The Gravity Is in the Ground

Gravity Is in the Ground

9:05

There's a pattern running underneath three separate AI headlines from the last 90 days, and if you're only reading them as individual stories, you're missing the one thing they all agree on.

Anthropic took operational control of Colossus 1 — xAI's 300-megawatt Memphis supercomputer complex, 220,000 NVIDIA GPUs — through a deal reportedly running $1.25 billion per month through May 2029. The same week, they doubled Claude Code's session limits and removed peak-hour throttling. NVIDIA H100 one-year rental contracts have surged nearly 40% since October 2025, from $1.70 to $2.35 per GPU per hour. On major cloud platforms the number is worse — AWS runs H100 instances at $6.88 an hour. Azure tops $12. Lead times for bare-metal capacity are running 5 to 6 months. Some commitments are stretching to 52 weeks. Private equity dropped $45.7 billion into US data centers in 2025 — the highest total in at least five years, representing 72% of all sector deal value. The largest data center acquisition in history was announced. BlackRock, MGX, and others built consortium packages around facilities nobody was talking about two years ago.

Three stories. Three different players. One pattern.

Everyone is moving to own something physical, scarce, and structural — before the window closes.

The Stack Is Stratifying

For most of the last decade, AI was described as a software story. Algorithms. Models. APIs. The infrastructure underneath was someone else's problem — cloud providers abstracted it away, and the prevailing assumption was that compute would follow Moore's Law into irrelevance as a strategic concern. Spin up more instances. Bring costs down. Keep shipping. That assumption is now visibly wrong, and the people with capital are repositioning around the correction.

What's actually happening is that the AI stack is hardening into distinct layers, each with its own pricing power, its own supply constraints, and its own strategic leverage:

The chip layer. NVIDIA controls the rate-limiting input for the entire stack. The H100 is the workhorse. The B200 ramp is underway but hasn't broken the pricing dynamic yet. Until custom silicon from the hyperscalers achieves meaningful displacement at scale — Google's TPUs, Amazon's Trainium, Meta's MTIA — NVIDIA's pricing power is structural, not cyclical. They sell the shovel. Everyone else is mining.

The facility layer. Power, cooling, land, and physical infrastructure. The hardest to build quickly, the most capital-intensive, and the most defensible once established. PE figured out in 2025 that this layer generates long-term contracted revenue from hyperscalers and AI labs desperate for capacity, with utility-like yield characteristics and an AI upside kicker. You don't need to pick the winning AI company. You need to own the building they all have to rent.

The compute access layer. Where AI labs and hyperscalers sit — controlling the allocation of GPU capacity to developers, enterprises, and end users. This layer has enormous leverage when supply is constrained, and enormous vulnerability when the layer below it — the facility and chip layers — can reprice or withhold.

Anthropic's Colossus move is a play to climb one layer up in this stack. It didn't escape NVIDIA's pricing — those 220,000 GPUs still cost what they cost. But it locked in supply at a known rate, removed spot market exposure, and bought the kind of infrastructure stability that lets you make enterprise promises without hedging everything on whether AWS reprices its reserved instances next quarter.

The Utility Parallel

The correct historical analogy here is not the dot-com buildout. It's not even the cloud infrastructure wave.

It's the early electrical grid. In the 1890s and early 1900s, the question of who controlled power generation was understood to be the question of who controlled industrial civilization. Municipalities fought over it. Robber barons structured monopolies around it. J.P. Morgan built General Electric in part to control the generation infrastructure that everything else depended on. What's happening now is structurally similar. AI compute is becoming a utility input — something that every economically meaningful activity will eventually depend on. And like every utility before it, the entities that control the infrastructure layer are accumulating a kind of leverage that is very difficult to dislodge once it's established.

PE understands this. That's why $45.7 billion moved into data centers in a single year. The investment thesis isn't speculative. It's the wireless tower trade, applied to AI: own the infrastructure that every player in the market has to lease, collect the rent, don't pick winners at the application layer.

NVIDIA understands this. Their pricing isn't irrational — it's a rational extraction of rent from a supply-constrained market where their product is the non-substitutable input. They're not being greedy. They're being accurate about their position. Anthropic understands this. The Colossus deal looks like a compute partnership. It's actually a supply chain sovereignty move. The product improvement — doubled Claude Code limits — was downstream of the infrastructure decision. Compute first. Capability second. That sequence reveals the strategic logic.

What Moves First in an Infrastructure Race

History is consistent on this point: in any industry undergoing infrastructure consolidation, the entities that move first to control physical chokepoints set the terms for everyone who comes later.

The railroad barons didn't win because they had better trains. They won because they controlled the routes. The telecoms that built fiber in the 1990s — even through the dot-com collapse — ended up holding infrastructure that the entire internet economy eventually ran on. The wireless tower companies that built out coverage in the 2000s are now collecting rent from every carrier in every market. The pattern is the same. Move early. Own the physical layer. Let the application market above you consolidate around your infrastructure.

Right now, three categories of player are executing this playbook simultaneously in AI: Chip manufacturers holding pricing power while custom silicon alternatives are still years from displacement. Infrastructure funds acquiring data center capacity before the AI demand curve makes it unaffordable. And AI labs — the ones with enough capital and relationships — quietly climbing one layer down in the stack to secure the compute foundation their product promises depend on.

The labs that don't have that capital or those relationships are building on ground they don't own. That's a workable position while the infrastructure is abundant and cheap. It's a vulnerable one when the infrastructure is scarce and getting more expensive by the quarter.

The Altitude Reading

From high enough up, this isn't really a story about Anthropic, or NVIDIA, or private equity. It's a story about what happens when a general-purpose technology transitions from the experimental phase to the infrastructure phase. In the experimental phase, software eats everything. Ideas move fast. Capital flows to the brightest algorithm. The infrastructure is someone else's problem.

In the infrastructure phase, the physical layer asserts itself. Capital flows to whoever controls the scarce, hard-to-replicate inputs. The entities that built on abstracted infrastructure discover they're tenants in someone else's building. The gravity shifts to the ground. AI crossed that threshold sometime in the last 18 months. The infrastructure phase is here. The consolidation is underway. The pricing signals are already in the data — GPU rental rates, facility acquisition multiples, compute deal structures — for anyone paying attention. The question worth sitting with is not which AI model wins the benchmark. The question is who owns the ground it runs on.

---

Rich Washburn is a technologist and strategist working at the intersection of AI, cybersecurity, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital, and CIO of Data Power Supply.