On November 1, 2025, Bernstein analyst Stacy Rasgon published an in‑depth report detailing the real cost to build 1GW of AI data‑center capacity: about $35 billion. This figure, based on industry interviews and third‑party infrastructure data, is significantly lower than the $50–60 billion cited by NVIDIA during its fiscal Q2 2026 earnings call. Rasgon argues NVIDIA's estimate is “forward‑looking,” accounting for rising rack and GW‑level costs in future GPU cycles. The report highlights that a GB200/NVL72 rack costs about $5.9 million (compute hardware $3.4M, physical infrastructure $2.5M), aligning more with Broadcom and AMD market‑opportunity estimates of $15–20 billion/GW. Amid the AI boom, this cost breakdown not only exposes the data‑center “cash burn” but also points to upstream suppliers (foundries, memory vendors) as key beneficiaries. The report finds GPUs dominate spending (39%), NVIDIA gross margin accounts for 29% of total cost; mechanical and electrical systems and networking each take roughly a third; storage is only 1–1.5%. Looking ahead, as rack power density soars (with 800V high‑voltage designs by 2027), the cost structure will shift dramatically, making power bottlenecks and the supply chain critical investment areas.
Cost Breakdown: $5.9M per Rack, GPUs Capture 39% of Spend
Bernstein's $35B/GW estimate represents the current GB200‑architecture “near‑term cost,” far below NVIDIA's “future outlook.” The report divides costs into two main categories: compute hardware and infrastructure:
- Compute hardware (about 57%): The largest expense is GPUs, accounting for 39% of total capital expenditure (capex). For NVIDIA's GB200 series, single‑rack hardware costs are $3.4M, with GPU chips and accelerators making up the bulk. Rasgon calculates that even with NVIDIA gross margins of 70%, NVIDIA's margin still represents 29% of total AI data‑center spend. That implies NVIDIA captures roughly $10 billion in profit per GW data center.
- ASIC alternatives: Using ASIC‑based racks could save about 19% of capex, but accelerated compute remains the largest share. CPUs account for only 3%, often bundled as a “freebie” with GPUs and comparable to switches.
- Networking and storage: Networking is 13%, including switches (3%), cabling, accelerators and connectors (2–3% each). Storage is negligible at just 1–1.5%—a typical rack has only ~2PB of storage with hardware costs around $20K. The report quips: “Storage has become 'scrap' in the AI era.”
On the infrastructure side, mechanical and electrical (M&E) systems make up nearly a third of costs:
- Power generation and delivery: Generators/turbines 6%, transformers 5%, uninterruptible power supplies (UPS) 4%.
- Cooling: Air plus liquid cooling is 4%, though liquid cooling's share is expected to rise to handle the surge in power density.
- Upstream supply chain: Foundries, memory and wafer‑equipment suppliers each account for 3–4% (about $1.0–1.2 billion per GW). Under an ASIC architecture, these shares are higher—under GB200, foundries are 2.5–3% (~$1.1B), with CPUs adding another 1%.
Rasgon emphasizes: “The true economic cost skews toward servers and networking because of shorter depreciation cycles and limited operating expenses.”
Operating Costs: Annual Power ~$1.3B, Staffing 'Negligible'
While capex is high, operating expenses (opex) are relatively modest. A 1GW data center's annual power bill is about $1.3 billion (assuming $0.15/kWh), the largest opex item. Staffing needs are very low: a 20GW complex might require only 8–10 people, making labor costs effectively “negligible.”
Power has become the key bottleneck: OEMs like Siemens Energy and GE Vernova are expanding turbine production to meet data‑center demand. The report warns that utility‑scale power shortages could raise costs, and future high‑voltage designs (800V) could increase rack power‑component costs by 7–8x.
NVIDIA vs. Bernstein: Why a $15–25B Gap?
NVIDIA's $50–60B/GW figure is “cycle‑forward,” accounting for successors to Blackwell and higher‑density, more expensive racks. Bernstein's $35B is more “grounded,” matching Broadcom/AMD market‑opportunity estimates of $15–20B. Rasgon argues: “NVIDIA is pricing in future GPU cycles where per‑rack and per‑GW costs rise.”
Even if ASIC margins fall to 50%, capex could fund about 19% more racks, significantly benefiting upstream suppliers (e.g., TSMC). The report predicts value will shift toward bottleneck positions: high‑voltage components, liquid cooling, and network accelerators.
Investment Takeaways: Upstream Suppliers and Power Players in Focus
This cost structure highlights investment targets:
- NVIDIA dominance: GPU margin at 29% means NVIDIA is indispensable for AI data centers.
- Networking and cooling: Together ~17%; switch and connector vendors like Arista and Broadcom stand to benefit.
- Power and M&E: Around 30%; demand for turbines from Siemens and GE Vernova will surge.
- Foundry/memory: 3–4%; TSMC and Samsung benefit from an ASIC wave.
Rasgon concludes: “Suppliers positioned higher up the stack will capture outsized value, especially if rack power jumps 7–8x.”
Industry Context: AI Data‑Center 'Arms Race' Accelerates
In 2025, AI training demand is driving explosive growth in data‑center capacity: giants like Meta, Google and Microsoft have annual capex in the hundreds of billions. One GW equates to millions of GPU hours, supporting trillion‑parameter models. The report implicitly warns that high costs could slow expansion, with power and chip shortages as choke points.
In sum, Bernstein's $35B/GW “reality check” brings AI infrastructure out of the black box. NVIDIA's higher estimate reflects optimism about future cycles, but current build costs are more accessible. Investors should watch supply‑chain bottlenecks: whoever controls power and high‑voltage components holds the AI era's “gold mine.”
▶️ Enjoyed this article? Creating great content takes effort!
Subscribe to our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares mean the world to us! 🙏
For more AI news, visit:
https://www.airosetta.com