Every AI pitch deck talks about the model. Almost none talk about what it takes to run that model at scale. That's a problem—because the infrastructure layer is where the real constraints are emerging, and those constraints shape everything from unit economics to competitive dynamics.

Accel's November 2025 Globalscape report puts numbers to the problem: an estimated $4 trillion in AI-related capital expenditure through 2030. The hyperscalers—Microsoft, Google, Amazon, Meta—are each spending tens of billions annually. Data center construction is accelerating globally. And still, demand outstrips supply.

What This Means for Founders

If you're building an AI company, your compute costs are likely your largest variable expense. GPU demand continues to outstrip supply. NVIDIA's dominance in training hardware remains largely unchallenged. Lead times for advanced chips extend for months. This isn't just a procurement headache—it's a strategic constraint that affects your burn rate, your timeline, and your ability to iterate.

The smart founders are thinking about this from day one. Can you build on inference rather than training? Can you optimize for smaller models? Can you secure compute access before you need it? These aren't technical questions—they're business model questions.

"The race for compute is not just about chips—it's about the entire stack: power generation, cooling infrastructure, data center real estate, and the supply chains that connect them all."

— Accel Globalscape Report, November 2025

What This Means for Investors

For investors evaluating AI companies, compute strategy is now a due diligence item. How does this company access compute? At what cost? What happens when they need 10x more? A startup with a great model and no compute strategy is a startup with a hidden risk factor.

The infrastructure layer itself is also an investment thesis. The opportunity extends beyond semiconductors—though that's real—to data center operators, power infrastructure, and cooling technology. AI data centers require 10-30x more power per rack than traditional facilities. Tech companies are signing direct power purchase agreements and investing in generation capacity. Nuclear is seeing renewed interest for reliable baseload power. These aren't AI investments in the conventional sense, but they're AI-adjacent investments with potentially better risk-adjusted returns.

The Takeaway

The AI story that gets told in pitch decks is about models and data. The AI story that determines outcomes is about infrastructure. For founders, understanding your compute constraints is as important as understanding your market. For investors, asking about compute strategy separates the sophisticated from the naive.

Source: Accel, "Globalscape: The Race for Compute," November 2025.