Nvidia pushes back Kyber rack to 2028: the physical wall of AI compute is approaching

Context

Nvidia's post-Blackwell roadmap initially planned Rubin for 2026, Rubin Ultra for 2027, and Kyber (NVL576, 576 GPUs per rack in liquid cooling) for 2027. According to a report by The Information relayed by Seeking Alpha on July 6, 2026, Kyber is postponed to 2028 due to manufacturing issues: advanced packaging, 800V DC intra-rack power supply, direct-to-chip cooling. This is to be linked to the dynamics of mega-cloud GPU contracts (cloud-gpu-specialized-contracts, pub #834, #587).

The Data

• Kyber (NVL576): rack designed for 576 Rubin Ultra GPUs, ~1.1 MW TDP - initial target 2027, now 2028.
• Blackwell NVL72 (2025): 132 kW/rack. Rubin Vera Ultra (2026): ~600 kW. Kyber: ~1.1 MW.
• Nvidia data center revenue CY25: ~$130B (JPMorgan, May 2026).
• TSMC CoWoS-L: capacity 60k wafers/month end 2026, advanced packaging bottleneck.

Analysis

Kyber is tackling three walls simultaneously: (i) 3D packaging with bump pitch < 20 µm and next-generation CoWoS-L; (ii) 800V DC intra-rack power supply replacing the standard 48V; (iii) direct-to-chip liquid cooling with flow rates > 200 L/min. Each generational leap increases the TDP per rack by ×1.7. The industrialization difficulty does not come from the chips but from the systems (busbars, cold plates, liquid PDU). This is a physical wall, not a silicon yield problem.

Probabilized Scenarios

• (50%) Confirmed delay to mid-2028: Rubin Ultra NVL144 extended, window open to AMD MI400X/MI450X.
• (30%) Partial delay to Q1 2028 with a "light" Kyber version of 288 GPUs.
• (20%) Nvidia bypasses with an intermediate Rubin Ultra "Wide" architecture in 2027.

Portfolio Implications

Hyperscalers (MSFT/META/GOOGL/AWS): re-planning capex 2027, arbitrage between waiting for Kyber and over-ordering Rubin. Nvidia: smoothing revenue, sustained cash flow (no cancellations). AMD: temporary winner on MI400X/MI450X that nibble at the 2027-mid-2028 window. Specialized cloud GPUs (CoreWeave, Applied Digital): mega-fixed contracts (Argentum $4.1B, pub #587) now align with an extended Rubin Ultra cycle - less risk of obsolescence at 18 months. Cooling/power ecosystem: 1.1 MW orders delayed (Vertiv, Schneider, Delta, Eaton).

Risks & Blind Spots

A Kyber delay can free up CoWoS-L capacity for ASIC competitors (Broadcom TPU, AWS Trainium 3, MTIA Meta) - acceleration of AI compute fragmentation. Symmetrical risk: rebound in Blackwell/Rubin orders that extends cash flow. Contrarian angle: specialized cloud GPUs benefit, the slower roadmap stabilizes the depreciable value of existing fleets.

To Watch

• Official roadmap CES 2027 (January) - Kyber 2028 confirmation.
• Hyperscaler capex guidance FY2027 (Q2 results end of July).
• TSMC CoWoS-L volumes - industrial tension indicator.
• AMD MI400X deliveries - T4 2026.