NVIDIA’s Jensen Huang framed the recent tidal wave of AI investment as real demand, not a reprise of dot-com excess — telling CNBC that “this isn’t Pets.com” and arguing the AI buildout is backed by trillions of dollars of productive business. That defense arrives amid two seismic infrastructure deals: NVIDIA’s letter of intent to invest up to $100 billion to support OpenAI’s plan to deploy at least 10 gigawatts of NVIDIA systems, and AMD’s separate multi‑year agreement to supply OpenAI with up to 6 gigawatts of Instinct GPUs (together with warrants that could give OpenAI up to a 10% stake in AMD). Those deals have crystallized two competing narratives: one of an industrial‑scale shift toward GPU‑driven “generative AI” infrastructure, and another warning that a circular economy of cross‑ownership, long‑dated supply contracts, and sky‑high valuations could create systemic risk if demand or economics falter. The reality sits between both extremes — enormous technical and market momentum, and real financial and operational fragilities that require sober scrutiny.
The headlines have three interlocking components: the compute commitments, the capital flows, and the market reactions.
Key technical arithmetic cited by NVIDIA and on‑camera by Jensen Huang: 10 GW equates to roughly 4–5 million GPUs (company estimates depend on the GPU generation and rack density), and NVIDIA said building a single gigawatt of “AI factory” — including land, powered shell, networking and compute — can cost on the order of $50–$60 billion. Huang emphasized capital will come via a mix of revenue, equity and debt, not all in one upfront payment. Those figures are company statements and should be understood as high‑level guidance rather than audited cost breakdowns.
But the winning path will not be a simplistic “more GPUs = more revenue” equation. It requires disciplined financing, diversified supply strategies, verifiable off‑takers, and careful engineering of data‑center ecosystems. Journalists, CIOs and investors should read the fine print: deployment timetables, milestone triggers, vesting schedules for equity instruments, and the practicalities of power and cooling. Those details will determine whether the AI era becomes a productive industrial revolution — or a capital‑intensive boom that produces painful corrections for overextended players.
At the moment, the balance of evidence supports continued investment and deployment — provided commercial monetization of AI services keeps pace with the scale of infrastructure being ordered. The next 18–30 months of first‑gigawatt turn‑ups, utilization rates and enterprise billing will be the truest test of whether “this isn’t Pets.com” is the correct verdict — or simply the most persuasive marketing line yet from a supplier at the center of the story.
Source: Windows Central Why NVIDIA’s CEO thinks the AI boom won’t go bust — and how it’s different from 2000
Background
The headlines have three interlocking components: the compute commitments, the capital flows, and the market reactions.- NVIDIA and OpenAI announced a strategic partnership to deploy at least 10 gigawatts of NVIDIA systems across multiple data centers, with NVIDIA indicating it intends to invest up to $100 billion progressively as infrastructure comes online. The first gigawatt is targeted for deployment in the second half of 2026.
- AMD announced a 6‑gigawatt agreement to supply OpenAI with AMD Instinct GPUs beginning with a 1‑gigawatt deployment planned for the second half of 2026; the contract includes warrants enabling OpenAI to acquire up to roughly 10% of AMD under milestone and stock‑price conditions.
- Commentary from skeptical academics and investors — notably Gary Marcus and others — has highlighted fragility in some headline-sized agreements (for example, the reported OpenAI–Oracle arrangement with an eye‑watering figure cited in market commentary), arguing that not all the dollars pledged on paper translate to real, timely funding and that infrastructure supply cannot always be monetized as assumed.
What the deals actually commit to
NVIDIA + OpenAI: the 10‑gigawatt LOI
NVIDIA’s public announcement and OpenAI’s corroborating statement describe a letter of intent for OpenAI to build and run at least 10 GW of NVIDIA‑based systems, with NVIDIA committing to progressive investment — described as “up to $100 billion” — tied to deployments and milestones. The companies have stated the first phase will use NVIDIA’s next‑generation Vera Rubin systems and is expected online in 2H 2026. NVIDIA and OpenAI framed this as a strategic, preferred‑supplier relationship intended to scale OpenAI’s frontier training and inference capacity. The announcement explicitly characterizes the $100 billion as an intent to invest progressively — not a single cash transfer — and includes typical forward‑looking caveats.Key technical arithmetic cited by NVIDIA and on‑camera by Jensen Huang: 10 GW equates to roughly 4–5 million GPUs (company estimates depend on the GPU generation and rack density), and NVIDIA said building a single gigawatt of “AI factory” — including land, powered shell, networking and compute — can cost on the order of $50–$60 billion. Huang emphasized capital will come via a mix of revenue, equity and debt, not all in one upfront payment. Those figures are company statements and should be understood as high‑level guidance rather than audited cost breakdowns.
AMD + OpenAI: 6 gigawatts and equity alignment
AMD’s press release and OpenAI’s statement confirm a 6 GW framework with the first 1 GW based on AMD’s forthcoming Instinct MI450, and a structured warrant package giving OpenAI the option to acquire up to 160 million shares (roughly 10% of AMD) that vest with staged deployments or share‑price milestones. AMD expects the deal to generate tens of billions in revenue over time and positions the company as a significant second supplier in an ecosystem heavily dominated by NVIDIA hardware. AMD’s language is similarly forward‑looking and conditioned on product delivery and market adoption.Why leaders like Jensen Huang say “this isn’t Pets.com”
Jensen Huang’s central argument rests on three linked assertions:- Scale and economic grounding: Today’s hyperscalers and enterprise cloud workloads already represent multi‑trillion‑dollar industries tied to revenue‑generating business — far larger and more materially grounded than the limited consumer dot‑com plays of 1999–2000. Huang quantified current hyperscaler business measured in the trillions versus the internet survivors of the dot‑com era.
- Productivity and revenue‑bearing AI: He argues generative AI has moved from entertaining or experimental models to systems that reason, search and perform tool‑enabled tasks that businesses will pay for (for example, increased developer productivity, automated research, and enterprise copilots). Huang claims token‑based pricing economics are now profitable because the models deliver measurable value per query. That transition from novelty to paid utility is key to his claim that demand is structural, not speculative.
- A multiyear industrial buildout: The practical reality of deploying GW‑scale data centers — land, permits, power, cooling, networking and the long procurement timelines for advanced GPUs — means buildout will occur over years, providing a multi‑year revenue runway for chip and systems vendors. Huang emphasizes that this multi‑year cadence is a structural buffer against a dot‑com‑style immediate collapse.
Why skepticism persists
1) Circularity and stake‑swapping create fragile incentives
The deals include a pattern: vendors commit hardware and capital while the buyer receives equity or warrants, creating a circular flow where the AI buyer gets upside in its supplier and the supplier gets locked‑in demand. That can look like a feedback loop where the same economic value circulates among a small set of players without underlying end‑market proof. Skeptics point out that if revenues from end customers don’t materialize at scale, suppliers could be left with specialized, hard‑to‑redeploy racks and data centers. Gary Marcus described several agreements as symptomatic of a “peak bubble” dynamic, arguing that headline valuations and contractual totals sometimes outstrip realistic revenue streams and the buyer’s ability to pay for long‑dated commitments.2) Financing gap and timing risk
OpenAI is widely reported to have expanded revenues but also to be burning cash heavily to fund rapid growth. Critics note that deals touted at hundreds of billions on paper depend on phased financing, third‑party capital, and optimistic projections of revenue and profit timelines. If fundraising stalls, if enterprise adoption is slower than expected, or if macro conditions curtail debt or equity markets, projects can be delayed or cancelled — leaving vendors or co‑investors exposed. Multiple analysts have flagged Oracle’s large‑figure commercial arrangement with OpenAI (reported in various outlets as in the hundreds of billions in total value from framework commitments) as difficult to reconcile with OpenAI’s immediate cash flow and profitability timeline. The practical implication: contractual face value is not the same as committed, liquid funding available at the time projects must be paid for.3) Overbuild and stranded capital risk
The dot‑com analogy’s core lesson is stranded capacity: empty data centers, write‑downs and long recovery periods. Today’s AI buildout is capital‑intensive and geographically concentrated: it requires power, distribution infrastructure and specialized cooling. If demand for high‑end GPU cycles slows — for example because better algorithmic efficiency reduces compute needs, or competing architectures erode the GPU advantage — then entire facilities could be underutilized. Analysts have warned about this overbuild risk in multiple industry reports during 2025.Technical realities: compute, power, and tokens
How to read “gigawatts” and GPU counts
- “Gigawatt” in these deals is shorthand for the aggregate power envelope of a fleet of racks plus support infrastructure. It’s not only raw GPU power; it includes power distribution, cooling and facility overhead.
- Public company estimates translating 1 GW into GPU counts vary by GPU generation and rack density. NVIDIA suggested 10 GW could equate to 4–5 million GPUs, which would imply roughly 400k–500k GPUs per gigawatt at certain assumed densities — numbers useful for scale but dependent on architecture and data center design. These are vendor estimates and should be treated as high‑level capacity proxies rather than fixed engineering conversions.
Energy and geographic constraints
Large GW‑scale clusters require substantial and often bespoke electrical and cooling infrastructure. Power availability and permitting are political and logistical bottlenecks in many regions. That means hyperscalers and data‑center builders are chasing not just chips but grid interconnections and long‑term power purchase agreements — another layer that increases capital intensity and calendar lead times. Multiple reporting threads and technical briefings have highlighted the months‑to‑years timescale for fully provisioning a GW‑class facility.Tokens, pricing and the unit economics of AI compute
“Tokens” are the unit of work for LLMs; each token processed during inference or training consumes compute and memory resources. Jensen Huang and several industry leaders have argued that useful, tool‑enabled AI (models that can reason, fetch documents, or execute code) generate higher customer willingness‑to‑pay per token than earlier generations, improving the revenue per GPU‑hour calculation. That shift in marginal economics — from “experiments” to productive services — underlies the vendor case for sustained demand. However, token economics are model and use‑case dependent: high‑value enterprise workloads (e.g., legal research, drug discovery) pay more per token than consumer chat-style interactions, so aggregate revenue per GPU will vary across customers and deployments.Market structure and competitive implications
- NVIDIA remains the dominant supplier of high‑end accelerators and ecosystem software (CUDA, cuDNN, NCCL), creating a moat of hardware plus software that raises migration costs for large AI stacks. That dominance has driven both hyperscalers and chip rivals to secure multi‑year relationships and to diversify supplier risk.
- AMD’s deal with OpenAI is strategically significant: it validates AMD’s Instinct roadmap for large inference workloads and injects competition into a market where vendor concentration has been a structural weakness. Multiple independent reports show the market reaction lifting AMD’s stock on the announcement while prompting analysts to reassess competitive dynamics.
- Hyperscalers and cloud providers are reacting by pursuing their own silicon and vertical integration strategies (Microsoft’s Maia, custom accelerators and Cobalt CPU efforts are examples documented by industry reporting). This creates a multi‑track supply environment where vendors, hyperscalers, and emergent chipmakers will coexist — but not without supply‑chain stress and pricing dynamics that remain difficult to model precisely.
Practical implications for enterprises and Windows users
The AI infrastructure boom touches Windows users and enterprise IT teams in several concrete ways:- Expect deeper integration of AI capabilities into Windows and Microsoft 365 products as cloud providers scale capacity and productize inference services. This is already visible in Copilot integrations and on‑device feature experiments. Those services will draw from the same backend capacity trends underpinning these deals.
- For IT procurement, multi‑vendor compute strategies become more attractive to reduce supplier concentration risk. Enterprises building private or hybrid AI platforms should account for lead times in procuring racks and specialized GPUs, and plan for energy and facility constraints.
- On‑device AI (e.g., Copilot+ PCs with NPUs) and offline models may partially ameliorate cloud costs and latency for some workloads, but large foundation model training and many inference workloads will remain cloud‑bound for the foreseeable future. That split reshapes licensing, telemetry and data‑governance choices for Windows environments.
Four scenarios to watch (ranked by probability and impact)
- Stabilized scaling with mixed vendors (most likely, high impact). Multiple vendors (NVIDIA, AMD, custom cloud silicon) coexist; deployments scale over 3–6 years; suppliers realize strong but uneven revenue growth while hyperscalers internalize parts of the stack. This yields sustained industry growth without wholesale stranded assets.
- Slowdown and selective repricing (plausible, medium impact). Buildouts continue but at slower pace; contract terms are renegotiated or delayed; suppliers face margin compression. Some smaller data‑center projects are shelved, but core capacity for high‑value workloads remains.
- Asset stranding and write‑downs (less likely but material). Widespread overbuild or an algorithmic efficiency breakthrough reduces GPU‑hour demand; specialized GPU racks and facilities face underutilization, causing write‑downs and credit stress among providers and co‑investors. This mirrors dot‑com-era stranded capacity but on a larger industrial scale.
- Rapid technical disruption (low probability, high upside). A new architecture or algorithmic approach dramatically lowers compute per model (e.g., orders‑of‑magnitude efficiency), cutting capital requirements and reshaping vendor economics overnight — a deflationary scenario for GPU capex but a boon for democratized AI deployments. This is the polar case many technologists watch for.
What the numbers mean for investors and CIOs
- Cost per gigawatt: public comments and filings place the fully‑loaded cost of a 1‑GW data center (land, enabled shell, networking, racks and compute) in the $50–$60 billion range — a reminder that GW‑scale deployments are projects the size of major infrastructure investments, not typical IT purchases. That a single GW can carry nine‑figure or low‑ten‑figure price tags means financing structure, off‑take agreements, and project de‑risking are central.
- Vendor concentration risk: NVIDIA’s ecosystem lock‑in via CUDA and system dominance creates switching costs that favor incumbent advantage, but high prices and supply constraints create openings for strong competitors (AMD, Microsoft’s Maia, in‑house silicon from clouds). Diversification strategies and migration pathways (e.g., using abstraction layers or portability tools) are a must for large model operators.
- Real vs. headline value: The face value of long‑term framework deals (often quoted as aggregate totals reaching into the hundreds of billions) must be parsed: many figures represent long‑dated potential purchases, not immediate cash flows. Due diligence must focus on staging, milestones, and the counterparty’s credible funding path. Skeptics’ warnings about circular ownership and unbacked headline totals should be treated as a prompt for careful contract design and stress testing.
Strengths, risks and the balanced verdict
Strengths
- Real technical demand: Enterprises are actively seeking productivity gains and AI automation that materially improve workflows; that creates paying customers for many inference services.
- Ecosystem momentum: Hyperscalers, chip vendors and cloud firms are aligning product roadmaps and supply chains to deliver large‑scale models, which reduces implementation friction for end customers.
- Competition improving resilience: AMD’s large deal with OpenAI signals stronger competition in accelerators, which can lower long‑term costs and reduce single‑vendor risk.
Risks
- Circular financing and headline inflation: Equity swaps and warrants tied to hardware deals can create headline totals that obscure timing and funding gaps; careful contract review is necessary.
- Stranded asset and cash‑flow risk: The capital‑intensive nature of GW‑scale buildouts means delays or demand drops can produce large write‑downs.
- Execution and supply chain constraints: Chip production speed, packaging, cooling solutions and grid access remain practical constraints that can delay deployments and push costs higher in the short term.
Immediate watchlist: four concrete indicators to track
- Deployment milestones and first‑gigawatt turn‑ups. The industry will test these agreements when the first 1 GW installations (NVIDIA/OpenAI Vera Rubin and AMD/OpenAI MI450) go live in 2026. Successful, timely deployments reduce counterparty risk.
- OpenAI’s cash flow and fundraising cadence. Quarterly revenue growth, margins on enterprise contracts, and any external funding rounds will indicate whether OpenAI can service long‑dated procurement commitments without excessive dilution.
- Utilization and pricing of GPU instances. Whether GPU‑hour spot prices and long‑term contracts stay elevated or compress will show whether token economics and enterprise willingness‑to‑pay support the capital base.
- Regulatory and energy constraints. Local permitting, grid access and environmental permitting trends will shape where and how fast GW‑scale facilities can be built. Delays here increase the probability of project repricing or cancellation.
Conclusion
The headlines about $100‑billion investments and multi‑gigawatt contracts are neither pure carnival nonsense nor guaranteed long‑term prosperity. They reflect a genuine inflection: AI has moved from academic and research curiosity to commercial infrastructure with real, high‑value use cases. That migration demands racks, GPU fleets and industrial‑scale power — and that creates multi‑year opportunities for suppliers and hyperscalers.But the winning path will not be a simplistic “more GPUs = more revenue” equation. It requires disciplined financing, diversified supply strategies, verifiable off‑takers, and careful engineering of data‑center ecosystems. Journalists, CIOs and investors should read the fine print: deployment timetables, milestone triggers, vesting schedules for equity instruments, and the practicalities of power and cooling. Those details will determine whether the AI era becomes a productive industrial revolution — or a capital‑intensive boom that produces painful corrections for overextended players.
At the moment, the balance of evidence supports continued investment and deployment — provided commercial monetization of AI services keeps pace with the scale of infrastructure being ordered. The next 18–30 months of first‑gigawatt turn‑ups, utilization rates and enterprise billing will be the truest test of whether “this isn’t Pets.com” is the correct verdict — or simply the most persuasive marketing line yet from a supplier at the center of the story.
Source: Windows Central Why NVIDIA’s CEO thinks the AI boom won’t go bust — and how it’s different from 2000
