Part 2 | Anthropic, the Physical Cost of AI: Compute, Memory and Energy

The first section demonstrated that announced annualised revenue growth has outpaced the valuation. The next step concerns the structure required to sustain that trajectory: compute, high-bandwidth memory, cloud and data centres demand commitments characterised by a degree of rigidity that far exceeds a software budget.

The central question is how much of the reserved capacity will actually come online, and at what utilisation rate.

Compute enters the financial structure

The $65 billion raised in the Series H represents only the most visible component of growth financing. In June 2026, Apollo, Blackstone and a group of banks joined Broadcom in constructing an initial $35 billion infrastructure solution designed to accelerate new computing capacity, part of which is also used by Anthropic.19

The financing takes a form distinct from conventional corporate debt recorded directly on Anthropic's balance sheet. The structure employs an infrastructure platform and dedicated vehicles that acquire and finance hardware and facilities, which are subsequently made available for use. This distinction will be material in any IPO prospectus, as economic risk may exist even when a portion of the assets or debt formally remains off-balance-sheet.19

Added to this structure are more than $100 billion in technology commitments to AWS over ten years, as well as the contract with SpaceX. According to information contained in SpaceX's filing and reported by Reuters, Anthropic may pay approximately $1.25 billion per month for access to Colossus capacity. The agreement does, however, include an initial phase and the option to terminate with 90 days' notice, making it meaningfully different from a rigid financial obligation for its full theoretical duration.520

Equity capital, infrastructure financing, purchase commitments and capacity contracts each require separate analysis. They carry different maturities, risks and rights. Taken together, however, they reveal that Anthropic is building a genuine financial architecture for compute. Its competitive advantage will also depend on the ability to finance chips, data centres and energy while keeping the growth of contractual obligations below that of the margin generated by its customers.1920

Memory is the invisible bottleneck

On 22 June 2026, Anthropic and Micron announced a strategic agreement encompassing joint memory and storage architecture design, a multi-year supply agreement, the adoption of Claude within Micron, and a strategic investment in the Series H.21

Memory is a central component of the system. HBM, DRAM and SSD units determine the speed at which data reaches accelerators, the energy efficiency of the infrastructure, and the number of requests that can be processed with a less-than-proportional increase in chip count. A highly capable GPU loses a portion of its value in the presence of insufficient memory or when data transfer becomes the bottleneck.21

In the following image, we observe how HBM memory is placed alongside the GPU to increase bandwidth, available capacity, and power efficiency.

For Anthropic, the issue relates directly to inference costs. Better integration between accelerators, memory, and storage can reduce per-token energy consumption, increase facility throughput, and improve the return on purchased capacity. For Micron, the agreement reinforces its transition from a mere component supplier to an industrial partner in infrastructure design.21

The collaboration gives Anthropic greater visibility into the procurement of a scarce resource, while also deepening the interconnections between investors, customers, and strategic suppliers. Micron supplies memory, uses Claude in its own processes, and holds an equity stake in the company. The economic relationship thus spans multiple tiers of the same supply chain.21

Energy has become a profit-and-loss variable

In 2024, data centres consumed approximately 415 terawatt-hours, 1.5% of global electricity. The International Energy Agency estimates approximately 945 TWh by 2030, just under 3%: an increase of 128%.7 In 2025 alone, consumption grew by 17%. In the same year, the aggregate capital expenditure of five major technology groups exceeded $400 billion and, according to the IEA, could increase by a further 75% in 2026.8

In the following chart, we observe the estimated trajectory of global data centre electricity demand, which under the IEA's central scenario could more than double between 2024 and 2030.

The Lawrence Berkeley National Laboratory's updated analysis places U.S. consumption by 2030 in a range of 521 to 843 TWh, with a central case of 649 TWh, equivalent to 11.8% of national electricity demand.9 The range is wide because it depends on accelerator shipments, chip operational lifespan, server utilisation rates, and cooling performance. This uncertainty broadens the risk margin and illustrates how the outcome hinges on industrial variables that are still in the process of stabilising.

Available power capacity represents a distinct metric from annual energy production. A data centre requires grid connections, transformers, turbines, cooling systems, water supply, and continuity of service. The IEA estimates that, in the absence of measures to address bottlenecks, approximately 20% of planned projects are at risk of delays.7 For a laboratory, a delayed grid connection can leave already-ordered chips sitting idle and push expected revenues further out. For a utility, rising demand can expand the regulated asset base, but only if permitting and rate-setting frameworks protect the return on capital.

The American advantage shifts across the value chain

The comparison between the United States and China varies depending on the metric examined. In 2025, U.S. private investment in AI reached $285.9 billion, compared with $12.4 billion in China: a ratio of approximately 23 to one. In the same year, the United States produced 59 models deemed significant, against 35 from China: approximately 1.7 to one. By March 2026, however, the leading U.S. model was ahead of its Chinese counterpart by only 2,7% on the measure reported by the Stanford AI Index.101112

In the following chart, we observe the scale of the gap in private investment in artificial intelligence: in 2025, the United States remains well ahead of both Europe and China.

In the following chart we observe the opposite dynamic: the advantage of the best American models over their Chinese counterparts remains far more modest than the gap recorded in private investment.

Available capital is far more concentrated in the United States than the ultimate capacity of the models. This runs counter to the thesis that compute alone determines the outcome. Chips and infrastructure remain indispensable, but talent, algorithmic efficiency, distillation, pricing and speed of adoption can partially offset lower availability.

Physical concentration remains both favourable and fragile at the same time. The United States is home to 5,427 data centres, more than ten times any other country. According to Stanford, TSMC manufactures virtually all leading-edge AI chips.10The American advantage in semiconductor design therefore coexists with a concentration of manufacturing capacity in Taiwan. A geopolitical or operational incident would ripple from the supplier through the entire value chain.

# The Six-Layer Supply Chain

Competition extends beyond the laboratories. The value chain begins with energy, networks and materials; runs through machinery, foundries, accelerators, memory and storage; and ultimately reaches cloud infrastructure, data centres, models and applications. Each layer adds capability, but also introduces a potential industrial or financial bottleneck.

The dependency runs from the bottom up. A more capable model drives greater demand for inference — that is, the execution of the model in response to user requests. Inference requires accelerators; accelerators depend on high-bandwidth memory, foundries, packaging, and tooling; data centres require power, networking, water, and cooling systems. Software remains the customer-facing layer, but the constraint on growth may lie in an electrical substation, a semiconductor fabrication plant, or the availability of HBM.

Incentives diverge. ASML may wish to preserve the Chinese market for its machinery. A US laboratory may benefit from restrictions that slow down a Chinese competitor. A hyperscaler — that is, a cloud operator at global scale — seeks greater compute demand and may benefit from the presence of more competing models. Industrial policy redistributes value across layers rather than favouring them all in equal measure.

"Content (text and/or images) produced with the assistance of artificial intelligence, under the editorial responsibility of the newsroom."

The Financial Spectator – Independent Analysis of Global Markets and the Economy