Key Moments
The $10 Trillion Token Economy — Alex Atallah, OpenRouter & Anjney Midha, AMP
Latent Space PodcastKey Moments
AI token economy to hit $10T in 10 years, but new fraud types require robust security infrastructure like Stripe's acquisition of OpenRouter.
Key Insights
The AI token economy is projected to reach $5 trillion in the next 5 years and $10 trillion within the next decade, creating a massive target for malicious actors.
OpenRouter's auto-router functionality allows for dynamic model selection based on performance and cost, providing developers with flexibility previously unavailable.
The early AI landscape lacked a centralized platform for discovering and accessing various large language models, a gap filled by OpenRouter, differentiating it from Hugging Face.
Discord served as an early incubator for AI innovation, hosting communities for models like Midjourney and providing a testing ground for AI applications before dedicated platforms existed.
Mistral's 8x7B model release in late 2023 triggered an 80% price drop in AI model inference, marking a significant shift towards cost-effective open-weight models.
Stripe's acquisition of OpenRouter is driven by the need for robust security and fraud prevention infrastructure to manage the burgeoning $10 trillion token economy, especially with the rise of AI agents.
The burgeoning AI token economy and the threat of fraud
The discussion highlights the explosive growth of the AI token economy, projected to reach $5 trillion in five years and $10 trillion in a decade. This rapid expansion, analogous to the growth of online payments in the 1980s and 90s, creates significant value that attracts malicious actors. As the value of tokens increases, so does the incentive for bad actors to exploit the system. The conversation emphasizes that as any system scales and its payload becomes more valuable, more malicious entities will attempt to access that value. This necessitates the development of new infrastructure to handle the escalating fraud risks associated with this rapidly growing digital economy. The analogy to online payments, which required entirely new fraud-prevention solutions, sets the stage for the challenges in the AI token space. The sheer scale of potential fraud, with blocked financial volume ten times larger month-over-month, underscores the urgency and complexity of securing this new frontier.
OpenRouter's emergence as a multi-model routing layer
OpenRouter was founded to address the growing need for model diversity and a unified platform for accessing various AI models. Early in 2023, the AI landscape was dominated by a few players like OpenAI. The release of LLaMA and subsequent open-weight models like Alpaca demonstrated that high-quality models could be created more affordably. Alex Atallah, co-founder of OpenRouter, recognized the potential for an ecosystem of diverse models but found existing platforms like Hugging Face insufficient for discovering and utilizing them, especially closed-source models. OpenRouter aimed to be a neutral routing layer, offering an API experience combined with a marketplace that links to various models. Its auto-router feature dynamically selects the best model for a given task based on performance, cost, and developer preferences. This multi-model approach was initially met with skepticism, with some VCs dismissing it as a mere 'wrapper,' but it has since become industry consensus.
The pivotal role of open-weight models and developer experience
The rise of open-weight models, spurred by releases like LLaMA and Mistral, democratized AI development and fostered innovation outside of large labs. Anjney Midha, from his time at Discord and as an investor, saw firsthand the limitations of relying on closed-source models. When Discord sought to implement content moderation using OpenAI's GPT-3.5, they encountered refusals because their specific moderation policies, tailored to individual servers, often violated OpenAI's built-in safeguards. This highlighted the need for greater control over model behavior and the importance of open-weight models. Furthermore, the developer experience was often poor. Researchers would release models, but there was no easy way for developers to access, deploy, and manage them. OpenRouter stepped in to provide this crucial layer of infrastructure, offering a unified API, key management, and an easier path to production for these diverse models.
Discord as an AI incubator and the rise of community-driven innovation
Discord played a significant role as an early incubator for AI applications. Communities around projects like Axie Infinity and Midjourney thrived on the platform, providing a space for users to interact with AI models, share prompts, and learn from each other. Midjourney, in particular, leveraged Discord to overcome the poor retention rates it experienced with its standalone web application. By allowing users to see others' creations and prompts within the Discord server, Midjourney fostered a sense of community and facilitated learning. This community-driven approach, where users could easily replicate and adapt prompts, proved highly effective. OpenRouter benefited from this ecosystem, observing how users shared links and demonstrating the demand for accessible AI tools. The success of AI applications within Discord highlighted the need for platforms that could manage and distribute these models effectively.
Mistral's impact and the AI model price wars
The release of Mistral's 8x7B model in late 2023 marked a turning point, offering performance competitive with leading models at a significantly lower cost. This event triggered an 80% price drop in AI model inference, sparking intense competition among model providers. Open-weight models became a viable alternative to proprietary ones, forcing major players to re-evaluate their pricing strategies. OpenRouter facilitated this price competition by providing a platform where developers could easily switch between models and access the most cost-effective options. The leaderboard feature on OpenRouter became a crucial tool for developers to discover the best models for their needs, driving further innovation and adoption of open-weight solutions.
Stripe's acquisition of OpenRouter and the future of AI security
The acquisition of OpenRouter by Stripe is a strategic move to bolster security and fraud prevention in the burgeoning AI token economy. Stripe, known for its robust payment infrastructure and fraud detection capabilities (Stripe Radar), recognized the escalating threat of fraud as the AI economy grows. The conversation draws parallels to the evolution of online payments, where new value flows necessitated new security measures. With the projected $10 trillion token economy, the potential for fraud by both human actors and increasingly sophisticated AI agents is immense. Stripe aims to leverage OpenRouter's platform and data to build a comprehensive security shield for the AI ecosystem, preventing misuse and ensuring trust. While OpenRouter will maintain its brand and product roadmap, the partnership will accelerate its development and security efforts, creating a more robust and trustworthy environment for AI applications.
The importance of focus and specialization in AI
Both speakers emphasized the critical importance of focus in the AI space. Anjney Midha highlighted how Anthropic's success was driven by a deliberate focus on programming tasks, despite other lucrative AI avenues like image and video models emerging. This laser focus allowed them to excel and build a billion-dollar company. Alex Atallah echoed this sentiment, explaining that while OpenRouter could have explored features like fine-tuning services or memory management, they chose to concentrate on their core mission: building a neutral, scalable marketplace for AI models. This focus allows them to serve their target customer—the developer—exceptionally well, building a strong brand recognized for its specific value proposition. Trying to be everything to everyone can dilute efforts and hinder product-market fit, especially in a rapidly evolving field.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Studies Cited
●Concepts
●People Referenced
Common Questions
The token economy refers to a new type of value unit being transmitted online. It's projected to reach around $5 trillion in the next five years and potentially $10 trillion within the next decade, driven by increasing value and demand.
Topics
Mentioned in this video
Co-founder of OpenRouter, discussed his product philosophy and the evolution of AI models.
Mentioned in the context of Affirm's business model, suggesting recurring patterns in value creation and infrastructure needs.
Founder of Affirm, mentioned in relation to the evolution of payment and security infrastructure.
Co-founder of OpenRouter, discussed his background at Discord and Anthropic, and his insights on AI development and market dynamics.
The founder of Stanford Review, mentioned in the context of the publication's origins.
Former director of AI at Tesla, mentioned as an influential figure who uses OpenRouter's leaderboards.
An early platform for accessing and sharing AI models, considered the closest alternative to OpenRouter at the time, but with limitations.
The company developing Devin, whose product lead previously created Dream Tavern.
Mentioned as one of the early players in the AI model landscape alongside OpenAI.
A company whose dominance in search is compared to the potential dominance of a single AI model provider. Also mentioned in relation to DeepMind's model deployment.
A payment processing service mentioned for comparison with Stripe's approach to fraud detection.
A leading AI research lab whose models like GPT-3.5 and GPT-4 were discussed in relation to their closed-source nature and impact on the LLM ecosystem.
A company that Anjney Midha invested in, mentioned in the context of the AI ecosystem.
Mentioned in the context of Google's rapid model deployment capabilities.
An interesting new model that attracted a different user base to OpenRouter, focusing on productivity and content creation.
A company mentioned for comparison regarding the recurring patterns of value creation and the need for security infrastructure.
An AI research company where Anjney Midha previously worked, and an early investor in, discussed in the context of LLM development and the need for open models.
An experimental model developed by OpenRouter that combined results from multiple LLMs.
A payment processing company that acquired OpenRouter, known for its fraud detection capabilities and developer-friendly approach.
A European payment provider mentioned alongside Stripe for its dominance due to fraud detection capabilities.
A company that Anjney Midha invested in, mentioned in the context of the AI ecosystem.
An application that adopted features like auto-routing and focused on agent memory and skill management.
A company mentioned as having a similar data-gathering model to Stripe's approach.
An AI model developed by Anthropic, discussed in the context of its potential for content moderation and the need for open models.
A powerful open-weight model from Mistral that significantly impacted the LLM market with its performance and cost-effectiveness.
A platform and marketplace for AI models, aiming to simplify model access and discovery for developers.
An AI software engineer product, discussed in relation to Anjney Midha's previous work and product ideas.
Mentioned as an example of an attempt to build programmable agent memory.
A company mentioned in relation to Anjney Midha's ventures.
An open-source text-to-image model that spurred innovation and the development of custom 'Midjourney-like' applications, highlighting the need for APIs.
A communication platform where Anjney Midha managed the platform and dealt with crypto and NFT-related security issues, and later integrated AI models.
An AI model that significantly improved coding capabilities, driving changes in application development on OpenRouter.
A foundational open-weight language model released by Meta, which spurred significant innovation and the development of many subsequent models.
An AI image generation model mentioned in comparison to Midjourney's user experience.
An earlier version of OpenAI's language models, discussed in comparison to LLaMA's performance.
An early open-weight LLM fine-tuned from LLaMA, demonstrating the feasibility of creating powerful models with limited resources.
An early AI model from OpenAI that Discord integrated internally, highlighting the limitations of closed-source models for content moderation.
A virtual friend bot for Discord, developed as an early use case for AI models.
An NFT marketplace whose transaction volume was significant for Discord, and with which integration was discussed.
OpenAI's flagship model, discussed as a benchmark against which Mistral 7B and other models were compared.
An early application that was successful on the OpenRouter platform.
A product principle discussed by Alex Atallah, framing products as a blend of data subscription and publication.
Distributed Denial of Service attacks, faced by Discord during its growth in the crypto space.
Discussed as a future source of malicious activity, analogous to human bad actors, necessitating robust security measures.
A growing problem in the token economy, involving various types of illicit activities, which OpenRouter and Stripe are working to combat.
More from Latent Space
View all 256 summaries
99 minRunway’s Bet Beyond Video: World Models, Robotics, and the Neural OS — Anastasis Germanidis
92 min🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)
122 min🔬 Google's AI Scientist Started as an Attempt to Automate Kaggle — John Platt, Google Fellow
143 minWhy We Made Jev — Diogo Almeida, TypeSafe Co-founder & CEO
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free