ARK • Disrupt

The proliferation of AI agents is likely to disrupt online commerce massively. On July 9, Perplexity launched Comet to browse and buy products on behalf of users,¹ and last week Shopify began to power OpenAI’s in-app checkout.² Now that traffic associated with generative AI is surging across retail sites, the era of agentic commerce is upon us.

Model Context Protocol (MCP) is a new internet layer powering this shift.³ Much like https standardized secure communication between browsers and websites in the early days of the internet, MCP is creating a framework for AI agent interaction with APIs, databases, and user interfaces, equipping agents with the context necessary to understand, navigate, and transact across the web precisely and reliably.

MCP’s impact on online commerce could be profound. Today, retailers manage a messy tangle of one-off integrations with sales channels like social media, ad networks, and marketplaces. MCP servers simplify those processes by allowing retailers to connect their back-end systems, providing a clearinghouse for commercial data like pricing, availability, and delivery windows. As a result, large language models (LLMs) can query and interpret real-time data and relay information to consumers through interfaces like voice assistants, search agents, smart TVs, and smart glasses, as depicted below.

Source: ARK Investment Management LLC, 2025. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

Importantly, this shift is creating a new paradigm of Agent-Oriented Optimization (AOO). In the internet era, search engine optimization (SEO) was necessary for relevance online. In the agentic era, MCP compatibility is key. AI agents will surface only products that they can access and understand. Retailers without MCP-compatible data could be out of luck.

Last week, after a headline-making hiring spree to kickstart its Superintelligence Lab, Meta disclosed details about its plans for the largest AI clusters on the planet. Scheduled for completion next year, Prometheus is likely to be the first AI data center to consume 1 gigawatt (GW) of power—more power than many small countries consume. Mark Zuckerberg explained that Meta is constructing a number of GW scale clusters, and that the length of the successor to Prometheus, Hyperion, will stretch toward that of Manhattan and scale to 5 GW.⁴

Source: Zuckerberg, M. 2025.⁵ For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security.

Zuckerberg suggested that Meta’s Superintelligence Lab will have “industry-leading levels of compute and by far the greatest compute per researcher.” In a world where scaling laws suggest that, all else equal, more compute equals better models, seeking to build the largest training cluster is becoming commonplace. Elon Musk made similar comments when raising money for xAI to build a “gigafactory of compute”⁶ that would surpass by multiples the largest clusters at the time. Based on xAI’s success in leapfrogging competitors to the frontier of LLM performance with Grok 4,⁷ Meta’s decision to poach top talent and fuel it with as much compute and data as possible is not surprising.

Now that it is intensifying the AI race, we wonder how Meta—once heralded as the open-source AI champion—fell behind xAI, which joined the race so much later. Dylan Patel of AI research firm SemiAnalysis recently detailed the failure of Llama 4.⁸ Contrary to the concept of scaling laws, companies must make tradeoffs in designing and training models. The challenge is akin to racing a Formula 1 car, always one misstep away from spinning out of control instead of riding a smooth curve up and to the right.

In Patel’s view, Meta made several sub-optimal decisions in evolving Llama 4—across algorithms, model design, data, and scaling strategies. Meta reportedly switched web data sources midway through the training run, for example, and failed to clean and deduplicate the new data source properly. Apparently, its engineers did not test algorithmic decisions enough to ensure that they would scale to larger model sizes. Dylan blames a lack of leadership.

With those issues in mind, Meta’s eyewatering acquisition of a 49% non-voting interest in Scale AI, a company that specializes in data and model evaluation, and the acquihire of its founder, Alex Wang, are somewhat understandable. Other hires, including former GitHub CEO Nat Friedman and co-founder of Safe Superintelligence and former machine learning lead at Apple, Daniel Gross, are likely to address Meta’s leadership gap in the AI space.

Given successful acquisitions like Instagram and WhatsApp juxtaposed against its flop in the metaverse, Meta is cultivating the necessary talent, data, and compute to increase its probability of success in the new AI world.

Nearly two years after its launch, Base announced a major step in its mission to make the onchain economy accessible, fast, and inclusive.⁹ What began as an Ethereum Layer 2 chain has become an open stack with three components: Base Chain, Base Build, and Base App. Upgraded significantly with Flashblocks, Base Chain has reduced block times to 200 milliseconds, speeding transactions by 10-fold, while Base Build is supporting developers with new tools like a comprehensive dashboard for building and monetizing mini apps.

Base App is the most transformative update, a reimagined Coinbase Wallet that serves as a full-featured social and financial platform. Built on open protocols, Base App combines trading, content creation, social interaction, and decentralized identity into a single, user-owned experience. It also integrates Farcaster for decentralized social feeds, allowing creators to own and capitalize on their content, without brand sponsorships or large followings. Enabling monetization from day one, Zora will tokenize every post.

Now, users will be able to watch friends trade, discover new tokens, and engage with mini apps—games, finance tools, and creative experiences—embedded directly into their feeds. Among the notable apps are Remix for creating and playing games, Noice for tipping based on engagement, and Decentralized Pictures for co-creating films. The app also enables tap-to-pay USDC transfers and real-time trading, and currently pays 4.1% APY (annual percentage yield) on USDC.

XMTP enables communication through encrypted chat with features like AI-assisted trading and in-chat USDC transfers. Users can build communities, manage group conversations, and transact seamlessly. Supporting this experience is the Base Account, a smart wallet enabling universal, portable identity across the onchain ecosystem, with single-sign-on through the “Sign in with Base” functionality.

Now available to Shopify merchants, Base Pay also is introducing fast, borderless USDC checkout, supporting express payments with no foreign exchange (FX) fees and providing US customers with 1% cashback on USDC-based purchases. With this relaunch, Base is positioning itself to become one of the cornerstones of the open internet, providing decentralization, creator empowerment, and global economic participation. The Base App beta is rolling out now and has invited users and developers to impact the next chapter of the internet.