What do hyperscalers buy when they buy compute?

Compute is not one commodity

A GPU-hour in a congested region, a rack in a power-secure campus, and a custom accelerator attached to an internal model stack are not the same product.

The useful unit is constrained capacity: what workload can run, when it can start, how predictable it is, and what bottleneck it exposes next.

The purchase bundles time and risk

Hyperscalers buy chips early, sign power contracts early, reserve packaging and HBM supply, pre-build data centers, and build custom silicon programs because time risk is the real enemy.

If the model roadmap needs capacity in 2028, the procurement move often had to begin years earlier.

Owning compute changes product strategy

A company with cheap internal inference can bundle AI into products more aggressively. A company renting scarce capacity has to price, throttle, or narrow use cases.

That is how infrastructure decisions flow upward into applications. Product abundance begins as capacity abundance.

Depreciation turns speed into pressure

AI hardware depreciates quickly because each generation changes performance, memory, power, and software assumptions. A cluster has to produce useful work before its economics are overtaken.

This creates pressure to keep utilization high and to match hardware vintages to workloads that still fit them.

The strategic read is optionality

The hyperscaler wants the option to train the next model, serve the current product, host enterprise workloads, and redirect capacity when the bottleneck moves.

Compute buying is no longer back-office procurement. It is one of the main ways AI strategy becomes real.

Sovereign Infrastructure & The Compute Arbitrage: The SpaceX ($SPCX) S-1

The S-1 public filing of SpaceX (targeting a $1.75T to $2.00T post-IPO valuation) following its February 2026 consolidation of xAI and X/Twitter fundamentally redefines compute procurement. SpaceX is debuting as a consolidated, vertically integrated deep-tech and AI infrastructure conglomerate utilizing Starlink's high-margin cash cow ($4.4B in operating income on $11.4B in revenue, 38.6% operating margin) to internally subsidize both Starship orbital development and high-expenditure AI cluster construction.

The filing exposes an extraordinary new dynamic in compute strategy: Infrastructure Arbitrage. Under Elon Musk's direction, xAI constructed the Memphis-based Colossus 1 data center in under 120 days, drawing over 300 megawatts to house 220,000+ NVIDIA GPUs (H100, H200, and upcoming GB200 systems). Once xAI migrated its primary training runs for Grok onto the adjacent, homogeneous Blackwell-based Colossus 2 facility, SpaceX leased the entire Colossus 1 cluster to competitor Anthropic for a massive $1.25 billion per month ($15.0 billion ARR) under a 36-month lease starting May 1, 2026.

This transaction immediately flips a heavy capital sink into high-margin recurring cash flow, demonstrating how physical speed-to-build can be converted into enterprise cash. However, it exposes extreme revenue fragility: the contract contains a 90-day mutual termination clause. If Anthropic experiences a capital crunch or faces pressure from backers like Amazon or Google to shift workloads to AWS or GCP, SpaceX will be exposed to severe, unutilized compute depreciation on 220,000 GPUs.

• Sovereign Infrastructure Flywheel: Low-cost heavy orbital launch (Starship) enables rapid deployment of Starlink connectivity, generating internal cash flow that acts as a platform subsidy for building massive AI clusters.
• Scale Arbitrage: Turning mature, active infrastructure (Colossus 1) into a high-margin enterprise leasing engine for competitors while training internal models on next-generation hardware (Colossus 2).
• Fragility Risk: A 90-day mutual cancellation window creates high revenue concentration risk and grid dependencies on the local Tennessee Valley Authority (TVA) power grid.

Source: SpaceX Form S-1 Filing, May 20, 2026

Sovereign AI Clouds & Token-Value Economics: The E2E Networks Model

Sovereign AI clouds in emerging markets represent a highly capital-efficient, hyper-growth model of compute procurement. India's E2E Networks Limited (Q4 FY26 earnings) is a premier pure-play AI Cloud GPU provider demonstrating this structural shift. Reporting a massive 186% year-on-year revenue surge (INR 956 million) and expanding EBITDA margins to 60.7% for Q4 FY26, E2E demonstrates that specialized, localized cloud providers can capture immense operating leverage by moving faster than traditional slow-moving hyperscalers.

E2E is executing a massive hardware scale-up, deploying 2,048+ NVIDIA Blackwell B200 GPUs (with the first 1,024-GPU cluster going live in mid-May 2026, followed by a second 1,024 cluster) and targeting at least 6,000 GPUs under management by the end of FY27. To fund this rapid expansion, E2E is leveraging structured private credit and an asset-light MOU with Larsen & Toubro (L&T) to monetize GPU infrastructure L&T is building, proving that compute operators can scale without bearing 100% of the underlying hardware CapEx on their own balance sheets.

Crucially, E2E's management is pioneering a shift in metrics from standard hardware asset turns to "Token-Value Economics." As AI models increase in accuracy, the generated tokens become more valuable to enterprises. Rather than selling raw hardware cycles at a commodity discount, E2E runs a proprietary, vertically integrated software orchestration stack (TIR platform) to capture high-value enterprise tokens (e.g., complex transaction resolution instead of basic search queries). This allows them to sustain high utilization rates (80-85% across their fleet) and maintain price stability or even premium realizations despite rapid generational hardware cycles.

• Asset-Light Co-Procurement: Partnering with heavy industrial players (L&T) to monetize external GPU clusters under a unified software platform, bypassing balance sheet limitations.
• Aggressive Depreciation: Writing down hardware over a compressed 6-year schedule creates paper net losses (e.g., INR 156M net loss for FY26) but hides highly profitable, cash-generative underlying EBITDA.
• Token Value Capture: Orchestrating bare-metal and container fleets via the TIR platform to prioritize high-value enterprise tokens, unlocking higher ROI for compute buyers than simple raw time-rentals.

Source: E2E Networks Q4 & FY26 Earnings Conference Call, April 20, 2026

Compute scarcity is procurement, not supply

A common framing in 2026 is that compute is scarce because chips are scarce. Jensen Huang pushed back hard on that at Stanford CS153: "it is just not true that people are giving me orders and we are not delivering chips. You got to place orders." The constraint is not silicon output.

The actual constraint is institutional procurement structure. Universities do not have $1 billion compute budgets because every research lab buys its own laptops and every grant funds its own small cluster. Mid-tier enterprises do not commit multi-year capacity contracts because their CFOs cannot underwrite the depreciation curve. Startups cannot place hyperscale orders because they have no collateral. The dollars exist; the willingness to organise them at the scale required does not.

For Pere, the implication is twofold. First, "we cannot get compute" is usually a procurement problem dressed up as a supply problem. Second, the buyers who solve procurement before their competitors get capacity at favourable terms that compound over multiple generations.

Source: Jensen Huang, Stanford CS153 Frontier Systems lecture, April 30, 2026 (https://cs153.stanford.edu/)