Circle’s IPO Filing Offers Investors Valuable Insights Into The Stablecoin Giant, & More
Want more research from the ARK Team? Have feedback on our publications? Click here to help inform our content creation.
1. Circle’s IPO Filing Offers Investors Valuable Insights Into The Stablecoin Giant
By: Lorenzo Valente | Director of Research, Digital Assets
Last week, in its second attempt after a terminated Special Purpose Acquisition Company (SPAC) merger valued at $9 billion in 2022, Circle filed for its long-awaited IPO, this time at a rumored $4-6 billion valuation.[1]
For 2024, Circle reported $1.7 billion in revenue, up 15% year-over-year, though net income and adjusted earnings before interest, taxes, depreciation and amortization (EBITDA) dropped 42% and 28%, respectively, to $157 million and $285 million. Meanwhile, USDC’s circulating supply surged 80% to $44 billion.[2]
Investors are debating the trajectory of Circle's distribution costs relative to its revenue. Out of $1.7 billion in revenue last year, Circle allocated $1 billion to distribution partners, of which Coinbase received $900 million,[3] as shown below.
In their revenue-sharing model, Coinbase is taking 100% of the interest on USDC held on its platform, and 50% on USDC held outside Coinbase.[5] This agreement highlights the close relationship between Circle and Coinbase.
Circle’s rumored IPO valuation implies an EBITDA multiple between 13x and 20x, comparable to fintech giants like Coinbase (16x), PayPal (10x), and Block (17.5x at the midpoint).[6] That said, some investors are cautious given Circle's recent drop in profitability and concentrated revenue stream.
Currently, Circle is valued like a typical crypto business—cyclical and interest rate-dependent with a concentrated revenue stream. How Circle frames its investment case will be key. Despite the concerns, Circle will be the only public pure play on stablecoins, the fastest-growing vertical in the digital assets space.
2. Anthropic’s Model Context Protocol (MCP) Is An Emerging Standard That Connects AI Agents
By: Frank Downing | Director of Research, Next Generation Internet
Last week, Sam Altman announced[7] that OpenAI would support Anthropic’s Model Context Protocol (MCP),[8] beginning with its agent software development kit (SDK) and then its ChatGPT desktop application and application programming interfaces (APIs).
MCP is a communication protocol enabling seamless integration between AI systems like chatbots, agents, and other applications, as well as the tools, data, and other external services that live on the web. Effectively, MCP helps bridge the gap between the human language of large language models (LLMs) and the computer code of web APIs.
Historically, developers had to customize integrations between their AI applications and external services to add functionality like web search or database lookups. Prior to MCP, every application developer had to integrate each application, increasing complexity and slowing down the time-to-market.
MCP is a shared standard that developers can use and re-use. The diagram below, based on data from Camel-AI,[9] illustrates the point. Now, for example, AI application developers can add Perplexity’s web search to their apps with little integration.
As an open standard, developers have adopted MCP rapidly. As measured by the website MCP.so,[11] more than 6,000 MCP servers have been created and shared.
Interestingly, the adoption of MCP as an open standard is evolving an ecosystem more like the internet protocol, or the open web, than mobile applications. In contrast to the economics in the open web, owners of mobile operating systems have been extracting high rents from anyone deploying applications in their walled gardens.
Enabling connections between tools/data sources and AI applications, MCP will be able to select the right tools with the right data sources for the job, possibly benefiting the creators of the smartest models that are delegating and synthesizing data most efficiently and effectively.
3. OpenAI’s New Image Generation Capabilities Inundate The Internet
By: Brett Winton | Chief Futurist
At the end of March, OpenAI launched[12] a new image generation model though ChatGPT that has swamped social media with AI-modified content. Here and in the chart below, we highlight five takeaways from that burst of activity:
With its image generation release, OpenAI has strengthened its lead in distribution, as shown below. OpenAI enjoys ~80% share of the endpoint chatbot market, having increased its daily active user (DAU) base by 1.7 million per day, or 14% in the week after releasing its new image generation tool.[13] Its distribution advantage could allow OpenAI to compete in situations where its product capabilities are good enough but not best-in-class. Recently, for example, Google’s Gemini and xAI’s Grok released tunable image generation models that performed at least as well as OpenAI’s. Despite its slightly “late” release, OpenAI won the latest user-acquisition burst, illustrating how a distribution advantage can compound.
Recommended by LinkedIn
The release of ChatGPT's image-generation feature has increased demand for GPUs significantly, overwhelming OpenAI's capacity. CEO Sam Altman noted that the surge led to temporary rate limits on image generation requests and exclaimed, "Our GPUs are melting.”[14] This trend underscores the need for additional AI compute to manage the increased number of users and higher utilization rates, the reported reasons for OpenAI’s Softbank-led $40 billion financing round.[15]
ARK’s research suggests that the demand for AI compute will increase 4.5-fold through 2030, as shown below. OpenAI’s demand supports our forecast. If this use case—transforming memes into Studio-Ghibli-style illustrations—seems silly, the demand it has generated is not: it is a step-change created by the underlying tool.
The popularity of AI-generated images styled like Studio Ghibli raises intriguing IP considerations. Although Studio Ghibli's co-founder, Hayao Miyazaki, has criticized AI-generated art,[16] Japan's copyright law allows AI training on copyrighted content without explicit consent. IP protections are supposed to reward creators but, in this case, Studio Ghibli seems to be benefiting incrementally from AI-driven exposure: interest in My Neighbor Totoro, for example, rose 40% after the AI image-generation release, as shown below. Just as media properties learned that making content available to search was key to their survival in the internet age, making ideas and art available to AI could prove critical to extending IP into the AI age.
More practically, these image generation systems could transform personalized marketing. An advertisement for a dress should be more effective—and merchandise return rates lower—if it depicts the user herself—not some generic model—attractively in the dress. Truly personalized ads now seem within reach and should help online commerce take more share from physical retail stores.
More broadly, this advance in image generation should impact other beneficiaries of AI. Robotaxi developer teams and humanoid robotics manufacturers, for example, use image generation to help with path planning, simulation, and scene interpretation. An advance in the state of the art should translate into better robot and robotaxi capabilities.
Not so long ago, experts claimed that AI systems would never be able to connect datapoints beyond those on which they had trained. A case in point was the inability of AI systems to generate an image of a horse riding an astronaut. Clearly, as shown below, the new models have crossed that threshold, a silly example perhaps but one possible because of a profound advance in AI capabilities. As has been the case consistently with AI, any task that naysayers claim is impossible eventually becomes a waystation along the route of AI’s accelerating advance.
[1] UNITED STATES SECURITIES AND EXCHANGE COMMISSION. 2025. “FORM S-1, CIRCLE INTERNET GROUP, INC.”
[2] Ibid.
[3] Ibid.
[4] Valente, L. 2025. “0/@circle just announced its IPO filing yesterday...” X. UNITED STATES SECURITIES AND EXCHANGE COMMISSION. 2025. “FORM S-1, CIRCLE INTERNET GROUP, INC.”
[5] Ibid.
[6] See Valueinvesting.io. 2025. “COIN EV/EBITDA.” Alphaspread. 2025 “PayPal EV/EBITDA.” Gurufocus. 2025. “Block (XYZ) EV/EBITDA.”
[7] Altman, S. “people love MCP…” X.
[8] Anthropic. 2024. “Introducing the Model Context Protocol.”
[9] Camel-AI. 2025. “How to Use MCP with the OWL Framework: A Quickstart Guide.”
[10] Ibid.
[11] MCP.so. 2025. “Explore MCP Servers and Clients.”
[12] OpenAI. 2025. “Introducing 4o Image Generation.”
[13] ARK Invest estimates based on data from SimilarWeb and SensorTower as of April 4, 2025.
[14] Welch, C. 2025. “OpenAI says ‘our GPUs are melting’ as it limits ChatGPT image generation requests.”
[15] The Guardian. 2025. “OpenAI raises $40bn in deal with SoftBank that values it at $300bn.”
[16] r/Fauxmoi. 2025. “A clip from 2016 of Studio Ghibli co-founder Hayao Miyazaki is trending due to his reaction of seeing AI-generated animation:…’I am utterly disgusted…’ ‘…I strongly feel that this is an insult to life itself…’” reddit.