The Cost of Blackwell

B200/GB200 elemental composition — unknown. No peer-reviewed ICP-OES or equivalent teardown of any Blackwell GPU exists as of May 2026. All elemental claims use Falk et al. 2026 A100 data as proxy. Blackwell dies are larger, packaging more complex (NVL co-packaging), and the rack includes Grace CPUs and NVLink ASICs not present in A100-era systems.

Power draw and PUE — more uncertain than modelled. Oracle confirms “over 120 kW at peak” but sustained draw during typical workloads may be lower. PUE 1.2 assumes efficient liquid cooling, but an Nvidia internal email (Dec 2025) described Microsoft’s GB200 facility-level cooling as “wasteful” — the setup uses a second air-cooling layer on top of in-rack liquid cooling, trading water efficiency for energy use. Real-world PUE may be 1.3–1.4 at some facilities, increasing carbon by 8–17%. [I]

Rack ASP — sell-side estimate only. Analyst Vinh (Benzinga) estimates $1.5M–$2M per rack [D]. Nvidia hardware cost to OEMs not publicly disclosed. Our previous $3–5M estimate was likely 2–3× too high.

Reserved contract pricing — not public. $3.30–3.50/GPU-hr sourced from CloudSyntrix analysis of private cloud Blackwell contracts [A]. Actual CoreWeave contract rates are confidential. On-demand spot ($10.50–$27/hr) is observable but represents a different risk structure.

Legal liability gap — cobalt. No company in the cobalt supply chain is legally liable for DRC extraction or labor conditions. ASM accounts for ~20% of DRC cobalt. Nvidia’s Responsible Minerals Policy covers disclosure, not remediation.

Cooling water — deployment-dependent, not a fixed figure. Our ~5,000 L/day estimate may be misleading. Supermicro’s GB200 NVL72 datasheet lists multiple CDU configurations including 250 kW, 240 kW, and 180 kW liquid-to-air solutions — some configurations require no facility water at all (closed-loop/liquid-to-air). An Nvidia internal email described Microsoft’s GB200 cooling as “wasteful due to the size and lack of facility water use” — suggesting Microsoft’s design uses more energy but less water than an evaporative design would. [I] The correct statement is: cooling water varies from near-zero (closed-loop dry cooler) to thousands of litres per day (evaporative), depending entirely on facility design. TSMC fab water (~1,173 L per B200 IC, ~118,000 L for the rack’s chips) is separately accounted. [J]

All figures are lower bounds. Falk et al. 2026 explicitly state their ICP-OES results exclude losses during refining, processing, and component manufacturing. Logistics carbon, NVLink tray footprint, and Grace CPU materials are not independently quantified.

Carbon price absent. ~974,000 kg CO₂e over 2 years. At $50/tonne = ~$48,700/rack — unbilled on any invoice. No mandatory carbon price applies to US data center operations.

Jevons’ Paradox — Nvidia’s own marketing confirms it. Nvidia’s sustainability page claims GB200 NVL72 delivers “25× greater performance per watt vs prior generation” and Blackwell Ultra delivers “50× more energy efficiency per token.” [K] These per-token efficiency gains are real and measurable. But they say nothing about total power. GB200 NVL72 draws ~120 kW; the H100 DGX it replaces drew 10.2 kW. “25× performance per watt” + “10× more absolute power” = 2.5× more total power per rack, not less. More efficient per inference does not mean less total environmental burden — it means the ceiling for total deployment rises. Nvidia also claims AI tasks use “100,000× less power than a decade ago” with no methodology cited — this compares best-case 2024 inference per-token against 2014 training compute and is not a statement about total system power consumption. [K]

Obsolescence timeline uncertain. Hyperscalers depreciate GPUs over 4–6 years for accounting purposes. Technological useful life may be 1–2 years given Nvidia’s annual cadence [G]. H100 spot rates fell 70% within one product cycle [F]. Reserved contracts transfer this risk to the customer.

Manufacturing water — upgraded from estimate to modelled figure. TechInsights (Aug 2025) applied the Murphy yield model, 90% fab utilisation, 98% package yield, and location-based Scope 2 accounting to Blackwell. Result: ~310 gallons (1,173 L) of water per Blackwell IC. Our previous figure of ~2,000 L/wafer overstated by ~41%. Full rack: ~118,000 L for IC manufacturing alone (72 B200 + 36 Grace + NVLink ASICs estimated), equivalent to 833 Americans’ daily water use. This is still IC-only — excludes packaging, assembly, cooling, and operational water. Architecture matters: Blackwell has ~20 cm² silicon vs MI300X >40 cm²; produces ~20% less total CO₂e in manufacturing but is ~2× more carbon-intensive per cm². Scope 2 (grid electricity) = 59% of Blackwell manufacturing emissions — meaning fab location is the single largest decarbonisation lever. [J]

Deployment lag not modelled. The Information (May 2026) reports that OpenAI, Oracle, Meta, Microsoft, and Amazon all experienced significant delays getting GB200 NVL72 up and running. Early 2025 chips failed testing; customers reduced orders or reverted to Hopper. Nvidia’s internal email on a Microsoft deployment noted validation documentation required extensive rewriting and handover processes needed “a lot more solidification.” Our P&L assumes revenue begins immediately. A 3-month lag on a $2M rack costs ~$106k in dead debt+colo and compresses the already thin profit window significantly. [H, I]