Sivaram proposes AI power throttling to balance energy grids
- AI data centers can balance power grids by pausing non-urgent workloads during peak electricity demand.
- Software load-throttling unlocks 100 gigawatts of US grid capacity without raising consumer electric rates.
- Amazon will spend $220 billion on AI hardware in 2026 while bypassing grid bottlenecks with nuclear deals.
The primary bottleneck for artificial intelligence is no longer silicon. On October 8, 2026, Emerald AI CEO Varun Sivaram argued on This Week in AI that waiting a decade for grid hookups is choking deployment.
Building new transmission lines for rare summer demand spikes inflates utility bills and stalls expansion. Software dynamic throttling allows compute facilities to act as grid shock absorbers instead. By pausing background model training or shifting inference jobs across regions during peak hours, operators tap existing infrastructure without degrading end-user response times.
Tests in London and Phoenix proved that brief automated load reduction lets data centers ride through grid spikes safely. Duke University and McKinsey estimates show this approach could unlock 100 gigawatts of available US grid capacity with curtailments required for less than two percent of the year.
The strategy addresses an urgent national disparity. The United States aims to add 70 gigawatts of AI data center capacity by 2028 but is projected to deploy only 30 gigawatts. Meanwhile, China added 500 gigawatts of total power capacity last year and is projected to hold 400 gigawatts of peak spare capacity by 2030.
To bypass public utility queues entirely, hyperscalers are taking extreme measures. AWS CEO Matt Garman revealed on The a16z Show that Amazon will spend $220 billion in capital expenditure in 2026 alone. To supply its two-million Nvidia GPU purchase, AWS is directly financing 20-year nuclear, solar, and behind-the-meter energy projects.
Sivaram warned that going off-grid severs data centers from public utilities, depriving local grids of revenue and leaving ratepayers stuck with higher bills. Raising overall US grid utilization by ten percent from its current 50 percent average would yield massive compute capacity without building expensive new transmission lines.
Emerald AI is now partnering with Nvidia and Digital Realty on a 100-megawatt flexible data center in Virginia as a reference design. With Texas Governor Greg Abbott pausing grid connections through the midterm election season, flexible software orchestration provides the only immediate path to keep American AI online.
Source Intelligence
- Deep dive into what was said in the episodes
Data Centers Aren't Killing the Grid. They Could SAVE It. • Oct 8
- Varun Sivaram argues that data centers must become flexible grid assets to accommodate rising energy demands. By dynamically throttling workloads or shifting compute to other regions during peak demand, data centers can integrate into the existing grid without raising community rates.
- Emerald AI demonstrated dynamic power throttling in London by coordinating with the national grid to reduce data center power during a sports halftime. The halftime pause mitigated a sudden one-gigawatt power spike caused by citizens turning on electric tea kettles.
- Studies from Duke University and McKinsey show the US grid holds 100 gigawatts of available power that could be unlocked with minimal curtailment. Data centers would only need to reduce power usage for less than two percent of the year.
- The United States aims to add 70 gigawatts of AI data center capacity by 2028 but is projected to deploy only 30 gigawatts. Meanwhile, China added 500 gigawatts of power capacity last year compared to 50 gigawatts in the United States.
- Varun Sivaram projects China will possess 400 gigawatts of peak spare capacity by 2030 to fuel its artificial intelligence infrastructure. This growth is driven by massive state investments in nuclear, solar, wind, coal, and battery power.
- The United States power grid operates at an average utilization rate of 50 percent or less. Increasing overall grid utilization by a stretch target of 10 percent would unlock significant capacity without building new transmission lines.
- Emerald AI is partnering with NVIDIA and Digital Realty to launch a 100-megawatt flexible data center in Virginia. The project will serve as a reference design for future power-flexible AI factories.
- Texas Governor Greg Abbott has paused the state's batch zero process for connecting new data centers to the grid. Varun Sivaram notes this pause is likely to persist through the midterm election season.
Also discussed on this episode: (6)
Chips (1)
- NVIDIA hardware efficiency has increased exponentially over the last decade. The Rubin GPU generation produces one million times more tokens per watt of input energy than the Kepler generation released in 2012.
Energy (1)
- The US electric grid lacks a standardized digital communication system across its 3,000 distinct utilities. Many operators still call data centers on the phone to request power reductions during peak load emergencies.
Agents (1)
- Vibhav Viswanathan co-founded Pascal AI to build context-aware AI agents for hedge funds and asset managers. The startup's current customer base represents an accumulated half-billion dollars in assets under management.
Models (1)
- Pascal AI integrates proprietary internal data, licensed market feeds, and web data to analyze investment theses. Viswanathan explains that accessing a fund's internal memos and portfolio history provides a proprietary edge that frontier AI models cannot replicate.
Enterprise (1)
- Pascal AI sells its software to secretive financial institutions, including a top-ten global hedge fund and a 40-billion-dollar Tiger Cup fund. These institutions use the platform to maintain a real-time, context-aware view of their portfolios.
Labor (1)
- Vibhav Viswanathan claims AI adoption in hedge funds is shifting analyst roles toward manual, boots-on-the-ground research. Because software automates document analysis, analysts are traveling more and conducting physical channel checks to discover investment alpha.
Building the Cloud for an Agentic World | AWS CEO Matt Garman • Oct 8
- AWS now finances solar, nuclear, and grid transmission projects to secure enough electricity for its massive data centers. Matt Garman notes that power availability requires the company to plan infrastructure footprints up to 20 years in advance.
Also discussed on this episode: (14)
Big Tech (2)
- Matt Garman states that AWS generates roughly $169 billion to $170 billion in annual revenue. He notes that the cloud business is still in its early stages because a massive volume of global workloads remains on-premise.
- Amazon plans to spend $220 billion in capital expenditure, driven heavily by the infrastructure demands of generative AI. Matt Garman states that AWS will not slow down this historic buildout because consumer demand remains massive.
Startups (1)
- Matt Garman estimates that companies starting out on AWS account for 30% to 40% of the platform's current total revenue. This high-value trajectory justifies the company's sustained business focus on two-person, early-stage startups.
Agents (4)
- Matt Garman observes that software agents care deeply about tail latencies like P99 S3 latency because high latencies block agentic execution. Optimizing for these specific performance metrics makes agentic workflows perform significantly better on AWS.
- AWS is testing a metadata tool called AWS Context to build a semantic layer across different storage services like Aurora and S3. Matt Garman says this allows autonomous agents to easily query data across disparate corporate databases.
- AWS is rolling out a frictionless sign-up process allowing users or agents to open an account in under 30 seconds using a Gmail address. Matt Garman explains this bypasses traditional friction like configuring VPCs and inputting credit cards.
- AWS deployed Amazon Quick across its internal teams, enabling non-technical staff in finance and human resources to build custom automated agents. Matt Garman reports this deployment has significantly accelerated internal software development and product release cycles.
AI Infrastructure (1)
- Matt Garman notes that many AI sandbox startups rely on Firecracker, a micro-virtual machine technology AWS developed a decade ago. Firecracker provides the rapid spin-up times and secure isolation boundaries required for running untrusted agentic code.
Chips (3)
- AWS plans to purchase two million Nvidia GPUs over the next couple of years. Matt Garman emphasizes that AWS intentionally reserves GPU capacity for startups to foster the ecosystem, rather than selling all inventory to frontier labs.
- Matt Garman traces AWS silicon back to its acquisition of Annapurna, which led to the Nitro card and Graviton arm processors. He states that Graviton represents the highest volume of server chips AWS deploys annually.
- Matt Garman asserts that Trainium has become highly efficient for inference due to its competitive architecture and pricing. AWS is currently sold out of Trainium capacity through late next year, with Bedrock driving the majority of its inference.
Society (1)
- Matt Garman highlights that local tax contributions from AWS data centers can save individual county residents up to $5,000 annually in property taxes. He argues the industry must do a better job highlighting these localized economic benefits.
Enterprise (2)
- Amazon Bedrock guarantees that enterprise data remains entirely within the customer's virtual private cloud. Matt Garman explains this architecture prevents proprietary data from returning to third-party model providers, satisfying crucial security requirements for large enterprises.
- AWS launched a service named Continuum that utilizes frontier artificial intelligence models to scan customer environments for software vulnerabilities. Matt Garman says the service contextualizes and prioritizes risks so security teams can remediate them at machine speed.