As Nvidia gears up to release its quarterly results on August 26th, the market's attention has shifted from the headline numbers to two more complex narratives. The primary focus is now on the production ramp-up trajectory of the next-generation Rubin architecture, alongside the intricate financial mechanics behind a multi-year, hyperscale compute agreement with OpenAI. Jefferies notes that while Nvidia's current quarter data appears strong, with upside expectations largely factored into the market, the pace of Rubin's volume ramp and the accounting treatment of the OpenAI deal are set to be the core topics on the earnings call.
Jefferies analysts project Nvidia's July quarter revenue to hit $95 billion, surpassing the consensus estimate of $91.9 billion, which represents a sequential growth of roughly 16%. For the October quarter, they anticipate guidance of $108 billion, about 4% higher than the $103.7 billion consensus. The firm forecasts that the Rubin series (VR/R200) will contribute approximately 12% of GPU revenue in F3Q27, climbing to over 40% by F4Q27. They expect F1Q28 to be the inflection point where Rubin overtakes Blackwell as the dominant revenue source, with projected shipments of 2 million units.
In parallel, Nvidia announced a deal on August 17th with SB Energy to develop the PORTS-Pike computing campus in Ohio. OpenAI holds a 20-year lease at the facility, with Nvidia serving as the exclusive compute provider and investing $1.5 billion into SB Energy. Management has indicated that OpenAI's existing and planned commitments now total approximately 12GW of Nvidia compute, scalable to 16GW. This corresponds to cumulative revenue of around $600 billion through 2030, marking the largest single-customer commitment disclosed to date. The accounting treatment of this agreement and the potential disclosure of contingent liabilities are being watched closely by the market as significant risk variables.
Rubin Ramp-Up: Simplified Design Accelerates Capacity Release
Jefferies believes Rubin's platform transition will be considerably more efficient than the initial ramp of Blackwell. They project Rubin rack deployments to exceed 13,000 units in C26 and surpass 120,000 units in C27. The Vera Rubin NVL72 retains the 72-GPU Oberon chassis format established by GB200/GB300, preserving the same physical footprint and manufacturing/deployment base. Its 45-degree Celsius warm-water cooling design is directly compatible with existing liquid-cooled data centers. At the assembly level, Nvidia has shifted to a modular design for compute and NVLink switch trays, eliminating cables, hoses, and fans. A central PCB replaces the extensive manually installed NVLink cabling of the Blackwell era, a move management says reduces compute tray assembly time from roughly two hours to just five minutes.
On the supply chain front, Nvidia describes the Rubin ecosystem as roughly twice the scale of Grace Blackwell, spanning 30 countries, 350 factories, and hundreds of partners. Volume shipments are slated to begin in F3Q27, with continued ramping in F4Q27. F1Q28 is being characterized by management as a "very large" quarter, which Jefferies models at 2 million units shipped. Notably, all frontier labs are expected to adopt Vera Rubin from the outset, a contrast to the initial phase of Blackwell adoption.
Vera CPU: Structural Opportunity Behind the $20 Billion Target
Management has set the total addressable market (TAM) guidance for CPUs at $200 billion, with a full-year F27 CPU revenue target of $20 billion, encompassing both the Grace and Vera product lines. Based on a bottom-up model, Jefferies calculates companion CPU revenue of approximately $8.65 billion in F27. This implies standalone Vera CPU revenue of around $11.35 billion, with an exit annualized revenue run-rate of about $32 billion in F4Q27. The Vera ASP is priced near $4,000, based on a monolithic die at the photolithography limit of the 3nm process, with gross margins consistent with Nvidia's overall AI business. On the customer front, SPCX has committed to an "all-Nvidia" architecture, Oracle has indicated plans to deploy hundreds of thousands of Vera CPUs, while Microsoft's commitment appears more moderate. OpenAI and Anthropic are still in the evaluation stage. Jefferies points out that the ACIE (AI and Cloud Infrastructure Enterprise) customer segment will bear the primary weight of Vera CPU revenue growth. In terms of competitive landscape, AMD's Venice remains the clear current market leader and is already in its ramp phase in 2H26. Intel has some capacity flexibility, but in Jefferies' view, remains at a technological disadvantage due to the lack of Spatial Multi-threading technology.
The OpenAI Agreement: The Dual Nature of a $600 Billion Commitment and Contingent Liabilities
On August 17th, Nvidia announced a collaboration with SB Energy on the PORTS-Pike technology campus in Ohio, built on the former DOE Portsmouth site. OpenAI holds a 20-year lease, and Nvidia acts as the exclusive compute provider. Nvidia provides credit support for 4.25GW of IT capacity and holds options for the remaining 3.75GW (totaling 8GW), while also investing $1.5 billion in SB Energy. According to Nvidia's disclosures, each generation of systems deployed at this campus could correspond to roughly 1.5 million GPUs, generating between $150 billion and $200 billion in revenue (approximately $100,000 to $133,000 per GPU), and can undergo multiple upgrade cycles over the 20-year term.
Management also disclosed that OpenAI's combined existing and planned commitments total roughly 12GW of Nvidia compute, expandable to 16GW. The cumulative revenue through 2030 is around $600 billion, the largest single-customer commitment disclosed to date. On a 16GW basis, this equates to roughly $37 billion per GW, broadly consistent with management's prior CapEx-per-GW metrics. Jefferies notes this partnership marks a significant shift on Nvidia's balance sheet. Nvidia is not bearing the full cost of the campus but rather specific portions of lease payments and electricity costs, along with specific residual value guarantees. These obligations phase in as data centers come online between 2028 and 2030 and decrease as OpenAI makes rental payments. However, the report specifically highlights that this move comes just one week after a $50 billion third-party financing platform was characterized as taking "customer financing off-balance-sheet," which could reignite investor concerns about circular financing. The core questions for the upcoming earnings call are the accounting treatment and disclosure scope of the contingent liabilities, whether this agreement structure can serve as a template for other frontier labs, and how Nvidia views the residual value risk of a 20-year single-tenant asset. The answers will directly determine how much of ACIE growth is driven by Nvidia itself and what valuation multiple the market will assign.
Networking and CPO: Kyber Delay Disrupts Scale-Up Roadmap
Jefferies believes Nvidia's publicly committed Scale-Up CPO plans from this year's GTC are facing timeline pressure. Based on supply chain data, the Kyber rack form factor has confirmed delays, potentially pushing Scale-Up CPO implementation from C28 to C29 or later. During this transition period, there is a view that Nvidia might advance NPO-type multi-rack interconnect solutions, such as an NVL576 configuration formed by eight Oberon NVL72 racks. Jefferies considers this path plausible and likely on the roadmap, but due to high technical complexity, it is not expected to become a mainstream shipping form. In the near-term Scale-Up direction, Nvidia is introducing NVLink 6 with the VR200 NVL72, increasing the number of NVLink switches from 16 in the previous NVL72 generation to 32. On the Scale-Out CPO side, Spectrum-6 CPO switches have entered volume production, with capacity expected to expand through 2H26. Early adopters include CRWV, Lambda, META, MSFT, and ORCL. Jefferies maintains its view that Broadcom (AVGO) and TH6 remain the dominant players in this segment, and near-term demand for Scale-Out CPO is expected to remain moderate.
Power Bottleneck: The Hidden Constraint on AI Deployment Pace
The question of whether power supply can keep pace with chip demand growth is becoming the central constraint on AI infrastructure expansion. Major hyperscalers (AMZN, MSFT, GOOG, META) are primarily driving the construction of dedicated gas power plants and transmission lines by paying utility companies. In contrast, frontier labs (OpenAI/Anthropic/xAI) and new cloud service providers are increasingly turning to "behind-the-meter" (BTM) power solutions to bridge the gap between grid readiness and XPU deployment needs. Jefferies estimates the gap between XPU deployment and power supply is roughly 30GW in C27, widening to 53GW in C28. However, the actual net gap is expected to be in the tens of GW range, well below the nominal 566GW peak difference, due to overlapping calculations between some new capacity and data center construction plans, as well as timing differences between power capacity calculation and actual data center power draw.
Among the major power sources under construction, GE Vernova (GEV) gas turbine capacity for C26/27/28 is approximately 20/22/24GW. Siemens (SIE) has about 15.5GW capacity in C26, progressing to 19/21GW in C27/28. For reciprocating gas generators, CAT's order backlog has extended beyond 24 months, and even with doubled capacity, it remains sold out through C27. Bloom Energy's solid oxide fuel cell annual capacity is expected to reach 2GW by the end of 2026. FTAI's retired aircraft engine conversion solution offers 25MW per unit. Nuclear power and small modular reactors (SMRs) have a low likelihood of commercial deployment before 2030. On the regulatory front, resistance to data center construction is growing in several states. Jefferies remains optimistic about the long-term AI outlook but believes the current power supply situation warrants cautious investor attention, and they look forward to management providing clear commentary on these potential headwinds during the earnings call.
Comments