The Return of Vertical Integration

The Pendulum Swings Back

For two decades, the tech playbook was simple: unbundle everything. Rent, don't buy. Cloud over on-prem. SaaS over software. Outsource infrastructure, focus on product. The specialists won—AWS crushed on-prem data centers, Stripe owned payments, Twilio handled communications.

That era is over.

The smartest companies in AI are doing the opposite. They're going vertical—and fast.

OpenAI is building massive data centers. Anthropic is designing custom silicon. xAI just broke ground on a facility that will house 100,000 GPUs under one roof. Meta, Google, and Microsoft have been doing this for years, but now the startups are following suit.

Why? Because the old rules don't apply when you're operating at the edge of what's physically possible.

When Horizontal Breaks Down

Horizontal integration works beautifully when:

• Markets are mature and standardized
• Marginal costs decrease with scale
• Commodity providers can meet your needs

AI training violates all three.

1. The market isn't mature. There is no "standard" for training a frontier model. Every architecture is experimental. Every training run is a moonshot. You can't outsource R&D when the entire field is R&D.

2. Costs don't decrease—they explode. GPT-4 reportedly cost over $100M to train. Gemini Ultra? Probably more. At that scale, cloud providers aren't giving you economies of scale—they're charging you a coordination tax. When you're the single largest customer, you might as well own the infrastructure.

3. No one can meet your needs. Frontier AI companies need power delivery measured in gigawatts, networking fabrics that don't exist yet, and cooling systems borrowed from nuclear plants. AWS and Azure are great—but they're optimized for enterprise SaaS, not for melting sand into superintelligence.

So you build it yourself.

The New Vertical Stack

Here's what vertical integration looks like in 2026:

Energy: Companies are signing deals directly with power plants, even exploring nuclear SMRs (small modular reactors). When your training run needs 50MW continuous, you can't rely on the grid—you are the grid.

Data Centers: Purpose-built facilities optimized for GPU density, not general compute. Liquid cooling, custom power distribution, fiber backhaul measured in petabits. These aren't "cloud regions"—they're weapons.

Silicon: Google has TPUs. Tesla has Dojo. Meta has MTIA. Designing your own chips used to be insane. Now it's table stakes. When you're spending billions on training, a 20% efficiency gain from custom silicon pays for itself in months.

Networking: Companies are laying their own fiber between data centers. When you're moving petabytes of gradient updates between nodes, you don't want to share bandwidth with Netflix traffic.

This isn't just infrastructure—it's a moat.

Why This Matters Beyond AI

The rebundling trend isn't limited to AI. It's a signal of a deeper shift.

When competitive advantage lives in the stack, you own the stack.

Stripe started as API-only payments. Now they're building banking rails, issuing cards, handling Treasury. Why? Because payment processing was commoditized. The alpha moved down the stack.

Shopify started as hosted e-commerce. Now they run warehouses, offer financing, and compete directly with Amazon logistics. The margins migrated from software to operations.

Tesla didn't just make EVs—they built Gigafactories, battery tech, and charging networks. The car was never the moat. Vertical integration was.

The pattern is clear: When you're pushing the frontier, you can't rent your edge.

The Downside (And Why It's Worth It)

Vertical integration is expensive, slow, and risky.

You trade agility for control. You trade flexibility for performance. You commit capital that could have gone into product or sales.

For most companies, this is a terrible trade. If you're building a SaaS tool, a marketplace, or a social app—stay horizontal. Rent everything. Move fast. Optimize for iteration speed.

But if you're building at the edge of the possible—training frontier models, running real-time inference at global scale, designing new compute paradigms—the calculus flips.

The companies that win the next decade won't be the ones with the best APIs. They'll be the ones with the best infrastructure.

What This Means for Builders

If you're starting a company today, ask yourself:

Is your competitive advantage in software, or in the stack below it?

If it's software—great. Use Vercel, Supabase, Stripe, and every other best-in-class tool. Build fast, stay lean, iterate.

But if your edge is in performance, scale, or novel compute—start thinking vertically. Own the parts of the stack that matter. Rent the rest.

The pendulum has swung. Vertical integration is back.

And this time, it's not about protecting a dying business model. It's about building things that literally couldn't exist any other way.

— Jens

---