Nvidia Earnings Highlights: Rubin Ramp-Up Pace and OpenAI Compute Commitment

```

NVIDIA will release its quarterly earnings report on August 26. The market’s focus has shifted from the results themselves to two deeper narrative threads: the ramp-up speed of the new Rubin architecture, and the financial logic implied by NVIDIA’s multi-year, ultra-large-scale computing commitment with OpenAI. Jefferies notes that NVIDIA’s data is overall strong this season, and upside expectations are well priced in, but the pace of Rubin’s rollout and the accounting treatment of the OpenAI agreement will be the core topics of this quarter’s earnings call.

Jefferies analysts estimate NVIDIA’s July quarter revenue at $95 billion, above the market consensus of $91.9 billion, up about 16% quarter-on-quarter; guidance for the October quarter is expected to be $108 billion, about 4% higher than the consensus estimate of $103.7 billion. The firm forecasts that the Rubin series (VR/R200) will account for about 12% of GPU revenue in F3Q27, rising to over 40% in F4Q27, with F1Q28 likely to be the inflection point when Rubin surpasses Blackwell as the main revenue source, expected to reach shipments of 2 million units.

Meanwhile, on August 17, NVIDIA announced an agreement with SB Energy to build the PORTS-Pike computing park in Ohio. OpenAI holds a 20-year lease, NVIDIA is the sole computing power provider, and will invest $1.5 billion into SB Energy.

Management has disclosed that OpenAI’s current and planned commitments have reached about 12GW of scalable NVIDIA computing capacity, expandable to 16GW, corresponding to about $600 billion in cumulative revenue before 2030. This is the largest single customer commitment disclosed to date. The accounting treatment and disclosure of contingent liabilities under this agreement will be one of the most closely watched risk variables in the market.

Rubin ramp-up: Design simplification enables faster capacity release

Jefferies believes the platform transition efficiency of Rubin will significantly exceed Blackwell’s initial ramp, and estimates Rubin rack deployments could reach over 13,000 in C26, exceeding 120,000 in C27.

Vera Rubin NVL72 continues the 72-GPU Oberon chassis form factor established by GB200/GB300, retaining the same physical footprint and manufacturing/deployment base. The 45°C warm-water cooling design is directly compatible with existing liquid-cooled data centers. On the assembly level, NVIDIA has made the compute tray and NVLink switch tray cableless, hoseless and fanless modular designs. A central PCB replaces the labor-intensive manual installation of NVLink harnesses from the Blackwell era. Management claims this has reduced compute tray assembly time from about 2 hours to just 5 minutes.

On the supply chain side, NVIDIA describes Rubin’s supporting ecosystem as twice the size of Grace Blackwell, covering 30 countries, 350 factories, and hundreds of partners.

Mass production shipments are scheduled to begin in F3Q27, continuing to ramp in F4Q27, with F1Q28 defined by management as a "very large" quarter. Jefferies’ model predicts shipments at 2 million units. Notably, all frontier labs are expected to adopt Vera Rubin from the outset, which differs from the early Blackwell rollout.

Vera CPU: The structural opportunity behind the $20 billion target

Management has indicated the CPU addressable market (TAM) is expected to reach $200 billion, with an F27 full-year CPU revenue target of $20 billion, covering both Grace and Vera product lines.

Jefferies’ bottom-up model estimates F27 accompanying CPU revenue at about $8.65 billion, and infers standalone Vera CPU revenue at about $11.35 billion; F4Q27 exit annualized revenue rate is about $32 billion. Vera ASP is priced at about $4000, uses a near-lithography-limit 3nm monolithic die process, and gross margin is consistent with NVIDIA’s overall AI business.

On the customer side, SPCX has committed to an "all-NVIDIA" architecture, Oracle plans to deploy hundreds of thousands of Vera CPUs; Microsoft’s commitment is relatively moderate; OpenAI and Anthropic are still in evaluation.

Jefferies points out that ACIEs (AI and cloud infrastructure enterprises) will bear the main weight of Vera CPU revenue growth. In terms of competition, AMD’s Venice is currently the clear market leader, already ramping in 2H26; Intel has some flexibility in capacity but, in Jefferies’ view, remains at a technological disadvantage due to lack of Spatial Multi-threading technology.

OpenAI agreement: The duality of a $600 billion commitment and potential contingent liabilities

On August 17, NVIDIA announced cooperation with SB Energy on the PORTS-Pike technology park in Ohio, built on the site of a former DOE Portsmouth facility. OpenAI holds a 20-year lease, NVIDIA is the sole compute provider, offering credit support for 4.25GW of IT capacity, holding an option for the remaining 3.75GW (8GW in total), and investing $1.5 billion into SB Energy.

According to NVIDIA, each generation of systems deployed in this park can correspond to about 1.5 million GPUs, generating $150–200 billion in revenue (about $100,000–$133,000 per GPU), and can undergo multiple upgrade cycles in 20 years.

Management also disclosed that OpenAI's current and planned commitment totals about 12GW, expandable to 16GW of NVIDIA computing, with cumulative revenue by 2030 of about $600 billion—the largest single customer commitment disclosed to date. At 16GW, that is approximately $37 billion per GW, generally matching management's previous CapEx/$GW estimates.

Jefferies points out this cooperation marks a significant change to NVIDIA’s balance sheet: NVIDIA does not bear the full cost of the park, but only a specific portion of lease and electricity costs, as well as some specific residual value commitments, which will phase in as the data centers start operations between 2028 and 2030, and decrease as OpenAI pays rent. However, the report also notes that this was announced only a week after a $50 billion third-party financing platform was characterized as "moving customer financing off-balance-sheet," which could rekindle investor concerns about circular financing.

The core issues for this earnings call will be: how contingent liabilities are accounted for and disclosed; whether this agreement structure can become a template for other frontier labs; and how NVIDIA assesses the residual value risk of a single-tenant 20-year asset. The answers to these will directly determine to what extent ACIE growth is driven by NVIDIA itself, and what valuation multiples the market will assign.

Networking and CPO: Kyber delays interrupt Scale-Up roadmap

Jefferies believes the Scale-Up CPO plan publicly committed to by NVIDIA at this year’s GTC now faces schedule pressure. According to supply chain data, the Kyber rack form factor has been confirmed delayed, and the Scale-Up CPO implementation may be pushed back from C28 to C29 or later.

During the transition period, some believe NVIDIA may advance NPO-type multi-rack interconnect schemes, such as creating an NVL576 from eight Oberon NVL72 racks. Jefferies considers this plausible and possibly on the roadmap, but due to technical complexity, it will likely not become the mainstream shipping solution. In recent Scale-Up moves, with VR200 NVL72, NVIDIA introduced NVLink 6, doubling the number of NVLink switches per NVL72 from 16 to 32 compared to the previous generation.

For Scale-Out CPO, the Spectrum-6 CPO switch has entered volume production, with capacity expected to ramp further in 2H26. Initial users include CRWV, Lambda, META, MSFT, and ORCL. Jefferies maintains: Broadcom (AVGO) and TH6 remain dominant in this segment, and Scale-Out CPO demand is expected to remain moderate in the near term.

Power constraints: the biggest hidden constraint in AI deployment pacing

Whether power supply can keep up with chip demand is becoming the core constraint variable for AI infrastructure expansion. The largest hyperscale cloud providers (AMZN, MSFT, GOOG, META) mainly drive the construction of dedicated gas power plants and transmission lines by paying utilities; frontier labs (OpenAI/Anthropic/xAI) and new cloud service operators are increasingly using "behind-the-meter" (BTM) schemes to bridge the gap between grid readiness and XPU deployment needs.

Jefferies estimates that the gap between XPU deployment and power supply will be about 30GW in C27, expanding to 53GW in C28, but the actual net gap is expected to be in the tens-of-GW range, well below the nominal 566GW peak difference. This is because some new capacity shares overlap with planned data center builds, and there are timing mismatches between power capacity calculations and actual data center deployment power.

Main new power sources under construction are: GE Vernova (GEV) gas turbines at about 20/22/24GW in C26/27/28; Siemens (SIE) at 15.5GW in C26, ramping to 19/21GW in C27/28; for reciprocating gas generators, CAT’s order backlog now extends over 24 months, and even with doubled capacity, it is still sold out into C27; Bloom Energy’s solid oxide fuel cell annual capacity is expected to reach 2GW by the end of 2026; FTAI’s repurposed aircraft engine solution provides 25MW per unit; nuclear and small modular reactors (SMRs) are unlikely to achieve commercial rollout before 2030.

On the regulatory front, opposition to data center construction is rising in multiple states. Jefferies remains optimistic on AI’s long-term outlook but believes current power supply trends warrant investor caution, and looks for management to address these potential obstacles clearly in the earnings report.

Risk Warning and DisclaimerMarkets contain risks, investment should be prudent. This article does not constitute personal investment advice and does not take into account the individual investment objectives, financial circumstances or needs of particular users. Users should consider whether any opinions, views or conclusions in this article are suitable for their specific circumstances. They invest at their own risk. ```