AI Infrastructure Mid-Year Update: Nvidia Kyber Delayed & Arista Stock Surging

By: Network-Switch.com | July 8, 2026

Table of contents

 

Nvidia Kyber NVL144 Delayed 12+ Months to 2028 - 78-Layer PCB Midplane Can't Be Built at Scale

On July 5-6, CNBC reported - citing semiconductor research firm SemiAnalysis - that Nvidia's next-generation Kyber NVL144 rack-scale architecture has been delayed by more than 12 months, slipping from its planned 2027 availability to 2028. The cause is a single manufacturing component: a specialized 78-layer printed circuit board midplane that connects eight Oberon compute racks to the NVSwitches inside the Kyber system. SemiAnalysis stated the PCB midplane "remains challenging from a manufacturability standpoint" - meaning Nvidia cannot yet build it reliably at production volumes. The Kyber system is designed to pack 144 of Nvidia's Rubin Ultra chips into a single cabinet, making them behave as one unified computing system for the largest possible AI training jobs.

The delay carries two additional consequences beyond the primary slip. First, a stopgap rack design called the NVL72x2 - which would have paired two existing Oberon racks back-to-back using copper NVLink - was also scrapped after cloud service providers and hyperscale data center operators rejected it due to "unconventional design and heavy operational burden." With the NVL72x2 canceled, Nvidia currently has no proven alternative for scaling Rubin Ultra at rack scale until Kyber arrives. Second, SemiAnalysis confirmed that the originally planned four-die Rubin Ultra GPU - which would have used four compute chiplets and 1 TB of HBM4E per package - has been canceled due to "manufacturing execution concerns," replaced by a two-die design that delivers roughly half the per-package compute. Nvidia's brief response to Tom's Hardware was four words: "Our roadmap is intact." The company has neither confirmed nor denied the specifics. Nvidia stock fell approximately 1.4% to $194.83 on the report, and Asian technology supplier stocks - several of which depend on Kyber component orders - slid more sharply. The delay does not affect the Rubin NVL72 (see story 2), which is on schedule.

Rubin NVL72 Enters Full Production - Fall 2026 Deliveries Confirmed for 8 Cloud Providers

Separate from the Kyber delay story, CNBC confirmed on July 6 that Nvidia's Vera Rubin NVL72 - the first-generation Rubin rack, using the existing Oberon rack structure rather than the new Kyber architecture - has entered full production and that deliveries to eight major cloud customers are scheduled to begin in fall 2026. Confirmed recipients include Amazon Web Services, Microsoft Azure, and Google Cloud, with five additional hyperscale and neocloud operators expected in the same fall delivery window. This delivery schedule is consistent with Dell Technologies' June 1 shipment of the world's first Rubin NVL72 to CoreWeave (covered in our previous edition) and with Nvidia's GTC Taipei announcement (May 31) that Vera Rubin had entered full production.

The NVL72 is an important commercial distinction from the delayed Kyber NVL144: it delivers 10x the agent throughput of Blackwell at 10x lower cost-per-token for agentic AI inferencing, using 72 Rubin GPUs per rack. It does not attempt the ultra-dense 144-GPU-per-cabinet architecture of Kyber - which requires the problematic PCB midplane - and instead uses the proven Oberon rack design with NVLink-6 interconnect. For enterprise buyers and OEMs, this distinction matters: Rubin NVL72-based systems from Dell, HPE, Lenovo, and Supermicro remain on track for H2 2026 general availability. It is the Rubin Ultra generation - the follow-on that would have required Kyber to scale beyond 72 GPUs - that has been affected by the delay.

Arista Stock Surges 7% on July 8 as Multiple Banks Raise Targets to $200 Ahead of Q2 Earnings

On Wednesday July 8 - today - Arista Networks (NYSE: ANET) is trading up approximately 7.06%, driven by a combination of strong underlying AI networking demand signals and fresh analyst upgrades ahead of the company's Q2 2026 earnings report. KeyBanc Capital Markets, Morgan Stanley, and Bank of America all raised their Arista price targets in the past week, with the consensus moving toward $200 per share. KeyBanc specifically cited "exceptional demand" for Arista's hardware in XPU and AI inference workloads, characterizing earlier supply and deferred revenue issues as "largely temporary" and expecting them to normalize as silicon and optics supply conditions improve through H2 2026.

The timing is also driven by the imminent Q2 2026 earnings report and a schedule of AI-focused investor conferences. Arista guided Q2 2026 revenue at approximately $2.8 billion, with gross margin between 62% and 63% and non-GAAP EPS of approximately $0.88 - representing roughly 29% year-over-year revenue growth. The Bank of America Global Technology Conference earlier in June produced a notable strategic update: Arista management described the total addressable market for AI-driven networking as having expanded from $50 billion to $105 billion in the past year, with projections of $160 billion by 2030. The AI networking investment cycle was characterized as "multi-year or even multi-decade," with demand broadening beyond hyperscalers into tier-2 cloud providers and enterprise deployments. Today's move recovers much of the ground lost in the June 23 insider-selling selloff, and signals that the market is treating Arista's near-term fundamentals as intact.

Microsoft Makes Anthropic's Claude Available on Nvidia GB300 Blackwell Ultra GPUs in Azure Foundry

In a milestone that bridges AI model deployment and AI infrastructure hardware, Microsoft announced that Anthropic's Claude is now generally available in Azure AI Foundry running on Nvidia GB300 Blackwell Ultra GPUs - the current-generation flagship AI accelerators. The deployment marks the first time Claude has been made available on Azure's latest infrastructure tier, and positions Blackwell Ultra as the hardware substrate for enterprise-grade Anthropic model access through Microsoft's cloud platform. GB300 Blackwell Ultra features 288 GB of HBM3e per GPU module and is the primary driver of Nvidia's Q1 FY2027 data center revenue of $75.2 billion.

The commercial significance extends beyond the specific model: it signals that the world's major AI frontier models are actively deploying on the latest Nvidia hardware - not waiting for Rubin - and that the GB300 generation has sufficient volume and stability to host enterprise customer workloads at scale. For enterprise networking and infrastructure teams, this is a concrete reminder that AI inference workloads are arriving now, on GB300-based infrastructure, and that network fabric decisions made today (800G, InfiniBand, Spectrum-X, Arista Ethernet) need to support GB300 cluster densities - not just the Rubin or Feynman workloads projected for 2027 and 2028.

Foxconn Q2 2026 Revenue Up 40% on AI Server Demand - Largest Nvidia Assembly Partner Confirms Supply Ramp

Foxconn Technology Group, Nvidia's primary server assembly partner and the world's largest contract electronics manufacturer, reported a 40% quarterly sales increase in Q2 2026, driven by AI server assembly demand - principally for GB200 and GB300 Blackwell systems and, beginning in Q3, Rubin NVL72 racks. The result confirms that Nvidia's component supply chain is ramping effectively at the assembly level, and that the constraint on AI server availability is not manufacturing throughput but upstream component availability - specifically DRAM, NAND, HBM memory, and switch silicon - consistent with what Dell COO Jeff Clarke flagged on the Dell Q1 FY2027 earnings call and what Micron CEO Sanjay Mehrotra disclosed in the Q3 FY2026 results (covered in our previous edition).

Foxconn's revenue signal is a useful bellwether for the broader AI server supply chain because it reflects actual shipment volumes, not just booked orders. A 40% increase in a single quarter at a company with Foxconn's scale represents a very large absolute dollar increase in AI server production. For enterprise buyers and OEM partners - including Dell, HPE, Lenovo, and Supermicro, all of which use Foxconn or comparable ODMs for assembly - the Foxconn data point supports the conclusion that the second half of 2026 will see significantly more AI server and GPU rack supply reaching the market than the first half, as Blackwell supply constraints gradually ease and Rubin NVL72 systems begin shipping to their first commercial customers.

Dell Pro Precision 7 R1 Now Available: 1U Rack-Mount AI Workstation with Nvidia RTX PRO Blackwell

On schedule with the availability date announced at Dell Technologies World in May, the Dell Pro Precision 7 R1 is now commercially available in July 2026. It is a 1U rack-mounted workstation that brings high-performance AI compute to space-constrained environments - machine rooms, trading floors, hospitals, broadcast production facilities, and other settings where a tower workstation is impractical but a full-scale GPU server rack is excessive. The system is equipped with Nvidia RTX PRO Blackwell Max-Q Workstation Edition GPUs and supports up to 64 TB of internal storage.

The Pro Precision 7 R1 sits at the intersection of Dell's "Deskside Agentic AI" strategy - announced at DTW 2026 to bring trillion-parameter model inference to the local environment - and the commercial workstation market that Dell has historically served through its tower and mobile Precision line. In practical terms, it enables enterprise customers to run production AI inference workloads (copilots, retrieval-augmented generation, code generation, medical imaging) on hardware that physically fits in a standard 19-inch rack alongside their existing networking and storage equipment - without the power infrastructure, cooling infrastructure, or capital cost of a dedicated AI server. For IT procurement teams evaluating edge AI deployments in space-constrained environments, the Pro Precision 7 R1 is the most capable 1U AI compute platform Dell currently offers.

Editor's Summary

This edition covers the most consequential AI hardware news since the Vera Rubin first-delivery milestone three weeks ago. The Kyber delay is significant because it pushes the highest-density AI compute system available to 2028 - meaning that for at least the next 18 months, the Rubin NVL72 (72 GPUs per rack) is the practical ceiling for enterprise AI infrastructure scale-up, and the Kyber NVL144 dream of 144 GPUs per cabinet remains out of reach. The four-die Rubin Ultra cancellation is the quieter but equally significant detail: the next-generation of Rubin Ultra will arrive with roughly half the originally projected per-package compute, reducing the performance leap from Rubin to Rubin Ultra and extending the competitive window for AMD and Google's own AI accelerators. Meanwhile, Arista's 7% intraday surge, combined with analyst price targets now clustering at $200, signals that the market's near-term concern about insider selling has faded and the focus has returned to Q2 earnings fundamentals. And Foxconn's 40% revenue jump provides the most concrete confirmation yet that AI server supply is ramping fast in H2 2026 - even as upstream component constraints at the memory and silicon level continue to be the binding limiter on how fast that supply reaches customers.

Network-Switch.com is a global professional distributor of networking equipment from Cisco, Huawei, Ruijie, H3C, and our own NS brand - including switches, routers, firewalls, wireless APs, optical modules, and fiber patch cables. Our CCIE, HCIE, H3CIE, and RCNP certified engineers deliver end-to-end technical support and complete network solutions for enterprise, campus, and large-scale infrastructure deployments worldwide.

Visit us: https://network-switch.com

Frequently asked questions (FAQs)

Why has Nvidia's Kyber NVL144 rack been delayed to 2028?

Nvidia's Kyber NVL144 rack-scale architecture has been delayed by more than 12 months to 2028 due to manufacturing difficulties with a specialized 78-layer printed circuit board (PCB) midplane that connects eight Oberon compute racks to the NVSwitches inside the cabinet. Research firm SemiAnalysis reported on July 6, 2026 that this component "remains challenging from a manufacturability standpoint," meaning Nvidia cannot yet produce it reliably at scale. Additionally, a stopgap design (NVL72x2) was scrapped after cloud providers rejected it, and the original four-die Rubin Ultra GPU has been canceled in favor of a two-die design due to manufacturing concerns.

Does the Kyber delay affect the Nvidia Vera Rubin NVL72 delivery schedule?

No. The Kyber delay only affects the Rubin Ultra generation and its NVL144 rack architecture, which was planned for 2027. The standard Vera Rubin NVL72 - using the existing Oberon rack design with 72 Rubin GPUs per cabinet - is on schedule. CNBC confirmed on July 6, 2026 that Rubin NVL72 has entered full production and that deliveries to eight major cloud customers, including Amazon Web Services, Microsoft Azure, and Google Cloud, are scheduled to begin in fall 2026. OEM availability from Dell, HPE, Lenovo, and Supermicro also remains on track for H2 2026.

Why is Arista Networks stock jumping in July 2026?

Arista Networks (NYSE: ANET) surged approximately 7% on July 8, 2026, driven by a combination of fresh analyst upgrades from KeyBanc, Morgan Stanley, and Bank of America - all raising price targets toward $200 - and positive AI networking demand signals ahead of Q2 2026 earnings. KeyBanc cited "exceptional demand" for Arista in XPU and AI inference workloads, characterizing supply constraints as temporary. Bank of America noted that Arista's total addressable market has expanded from $50 billion to $105 billion, with projections of $160 billion by 2030, and that the AI networking investment cycle is "multi-year or even multi-decade."

What does the Foxconn Q2 2026 revenue increase mean for AI server supply?

Foxconn Technology Group - Nvidia's primary server assembly partner - reported a 40% quarterly revenue increase in Q2 2026, driven by AI server assembly demand for GB200 and GB300 Blackwell systems and beginning Rubin NVL72 racks. This confirms that the AI server supply chain is ramping effectively at the assembly level in H2 2026, and that the primary binding constraint on availability is upstream component supply (HBM memory, DRAM, switch silicon) rather than manufacturing throughput. For enterprise buyers, this signals meaningfully more AI server supply entering the market in the second half of 2026 compared to the first half.

What is the Dell Pro Precision 7 R1 and who is it designed for?

The Dell Pro Precision 7 R1, now available in July 2026, is a 1U rack-mounted workstation equipped with Nvidia RTX PRO Blackwell Max-Q Workstation Edition GPUs and up to 64 TB of internal storage. It is designed for enterprise environments where a tower workstation is impractical but a full-scale GPU server rack is cost- or space-prohibitive - including hospital machine rooms, broadcast production facilities, trading floors, manufacturing sites, and edge data closets. It is part of Dell's "Deskside Agentic AI" strategy to enable production AI inference workloads (copilots, RAG, code generation, medical imaging) on hardware that fits in a standard 19-inch rack alongside existing networking and storage equipment.

 

Sources