Lock-in: Nscale and the $103 billion shift redefining AI infrastructure value

The market has stopped pricing computational capacity. It has started pricing contractual lock-in, and Nscale's valuation makes this regime change visible to everyone.

Nscale, a British AI infrastructure company founded two years ago, communicated to potential investors a revenue figure near $103 billion following the Anthropic agreement, according to TechCrunch on September 4, 2026^[1]. That number reflects projections based on leases signed by customers, distinct from current sales, as The Information clarifies.

This is a regime change. Value in the AI infrastructure layer is migrating from hardware to control over the duration and exclusivity of commitments. Anyone reading that valuation as a bet on GPUs is reading the data wrong.

Consensus has the wrong frame

Consensus looks at data centers, GPUs, megawatts. It measures installed power and concludes that advantage belongs to whoever owns the most silicon.

Consensus has the wrong frame. Computational performance descends along a steep cost curve, and whatever falls in price loses pricing power. The durable asset becomes the multi-year contract that locks in the customer before commoditization erodes margins.

Nscale raised its Series A in December 2024 for $155 million. The March 2026 Series B, led by Aker, raised $1.1 billion, called the largest Series B in European history. Now it is seeking an additional $3.5 billion in pre-IPO funding, with $1.5 billion in convertible notes and $2 billion expected from Nvidia, Bloomberg reports.

This progression tells a story of a race to lock in customers before the market understands where value truly resides.

The cost curve says everything

The AI compute cost curve follows a documented trajectory. The price per unit of performance collapses year after year.

The cost per token for a GPT-4 level model points toward $0.001 by the end of 2026, according to this desk's thesis on foundation model commoditization. The cost of inference has fallen orders of magnitude in eighteen months.

The historical parallel is solar. The IEA documents a cost decline near 90% between 2010 and 2020. Compute follows the same pattern: rising abundance, falling unit price, hardware margin compressed toward zero.

When input descends along this curve, pricing power abandons whoever sells raw capacity. It migrates toward whoever signed the customer on a multi-year horizon.

The cliff event: when the contract becomes the asset

Cliff event: price parity between proprietary compute and commodity compute will arrive, and at that point already-signed contracts become the only scarce asset.

Adoption of this logic jumps, rather than growing linearly. Every provider watches competitors lock in customers on multi-year leases and reacts by accelerating its own signatures. The result is a contractual race.

The Anthropic agreement, valued at approximately $45 billion, represents 44% of the $103 billion valuation. One customer, nearly half the company's perceived value. This concentrates risk and reveals the mechanism: a single customer's lock-in anchors the entire company's financial narrative.

The rush toward IPO confirms this reading. A company taking its books to market at this stage wants to crystallize the value of commitments before commoditization erodes it.

What changes for those deciding now

For the CTO and Chief Innovation Officer the question changes. The relevant choice concerns the duration of the binding commitment accepted, before it becomes obvious to everyone.

For Technology Procurement the danger is real. Signing today a multi-year contract on capacity destined to become commodity means paying a lock-in premium for a depreciating asset. The vendor sells future scarcity, the customer buys future abundance.

For Venture Capital the apparently impossible bet is clear: the value of compute providers resides in the contractual portfolio, valued as recurring revenue, distinct from the installed base of hardware.

For the Chief Strategy Officer a three-year plan that assumes compute as a competitive advantage assumes a dissolving world. Advantage shifts to the application layer with proprietary data moats.

Three categories that will change shape by 2028

• Pure compute providers
• Foundation model labs
• Traditional hyperscalers

Pure compute providers will see value migrate from capacity to the contract book, valued as a financial leasing company. Whoever owns long and exclusive leases commands the premium.

Foundation model labs will consolidate their infrastructure suppliers, replicating the Anthropic-Nscale agreement on industrial scale. Mutual dependency creates a web of bidirectional constraint.

Traditional hyperscalers will face competitors born two years ago with comparable valuations, built entirely on customer commitments rather than capital depreciation.

My position, and what would change it

My position is clear: Nscale's valuation prices contractual lock-in, the first visible case where the AI infrastructure market values binding duration above machine capacity.

What would change this reading. A compute price collapse slow enough to preserve hardware margins for a decade would dismount the thesis. In that case capacity would remain the asset, and contracts would return to accessory status.

A second counter-signal would be an easy exit clause in signed leases. Weak binding is worth little. The thesis strength depends on commitment rigidity, and that rigidity is today the true product.

Prediction

At least one other large-scale AI compute provider will be valued above $50 billion on a basis of signed contracts rather than current revenue by the end of 2027.

Confidence: high on direction, medium on timing. Horizon: end of 2027. Kill signal: Nscale completes its IPO at a valuation below $50 billion, a sign that the public market refuses to price contractual commitments at the private market level.