Key Moments
Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Infrasctructure, Enterprise AI, SaaS
Stanford OnlineWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
AGI is already here, but current enterprise adoption is minimal because AI lacks organizational context, not because the AI itself is insufficient. True impact requires human-led process rewiring.
Key Insights
Guest speaker Ali Ghodsi argues that Artificial General Intelligence (AGI) has already been achieved and is smarter than many people, contrary to popular belief.
Despite having AGI, most enterprises are seeing minimal AI adoption and productivity gains because AI models lack the nuanced organizational context that human employees possess.
Historically, technological revolutions, like the introduction of PCs and electric engines, took decades to show productivity gains due to the need for extensive organizational process rewiring.
Barriers to entry and switching costs in software are significantly decreasing, leading to increased competition and the need for software companies to innovate or face obsolescence.
The value accrual in technology stacks tends to move upwards; while infrastructure and models are currently dominant, applications are expected to ultimately capture the most value.
Open-source models are rapidly improving and applying pricing pressure, suggesting that the core business of providing proprietary frontier models will become a low-margin, economies-of-scale game akin to Amazon's book selling.
The premature pursuit of superintelligence is unwarranted
Ali Ghodsi, CEO of Databricks, begins by addressing the widespread anxiety surrounding AI, particularly the quest for 'superintelligence' and the fear of missing out on AGI. He argues that this stress is unwarranted and leads to tunnel vision. Ghodsi controversially asserts that AGI has already been achieved, stating it's 'smarter than many of the people that you interact with.' He draws a parallel to his time at UC Berkeley's AMPLab in 2009, where leading AI researchers agreed they had met the AGI benchmarks of that era. The current narrative of not having AGI, he suggests, is a shifting of goalposts. This perspective implies that the intense focus and massive investment in GPUs and data centers, driven by the perception of AGI being just out of reach, might be misplaced. The real challenge, according to Ghodsi, is not developing more advanced AI but integrating the existing capabilities effectively into real-world processes.
The context gap: Why AGI isn't transforming enterprises
Despite the existence of AGI, its impact within most organizations is negligible, resembling a "shadow of AI." Ghodsi explains this disconnect by highlighting the critical lack of organizational context. Humans possess vast amounts of implicit knowledge—the "context in our heads"—gathered through years of experience, relationships, and understanding of internal workflows. AI models, lacking this human context, make "stupid mistakes" and are therefore deemed "useless" in practical enterprise settings. He uses the analogy of a company's go-to person, the one who "knows everything," emphasizing that this critical, tacit knowledge is not captured or understood by current AI systems. To achieve meaningful impact, Ghodsi stresses the necessity of transferring this deep context from human experts to AI, transforming how old-school businesses operate by integrating these processes into AI agents.
Historical parallels of technological adoption
Ghodsi illustrates the slow adoption of transformative technologies with historical examples. He references a 1990 Stanford research paper on the transition from the "dynamo to the computer," which shows that it took decades for new technologies to impact economic productivity. When personal computers (PCs) were introduced, users often employed them as mere typewriters, printing documents and filing them, yielding no significant productivity gains. Similarly, the electric engine replaced steam engines in factories, but it took around 40 years (from 1880 to 1920) to see widespread productivity increases. This was because factories were initially rewired by simply replacing the steam engine with an electric one, without fundamentally redesigning the factory floor to leverage the new technology's capabilities, such as distributed power. This historical pattern suggests that the current lack of AI-driven productivity in enterprises is not due to AI's limitations but the inherent difficulty and time required for organizations to "rewire" their processes and adapt to new technological paradigms.
The evolving landscape of software and its moats
The discussion shifts to the state of software, with Ghodsi challenging the notion that "software is dead." He argues that companies like OpenAI and NVIDIA, which are built on software and advanced hardware, are thriving, indicating software's continued relevance. However, he acknowledges two significant changes: lower barriers to entry and reduced switching costs. It's now cheaper and easier to produce software, a capability accessible to everyone, including established players like Databricks. Simultaneously, features like conversational AI interfaces are eroding traditional switching costs associated with user interface learning curves and data migration. This increased competition necessitates that software companies operate more efficiently. Ghodsi points out that software is not the only "moat"; other factors like economies of scale (Amazon AWS), brand (Ferrari), trust, patents, and proprietary data also serve as crucial competitive advantages.
Innovation as the key differentiator for software companies
Ghodsi offers a stark warning to software companies that have not innovated. If a company's software remains unchanged after a decade while revenue has grown, it should be deeply worried. Such companies likely lack the innovation muscle and are vulnerable to newer, more agile competitors who can leverage lower barriers to entry to create superior solutions. However, companies that have a history of innovation or are now spurred by competition to adapt and build great AI applications can thrive. These incumbents often possess advantages like existing customer bases, data, and scale, which can be leveraged to their benefit, provided they effectively integrate AI and innovate their offerings. The ability to adapt pricing and cost structures will also be critical.
The jagged frontier of AI capabilities
Referencing Ethan Malik's "jagged frontier" concept, Ghodsi notes that AI excels in specific areas like customer support and software engineering but struggles significantly in others. Within Databricks' customer base of 20,000, AI is beginning to show success in some tasks, but many companies are still only partially benefiting. Support, often cited as an area for AI automation, is surprisingly difficult because it typically handles the complex issues humans themselves struggle with. Ghodsi argues that current AI models cannot fully automate complex support scenarios without the deep contextual understanding that Databricks' own support engineers possess. The majority of enterprises are operating at the "rough" end of the AI frontier because they lack this essential data and contextual integration.
Rewiring processes: The human challenge in AI integration
Ghodsi reiterates that the primary challenge in AI adoption is a human and process problem, not an AI one. He illustrates this with an example at Databricks concerning building connectors to various enterprise systems. While an individual could theoretically build a basic connector in two days using LLMs, the internal process, involving product managers, testing, and dedicated engineers, took three quarters (nine months). Even with AI assistance, the process was only compressed to seven and a half months. A breakthrough came when a first-principles approach was applied, leading to a complete rewiring of the process. This involved reducing requirements gathering time, outsourcing setup of external systems, and restructuring team workflows for parallel development across multiple connectors. This resulted in seven connectors being shipped in one quarter. This transformation had little to do with more advanced AI models and everything to do with human-led process change and organizational refactoring. This extensive rewiring of processes is what Ghodsi believes is necessary for broad AI impact.
Strategic investment and the future of value accrual
Ghodsi, while not an investor and declining to give financial advice, shares his perspective on where value will accrue in the AI supercycle. He believes applications will be the ultimate winners, suggesting a focus on early-stage startups. He draws parallels to the early internet, where seemingly niche ideas like Uber (taxi business), Amazon (selling books), Airbnb (renting rooms), and Twitter (short texts) eventually became multi-trillion dollar companies, disrupting established industries. He identifies healthcare and education as potentially massive future markets for AI-driven innovation, citing their significant GDP contribution, high propensity to pay, and current inefficiencies. These sectors, he argues, could host trillion-dollar companies if they leverage AI effectively, building advantages through data and economies of scale. He also notes that value consistently moves up the tech stack, from infrastructure to applications, and with open-source models rapidly improving, the proprietary model layer will likely become a low-margin business driven by economies of scale, similar to cloud providers.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Books
●People Referenced
Common Questions
The speaker contends that Artificial General Intelligence (AGI) already exists and is smarter than many people. The issue isn't the lack of AGI, but the failure of organizations to integrate it effectively due to a lack of contextual understanding within AI models.
Topics
Mentioned in this video
Cited as an example of a company with a significant moat due to economies of scale, demonstrating how large-scale operations can create a competitive advantage.
Mentioned as a future frontier model that users are eager to access, highlighting the demand for cutting-edge AI capabilities.
Mentioned as a leading proprietary model that users are eager to access, contrasting with the growing capabilities of open-source models.
A model released by Moonshot, described as the best model ever produced at the time of its release but quickly surpassed due to the rapid pace of AI development.
A DataBricks generative AI product that the speaker uses frequently for numerical and quantitative decision-making, such as ROI calculations and cost analysis.
Mentioned as a leading proprietary model that users are eager to access, contrasting with the growing capabilities of open-source models.
Used as an example to illustrate switching costs in software, where users are hesitant to switch due to UI unfamiliarity and data transfer.
A preferred AI coding tool used by the speaker, liked for its diff functionality. Its usage was discontinued after Elon Musk's acquisition of Twitter due to potential access issues.
Mentioned as an enterprise system that needs to be set up for testing connectors, highlighting the complexity beyond DataBricks' core offerings.
Mentioned as a leading company in the AI supercycle, involved in chip design and deployment powering AI, and discussed in the context of where value accrues in the AI stack.
Cited as an example of value accruing in virtualization software, which later also became commoditized, fitting the pattern of value moving up the tech stack.
A Chinese company that released the Qimi 2.6 model, which is described as the best model ever released if it had come out earlier in the year, highlighting the rapid pace of open-source development.
AI company mentioned in the context of the 'software is dead' debate, with its valuation questioned if software itself is obsolete. Also mentioned in relation to cursor access.
Cited as an example of a 'weird' but successful internet-era business that disrupted a traditional industry (taxis).
The company where the speaker works, discussed in terms of its evolution from a data business to a lakehouse to an AI business, and its internal processes for AI implementation and product development.
Used as an example of a disruptive business that could have emerged earlier but took time to manifest, highlighting the role of timing and individual insight.
A prominent AI research and deployment company, discussed in the context of the AI supercycle and its valuation despite being a software company.
Mentioned as a competitor in the AI space, alongside OpenAI, in the context of the rapid development and competition.
Mentioned as an example of a simple idea (sending short texts) that became a major platform, illustrating how disruptive innovations can emerge from unexpected places.
Mentioned as an example of an enterprise system for which DataBricks builds connectors, highlighting the time and process involved in creating them, and how AI can potentially speed this up.
Used as an example of a brand that constitutes a significant moat, indicating that strong brand recognition and reputation can be a powerful competitive advantage.
Used as a historical example of a company that once dominated the PC market but eventually saw its value commoditized as value moved up the tech stack.
Mentioned as an example of a company whose value accrued in the software layer following the commoditization of hardware (like IBM's PCs).
Started by selling books online, it grew into a massive company (AWS), demonstrating how a seemingly simple idea can leverage secular trends to immense success.
Mentioned in the context of early internet infrastructure concerns like routing and BGP, which the speaker's early work was related to.
Mentioned as a leading AI company that everyone is focused on, as part of the current AI supercycle.
Cited as a source for a statistic claiming 95% of PCs are failing, used to illustrate a potential disconnect between AI advancements and enterprise adoption.
A book recommended for understanding business moats beyond just software, covering concepts like economies of scale, brand, trust, patents, and data.
Author of 'Seven Powers', whose work on process power is recommended for investors, particularly for assignment option one.
Referred to as the 'god of AI' and the 'Michael Jordan of AI' in the context of his work at AMPLab in 2009, highlighting his prominent role in the field.
Implicitly referred to as 'Jensen' managing Nvidia, happy because value accrues to one side of the AI stack (likely infrastructure/chips).
The source of a chart discussed by the speaker, illustrating the 'jagged frontier' of AI capabilities across different tasks.
Used as an example of someone who made a long-term bet on the internet, starting with a modest business (books) that eventually led to Amazon's 'everything store' model.
More from Stanford Online
View all 86 summaries
48 minStanford CS547 HCI Seminar | Spring 2026 | Toward Ontological Multiplicity in AI and Computing
48 minStanford CS547 HCI Seminar | Spring 2026 | Just-in-Time Objectives for Specialized AI Interactions
97 minLive from Stanford Health AI Week
35 minStanford Robotics Seminar ENGR319 | Spring 2026 | Towards Trustworthy Autonomy
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