NVIDIA flags power bottleneck
What happened
- Nvidia executives said on May 26 AI demand had crossed a “critical threshold,” with growth now tied as much to power, land and utility delivery. - Jensen Huang told investors “tokens are now profitable,” while Nvidia’s own engineers wrote in March that “power is the ultimate constraint.” - NuScale and ENTRA1 are pitching behind-the-meter nuclear for data centers, while Bloom and Brookfield have a $5 billion AI infrastructure partnership.
Why it matters
Nvidia’s latest message is that the next limit on AI growth is no longer just chip supply. Jensen Huang, speaking on the company’s first-quarter fiscal 2027 earnings call, said AI had crossed a “critical threshold” because “tokens are now profitable,” tying new demand for compute directly to revenue generation rather than experimentation. Nvidia has also been describing the constraint in infrastructure terms. In a March technical post, Nvidia engineers wrote that “power is the ultimate constraint” for AI factories and said deployment now depends on access to land, power and a data-center shell, with performance per watt becoming the key operating metric. (benzinga.com) ### Why is Nvidia talking about power instead of chips? Jensen Huang’s May 26 remarks linked AI demand to economics at the application layer. He said agentic AI can now do “productive and valuable work,” and that model developers are racing to add capacity as AI output turns into revenue. Nvidia CFO Colette Kress said on the same call that the company’s GB300 platform delivers a 60% reduction in cost per token versus systems available six months earlier. (developer.nvidia.com) That matters because falling inference costs can make the limiting factor less about whether chips exist and more about whether operators can energize and house them. That inference is supported by Nvidia’s own March description of AI data centers as “token factories” operating within a fixed power envelope. (benzinga.com) ### What does “power bottleneck” mean in practice? Nvidia’s March post defined the issue as a hard operating limit. The company said every AI factory runs within a fixed power envelope and that revenue growth increasingly depends on maximizing tokens per watt, not simply adding more hardware. That shifts attention to the physical buildout around servers. (benzinga.com) Nvidia said AI data centers are tied to the energy ecosystem and that access to land, power and shell determines deployment. In practice, that means substation capacity, interconnection timing, cooling, and the readiness of a site can slow expansion even when demand for accelerators remains strong. The last sentence is an inference drawn from Nvidia’s description of deployment constraints. (developer.nvidia.com) ### Who is trying to solve the problem? NuScale and its commercialization partner ENTRA1 are pitching small modular reactors as a dedicated supply option for data centers. NuScale says long-term power purchase agreements are available through ENTRA1 for new and growing data centers, and its materials describe SMRs as suited to reliable, carbon-free power for AI infrastructure. NuScale has also been explicit about the behind-the-meter model. (developer.nvidia.com) In a May post, the company said behind-the-meter generation offers reliability, cost predictability and energy security, and said its reactor technology is approved to operate behind the meter. Bloom Energy is making a parallel case for on-site generation. Bloom and Brookfield announced a $5 billion strategic AI infrastructure partnership in October 2025, with Bloom as preferred onsite power provider for Brookfield’s AI factories, and Bloom separately signed an agreement with American Electric Power for up to 1 gigawatt of fuel-cell products to power AI data centers. (nuscalepower.com 1) (nuscalepower.com 2) ### Why are on-site and behind-the-meter deals getting attention now? Bloom said in January that data centers planned to reduce reliance on the grid, and CEO KR Sridhar said in April the company was “ushering in the era of digital power for the digital age.” Those statements line up with a market where operators want faster and firmer power delivery than utility buildouts alone may provide. The final clause is an inference based on Bloom’s positioning around on-site power. (investor.bloomenergy.com) NuScale’s materials make the same pitch from a nuclear angle. ENTRA1 says it can develop, own and operate plants for customers under long-term arrangements, aimed at hyperscale data centers seeking round-the-clock supply without building and operating the reactors themselves. ### What should readers watch next? Nvidia’s next public test will be whether its customers keep converting lower token costs into more infrastructure orders. (bloomenergy.com) The company said on May 20 that second-quarter fiscal 2027 revenue is expected to be about $91 billion, plus or minus 2%, giving investors a near-term marker for whether demand remains at the level Huang described. (interactive.nuscalepower.com) NuScale, ENTRA1, Bloom, Brookfield and utilities such as AEP are the named participants to watch on the power side. Their next milestones are likely to show up in project announcements, power purchase agreements, interconnection filings and data-center energy contracts rather than in chip launches alone. That framing is an inference from the companies’ current announcements and positioning. (nuscalepower.com) (investor.nvidia.com)
Key numbers
- Nvidia executives said on May 26 AI demand had crossed a “critical threshold,” with growth now tied as much to power, land and utility delivery.
- Jensen Huang, speaking on the company’s first-quarter fiscal 2027 earnings call, said AI had crossed a “critical threshold” because “tokens are now profitable,” tying new demand for compute directly to revenue generation rather than experimentation.
- Jensen Huang’s May 26 remarks linked AI demand to economics at the application layer.
- Nvidia CFO Colette Kress said on the same call that the company’s GB300 platform delivers a 60% reduction in cost per token versus systems available six months earlier.
What happens next
- Nvidia’s latest message is that the next limit on AI growth is no longer just chip supply.
- Jensen Huang’s May 26 remarks linked AI demand to economics at the application layer.
- (developer.nvidia.com) In a May post, the company said behind-the-meter generation offers reliability, cost predictability and energy security, and said its reactor technology is approved to operate behind the meter.
Quick answers
What happened in NVIDIA flags power bottleneck?
Nvidia executives said on May 26 AI demand had crossed a “critical threshold,” with growth now tied as much to power, land and utility delivery. Jensen Huang told investors “tokens are now profitable,” while Nvidia’s own engineers wrote in March that “power is the ultimate constraint.” NuScale and ENTRA1 are pitching behind-the-meter nuclear for data centers, while Bloom and Brookfield have a $5 billion AI infrastructure partnership.
Why does NVIDIA flags power bottleneck matter?
Nvidia’s latest message is that the next limit on AI growth is no longer just chip supply. Jensen Huang, speaking on the company’s first-quarter fiscal 2027 earnings call, said AI had crossed a “critical threshold” because “tokens are now profitable,” tying new demand for compute directly to revenue generation rather than experimentation. Nvidia has also been describing the constraint in infrastructure terms. In a March technical post, Nvidia engineers wrote that “power is the ultimate constraint” for AI factories and said deployment now depends on access to land, power and a data-center shell, with performance per watt becoming the key operating metric. (benzinga.com) Why is Nvidia talking about power instead of chips? Jensen Huang’s May 26 remarks linked AI demand to economics at the application layer. He said agentic AI can now do “productive and valuable work,” and that model developers are racing to add capacity as AI output turns into revenue. Nvidia CFO Colette Kress said on the same call that the company’s GB300 platform delivers a 60% reduction in cost per token versus systems available six months earlier. (developer.nvidia.com) That matters because falling inference costs can make the limiting factor less about whether chips exist and more about whether operators can energize and house them. That inference is supported by Nvidia’s own March description of AI data centers as “token factories” operating within a fixed power envelope. (benzinga.com) What does “power bottleneck” mean in practice? Nvidia’s March post defined the issue as a hard operating limit. The company said every AI factory runs within a fixed power envelope and that revenue growth increasingly depends on maximizing tokens per watt, not simply adding more hardware. That shifts attention to the physical buildout around servers. (benzinga.com) Nvidia said AI data centers are tied to the energy ecosystem and that access to land, power and shell determines deployment. In practice, that means substation capacity, interconnection timing, cooling, and the readiness of a site can slow expansion even when demand for accelerators remains strong. The last sentence is an inference drawn from Nvidia’s description of deployment constraints. (developer.nvidia.com) Who is trying to solve the problem? NuScale and its commercialization partner ENTRA1 are pitching small modular reactors as a dedicated supply option for data centers. NuScale says long-term power purchase agreements are available through ENTRA1 for new and growing data centers, and its materials describe SMRs as suited to reliable, carbon-free power for AI infrastructure. NuScale has also been explicit about the behind-the-meter model. (developer.nvidia.com) In a May post, the company said behind-the-meter generation offers reliability, cost predictability and energy security, and said its reactor technology is approved to operate behind the meter. Bloom Energy is making a parallel case for on-site generation. Bloom and Brookfield announced a $5 billion strategic AI infrastructure partnership in October 2025, with Bloom as preferred onsite power provider for Brookfield’s AI factories, and Bloom separately signed an agreement with American Electric Power for up to 1 gigawatt of fuel-cell products to power AI data centers. (nuscalepower.com 1) (nuscalepower.com 2) Why are on-site and behind-the-meter deals getting attention now? Bloom said in January that data centers planned to reduce reliance on the grid, and CEO KR Sridhar said in April the company was “ushering in the era of digital power for the digital age.” Those statements line up with a market where operators want faster and firmer power delivery than utility buildouts alone may provide. The final clause is an inference based on Bloom’s positioning around on-site power. (investor.bloomenergy.com) NuScale’s materials make the same pitch from a nuclear angle. ENTRA1 says it can develop, own and operate plants for customers under long-term arrangements, aimed at hyperscale data centers seeking round-the-clock supply without building and operating the reactors themselves. What should readers watch next? Nvidia’s next public test will be whether its customers keep converting lower token costs into more infrastructure orders. (bloomenergy.com) The company said on May 20 that second-quarter fiscal 2027 revenue is expected to be about $91 billion, plus or minus 2%, giving investors a near-term marker for whether demand remains at the level Huang described. (interactive.nuscalepower.com) NuScale, ENTRA1, Bloom, Brookfield and utilities such as AEP are the named participants to watch on the power side. Their next milestones are likely to show up in project announcements, power purchase agreements, interconnection filings and data-center energy contracts rather than in chip launches alone. That framing is an inference from the companies’ current announcements and positioning. (nuscalepower.com) (investor.nvidia.com)
