Shane Greenstein on Co-Invention and the Geography of AI Innovation

Episode Summary

Executive Summary: Scott Walston and Shane Greenstein examine the AI data center boom through the lens of location, demand, and co-invention. Greenstein argues that value from digital infrastructure depends on adaptation after invention, while cautioning that hyperscale AI spending may outrun near-term revenue. He highlights shifting demand, agglomeration in places like Ashburn, and the role of policy, subsidies, and domain expertise in turning technical advances into value.

Main Topics: Data center location drivers (Priority: 5/5): Greenstein outlines the enduring factors in site selection: customer proximity, electricity, internet connectivity, water, and land. He emphasizes that the relative importance of these inputs has changed over time as use cases evolved. Why Ashburn/Loudoun County became a data center hub (Priority: 5/5): The conversation explores why network interconnection and trunk-line geography helped Ashburn emerge as the main U.S. agglomeration, while similar effects did not appear in places like Palo Alto, Dallas, or Chicago. AI hyperscale buildout and uncertainty of returns (Priority: 5/5): Walston presses on whether massive AI infrastructure investments will generate enough revenue. Greenstein distinguishes between firms seeking new revenue and incumbents protecting existing revenue, while remaining skeptical that all capex will pay off quickly. Changing demand for cloud and compute (Priority: 4/5): The discussion contrasts older demand for close-by, heavily controlled data storage with today’s greater willingness to outsource to cloud providers for security, scale, and operating-budget advantages. Co-invention and the geography of innovation (Priority: 5/5): Greenstein explains his core framework: the most valuable innovation often comes after the core invention, when firms adapt technology to real use cases. Incremental co-invention is distributed; novel co-invention clusters in places with both technical and domain expertise. Policy, subsidies, and local incentives (Priority: 4/5): The speakers discuss local tax abatements and federal subsidy debates around AI infrastructure. Greenstein notes local construction benefits and year-round tax base changes, but warns economists are skeptical of industrial policy and tariffs. Need for domain experts in AI applications (Priority: 5/5): Greenstein argues that AI value creation requires subject-matter expertise—e.g., architects for architecture and doctors for medical applications—because technical models alone do not create scalable, workflow-integrated value.

Key Arguments: Data center site selection still depends on the same categories as before, but firms now prioritize them differently based on the application and service requirements. Ashburn/Loudoun County’s success came from being a key interconnection point where major fiber trunks met, rather than from a generic attraction to data-center economics. The current AI infrastructure boom is partly demand-driven, but also supply-driven by transformers, GPUs, and new engineering methods that reward scale. Not all AI spending is aimed at creating new revenue; incumbents may be investing to protect current revenue streams or improve existing products. Cloud adoption increased because third-party providers offer better security, more scalable services, and a more manageable operating-budget treatment than owning IT infrastructure internally. The budget classification matters: paying a cloud provider often sits in operating expense, whereas self-built infrastructure is capital expense, and organizations strongly prefer the former. AI value creation will depend on co-invention, meaning adaptation work after the core invention; the most valuable uses will require both technical skill and domain knowledge. Novel co-invention tends to cluster in cities with dense technical and industry expertise, especially Seattle, San Francisco, New York, and Los Angeles. Most local data-center benefits are short-term construction jobs, while long-term jobs are relatively few but can still matter in rural places with otherwise limited economic activity. Economists are generally skeptical of industrial subsidies and especially tariffs, though Greenstein allows that strategic national-security arguments could justify some support for frontier AI capacity.

Data Points: Podcast date: Wednesday, November 5th, 2025 - Opening introduction by Scott Walston Reliability standard: 99.9% to 99.999% uptime - Discussion of 4.9 and 5.9 reliability expectations for data centers Hurricane anecdote: One Houston data center stayed open during a hurricane - Illustrates redundancy and backup power planning; the larger issue became employee access and food logistics Current frontier AI workers: Less than 10,000 people - Greenstein estimates the number of people directly doing the core frontier model work is small ChatGPT 3.5 timing: November/December 2022 - Used as the datable market turning point for consumer demand in LLMs Copilot investment: $1 billion - OpenAI access purchased by Microsoft, described as yielding strong value for GitHub Copilot Later OpenAI-style investment: $10 billion or whatever they've put in - Reference to much larger follow-on commitments for future options value Construction timeline: 18 months to 2 years max - Estimated period for construction-related local economic benefit from a data center build Seattle and San Francisco: Major sources of novel co-invention - Cities identified as especially important for valuable adaptation work New York and Los Angeles: Also major sources of novel co-invention - Additional media-capital cities that generate novel applications and value Washington, DC: Distant fifth in novel co-invention - Not a top source, but still relevant in Greenstein’s geography of innovation discussion Sales tax savings: 5% or 6% in Illinois (or comparable tax states) - Example of local abatements on server and rack purchases Uptime nomenclature: 4.9 / 5.9 reliability - Industry shorthand for near-continuous availability requirements

Pivotal Quotes: "the innovation that happens after technology exists when companies adapt it to new uses and realities" — Narrator explaining Shane Greenstein: Definition of co-invention and the podcast’s central analytical frame "the frontier is uniquely affiliated with the hyperscalers" — Shane Greenstein: Discussion of why some argue the federal government may want to subsidize AI infrastructure domestically "What we measure tends to get more attention than what we don't measure" — Shane Greenstein: Reflection on why co-invention is under-studied despite its importance

Implications: AI infrastructure, cloud adoption, and local tax incentives will matter only if firms can convert technical scale into real applications. The winners will likely be places and organizations that combine compute, domain expertise, and adaptation skills—not just capital.