The AI market is moving beyond a chip-centric growth cycle into a full-stack infrastructure buildout. Electricity, cooling, networking, memory, uptime, permitting, and capital are becoming as consequential as accelerator demand itself. The constraint is no longer simply whether customers want more compute. It is whether the relay chain—from generation and grid interconnection to an operating AI factory—can transmit dependable power and convert it into useful output.
The scale of announced demand is substantial. Industry trackers place announced or tracked data-center pipelines at approximately 250 GW to more than 330 GW 16, while Texas alone has accumulated requests totaling 474 GW 23,81,85. Individual proposals include a 10 GW Ohio hub 31,44,49,77,100, a 2.67 GW West Texas campus 88, a 1.8 GW Paducah campus by 2032 35, Amazon’s proposed 7.65 GW generation facility 71,83, and a 2 GW Chinese campus designed for one million accelerators 87. These figures are not equivalent measures of operating AI load. They do, however, establish the central fact: NVIDIA’s addressable market is increasingly governed by the availability and quality of physical infrastructure, not accelerator demand in isolation.
The investment implication is two-sided. Sustained AI demand could support a multiyear expansion in NVIDIA accelerators, networking, systems, and software, with total compute market potential cited at $2 trillion 46. Yet announced gigawatts, nameplate generation, utility requests, and critical information-technology load are frequently conflated 58. A 500 MW utility request does not imply 500 MW of continuous accelerator demand 58. The relevant analytical task is therefore to distinguish committed, dependable, revenue-generating compute capacity from aspirational project announcements.
Key Insights
Power scale is real; headline capacity is not a revenue proxy
The evidence broadly confirms that AI infrastructure is becoming an electricity-scale industry. Global electricity demand associated with data centers, AI, and cryptocurrencies is estimated at approximately 800 TWh in 2026 39, while a broader Siemens scenario assumes 116 GW of data-center additions by 2030 78. The AI transition is described as a ten-year shift, materially faster than prior computing revolutions 50, and one estimate projects compute demand doubling every six months 76. Frontier-training compute is also estimated to grow at approximately 1.31 per year 48. These indicators support a constructive long-term backdrop for NVIDIA, but they remain forecasts rather than evidence of contracted shipments.
At the project level, the range of proposals is equally large: a 1.2 GW power project with potential expansion to 2.4 GW 53; a 1.2 GW Meta campus that could become future demand for Oklo 64; a 1 GW Hut 8 Beacon Point campus 97; a proposed 1.6 GW Anthropic Texas campus 2; a 1.4 GW project 73; a 150 MW British Columbia sovereign-AI cluster by 2032 86; 300 MW projects in Castile-La Mancha 36; a 300 MW Bell Regina project 86; and India’s projected increase from 1.1 GW to 4.2 GW of data-center capacity between fiscal 2025 and 2030 15. Korea’s planned AI factories require more than 2 GW and are compared with the consumption of 1.5 million homes 44,49. Elsewhere, 2 GW is similarly compared with 1.5 million homes 49, while a 1 GW data center is likened to the demand of a city the size of Boston 88 or several hundred thousand to more than two million homes, depending on usage 42.
These comparisons are useful as measures of physical scale, but they also demonstrate why power capacity cannot be converted mechanically into NVIDIA revenue. The relay has several intermediate stations: interconnection, generation, transformers, switchgear, cooling, networking, commissioning, utilization, and customer economics. A failure at any one of them leaves nominal capacity unable to produce tokens.
The OpenAI-related projects illustrate the distinction particularly well. The reported infrastructure program is estimated at more than $500 billion in total 5, with an initial 800 MW phase not expected until 2028 5. Southern Ohio is variously described as a 10 GW project 5,41,89,95,99, although one account specifies an initial 800 MW phase 31. The reported $1.4 trillion spending figure represents aggregate multiyear maximum commitments rather than immediate liabilities 6, and infrastructure agreements span four to ten years 6. For NVIDIA, this means that a large customer-commitment headline may indicate a substantial eventual accelerator opportunity while saying considerably less about purchase cadence, financing, utilization, or deployment timing.
Texas provides a similar warning. More than 70 GW of dedicated-power capacity is planned or associated with the buildout 10, yet a 2.67 GW West Texas project was delayed by one year 88. Power delivery moved from 2027 to 2028 after additional engineering work 88, and the project requires 99.999% reliability to satisfy Microsoft 21,88. This is an important industry signal. Hyperscalers may prefer to delay revenue-generating capacity rather than compromise power quality. NVIDIA benefits when high-performance workloads require premium infrastructure, but delays can push accelerator deployments into later periods and increase volatility in system orders.
Inference economics are moving beyond PUE
For large-scale inference, the more informative metric is increasingly output per unit of constrained power. “Tokens per Watt” is emerging as a headline measure 20,94, linking AI output to electricity consumption and environmental footprint 20. Performance per watt remains a critical accelerator-market metric 60, particularly because data-center power envelopes are often fixed before chip orders are placed 11. AI-factory output is measured in inference tokens 94, and operating success is increasingly framed through Time to Token and Tokens per Watt 94. A 15% improvement in Tokens per Watt would produce 15% more output from the same power feed without additional energy cost 94.
This framework favors a platform capable of optimizing silicon, memory, networking, software, and power utilization together. It also raises the competitive bar. Power Usage Effectiveness (PUE), which measures facility overhead, remains useful but is insufficient as a standalone measure of AI-facility efficiency 94. A facility with a 1.10 PUE that produces no tokens may be economically inferior to a 1.30 PUE facility producing millions of tokens 94. The relevant question is not simply whether a facility consumes less energy in aggregate, but how much useful output it produces per governed outcome 38. Claims of roughly 30% lower total cost of ownership 94 and of lower cost per AI operation as efficiency improves 65 are directionally supportive, but should not be treated as independently verified.
The implication is that NVIDIA’s moat increasingly depends on system-level efficiency and utilization. Hardware selection for AI-agent inference is primarily a memory-capacity and memory-bandwidth problem rather than a raw-compute or teraFLOPS problem 19,32. AI-factory network efficiency becomes critical as accelerator counts rise 56, with front-end networks typically operating at 100G–400G Ethernet 94. Fiber pathways may require 288-fiber trunks even when current demand is 144 fibers, and hierarchical identification and TIA-606-D labeling become operational requirements when thousands of fibers run through each row 94. The opportunity therefore extends into networking, interconnect, memory, rack-scale systems, and deployment services—precisely the stations at which NVIDIA’s integrated platform strategy can add value.
Cooling is a similarly hard constraint. Modern AI-factory racks range from approximately 40 kW to more than 140 kW 94, with roughly 100 kW per rack cited as a current benchmark 40. A 120 kW rack produces approximately 120 kW of heat 15, and air cooling becomes ineffective above approximately 30 kW of rack density 15. Ultimately, approximately one watt of heat must be removed for every watt entering the rack 51. A future rack could consume as much electricity as hundreds of homes 51. This supports demand for liquid cooling, power delivery, thermal management, and data-center design around NVIDIA deployments, while simultaneously increasing project complexity and the risk that nominal accelerator capacity cannot be operated at full utilization.
Uptime, utilization, and delivery timing determine realized economics
Nameplate capacity is only productive when the facility is operating. Some facilities are reported to run near 80% uptime 88, while investors and lenders question whether projects modeled for continuous operation can achieve high uptime 88. An 80% rather than 100% operating assumption materially reduces asset utilization and project returns 88. Daily AI-usage targets may also be set below full utilization deliberately. Thinkific, for example, uses 90% to account for training, vacations, illness, and other absences 33. These observations challenge the assumption that every contracted megawatt becomes continuously active accelerator capacity.
Time to Token—the period from project concept to an operational facility producing tokens 94—is consequently as important as nameplate capacity. Supercomputers require substantial capital, energy, cooling, networking, and construction resources 28, while project economics vary with utilization and token pricing 47. The assumed five-year life of computing equipment 15 increases the importance of deployment speed and refresh cycles. NVIDIA’s financial outlook should therefore be assessed against delivered, commissioned, and utilized systems rather than announced power reservations.
Former crypto miners create optionality—and competition
Former crypto miners are attempting to convert their power positions and sites into AI infrastructure. Hut 8’s stated strategy is to use long-dated AI power contracts and its Bitcoin treasury to move from traditional mining toward an AI-infrastructure platform 96. Hut 8 has characterized project-level margins on contracted AI capacity at approximately 99%–100% 54, while TeraWulf signed a 20-year, 401 MW lease with Anthropic 75. TeraWulf, Cipher Mining, IREN, and possibly MARA are all pursuing transitions from Bitcoin mining toward AI data-center or compute businesses 43.
A proposed TeraWulf–Cipher combination would bring together contracted sites across New York, Kentucky, and Texas 74, TeraWulf’s data-center assets, Cipher’s ERCOT presence, and relationships with major AI and cloud counterparties 74. The combined entity could become a larger, more systemically exposed AI-infrastructure platform 74, with TeraWulf contributing low-cost, zero-carbon repurposed assets and contracted relationships 74. These developments may create customers and partners for NVIDIA, but they also introduce infrastructure competitors that could capture more of the value between power procurement and delivered compute.
The valuation discipline is straightforward: Bitcoin-mining power capacity is not automatically AI-ready capacity 54, and mining capacity is distinct from AI-ready data-center capacity 54. Uncontracted optionality should be discounted until interconnection, transformers, switchgear, cooling, fiber, customer contracts, and commissioning are demonstrated. The infrastructure supporting the WULF/CIFR developments includes power capacity, transformers, switchgear, and contracted sites 74, but the claims do not establish that all announced capacity is operational or suitable for NVIDIA’s highest-density deployments.
Other pipeline claims require the same mechanical test. Applied Digital is described as having a 5 GW pipeline 43, although the figure is explicitly characterized as unverified 43. Nebius has a year-end capacity target above 4 GW 68, while UPRISE has a stated 5 GW objective 57. Aethir’s claimed compute-capacity base exceeds $400 million but lacks independent substantiation 1. These are relevant signals about the direction of the market, not dependable inputs to NVIDIA revenue forecasts.
Terafab is a strategic possibility, not a base-case supply-chain assumption
The proposed Tesla/SpaceX Terafab represents a vertically integrated response to supply insecurity 66. Its concept combines chip design, wafer fabrication, extreme ultraviolet (EUV) lithography, memory, advanced packaging, and testing under one roof 67. The principal proposed sponsors or users are SpaceX and Tesla 69, with potential demand from satellite-network data centers 69 and long-term platforms including Optimus, Cybercabs, and space data centers 66. The facility is proposed for Grimes County, Texas 10,66, with an initial construction investment reportedly of $16.8 billion 69, a scale potentially up to 50 times the Pentagon 66, and at least 3,000 jobs 66 against a stated requirement of 1,800 jobs 10. Public support includes a $30 million Texas incentive 10 and a 100% county tax abatement conditional on at least $5 billion of investment by 2030 and 1,800 full-time jobs by 2035 10.
The proposed manufacturing model would treat lithography light as a shared utility 67,70, replacing individual EUV sources with centralized accelerator-driven free-electron lasers (FELs) 67. The concept targets approximately 10 kW of EUV power 69, potentially serving 10–20 scanners 69 and spreading the cost of one expensive FEL across ten or more scanners 70. It would begin with standard lithography and later use a drop-in FEL conversion 69, with theoretical benefits including higher scanner throughput and increased wafer output 63,69. Electron-beam energy recovery is intended to improve efficiency 67, while redundant accelerators could allow scanners to operate during maintenance 67. The design also aims to eliminate tin contamination and destructive plasma debris 67.
For NVIDIA, the proposal would become strategically relevant if it produced advanced accelerators at volume and reduced dependence on conventional foundry capacity. A bullish Intel scenario explicitly assumes Terafab will deliver volume 91. The evidence, however, remains overwhelmingly preliminary. The FEL interpretation is an observer inference rather than a fully documented technical disclosure 61. The 10 kW target remains theoretical and unproven for semiconductor mass production 69,70. The project’s one-terawatt annual AI-computing target is undefined 69. Wafer throughput and yield remain uncertain 62, as do unresolved optical challenges 67, photon-flux gains, stochastic-defect suppression, and reliable operation at a 6.x-nanometer wavelength 67. The architecture depends on xLight-related technology 69 and advanced accelerators, magnets, electron-beam controls, and high-power systems 62.
The shared-light design also introduces concentration and operational-dependency risk because multiple scanners would rely on centralized light generation 67. If technically realized, it could transform lithography from isolated tools into shared industrial infrastructure 67. But the facility’s electrical load would include more than the FEL: accelerators, beam systems, cooling, vacuum equipment, fabrication tools, clean rooms, water systems, and AI-compute infrastructure would also contribute 69. A 10 kW EUV source is not the facility’s total electrical load 69. Economics depend on sufficient demand, utilization, and beam-distribution reliability 70, while the broader project faces capital, construction, permitting, energy-price, environmental, and uncertain-return risks 10,69. Terafab should therefore be treated as a speculative supply-chain scenario, not a base-case threat to NVIDIA’s foundry access or a dependable source of incremental chip supply. Its proposed scale is explicitly aspirational 63, and the initiative remains an unproven, capital-intensive infrastructure thesis 62,67 rather than an operating-company valuation case 67.
Energy, water, emissions, and permitting will determine deployment geography
Dedicated generation is becoming a recurring feature of the largest proposals. SpaceX’s Terafab plans dedicated electricity generation rather than relying solely on the public grid 10, and the facility is reportedly expected to use natural gas rather than Tesla solar power 7. That configuration may increase direct emissions and local pollution relative to renewable power 10, creating potential pollution-related liabilities 10. Battery storage could mitigate some effects, but its capacity, duration, chemistry, lifecycle emissions, and operating role are unspecified 10. A 1 MW thermal-powered data center is estimated to emit approximately 8,760 metric tons of CO2 annually, equivalent to roughly 1,900 cars 79.
Other projects propose solar additions, including a 150 MW facility for the Jay data-center project 37 and a 150 MW solar facility proposed by JGT2 37. Tesla reportedly targets 100 GW of annual U.S. solar manufacturing within three years 55. That target would be extremely large relative to current U.S. capacity and would require major capital, equipment, supply-chain, permitting, workforce, technology, and offtake execution 55.
Integrated energy systems may improve project economics at the margin. One proposed model combines AI data centers with small modular reactors (SMRs), renewable baseload systems, and controlled-environment agriculture 25, treating AI waste heat as a productive input 25. A related infrastructure and policy-finance model integrates hyperscale data centers, energy systems, waste-heat recovery, and vertical farming 25, with AI HVAC waste heat feeding controlled-environment food production 24,25. These concepts may extend the infrastructure ecosystem, but their efficiency case concerns relative output per outcome rather than guaranteed reductions in aggregate energy use 38.
Water and permitting are equally material. A roughly 1 GW IT load distributed across 31.6 million square feet could use approximately three million gallons of water per day 34. The Terafab proposal could face water-use, emissions, environmental-permitting, construction-approval, grid-interconnection, radiation, workplace-safety, and high-power particle-accelerator requirements in Texas 69. Natural-gas generation and the project’s scale would intensify local scrutiny. More broadly, the Federal Energy Regulatory Commission (FERC) has issued an order directed at regional grid operators concerning requirements relevant to AI data-center development 81. For NVIDIA, regulatory friction can delay customer deployments. It can also reinforce the value of efficient accelerators, integrated power management, and high-output-per-megawatt systems.
Efficiency may expand total demand rather than reduce it
The cluster contains several indications that efficiency gains could enlarge the market rather than contract it. VORTIQ-X is described as reducing wasted inference rather than reducing demand for AI infrastructure 38, and its benchmark is an efficiency and outcome-quality claim, not evidence that customers need fewer AI systems 38. AI-enabled supply-chain planning may shorten planning cycles by up to 80% 27, while task-level productivity gains range from approximately 6% to 88%, with the highest gains in coding 93. AI can reduce staffing requirements for a given team 80, but the societal benefit of HFT-funded infrastructure depends on whether high-frequency-trading systems can be applied and scaled to AI workloads 72. A potential AI-demand collapse is characterized as a one-in-ten event 98, indicating meaningful downside risk despite a favorable base case.
Additional demand signals include machine-generated internet traffic potentially increasing 1,000-fold over five years 22; state integration of privately developed commercial AI into intelligence and defense systems 84; Japan’s expansionary fiscal policy and Takaichi government investment in AI 4; and expectations that quantum computing, personalized AI agents, and nuclear power will be present by 2035 92. OpenAI is reportedly associated with a valuation of approximately $1 trillion despite requiring continued exceptional growth 6, while SK Group’s chairman stated that OpenAI would require massive computing capacity 3. These claims suggest that infrastructure demand is broadening across enterprise, government, communications, robotics, and defense. Several are single-source or opinion-based, however, and should not be treated as firm forecasts.
The decisive distinction for NVIDIA is between efficiency that lowers the cost of each inference and demand that grows faster than efficiency improves. A more efficient accelerator can reduce operating cost, increase tokens within a fixed power envelope, and improve the economics of constrained sites. At the same time, agentic workloads, generated traffic, robotics, and sovereign AI can increase aggregate compute requirements. HBF is explicitly aimed at addressing AI-inference power limits 52 and is being considered for models scaling from trillions to tens of trillions of parameters 82. Extropic has proposed up to $75 million for thermodynamic sampling units and energy-efficient computing 45. These technologies may expand the competitive set, but they also confirm that power efficiency is becoming a central purchasing criterion.
Implications for NVIDIA
The central monetization opportunity is scarce, dependable compute capacity. NVIDIA is no longer selling accelerators into a fixed data-center footprint. As AI factories scale from tens to hundreds of megawatts, customers require a coordinated architecture spanning accelerators, high-bandwidth memory (HBM) and memory bandwidth, networking, rack design, cooling, software utilization, and power management. NVIDIA is well positioned to capture systems-level spend if its value proposition is measured in Tokens per Watt, Tokens per Megawatt, Time to Token, and revenue per megawatt rather than peak FLOPS alone. Revenue per megawatt measures how effectively an AI infrastructure business converts scarce electricity into revenue 13, while wasted power represents lost productive capacity and unavailable token production 94.
The buildout also supports a broader ecosystem opportunity. TeraWulf’s Anthropic lease 75, Bitdeer–Volta’s 121 MW of critical IT capacity 90, the Anthropic–Volta Infra agreement for a 133 MW Norwegian data center 17, and a Stockholm AI data center expected to receive first power in early 2027 29 show that demand is increasingly organized through long-dated capacity contracts. OpenAI infrastructure commitments, the Ohio and Texas campuses, and Meta’s potential 1.2 GW campus provide further evidence of hyperscaler and model-developer appetite. NVIDIA should nevertheless be evaluated on actual GPU shipments, system deployments, networking attach rates, and software monetization—not on every announced gigawatt.
Competitive risk is rising along three paths. Specialized inference hardware and memory-centric architectures could pressure accelerator pricing if they deliver materially better Tokens per Watt. Vertically integrated customers may seek greater control over chip supply, as reflected in the Terafab concept and Tesla’s broader solar and energy ambitions. Infrastructure developers and former miners may also capture more of the economics by offering power, sites, and contracted capacity as an integrated service. Yet the claims that Bitcoin capacity is not automatically AI-ready and that power requests do not equal critical IT load point to the same conclusion: execution, not announcement volume, will determine competitive outcomes. NVIDIA’s strongest defense is to make its platform indispensable to delivered performance, uptime, and utilization.
The principal financial risk is a mismatch between accelerating capacity announcements and slower physical realization. West Texas delays, 80% site uptime, uncertain project utilization, five-year hardware lives, and the difference between maximum commitments and immediate liabilities all support a staged forecasting framework. In the base case, continued AI demand and fixed power envelopes support durable accelerator and networking growth. In a downside case, grid bottlenecks, permitting delays, lower uptime, weaker token pricing, or an AI-demand shock could defer orders and expose customers with overbuilt capacity. A separate technology-risk case would involve successful alternative architectures, including FEL-enabled manufacturing, HBF, thermodynamic computing, or other inference-specific solutions. The Terafab evidence, however, is presently too speculative to justify a near-term change in NVIDIA earnings assumptions.
Valuation also requires discipline. T1’s reported $245 million–$255 million of sales on 835 MW imply approximately $0.293–$0.305 per watt 59, while a $0.01-per-watt spread is estimated to be worth $31 million–$42 million 59. These figures illustrate the sensitivity of infrastructure economics to small pricing and margin differences, but they are not directly transferable to NVIDIA’s higher-value accelerator and platform model. Similarly, contracted AI-capacity margins characterized at 99%–100% 54 may exclude financing, depreciation, construction, maintenance, and customer-specific costs. NVIDIA’s valuation should emphasize sustainable gross profit, system attach, recurring software revenue, and cash conversion rather than extrapolating project-level infrastructure margins.
Several additional claims are useful as secondary topic signals but carry limited direct valuation weight. LG’s entry-level AI-home system-on-chip is rated at 1.5 TOPS 8; Seeed Studio’s reComputer Mini J5011 reaches 200 TOPS 30; one petaFLOP equals 10^15 operations per second 12; and a 4U AI server may require significant cooling 14. A single chip integrating powerful terahertz systems could improve portability, scalability, and deployment flexibility 9,26. Nokia–NVIDIA AI-RAN is expected, according to one claim, to double network capacity by 2028 without adding towers 18. These developments reinforce NVIDIA’s exposure to edge, communications, and specialized computing, but they are not yet evidence of material incremental earnings.
Conclusion
The most robust conclusion is that AI demand is becoming constrained by power and infrastructure. The announced gigawatts are large enough to establish a new industrial category, but dependable IT load, uptime, commissioning, utilization, and customer economics—not headline capacity—should anchor NVIDIA forecasts.
NVIDIA’s opportunity is to turn scarce electricity into predictable compute output. That requires more than peak accelerator performance. It requires memory bandwidth, efficient networking, liquid-cooling-aware rack design, software utilization, and mechanical control of the complete signal path. Tokens per Watt, Tokens per Megawatt, Time to Token, and revenue per megawatt are therefore becoming more informative than PUE or peak FLOPS alone.
Former crypto miners and vertically integrated projects such as Terafab expand both the supply of potential infrastructure and the competitive field. Their uncontracted capacity and theoretical manufacturing capabilities should receive substantial discounts until interconnection, technical performance, customer contracts, and commissioning are demonstrated. The relay test remains decisive: a proposed tower is not a functioning relay, and a reserved megawatt is not yet a productive accelerator.
The long-term view remains constructive, but the monitoring framework should be precise. Grid interconnection, project delays, customer concentration, uptime, alternative inference architectures, and the conversion of announced capacity into actual NVIDIA deployments will determine whether the current pipeline becomes durable revenue or merely another procession of unbuilt towers.