NVIDIA, OpenAI, and the New Economics of Compute: What the “AI Factory” Era Really Means
NVIDIA, OpenAI, and the New Economics of Compute: What the “AI Factory” Era Really Means
By Zion Zhao Real Estate | 狮家社小赵
Executive summary
The AI industry has crossed a structural threshold. The conversation with Jensen Huang on BG2 podcast with Brad Gerstner and Clark Tang crystallizes three big ideas: (1) general-purpose computing is giving way to accelerated computing and AI-native “factories,” (2) inference is no longer a one-shot answer but a multi-stage process of “thinking” (which radically increases compute demand), and (3) the winners won’t just ship chips—they will deliver full, co-designed systems at the scale of nations. In this essay I unpack those claims, test them against outside evidence, flag where the hype is ahead of the facts, and explore the policy and macro consequences—from power grids to immigration and the “Invest America/Trump Accounts” push to broaden capital ownership.
1) From general-purpose computing to accelerated “AI factories”
Huang’s central thesis—that traditional CPU-centric computing has run out of road and will be replaced by accelerated computing and AI—maps to a well-documented slowdown in classical Moore’s-law style gains and to the rise of domain-specific acceleration (GPUs, DPUs, AI NICs). International Energy Agency modeling shows data-center electricity demand is entering a steep up-curve through the mid-2020s, with AI the principal driver; that trend is consistent with a large, durable capex cycle in accelerated compute, networking, and power infrastructure (IEA, 2024). Fabricated Knowledge
NVIDIA’s product roadmap reinforces the “systems not chips” framing. Since Blackwell (2024), the company has moved to an annual cadence—Blackwell → Blackwell Ultra (GB300, 2025) → Vera Rubin (2026) → Rubin Ultra (2027) → Feynman (2028)—with each generation paired to networking (InfiniBand and Spectrum-X Ethernet) and software stacks (CUDA, cuDNN, Triton, NIMs) designed to squeeze more tokens per watt out of the same power “shell.” Independent coverage and NVIDIA releases confirm this cadence and networking push (Reuters, 2025; AP, 2025; NVIDIA, 2023, 2025). NVIDIA Investor Relations+3Reuters+3AP News+3
Implication. In practical terms, hyperscalers optimize at the factory level: chips, boards, NVLink/PCIe, top-of-rack switching, optics/photonic links, schedulers, and datacenter layout. If your limiting reagent is power (and it is), you buy tokens per watt per square meter, not chips per se.
2) Inference is “thinking,” not “answering”—and that multiplies demand
The podcast highlights three “scaling laws”: (i) pre-training, (ii) post-training (RL/finetuning/“practice”), and (iii) inference “thinking” (multi-step reasoning, tool use, research) that trades latency for quality. The academic literature long ago showed that loss scales predictably with model/data/compute (Hestness et al., 2017; Kaplan et al., 2020). What’s new is that inference itself is becoming compute-intensive via chain-of-thought/agentic workflows—evident in recent reasoning-model announcements that explicitly increase inference compute to improve accuracy. OpenAI’s 2024–2025 technical notes on “reasoning models” align with this shift, as do usage statistics showing extraordinary growth in tokens generated per user session. Wikipedia+1
On the demand side, ChatGPT now reports ~700–800 million weekly active users (OpenAI/Harvard working paper; OpenAI blog; major-press summaries)—and usage keeps compounding as agentic features roll out (OpenAI, 2025; Barron’s, 2025). If “thinking” adds multiple passes of planning, retrieval, tool calls, and self-critique, required compute per query rises non-linearly. Barron's+3OpenAI+3NBER+3
Implication. The “glut soon” narrative underestimates the second exponential: even at a fixed user base, per-request compute is climbing as products shift from chatbots to agents.
3) Is the OpenAI–NVIDIA–Oracle triangle “circular”—or just industrial finance?
Critics worry about “round-tripping” (revenue inflated by reciprocal deals). The historical red flag is early-2000s telecoms that booked round-trip transactions with little substance; the SEC subsequently issued guidance and pursued cases to curb such practices. But the present arrangement differs in economic substance: OpenAI sells services to consumers and enterprises at massive scale; hyperscalers lease capacity to paying tenants; NVIDIA ships hardware/software that runs those workloads. If contracts are non-contingent and priced at arm’s length, the accounting is fundamentally different from “you buy my dark fiber, I buy yours” shells of 2000–2001 (U.S. SEC; see also contemporaneous reporting on Nortel/Cisco-era round trips). SEC+1
What is new is scale. Multiple reputable outlets report NVIDIA’s plan to invest up to $100 billion in OpenAI to accelerate a ~10 GW self-build called “Stargate,” alongside Azure, Oracle Cloud Infrastructure (OCI), and CoreWeave expansions. This has been corroborated by Reuters and others; CoreWeave’s own OpenAI commitment has been revised upward multiple times (from ~$12B equivalent to billions more in 2024–2025)—a sign of demand, not purely financial optics (Reuters, 2025a; 2025b). SEC+2fintracadvisors.com+2
Implication. The structure resembles automotive and industrial finance: vendors co-invest with anchor customers, but revenue is still tied to third-party demand. The key risks are execution (power, supply chain, software readiness) and regulatory/political constraints, not accounting per se.
4) Will there be a glut? The physics and the economics say “not yet”
Skeptics argue: “After the 2024–2026 buildout, demand plateaus.” But three forces work the other way:
Refresh & substitution. Hyperscalers are still migrating recommender/search/video pipelines from CPU-centric designs to accelerated AI inferencing. This alone is a multi-year refresh cycle (NVIDIA/Reuters coverage of the roadmap suggests sustained cadence). Reuters
Token-per-watt race. If power caps bind, the only way to expand output is new generations with higher perf/W (NVIDIA and networking press releases detail Spectrum-X/photonic plans to lift throughput per watt and per port). NVIDIA Investor Relations+1
New workloads. Multimodal generation, AI video, and physical AI (robotics) are early but real; AP’s GTC reporting and NVIDIA’s robotics stack announcements suggest these will be next-wave demand sinks. AP News
IEA’s energy outlook is the sober check: grids must add capacity and flexibility. Yet adoption curves and the policy push for more generation (including nuclear life-extensions, gas turbines, and new interconnects) imply supply can grow with time (IEA, 2024). The bottleneck is near-term power siting and transmission, not lack of AI workloads. Fabricated Knowledge
5) Competitive dynamics: GPUs vs. ASICs is the wrong frame
Yes, Google’s TPU proves a focused vertical can scale, and more companies are taping out accelerators. But Huang’s point—borne out by the last two years—is that the competitive unit is the factory: compute + memory + IO + networking + software + orchestration. Independent reports confirm NVIDIA’s disaggregation strategy (e.g., specialized CPX parts for specific inference phases, orchestration layers like DynamO/NVLink-Scale/Spectrum-X) and an annual whole-stack refresh (Reuters, 2025; NVIDIA releases). If your capex is dominated by land/power/shell, a 20–30× jump in tokens per watt beats free silicon that can’t utilize the same facility as efficiently (Reuters, AP; NVIDIA). Reuters+2AP News+2
Reality check. There will be substantial ASIC success—especially for stable, high-volume sub-tasks (video transcode, embedding pre-compute, KV-cache handling). But for the heterogeneous, fast-changing soup of agentic inference and post-training, programmability and networked scale dominate.
6) Sovereign AI, export policy, and the China question
Huang argues every country will need sovereign AI—both to encode culture/language and to run critical infrastructure. That already shows up in national AI strategies and in regulation (e.g., the EU AI Act), but the deeper point is industrial capacity: compute, data, networks, and energy. Policy trade-offs are hard—e.g., U.S. export controls aimed at slowing advanced AI in China have side effects, including fueling domestic Chinese champions (Huawei) behind a protected wall. Public reporting in 2024–2025 documents iterative U.S. rules, continual Chinese design changes, and tightening around high-end interconnects. This is less a clean decoupling than a slow technology knife-edge.
On the U.S. side, two parallel moves in 2025 matter:
Licenses & exports. The Commerce Department has emphasized a “small yard, high fence” approach—restricting the most sensitive capabilities but not broad AI exports. That nuance now appears to be loosening to speed “American stack” diffusion abroad, according to administration briefings and official fact sheets (Commerce/White House materials in 2025). The White House+1
“Trump Accounts” / Invest America.** As part of the “One Big Beautiful Bill,” the administration promoted $1,000 seeded accounts for newborns (“Trump Accounts”) with optional private top-ups—framed as widening capital ownership. White House publications and major-press reporting describe the policy architecture, though critics debate distributional effects and long-term efficacy (White House; Washington Post; The Guardian). Note that Cory Booker/Ayanna Pressley’s “Baby Bonds” is a distinct proposal with larger, progressive contributions; it provides useful policy research context but is not the same program (Congress.gov; CAP). Center for American Progress+4The White House+4The Washington Post+4
Implication. Sovereign AI is simultaneously a national security project and a development project. The strategic game is: can the U.S. export a compelling, integrated stack (models + hardware + cloud + tooling) fast enough, to enough friends, to set de facto standards?
7) Talent, H-1B fees, and the American Dream
Huang is blunt that U.S. leadership hinges on immigration of top technical talent. In September 2025, the U.S. government announced a new $100,000 fee for H-1B petitions, with USCIS guidance spelling out exemptions and DV draw rules. The policy has sparked significant debate about impacts on startups versus incumbents (White House; USCIS). forensicrisk.com+1
From a labor-economics standpoint, the evidence base is clear: frontier technology productivity gains diffuse faster when skilled immigration pipelines are open. Randomized and quasi-experimental studies (e.g., Brynjolfsson et al., 2023/2024 on AI assistance; broader high-skill immigration literature) show meaningful productivity and wage complementarities for complementary domestic workers. The question is not if to recruit global talent, but how to structure fees/quotas so they don’t unintentionally tilt against early-stage companies.
Implication. If the U.S. aims to be the world’s AI “workshop,” immigration and grid build-out are not side dishes; they are the main course.
8) What to watch next (and how to invest with discipline)
Power, not paper. Follow interconnect approvals, substation builds, and firm power PPAs, not just chip SKUs. IEA’s updates, utility IRPs, and hyperscaler energy disclosures are leading indicators. Fabricated Knowledge
Tokens-per-watt and networking. Keep an eye on Spectrum-X and silicon photonics milestones; they translate directly into effective capacity without new land. NVIDIA Investor Relations+1
Agentic use-cases. The fastest-growing compute sink is agentic inference in enterprise workflows—retrieval + tools + verification. Adoption will track compliance & audit tooling, not just model benchmarks.
OpenAI’s monetization. With 700–800M WAU and growing enterprise attach, even small ARPU steps (ads, API minimums, enterprise seats) materially de-risk long-term offtake. OpenAI+1
Policy drift. H-1B, export license tempo, and “Trump Accounts” implementation will shape where value accrues (startups vs. incumbents; U.S. vs. allies).
9) Bottom line
Huang’s bolder claim—that NVIDIA could be the first $10T company—sounds sensational until you model capex as a function of global token generation and perf/watt. The strongest counter-argument is physics (power and heat) plus execution risk (annual co-design cadence is brutal). But the balance of evidence—usage, roadmaps, power planning, and policy momentum—supports the AI factory thesis: this is a multi-year, system-level buildout with real economic substance.
The American twist is whether we pair the buildout with broad-based participation—immigration that attracts the world’s best, and ownership programs that let more citizens share in the upside. That’s not a side narrative to AI. It is the American Dream applied to the compute age.
Your Edge in a World That’s Changing Faster Than Ever
If you believe great portfolios are built at the intersection of property, policy, and technology—I’m your agent.
Based in Singapore and serving clients across China, Southeast Asia, and global markets, I help ultra–high-net-worth families, family offices（家办）, institutional investors, and internationally mobile parents（陪读家长、留学规划） allocate into Singapore real estate with the same rigor you’d expect from a global macro fund.
Why work with me
Cross-asset intelligence, not tunnel vision. I track and write daily on AI’s “factory” economy (e.g., NVIDIA × OpenAI × Oracle), TikTok’s U.S. carve-out, FAANG+ technicals, H-1B policy shifts, and the evolving rules of speech and platforms—because capital flows follow technology, regulation, and geopolitics, not headlines.
Portfolio-first property advice. As a seasoned equity & crypto trader and macro analyst, I model how Singapore real estate complements risk assets—lower volatility, attractive rental yield (dividend-like cash flow), and disciplined capital appreciation potential—so your property strengthens the whole portfolio, not just fills a postcode.
On-the-ground legal fluency. Proficient in Singapore Land Law, Business Law, statutes and regulations, I structure transactions with precision, coordinate with counsel, and help you move decisively—compliantly.
Military-grade execution. As an Officer Commanding (SAF, Captain), I bring planning discipline, discretion, and reliability to every mandate.
What you can expect
Institutional diligence. I dedicate hours every day to publish essays and study macroeconomics, policy, and markets—so your decisions ride tailwinds, not noise.
Global-to-local translation. I convert AI, policy, and liquidity shifts into neighborhood-level strategy: where to buy, what to hold, how to finance, and when to rebalance.
Full-lifecycle solutions for international families. School districts, visas, family-office set-up（家办）, 陪读/留学 neighborhoods, and rental management—all with one point of accountability.
UHNW & institutional coverage. Discreet off-market sourcing, block purchases, development participation, and data-driven hold/exit frameworks.
Why add Singapore property now
Stability with yield. Blue-chip legal system, strong title security, and a landlord-friendly rental market provide income resilience and lower drawdowns versus public markets.
Policy clarity. Transparent regulation and prudent monetary/immigration frameworks create an investable runway for long-term capital.
AI-era beneficiary. As compute, cloud, and capital consolidate in Asia, Singapore remains a hub for talent, family offices, and sovereign flows—supporting demand in prime residential and commercial segments.
Let’s build a portfolio that earns in any cycle
Whether you’re deploying from Beijing, Shanghai, Shenzhen, Hong Kong, Jakarta, Bangkok, KL, Dubai, London, or New York, I’ll bring macro perspective + legal precision + local execution to your mandate.
咨询欢迎（中文/English）：
家办与资产配置｜新加坡置业与税务配合｜陪读与留学安家
制定跨资产与跨司法辖区的一致性方案，稳中求进。
Add a resilient, income-producing Singapore property sleeve to your portfolio—backed by deep research, daily diligence, and uncompromising execution.
👉 Message me to schedule a private strategy call (virtual or in person). Let’s align on objectives, risk, timelines, and convert today’s macro into tomorrow’s cash flow.
References (APA)
Barron’s. (2025, September 27). Nvidia CEO Jensen Huang says OpenAI will be the next ‘multi-trillion-dollar’ company.https://www.barrons.com Barron's
Brynjolfsson, E., Li, D., & Raymond, L. (2023). Generative AI at work (NBER Working Paper No. 31161). National Bureau of Economic Research.
Hestness, J., Narang, S., Ardalani, N., et al. (2017). Deep learning scaling is predictable, empirically. arXiv:1712.00409. acquirersmultiple.com
International Energy Agency (IEA). (2024). Electricity 2024: Analysis and forecast to 2026 (Data centers & AI section). https://www.iea.org Fabricated Knowledge
Kaplan, J., McCandlish, S., Henighan, T., et al. (2020). Scaling laws for neural language models. arXiv:2001.08361. Wikipedia
NVIDIA. (2023, May 28). NVIDIA launches accelerated Ethernet platform for hyperscale generative AI (Spectrum-X).https://nvidianews.nvidia.com NVIDIA Newsroom
NVIDIA. (2025, August 22). NVIDIA introduces Spectrum-XGS Ethernet to connect distributed data centers into giga-scale AI super-factories. https://investor.nvidia.com NVIDIA Investor Relations
OpenAI & Deming, D. (2025, September 15). How people are using ChatGPT (OpenAI Research / NBER Working Paper No. 34255). https://openai.com & https://www.nber.org OpenAI+1
Reuters. (2025, March 18). Everything Nvidia announced at its annual developer conference (GTC).https://www.reuters.com Reuters
Reuters. (2025, September 20). CoreWeave expands OpenAI deal again, adding $6.5B in capacity.https://www.reuters.com SEC
Reuters. (2024–2025). OpenAI “Stargate” buildout: Oracle, SoftBank, and 10 GW target. https://www.reuters.comfintracadvisors.com
The Associated Press. (2025, March 18). Nvidia CEO Jensen Huang unveils Blackwell Ultra and Vera Rubin at GTC 2025. https://apnews.com AP News
The U.S. Securities and Exchange Commission (SEC). (2003). Revenue recognition in transactions involving “round-trip” arrangements (Staff Accounting Bulletin references & enforcement actions). https://www.sec.gov SEC
The White House. (2025, September 17). Proclamation—Securing America’s future through high-skill immigration: H-1B modernization. https://www.whitehouse.gov forensicrisk.com
The White House. (2025, June 9 & August 29). Trump Accounts / “Invest America” fact sheets.https://www.whitehouse.gov The White House+1
U.S. Citizenship and Immigration Services (USCIS). (2025, September). H-1B program fee update and implementation guidance. https://www.uscis.gov TMS
The Washington Post. (2025, June 9). Trump pushes $1,000 “Trump accounts” for babies.https://www.washingtonpost.com The Washington Post
The Guardian. (2025, June 9). Trump announces $1,000 government-funded accounts for American babies.https://www.theguardian.com The Guardian
Congress.gov. (2023). S.441 — American Opportunity Accounts Act. https://www.congress.gov Congress.gov
Center for American Progress. (2025, February 20). Baby bonds: A worthwhile step to reduce the racial wealth gap.https://www.americanprogress.org Center for American Progress
Notes on method and fact-checks
OpenAI user counts. I used OpenAI’s own research (OpenAI/Harvard/NBER) and contemporaneous coverage to ground the 700–800M WAU range; third-party dashboards were used as corroborative, not primary, sources. OpenAI+2NBER+2
Stargate & $100B. Multiple reputable outlets (Reuters, Barron’s) reported details of NVIDIA’s investment intent and OpenAI’s 10-GW program; I treat the numbers as directional and subject to contract execution. SEC+1
Round-tripping. I referenced SEC guidance and historical cases to explain why today’s arrangements differ in substance from early-2000s telecom accounting. SEC
Policy statements. For H-1B fees and Trump Accounts, I anchored on White House/USCIS and leading outlets for verification; the “Baby Bonds” bill is separate and cited for context. Congress.gov+4forensicrisk.com+4TMS+4
Disclosure: This essay is for informational purposes only and reflects my opinions as the author. It is not investment advice.
Comments
Post a Comment