quickly that he has little opportunity or willingness to try Codex. And Bryan Bischof, head of AI at Theory Ventures, has personally spent six figures on tokens in a single month, mostly on Claude, yet his team sometimes turns to open-weight models, which are far cheaper.


Even Microsoft felt it: GitHub Copilot was the early leader among developers, yet Satya Nadella admitted in a recent interview with The Pragmatic Engineer that Claude Code caught Microsoft off guard by moving coding from the editor to the command line, and that within about six months the IDE “kind of faded away.” Fittingly, he chose to make the admission on The Pragmatic Engineer, a newsletter written by and for software engineers, which is exactly where the B2D audience lives.


Is Claude Code as sticky as CUDA? The evidence points both ways. On one hand, developers tend to stay with what they know: A 2025 survey by Menlo Ventures (itself an Anthropic investor) found that only 11% of teams switched model vendors over the previous year, while 66% simply upgraded within their existing provider. On the other hand, switching a coding agent is far easier than rewriting CUDA code. CUDA’s moat took nearly twenty years to build, while Copilot’s lead was overturned in about six months. OpenAI’s Codex grew to more than 5 million weekly active users by June (although it’s unknown whether these are users migrating from Claude Code or simply new users), and when Microsoft wanted to cut costs, it canceled most of its Claude Code licenses in one of its largest divisions and moved its engineers to its own GitHub Copilot CLI, even though many of them preferred Claude Code.


IMHO, the IPO was a public vote on exactly this question: Whether developers’ preference for Claude Code is a moat like CUDA, or a loyalty that lasts only until a cheaper deal or a better tool comes along. By making Anthropic the largest IPO on record, investors voted for the moat, at least for now.


Either way, the broader lesson is that B2B and B2C are not the only options. B2D, selling first to the developers who will later carry a product into their companies, is a third go-to-market strategy in its own right. It worked for Nvidia, AWS, GitHub, MongoDB, and now Anthropic. In hardware, as Raspberry Pi, Arduino, and Hailo show, it works only when the company owns its developer community, and those developers are also the ones who decide what goes into production.


The lesson for anyone bringing a new technology to market is that enterprises tend to buy what their engineers already trust. To a large extent, developers are often the bridge between a new technology and the enterprise budget.