Varun Sivaram

🚨Two big @EmeraldAI_ announcements! 1⃣World's first power-flexible AI Factory at commercial scale (96MW!) together with @nvidia, @digitalrealty, @EPRINews, @pjminterconnect at NVIDIA's Aurora facility in Virginia! 2⃣$42.5M in total funding to scale!🚀 🔗Links in thread below

@dr_alphalyrae Congratulations Vega and NVIDIA team!!

Today we launch NVIDIA’s Nemotron Super 3, a 120B param open model designed to run agentic AI systems across scientific, enterprise and industrial applications. Partners working with us include Dassault Systèmes, Palantir Technologies, Lila Sciences and Edison Scientific Key highlights: - Nemotron 3 Super uses a hybrid mixture-of-experts (MoE) architecture with three core innovations, delivering up to 5x higher throughput and 2x higher accuracy than the previous model. - Hybrid architecture: Combines Mamba layers for ~4× higher memory and compute efficiency with transformer layers for advanced reasoning. - MoE efficiency: Only 12B of 120B parameters are active during inference. Activates 4 expert specialists for the cost of one to generate the next token, improving accuracy. - Multi-token prediction: Predicts multiple future tokens simultaneously, enabling ~3x faster inference. Read the tech blog👇

We are pleased to announce Varun Sivaram, Founder & Chief Executive Officer at Emerald AI will be joining us as a speaker at #CERAWeek 2026. Register and engage in valuable discussions, learning & networking this March 23-27: ceraweek.com/en/register

@s0larsurf Completely agree with you. Critical to figure this out. And thank you!!

@vsiv Agree that the export driven AI economy will build aggregate US macro resilience. Just not sure how that value will get transmitted to the avg person. How much interchange and subs savings balance against a displaced salary? As an aside, big fan of your books and work!

Although this is scarily plausible, I was struck by the number of AI consequences that are framed as bad but could actually position the US as a much more resilient and globally competitive economy after wringing out inefficiencies going forward: -AI agents help consumers sidestep 2-3% credit card interchange fees (bad for credit card companies, great for everyone else). -AI helps consumers avoid zombie subscription renewals and forces companies to earn loyalty because stickiness is hard to achieve. Great outcome for consumers. -Oligopolistic platforms get disrupted because AI makes it easier to create and operate them, so ridesharing and food delivery apps can't price gouge. Seems good for the workers and the consumers. -AI infrastructure investment continues to boom in the US and data centers are built at home, even as business process outsourcing centers in countries around the world (e.g., India) are outcompeted by AI. It's true that the treadmill of US consumer spending drives domestic employment and equity value. But an America that leads the world in AI services, with the AI token factories sited at home, will be an export powerhouse whose economy depends less on inefficient domestic consumer spending. Americans will reduce wasteful spending and feel richer for it. This only works if we outcompete the rest of the world.

The US should export better and cheaper lawyers, bankers, doctors, teachers, accountants, all powered by AI. Building and controlling US AI infrastructure (and powering it!) enable the US economy to weather the coming AI shock and emerge even stronger as an export powerhouse.

Totally agree that's the point of the essay. My point is that if the US economy is driven less by US consumer spending and more by export-driven AI service revenue where the rest of the world pays US companies for extraordinary services that only our leading frontier models and extensive domestic AI infrastructure can deliver, then our economy will be on more resilient footing rather than depending on wasteful domestic consumer spending. US wages rise, US prices fall, US quality of living rises, and the rest of the world pays for it. In some sense, this is a China export strategy applied to the AI age.

@vsiv Don't think anybody is denying those initial consumer benefits. I believe the issue becomes the next step when these platforms get displaced, it leads to layoffs of large parts of the white collar workforce with no real options of where to go next, driving down consumer spending

@JEBistline Nicely put--thanks!

Rate design and flexibility: name a more iconic duo. EPRI's DCFlex is testing flexibility approaches in practice to make data centers grid assets, including recent evidence with Emerald AI and SRP published in Nature Energy: nature.com/articles/s4156…

A lot of debate about large new loads skips the key question: How do rate design and flexibility shape who pays? Figure 5 from EPRI's new Win-Win Watts paper shows why outcomes diverge sharply.

@ShanuMathew93 @IEA @fbirol Will be there Mar 2-5! Will text you

@vsiv @IEA @fbirol Looks like an amazing convening of industry leaders Pop over to London after and say hello! 👀

Delighted to join the @IEA Ministerial meeting today in Paris, hosted by @fbirol

@QuincyEdmundLee Great piece!

@BesxarFoundry @AviatorAshley @SpaceX @BricklayerAI @AdamVincent @CapeCellular @JohnDoyleCape @constanellis @CruxClimate @Alfredijohns0n @DEFCONAI 7) Deterrence: Automated energetics manufacturing. Founder: @dkrajendra. 8) EM1: Emergency response systems for cities. Founder: @TMFelous. 9) @EmeraldAi_: Flexible power demand for AI infrastructure. Founder: @vsiv.

“One thing’s for certain. No matter where you are around the globe, energy will be central to the next stage of the AI revolution,” says CFR energy and climate expert @vsiv. Returning from the World Economic Forum in Davos, Sivaram unpacks two risks and one opportunity at the intersection of energy and AI.

.@EmeraldAI_ co-founder & CEO @vsiv thinks we're at the "very beginning of a mega-boom in the demand for compute," he said at #AxiosHouse in Davos. "In the long run, my bet is that we have built less than 1% of all the compute that we will need as humanity."

Thanks @NatBullard for including @EmeraldAi_ in your incredible annual deck! nathanielbullard.com/presentations

@tylerhnorris @californiapuc Thanks to you, Tyler!

Since the early days of the industry, we’ve planned America's power system around the assumption that all new electrical loads are 100% inflexible. So deeply rooted is this planning paradigm that the @CaliforniaPUC just concluded – based on a substantial evidentiary record – that "parties are not aware of any standardized, scalable [flexible load planning] processes in the United States that California can look to for learnings." [1] 2025 was the year this started to change. Change rarely happens quickly in this industry, perhaps least of all in the realm of system planning. But necessity is the mother of invention, and as @EPRINews' CEO Arshad Mansoor put it in August, "We cannot do what we did in the 1960s and just build... we have to use the underused capacity we already have, and we have to think differently." [2] Because a growing number of people have been willing to think differently, we’ve seen remarkable progress in a short window, from @SPPorg's new transmission service options, to a unanimous bipartisan @FERC order on colocation, to a landmark proposed order on Dec. 24 from @CaliforniaPUC's president to create a standard flexible service connections offer for electrical loads (notably, all responding parties supported this step, including PG&E and SCE) [1]. To all who helped enable this momentum in 2025: thank you. We face big challenges, and this body of work won't be a silver bullet or avoid trade-offs. But our ongoing capacity to develop meaningful insights on complex, consequential topics and translate them into bipartisan policy gives me genuine hope. Onwards! [1] CPUC president's proposed decision: #DER" target="_blank" rel="nofollow noopener">docs.cpuc.ca.gov/PublishedDocs/… [2] Here at 6:10: youtube.com/watch?v=Rul-tm…

California is moving toward a landmark standardization of Flexible Service Connections (FSCs) for electrical loads. A new proposed decision by @CaliforniaPUC's president, issued Dec. 24, directs PG&E and SCE to replace ad-hoc pilots with a formalized standard offer, enabling customers to bypass long grid upgrade delays by connecting via a temporary "bridge" that matches power usage to existing grid capacity using a Limited Load Profile (LLP). Notably, all responding parties supported this step, including PG&E and SCE. As the proposed decision explains, "The concept of Flexible Connection is allowing the customer to match their site’s power levels to the amount of power that the grid can safely handle. The customer does this by adhering to a profile of values generated by IOU engineers... it is likely that circumstances exist where electric capacity can safely be shared between more than one customer located upon shared infrastructure and that increased utilization of existing infrastructure that results from sharing capacity may be cost-efficient." Also notable is how the evidentiary record concluded that no standardized process of this kind exists anywhere in the US (!): "Ultimately, parties are not aware of any standardized, scalable processes in the United States that California can look to for learnings." The Commission is expected to vote on this proposal at its 2/5/26 meeting. If approved, PG&E and SCE will have 30 days to launch the program. Proposed decision: #DER" target="_blank" rel="nofollow noopener">docs.cpuc.ca.gov/PublishedDocs/… Docket: apps.cpuc.ca.gov/apex/f?p=401:57::::::

"My hot take on space-based data centers: We will have tens of GW of flexible AI Factories unlocking terrestrial, grid-connected power capacity before we have the first gigawatt of space-based compute." -@vsiv

@tylerhnorris 👏👏

FT's big new piece on "The AI power crunch" appropriately includes a section on flexibility, and I appreciate them referencing our research.