Bernstein: 1GW AI Compute Capex at $35-40 Billion, Detailed Cost Breakdown by Architecture
nashnova research
Bernstein's latest supply-chain analysis puts the capital cost of a 1GW AI data center at $35 billion to $40 billion, with per-watt spending across architectures varying by less than $5 — and the overwhelming share goes to servers, storage, and networking, not power or labor.
How much does 1GW of AI compute cost — and how close are the architectures?
Bernstein's range spans from OpenAI Jalapeno at roughly $35B/GW to Nvidia Vera Rubin NVL72 at roughly $40B/GW, or $34.6 to $39.5 per watt.
This means → regardless of chip vendor, building the same scale of compute factory differs by less than 15% in total cost. Architecture choice reshapes the cost mix, not the order of magnitude.
Non-rack costs — power delivery, cooling, civil works — land closer to the low end of a prior $12B–$15B estimate. Overall capex tilts heavily toward server, storage, and networking hardware.
What goes into a single Vera Rubin rack — and where does the money go?
Bernstein cuts its Vera Rubin NVL72 per-rack cost from $9.1M to $7.5M, aligning with supply-chain signals of $7M–$8M. Drivers: lower NAND — flash storage — allocation per rack, more conservative HBM 4 — high-bandwidth memory — pricing, and revised networking and cooling estimates from Asia-channel checks.
Compute is nearly half the bill: each GPU at roughly $55,000, 72 per rack, totaling $3.96M. Vera CPUs add $180K (36 units at ~$5,000 each).
Memory and storage run about $1.8M: HBM 4 pricing drops from $48/GB to $30.6/GB, plus Nvidia's 10% markup, yielding ~$697K per rack in HBM alone. NAND is set at ~8TB per GPU (576TB per rack); CPU DRAM stays at 54TB. Networking adds roughly $1.2M — split across scale-up (~$600K), scale-out (~$500K), and front-end (~$80K).
How do AMD and OpenAI's designs differ from Nvidia's?
AMD MI455X costs roughly $7.3M per rack, close to Vera Rubin. The mix shifts: AMD's lower GPU margins mean less compute value, but up to 3 NICs per GPU pushes networking value markedly higher. AMD CPU value also exceeds Nvidia Vera's.
OpenAI Jalapeno comes in at about $5M per pod, the lowest of the four. In plain terms = Jalapeno is not a white-label Nvidia clone — it is a fundamentally different design. OpenAI wants to expand the scale-up domain, keep data local in HBM and DRAM, and reduce reliance on scale-out networking and NAND.
Google TPU v7 Ironwood is estimated at roughly $2.7M per pod, but Bernstein flags limited disclosure and low confidence, noting the figure will be revised substantially as Google expands external sales.
How much does it cost to actually run a 1GW data center day-to-day?
Even at a relatively high $0.15/kWh, annual electricity for a 1GW facility is only about $1.3B. Staffing is negligible — the largest sites reportedly need just 8–10 employees.
This means → compared to Vera Rubin's roughly $6.6B in annual depreciation on a six-year cycle, power and labor are rounding errors. The per-hour economics are dominated by hardware depreciation, not utilities or headcount.
This reflects a cost structure fundamentally unlike traditional infrastructure: short hardware depreciation cycles and rapid refresh rates concentrate supply-chain leverage on servers, storage, and networking — not on power and operations.
How much compute capacity is coming over the next two years?
Bernstein projects global shipment capacity of 36GW in 2026 and 43GW in 2027 — well above its U.S.-only new-capacity forecast of 18GW and 26GW for the same years.
The gap comes from international projects accelerating: U.S. delays in permitting, power availability, and policy are increasingly common, and AI labs are turning overseas. OpenAI's deal with Firmus, for example, secures capacity across Australia, Singapore, Indonesia, and Malaysia.
In plain terms = the bottleneck for compute buildout is no longer just chip supply — it is where power can be switched on, permits secured, and sites brought online. Whoever solves those problems first captures the capacity. As the Vera Rubin cycle ramps and AI-specific chips see broader adoption, available compute is set to leap again.
市场有风险，内容仅供研究参考，不构成投资建议。