In the year and a half since ChatGPT launched a new AI wave, a crucial shortfall has dominated headlines: the insatiable demand for GPU computing resources. But when we talk to cloud leaders and data center operators to understand why there’s a compute crunch, it’s clear that access to power is often a bigger bottleneck than access to NVIDIA’s coveted GPUs themselves.
Given our aging, overtaxed electric grid and the huge amount of power required to train state-of-the-art models on the latest chips, it’s incredibly difficult to bring new data centers online in the United States. This is the kind of hard, messy challenge that originally inspired me to study engineering. It’s also, ironically, a problem uniquely suited for AI to help solve.
Software may have eaten the world, but today it gets choked up any time high complexity meets little room for error — when interacting with safety-critical hardware, for example. In these cases, it’s often too hard to encapsulate all the possible edge cases in software, so we rely on the instincts of operators who have been working in these industries for decades, which doesn’t scale.
AI systems, on the other hand, are purpose-built to deal with complexity that’s extremely difficult to program deterministically, and are pretty good at looping in a human when they’ve reached the limit of what they “know” how to do.
This challenge of balancing complexity and criticality is not limited to the constraints of our power grid. The U.S. industrial base is facing a series of risks that compromises our ability to remain globally competitive. This non-exhaustive list includes:
We are at a unique moment in time. Thanks to advances in reinforcement learning, computer vision, and LLMs/VLMs, many industrial use cases once thought impossible are now nearly, if not totally, possible. For example: the ability to predict, simulate, produce, and test new advanced materials in a closed loop process with minimal human intervention (e.g., for use in satellites, or to replace materials we source from conflict zones); robots that help install solar panels in remote desert locations 1000x faster than humans can alone; autonomous controls that more efficiently manage energy-taxing industrial processes, such as cement production or chemical manufacturing; or autonomous factories to reshore production of goods currently too costly to manufacture in the United States. I believe that in 5 to 10 years, our physical world will be completely transformed by AI.
It takes a particular mindset and combination of skills to go after these gnarly problems, combining technical acumen, deep subject matter expertise, and high pain tolerance. We recognize that these deep-rooted issues are not neatly or quickly solved; their timelines are often slowed by hardware schedules, heavy implementations, and government bureaucracy. Often, companies building in these categories require support that looks different from traditional VC-backed software companies, whether related to an industry-specific go-to-market motion, building teams that blend expertise from traditional industry and fast-scaling tech, or interacting with regulatory bodies.
a16z started the American Dynamism practice because we’re in the weeds on these pressing problems, and optimistic that forward-thinking builders can tackle them. We’re deeply invested in easing the country’s operational and structural limitations with AI technology, including autonomous systems, controls, robotics, simulation and spatial computing, advanced chemical and material design, and much more — any use case in a critical industry that today relies on numerical solvers running on CPUs.
With the close of our new $600 million fund, we’re excited to support early stage teams building AI for the physical world. If you’re working in any of these areas, let’s talk!