NVIDIA CFO Colette Kress stated on Tuesday that the company's $500 billion bookings for Blackwell and Rubin AI chips through 2026 do not include any work related to ongoing negotiations with OpenAI for their next-phase agreement. She also revealed that the previously announced $100 billion investment deal between NVIDIA and OpenAI remains unsigned.

In September, NVIDIA disclosed a letter of intent to invest in OpenAI, committing to deploy at least 10 gigawatts of NVIDIA systems—enough to power over 8 million U.S. households. The agreement outlined potential investments of up to $100 billion as systems come online. This announcement had previously fueled a rally in AI infrastructure stocks and reinforced market expectations of a deep partnership.

During the UBS Global Tech and AI Conference in Arizona, when asked about progress on the 10GW computing commitment, Kress responded: "We haven't finalized an agreement with OpenAI yet. OpenAI continues on their path, and I believe our collaboration with them will never stop."

She emphasized that the $500 billion demand projection for Blackwell and Vera Rubin systems during 2025-2026 "doesn't include any work we're doing for the next phase with OpenAI." Currently, OpenAI's procurement still occurs indirectly through cloud partners like Microsoft and Oracle, rather than via the direct purchasing path outlined in September's letter of intent.

"OpenAI does want to shift to direct procurement," Kress noted, "but again, we're still working toward a final agreement." NVIDIA shares rose as much as 2.6% during Tuesday's trading before paring gains.

While NVIDIA has previously disclosed sales figures and market demand for Blackwell and Rubin chips, the company hadn't clarified the relationship between the $500 billion booking figure and the OpenAI collaboration until now. Last month, CEO Jensen Huang first revealed at Washington's second GTC conference this year that NVIDIA is experiencing rapidly growing demand for AI infrastructure, projecting accelerated sales growth for Blackwell and Rubin architecture chips over coming quarters, collectively expected to generate $500 billion in GPU sales over five quarters.

During NVIDIA's Q3 earnings call, Kress reiterated that the $500 billion figure would likely grow, noting that newly announced partnerships (including deals with Saudi Arabia and Anthropic) weren't fully reflected. She stated: "We certainly have opportunities to secure additional orders beyond the announced $500 billion."

**A Mega-Deal Not Yet Counted** Analysts find Kress's clarification particularly noteworthy regarding what Huang had called "the largest AI infrastructure project in history." However, her comments suggest the months-old framework remains more exploratory than concrete.

While reasons for the delay remain unclear, NVIDIA's latest 10-Q filing offers clues. The document explicitly states that investment arrangements—including the OpenAI deal, a planned $10 billion Anthropic investment, and a $5 billion commitment to Intel—carry no guarantee of completion under expected terms. In its risk factors section, NVIDIA detailed the fragile realities behind such mega-deals, noting their realization depends on constructing and powering data centers capable of running its systems. The company must order GPUs, HBM memory, and networking components over a year in advance via non-cancelable prepaid contracts, warning of potential "excess inventory," "cancellation penalties," or "inventory write-downs" if clients scale back. Past supply-demand mismatches have "severely harmed our financial results," the filing noted.

The biggest variables appear physical: NVIDIA identified "data center capacity, energy, and capital" availability as critical to clients deploying promised AI systems, describing power infrastructure as a "multi-year process" facing "regulatory, technical, and construction challenges." Delays or downsizing may occur if clients can't secure sufficient power or funding.

NVIDIA also acknowledged its accelerated architecture cycle—now yearly with Hopper, Blackwell, and Vera Rubin—complicates planning, potentially "amplifying demand forecasting challenges" and reducing demand for existing architectures. These disclosures align with bearish investors like Michael Burry, who warned of imbalanced investment cycles from extended chip lifespans, though Huang countered that even six-year-old chips remain fully utilized.

The company further referenced boom-bust cycles like cryptocurrency mining, cautioning that unpredictable AI workload fluctuations could flood gray markets with used GPUs.

**NVIDIA Maintains Confidence in Its Moat** Despite the pending agreement, Kress stressed NVIDIA's "very strong" decade-long relationship with OpenAI, which views NVIDIA as its "preferred partner." However, she clarified that NVIDIA's sales outlook doesn't depend on this unsigned mega-deal.

Addressing competition, Kress dismissed concerns about Google's TPUs potentially rivaling NVIDIA's GPUs on efficiency: "Absolutely not." She highlighted NVIDIA's moat lies not in individual chips but in its full-stack platform encompassing hardware, CUDA, and expanding software libraries—explaining why older architectures remain widely used even as Blackwell becomes the new standard.

"Everyone uses our platform—all models run on it, whether in cloud or on-prem environments," she concluded.