Answer extracted from the Invest Like The Best podcast — listen to the full episode below.
What regulatory and infrastructure barriers are preventing rapid scaling of compute capacity for AI development?
The real bottleneck is not a technology, capability, or capital problem—it's regulatory and community acceptance barriers. Building data centers requires convincing communities they are necessary, and making nuclear power competitive requires proving safety and enabling construction of Small Modular Reactors (SMRs). Nothing moves the needle at sufficient scale before 2030.
Why 2030 Is the Hard Deadline
Sarah spoke with a hyperscaler infrastructure leader who made a sobering assessment: no meaningful progress on compute capacity constraints will materialize before 2030. This timeline is not set by engineering capability or lack of capital—it reflects the regulatory and approval processes that govern data center construction and nuclear energy deployment.
The challenge is compounded by the fact that global AI development has become intensely competitive, creating pressure for faster scaling at a moment when infrastructure timelines are fundamentally constrained by political and social processes outside the control of AI labs.
The Real Bottleneck: Community and Regulatory Permission
Data center deployment is not blocked by construction technology or energy availability. Instead, the bottleneck is convincing local communities and regulators that new facilities are necessary. This is a governance and persuasion problem, not an engineering one.
Similarly, nuclear energy—the most plausible long-term solution for AI compute power consumption—requires two regulatory breakthroughs: proving safety protocols at scale and permitting construction of Small Modular Reactors (SMRs). Both depend on policy frameworks and public trust, not on advances in reactor design.
Sarah — Investor at Greylock, focused on AI, biology, defense, and robotics. Sarah co-founded or leads a venture fund with partners including Mike, Pranav, and Bella. She brings deep conviction about high-agency founders and ecosystem resilience, and tracks over 250 people at the frontier of AI innovation.
The implications are worth understanding in detail: this infrastructure constraint may reshape how labs allocate compute and could favor companies capable of navigating regulatory environments over those relying purely on engineering talent.
Key takeaways
- Compute availability and energy supply are critical constraints over the next five to ten years, not near-term solvable problems.
- The bottleneck is regulatory and alignment-based, not technological or financial.
- Data center scaling requires winning community and political support; it's a persuasion challenge, not an engineering challenge.
- Nuclear power (SMRs) is essential for long-term compute energy, but requires both safety validation and construction permitting—both politically constrained processes.
See also
What psychological challenges do frontier AI researchers face when working at scale in compute-intensive environments?
Sarah observes that many researchers feel disempowered by the massive scale and compute requirements of frontier labs. A large contingent believes either that only a handful of labs will lead AI development or that smaller, more agile teams offer genuine advantages in research velocity.
What concerns do investors have about research-heavy AI companies when assessing capability and impact?
Sarah expresses concern that many investors are making large-scale research bets without fundamental intuition or a grounded point of view on the business case, competitive moat, or real-world deployment path of the technology they are backing.
How should venture investors balance conviction-based decision-making with seeking external validation on investment theses?
Sarah describes her process as starting with instinctive conviction on people and ideas, immediately rating them 8 or 9, then spending days to weeks conducting deeper due diligence to test whether that initial signal holds up under scrutiny or should be revised.