Key Takeaways

  • AI adoption is spreading at a radically faster rate and wider scale than the early internet, but it faces a critical, fundamental difference: it's supply-constrained, not demand-constrained.
  • Major tech players like Google, Amazon, Microsoft, OpenAI, and Anthropic are "token constrained," meaning their ability to generate more revenue and deliver more intelligence is directly limited by the physical availability of computing power.
  • There's no unused compute capacity – “not a dark GPU in the world today,” as Gerstner puts it. Every available memory wafer, logic wafer, and kilowatt of power is already being deployed to produce AI tokens, signaling insatiable demand.
  • Despite these physical limits, the industry is preparing for parabolic growth: OpenAI and Anthropic, which started the year with 3 gigawatts of compute combined, are projected to reach 10 gigawatts by year-end and 20 gigawatts next year.
  • America's global leadership in AI, and thus its economic future, depends on maintaining a relentless pace in building out this compute infrastructure, especially in the face of data center moratoriums.

The New Scarcity Driving Intelligence

Forget the dot-com era's "dark fiber" — that abundance of unused internet infrastructure that waited for demand. Today, in AI, it's the exact opposite. Brad Gerstner lays it bare: "There's not a dark GPU in the world today. ...there's not a dark token in the world today." This isn't just about demand; it's about a physical bottleneck. The speed of AI intelligence growth isn't limited by how many users want it, but by how many silicon wafers we can stamp out and how much power we can generate.

Gerstner points out that major players are feeling this crunch. “What did they report on their earnings calls? Google was token constrained. They said if we had more tokens we'd be able to generate more revenue. Same for Amazon, same for Microsoft, same from OpenAI, same from Anthropic.” This isn't a hypothetical. These are the titans of tech, directly stating that their revenue growth is tied to the physical limits of memory, logic, and power. Your ability to create more intelligence, to serve more users, to build bigger models, is literally capped by the number of "tokens" you can produce. It's a hard limit, unlike the software-first scalability many founders are used to.

The Coming Compute Explosion

This scarcity, paradoxically, is fueling an unprecedented build-out. Gerstner details staggering numbers: “OpenAI and Anthropic combined to start the year had three gigawatts of compute. Three combined. They're going to end the year closer to 10 and end next year closer to 20.” That's more than a 6x increase in just two years for two companies. This isn't steady growth; it's parabolic. The fight for intelligence is a fight for compute, and the winners will be those who can secure and deploy it at scale.

This aggressive expansion is the engine behind the "parabolic growth in new models and intelligence" that Gerstner predicts for the next nine months. It's a stark reminder that while software eats the world, it first needs a massive, physical appetite for power and silicon. For founders, this means the landscape is shifting from purely algorithmic innovation to one where access to and efficient use of raw compute becomes a defining competitive edge.

America's Race for Compute Dominance

Gerstner pulls no punches when discussing the national stakes. The AI race isn't just about who builds the best models; it's about who builds the most data centers and secures the most energy. Data center moratoriums, for example, aren't just local planning issues; they're direct threats to the US economy and its leadership in AI. “I'm really worried about making sure that America stays foot on the accelerator, competing globally, and winning the AI race,” Gerstner stresses. This isn't just an investor's concern; it's a call to action for policy and infrastructure development.

For ambitious builders, this underscores that the macro environment — energy policy, infrastructure investment, even local zoning laws — directly impacts your ability to innovate. The future of AI isn't just in the cloud; it's on the ground, in the power lines, and in the factories producing chips. Understanding this physical layer of constraint and growth is key to navigating the next decade of intelligence expansion.

What to Do With This

Stop thinking of AI as infinitely scalable software. Your ability to grow hinges on compute access. This week, map your product's "token flow" – how many tokens does it consume per user, per feature? Then, research the current compute allocation strategies of major cloud providers and key investors. Position your startup not just on algorithmic genius, but on demonstrable token efficiency and a clear strategy for securing the ever-scarcer compute resources required for massive scale.