Labels

NVIDIA × OpenAI: Inside the “Biggest AI Infrastructure Project in History” — What 10 GW Really Means for Chips, Power, and the AI Economy

NVIDIA × OpenAI: Inside the “Biggest AI Infrastructure Project in History” — What 10 GW Really Means for Chips, Power, and the AI Economy

Author: Zion Zhao Real Estate｜狮家社小赵

Author's Note: For full transparency, I am a long time shareholder of AMD, NVDA and TSM, these are not financial advice and my views are bias. Please do your own due diligence.  

Executive overview

NVIDIA and OpenAI have announced a multi-year strategic partnership to deploy at least 10 gigawatts (GW) of NVIDIA systems for OpenAI’s next-generation AI infrastructure, with the first 1 GW slated for the second half of 2026on the new NVIDIA Vera Rubin platform. NVIDIA also intends to invest up to $100 billion in OpenAI, progressively, as each gigawatt comes online. OpenAI frames the buildout as essential to remove binding compute constraints on product velocity and research, while NVIDIA characterizes it as the largest computing effort ever undertaken—an “AI industrial revolution” moment that will push AI out of the lab and into every industry at scale (NVIDIA, 2025a; OpenAI, 2025; NVIDIA, 2025b). NVIDIA Blog+2OpenAI+2




What, exactly, has been announced?

  • Scale & timeline. OpenAI will deploy ≥10 GW of NVIDIA-powered data-center capacity, beginning with 1 GW in H2 2026 on Vera Rubin systems. The partnership spans millions of GPUs over multiple years and sites. NVIDIA’s investment plan (up to $100 billion) ties disbursements to delivered gigawatts—a mechanism that aligns capital with execution milestones (NVIDIA, 2025a; OpenAI, 2025; NVIDIA, 2025b). NVIDIA Blog+2OpenAI+2

  • Context with partners. OpenAI and NVIDIA position this as additive to existing contracted capacity (e.g., Microsoft Azure, Oracle OCI, CoreWeave) and to Stargate build-outs with Oracle and SoftBank, underscoring the magnitude of anticipated demand (OpenAI, 2025; YouTube/CNBC, 2025; OpenAI, 2025b). OpenAI+2YouTube+2

  • Demand signal. OpenAI points to extraordinary user traction—~800 million weekly active users—and argues that compute scarcity is delaying features and constraining the product roadmap (TechCrunch, 2025). TechCrunch

The Vera Rubin platform: why it matters for inference at scale

NVIDIA’s Vera Rubin NVL144 CPX platform is engineered for long-context, multimodal, and agentic inference at rack scale. Public technical materials describe a single rack integrating 144 Rubin CPX GPUs + 144 Rubin GPUs + 36 Vera CPUs to deliver ~8 exaFLOPs of NVFP4 compute, ~100 TB of fast memory, and ~1.7 PB/s of memory bandwidth—about 7.5× the AI performance of GB300 NVL72 systems (NVIDIA Developer Blog, 2025; NVIDIA Newsroom, 2025; CRN, 2025). This rack-scale design is a direct response to the economics of tokens-per-dollar and latency-per-query that dominate inference, particularly for million-token contexts and streaming video generation. By densifying compute and memory, Rubin CPX aims to lower serving costs while enabling new classes of workloads that were impractical on prior stacks. NVIDIA Developer+2NVIDIA Newsroom+2

Fact-check note. NVIDIA has publicly indicated that Rubin follows Blackwell (with HBM4, 3-nm TSMC processes) on a 2026 cadence—consistent with the H2-2026 milestone for the first 1 GW (Wikipedia summary corroborates the announced sequencing; company sites are the primary sources of record) (Wikipedia, 2025; NVIDIA, 2025b). Wikipedia+1

Why “10 GW”? Translating gigawatts into clusters, sites, and supply chains

A gigawatt in a hyperscale AI context bundles three realities:

  1. Semiconductor throughput. Millions of high-end accelerators and companion CPUs/NICs must be fabbed, packaged (advanced HBM), tested, and integrated. NVIDIA is the only vendor today with end-to-end platform control at this scale—silicon, systems, networking, software—and an ecosystem (CUDA, TensorRT-LLM, NeMo, DGX, MGX) tuned across training and inference. The phased “per-GW” investment links capital to real manufacturing and delivery milestones, a critical de-risking step for both parties (NVIDIA, 2025b; NVIDIA, 2025a). NVIDIA Newsroom+1

  2. Energy availability. Each 1 GW block implies multi-site, multi-PPA power planning (grid upgrades, renewables, potentially nuclear on longer horizons), with PUE-aware design to keep effective compute watts maximized for inference. OpenAI executives emphasize a world of compute scarcity, predicting that far more capacity will be needed as agentic use grows (YouTube/CNBC, 2025; OpenAI, 2025b). YouTube+1

  3. Operational readiness. To unlock usable tokens at scale, Rubin must arrive with mature software(compiler/runtime, KV-cache policies, quantization paths like NVFP4/FP8), rack-scale orchestration, and networking that avoids bottlenecks under MoE and multi-modal fan-out. Rubin CPX’s rack-level memory and bandwidth targets are purpose-built for these pressure points (NVIDIA Developer Blog, 2025; CRN, 2025). NVIDIA Developer+1

Strategic logic: scale, alignment, and option value

  • For OpenAI. The partnership provides a guaranteed lane on NVIDIA’s most advanced serving platform, tied to a financed build plan that scales with delivered GW. It is additive to OpenAI’s broader multi-cloud strategy (Azure, OCI, CoreWeave) and Stargate—raising the ceiling on product velocity and research throughput (OpenAI, 2025; YouTube/CNBC, 2025; OpenAI, 2025b). OpenAI+2YouTube+2

  • For NVIDIA. The arrangement converts roadmap leadership into long-dated demand visibility, while placing Rubin CPX as the reference rack for frontier inference. The per-GW investment is a capital-alignment tool that magnifies ecosystem pull-through (servers, networking, software, services) and signals confidence to power and construction partners (NVIDIA, 2025b; NVIDIA, 2025a). NVIDIA Newsroom+1

  • For the ecosystem. If the token economy compounds (longer contexts, ever-on agents, video-native interfaces), 10 GW will be a floor, not a ceiling. OpenAI’s public usage metrics and executives’ framing of “one GPU per person” as an aspiration illustrate the demand convexity if serving costs keep falling (TechCrunch, 2025; YouTube/CNBC, 2025). TechCrunch+1

What to watch: five execution milestones

  1. H2 2026 “1 GW on Rubin” landing. Evidence of stable latency-per-token and compelling tokens-per-dollar vs. prior platforms. NVIDIA Newsroom

  2. Rack-scale maturity. Live deployments of NVL144 CPX footprints with consistent memory-bandwidth utilization near spec in production. NVIDIA Developer

  3. Power deals and interconnects. Transparent PPAs and substation/intertie progress for multi-region sites (especially those linked to Stargate). OpenAI

  4. Developer surface area. Expansion of toolchains (e.g., TensorRT-LLM for multimodal/video, NeMo Guardrails) that reduce serving costs for million-token contexts. NVIDIA Developer

  5. Through-cycle financing. Continued cadence of per-GW investment alongside third-party capital (clouds, infra cos) to maintain build velocity. NVIDIA Newsroom

Risks & realities

  • Energy and siting. Grid lead times, PUE optimization, and local permitting can delay gigawatt-scale projects, even when chips are ready. Mitigation: multi-region build, diversified PPAs, modular racks. OpenAI

  • Supply chain. HBM4 and advanced packaging remain tight globally; synchronized capacity ramps with TSMC and memory vendors are non-negotiable (Wikipedia context on Rubin/HBM4; primary confirmations from vendors are pending detailed disclosures). Wikipedia

  • Software parity for new modalities. Achieving consistent latency/quality for long-context, streaming multimodal inference will require rapid iteration in compilers, schedulers, and runtime policies (NVIDIA Developer Blog, 2025). NVIDIA Developer

Final verdict

NVIDIA and OpenAI are attempting something categorically larger than a typical cloud rollout: a staged, capital-aligned sprint to 10 GW on a platform purpose-built for long-context, multimodal inference. Tethering investment to delivered gigawatts minimizes capital misfire while signaling to fabs, power providers, and builders that demand is real and dated. If Rubin CPX lands on time, with the promised memory/bandwidth density and runtime efficiency, this partnership will likely reset the unit economics of serving—and pull the broader ecosystem (apps, agents, safety tooling, edge offload) into a higher-productivity orbit. The first clear gate: 1 GW in H2 2026. After that, the question won’t be whether 10 GW is needed—but how soon the next 10 GW must be scheduled. NVIDIA Newsroom


Build Durable Wealth in the AI Decade—With a Cross-Asset Real Estate Strategist

When AI megaprojects like NVIDIA × OpenAI’s 10-GW buildout reshape capital flows, you deserve a property advisor who connects chips, power, rates, and rents—calmly and rigorously.

I’m a Singapore-based Real Estate Salesperson with a cross-asset lens. Drawing on my background in macroeconomics, equities, and crypto trading, plus training in Singapore Land & Business Law, I translate global shifts—AI infrastructure, energy supply, interest-rate cycles, and policy—into clear, property-first strategies. As an SAF Officer (OC, Captain), I bring discipline, integrity, and steady execution to every mandate.

Every day, I dedicate hours to research and writing—publishing analysis on AI infrastructure (e.g., NVIDIA × OpenAI 10-GW), international geopolitics, and markets—so your decisions are grounded in thorough due diligence, not headlines.

Why engage me—now

  • Cross-Asset Intelligence, Property Outcomes
    I connect the dots from AI compute and power constraints to Singapore’s leasing, pricing, and developmentdynamics—helping you time entries, structure financing, and select assets that compound.

  • Portfolio Construction Mindset
    Position real estate as your lower-volatility core—with dividend-like rental yields and long-run capital appreciation—complementing equities, fixed income, and digital assets.

  • Institutional-Grade Process
    From PSF vs breakeven and developer margin analysis to stress-tested cash flows and sensitivity to rates/vacancy, my process is transparent, numbers-first, and execution-ready.

  • Global Client Coverage
    Bespoke advisory for UHNW families, family offices, institutions, and international clients across Singapore, China, and Southeast Asia（含 陪读家长 / 留学 / 家办）.

What working together looks like

  1. Discovery: Goals, risk budget, horizon.

  2. Strategy: Sub-market selection, unit typologies, yield & appreciation targets.

  3. Numbers: Scenario models, financing options, downside controls.

  4. Execution: Shortlist → viewings → negotiation → documentation → asset management.

  5. Ongoing Intelligence: Regular updates on AI-driven demand, rates, and supply—so you stay ahead, not reactive.

A calm, humble promise

No hard sell. Just courteous, professional, diligent work. I show my sources, explain trade-offs, and protect your time and capital.


行动邀请（适用于中国与东南亚客户）

在 AI × 能源 推动的新周期里，全球资本与产业链正在重塑。新加坡房地产可作为您组合中的稳健核心资产：稳定租金回报＋长期增值。我每天投入大量时间研究与写作（含 NVIDIA × OpenAI 10GW 解析），以严谨尽调为您与家人（含陪读家长、留学、家办）提供定制方案与全流程服务。欢迎联系我，一对一私享咨询。



Ready to future-proof your portfolio?

Message me for a private consultation. Let’s add a stable, income-producing Singapore property to your multi-asset strategy—so you capture the upside of the AI decade without taking on unnecessary volatility.

In-text citations (APA-style)

  • NVIDIA. (2025a). NVIDIA and OpenAI announce “the biggest AI infrastructure project in history” (company blog). NVIDIA Blog

  • NVIDIA. (2025b). OpenAI and NVIDIA announce strategic partnership to deploy 10 GW of NVIDIA systems(newsroom release). NVIDIA Newsroom

  • OpenAI. (2025). OpenAI–NVIDIA systems partnership (press page). OpenAI

  • OpenAI. (2025b). OpenAI, Oracle, and SoftBank expand Stargate with five new AI data-center sites. OpenAI

  • NVIDIA Developer Blog. (2025). NVIDIA Rubin CPX accelerates inference performance and efficiency for 1M-token context workloads. NVIDIA Developer

  • NVIDIA Newsroom. (2025). NVIDIA unveils Rubin CPX: A new class of GPU designed for massive-context inference. NVIDIA Newsroom

  • CRN. (2025). Vera Rubin NVL144 CPX platform specs and positioning. CRN

  • TechCrunch. (2025). Sam Altman says ChatGPT has hit 800M weekly active users. TechCrunch

  • YouTube/CNBC. (2025). Interview with Jensen Huang, Sam Altman, Greg Brockman—deal details and context. YouTube

  • Wikipedia. (2025). Rubin (microarchitecture) (context on HBM4/TSMC and sequencing beyond Blackwell). Wikipedia

References (APA)

CRN. (2025, September 9). Nvidia packs “new class of GPU” inside Vera Rubin NVL144 CPX platform. CRN

NVIDIA. (2025a, September 22). NVIDIA and OpenAI announce “the biggest AI infrastructure project in history”. NVIDIA Blog

NVIDIA. (2025b, September 22). OpenAI and NVIDIA announce strategic partnership to deploy 10 GW of NVIDIA systems. NVIDIA Newsroom

NVIDIA Developer Blog. (2025, September 9). NVIDIA Rubin CPX accelerates inference performance and efficiency for 1M-token context workloads. NVIDIA Developer

OpenAI. (2025, September 22). OpenAI–NVIDIA systems partnership. OpenAI

OpenAI. (2025, September 23). OpenAI, Oracle, and SoftBank expand Stargate with five new AI data-center sites. OpenAI

TechCrunch (Bellan, R.). (2025, October 6). Sam Altman says ChatGPT has hit 800M weekly active users. TechCrunch

Wikipedia. (2025, September). Rubin (microarchitecture). Wikipedia

YouTube/CNBC. (2025, September). Nvidia plans to invest up to $100 billion in OpenAI as part of 10 GW deal. YouTube

Labels:

Comments

Post a Comment