The frontier model labs are reaching down, one layer at a time, into the silicon, power, and data centers their models actually run on — because they no longer trust the open market to supply those things on time.
Anthropic becomes anchor tenant of the Theseus data-center vehicle, backed by Macquarie-managed funds and GIC. OpenAI signs a twenty-year lease on the SB Energy campus in Ohio.
2025–2026 — Power
OpenAI explores an investment in or acquisition of Lancium — the power developer behind the Stargate campus in Abilene, Texas — last quarter, per The Information. Nvidia ultimately invests instead. OpenAI does not acquire Lancium.
2026 — Silicon
Anthropic hires Amir Salek, founder of Google’s TPU program. Signs a roughly $250 million deal with UK inference-chip startup Fractile — production silicon not expected until around 2027.
2026 — The Pincer Closes
Nvidia integrates up into power and land to protect chip demand. The labs integrate down into that same infrastructure. Both sides meet at the data center.
The Information reported last week that OpenAI explored taking a stake in, or acquiring, Lancium — the power developer that supplies the Stargate data-center campus in Abilene, Texas — at some point last quarter. It did not complete a deal. Nvidia ultimately invested in Lancium instead. On its face, that is a footnote: an exploration that went nowhere, attributed to one company. What makes it worth examining is what sits next to it.
Anthropic, in a separate and unrelated set of moves, has hired Amir Salek — the engineer who founded Google’s TPU program — and signed a deal with the UK inference-chip startup Fractile worth roughly $250 million, according to reporting we have covered previously. That silicon is not expected to be production-ready until around 2027; the Salek hire is groundwork, not a shipped product. Anthropic is also the anchor tenant of the Theseus data-center vehicle, a structure backed by Macquarie-managed funds and GIC — Anthropic is the tenant, not the owner; the infrastructure sits on Macquarie and GIC’s balance sheets. Meanwhile, Nvidia is guaranteeing up to $105 billion of OpenAI’s Ohio lease obligations — a contingent backstop, not cashOpenAI is spending — as part of an arrangement tied to the SB Energy campus.
These are separate hires, leases, financings, and one abandoned deal at two different companies. No lab has announced a unified infrastructure strategy. The synthesis below — that the frontier labs are forward-integrating into the physical stack — is the author’s analytical read across these data points, not a plan on file anywhere. With that stated, the through-line is hard to dismiss.
The key insight:OpenAI explored buying its power developer. Anthropic is building toward its own chips while anchoring a data-center vehicle it does not own. These are early, capital-heavy bets with real execution risk — power delays and grid constraints are the very problems driving the strategy — but the direction is consistent: model companies are reaching down into the infrastructure their models run on, because they have concluded that in an era of scarce compute, being only a model company is a strategic vulnerability.
The Structural Read
The logic here is classical vertical integration, applied to an unusual input. When a critical resource becomes scarce and the open market cannot reliably deliver it, the companies that depend on it start buying up the supply chain. Steel companies owned coal mines. Automakers owned rubber plantations. The difference in 2026 is that the scarce input is not a commodity with centuries of infrastructure behind it — it is the intersection of advanced chips, grid-connected power, and purpose-built real estate, all of which are being built, contested, and financed simultaneously.
The labs have reached the same conclusion through different financing structures. OpenAI leans on partners and Nvidia’s backstops — the $105 billion Ohio guarantee is Nvidia’s contingent commitment, not OpenAI’s balance-sheet spending, and the Lancium interest never became a transaction. Anthropic favors off-balance-sheet vehicles: the Theseus entity keeps data-center assets with Macquarie and GIC while Anthropic retains the tenancy, and a roughly $35 billion Broadcom structure finances its compute separately. Different ways to pay for the same push toward controlling supply.
The silicon layer is the most nascent piece. Salek’s hire is an intent signal, and the Fractile deal is a financing commitment toward chips that do not yet exist at production scale. That is worth holding carefully: the labs are not chip companies yet, and the gap between a $250 million deal and a reliable supply of inference accelerators at frontier scale is a large one. But it is the same direction as the power and data-center moves, just earlier in execution.
BE Framework — Vertical Integration Into the Physical Stack
Don’t Trust the Market — Own the Supply
When compute and power become binding constraints, “just buy from the market” stops being a strategy. The labs have each, independently, begun acquiring control — through ownership, tenancy, financing, and talent — over the silicon, electricity, and real estate underneath their models. The financing differs by lab. The goal does not. And layered on top of this is a pincer: Nvidia is integrating up the stack into power and land to protect demand for its chips, while the labs integrate down into that same infrastructure. They meet in the middle at the data center — which is why Nvidia ended up as the investor in Lancium that OpenAI had explored becoming.
Three Implications
IMPLICATION 1 — THE MARKET FOR COMPUTE IS TIGHTENING AT THE TOP
When frontier labs start exploring ownership of their power developers — even if those deals do not close — it signals that the spot market for grid-connected capacity feels insufficient at the scale they are planning for. OpenAI’s Lancium exploration and Anthropic’s Theseus tenancy are different instruments solving the same problem: securing supply before it becomes the bottleneck that leaves expensive chips idle. That the exploration failed and Nvidia stepped in is itself informative — it reveals who has the balance sheet and the strategic motive to own these assets outright.
IMPLICATION 2 — NVIDIA’S POSITION IS MORE COMPLICATED THAN IT LOOKS
Nvidia investing in Lancium — the same asset OpenAI explored — puts it inside the power infrastructure of a customer’s data-center campus. Nvidia guaranteeing up to $105 billion of OpenAI’s Ohio lease obligations backstops a lab it also supplies with chips. These are not neutral vendor relationships; they are structural entanglements that give Nvidia leverage and exposure simultaneously. As the labs integrate down into infrastructure, Nvidia integrates up into the same layer — and the two strategies are now intersecting in the same physical assets. That is a dynamic worth watching as the labs mature their own chip programs.
IMPLICATION 3 — THE EXECUTION RISK IS REAL AND BUILT INTO THE STRATEGY
Power delays, grid interconnection queues, permitting fights, and chip development timelines are not edge cases here — they are the central obstacles the integration strategy is designed to hedge. Fractile’s silicon is not production-ready until around 2027. Data centers take years to permit and build. Grid connections in high-demand regions can slip by similar amounts. The labs are betting that taking ownership of these inputs now reduces the probability of being supply-constrained later. That is the right analytical frame. Whether the execution follows the intent is the open question, and it will not close quickly.
Business Engineer Framework
The AI Value Chain — Where the Labs Are Moving
The AI Value Chain maps the full stack from raw silicon and power through model development, deployment infrastructure, and application layer — and shows which companies are moving across layers and why. The lab-to-infrastructure moves described here are a textbook forward-integration play in the value chain: model companies taking ownership of the inputs immediately beneath them because the market price and availability of those inputs has become a strategic risk. The pincer with Nvidia — integrating from opposite ends toward the same data-center layer — makes the competitive dynamics of that middle layer worth mapping carefully.
OpenAI did not buy Lancium. Anthropic does not own Theseus. The Fractile chips are not shipping yet. Hold all of that — and then notice that two of the most valuable model companies on earth are each, independently, reaching down into the power lines, the land, and the accelerator designs that sit beneath their models, because they have concluded that relying on the open market for those things at the scale they are planning is a risk they are no longer willing to carry. They are not infrastructure companies yet. The direction in which they are moving is no longer ambiguous.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
