NVIDIA CFO: OpenAI $100 billion order not finalized — what investors need to know

Key Points

NVIDIA CFO: OpenAI $100 billion order not finalized.

Quick headline — OpenAI LOI, not a signed deal

NVIDIA (Yīngwěidá 英伟达) announced in late September an agreement described as deploying “at least 10 GW” of systems for OpenAI.

That reported headline figure — about $100 billion USD — remains a letter of intent and has not been finalized.

The $100 billion USD figure would equal roughly ¥720 billion RMB ($100,000,000,000 USD) at an exchange rate of ¥7.20 per $1 USD.

NVIDIA’s CFO Colette Kress (科莱特·克雷斯 Kēlái tè Kèléisī) said revenue tied to that LOI is not included in company guidance because the agreement is unsigned.

Analysts have previously modeled that full deployment could imply roughly $400–500 billion USD of revenue for NVIDIA — about ¥2.88–3.60 trillion RMB at ¥7.20/USD — if the program were fully realized.

Other remarks from UBS Global Technology & AI Conference — concise summary

• Working with OpenAI: “We still haven’t signed a final agreement, but we are working with them,” Kress said.

• Anthropic commitment: NVIDIA’s commitment to Anthropic could be up to $10 billion USD (≈ ¥72 billion RMB at ¥7.20/USD).

• Revenue guidance context: NVIDIA indicated Blackwell and Vera Rubin systems could drive roughly $500 billion USD in demand across 2025–2026 (≈ ¥3.60 trillion RMB).

• Management outlook: NVIDIA guided next-year revenue in the $350–400 billion USD range (≈ ¥2.52–2.88 trillion RMB).

Q&A highlights — what Kress said and why it matters

1. Is there an AI bubble?

Kress said the industry is in a structural shift from CPU-driven data centers to GPU-accelerated computing.

She estimated that $3–4 trillion USD will be invested in data-center infrastructure over this decade, with roughly half tied to the CPU-to-GPU transition.

Key point: this is an expansion of compute capacity, not just a like‑for‑like swap of old equipment.

2. Is NVIDIA’s lead narrowing?

Kress was direct: “Absolutely not. Our lead has not narrowed.”

She framed NVIDIA’s advantage as a platform story — cross-chip system-level co‑design, multiple cooperating chips, plus a software stack centered on CUDA and industry libraries.

She added that workloads — cloud or on-premises — continue to run on NVIDIA’s platform, which she described as a durable moat.

3. Can NVIDIA’s customers make money?

Kress explained a “flywheel” dynamic.

As the market moves from training to inference and production, model output and token volumes rise, and end users become more willing to pay for reliable, high-performance inference.

That creates demand for more compute, which in turn drives model expansion and further investment.

4. Are preorders from large-model developers risky?

Kress noted many preorders reflect cloud providers moving to GPU infrastructure and long-term compute needs tied to capital availability.

NVIDIA focuses on customer purchase orders (POs) and capital readiness, and is working to ensure compute capacity and funding for models being developed now and over the next 12–24 months.

5. Vera Rubin and expected performance leap

Kress said the Vera Rubin architecture has completed tape-out and is slated for market launch in H2 next year.

She expects multi‑fold performance gains versus prior generations.

6. Gross margin outlook

Despite HBM memory cost concerns in the industry, Kress said NVIDIA expects to maintain gross margins in the mid‑70% range next year.

7. Inventory and purchase commitments surged

NVIDIA reported about a $25 billion USD increase in combined inventory and purchase commitments last quarter — roughly ¥180 billion RMB ($25 billion USD).

Kress described this jump as building supply for future growth, with much inventory in production and likely being shipped to customers.

8. Capital allocation priorities

With guidance of $350–400 billion USD for next year (≈ ¥2.52–2.88 trillion RMB), capital priorities are:

• First: internal expansion — supply, capacity and product development, such as launching Vera Rubin.

• Second: shareholder returns — buybacks and dividends funded by free cash flow.

• Third: strategic ecosystem investments and smaller M&A that bring engineering teams or R&D advantages, since transformational deals are hard to pull off today.

What this means for investors, founders, and operators

Keep these concise signals in your playbook.

• Contracts vs. guidance: Large LOIs can move sentiment, but only signed deals are guidance‑backed.

• Platform stickiness: CUDA and NVIDIA’s system co‑design are central competitive edges to watch for any GPU rival.

• Supply-side signals: A $25 billion USD jump in inventory and purchase commitments suggests customers and cloud builders are positioning for scale.

• Product cadence: Vera Rubin tape-out and Blackwell deployments are critical catalysts for performance and demand expectations.

• Margin resilience: Management expects mid‑70% gross margins despite memory-cost headwinds, which matters for free cash flow and buyback capacity.

Takeaway

Colette Kress’s message was cautiously optimistic.

Big potential contracts, including the OpenAI LOI, could materially change NVIDIA’s revenue trajectory, but they are not yet baked into guidance.

NVIDIA continues to position its hardware and software stack as the dominant platform for accelerated AI computing while preparing supply and product roadmaps to meet expected demand.

Keyword: NVIDIA CFO: OpenAI $100 billion order not finalized.

References

• NVIDIA CFO Interview — Cailian Press (财联社)
• 英伟达CFO访谈：OpenAI千亿大单尚未敲定 领先优势“绝对没缩小” – 东方财富 (Eastmoney)
• Investor Relations — NVIDIA