E2E Networks Ltd Q1 FY2027 Concall Summary

Key Takeaways

• Revenue explodes on B200 go-live: Revenue jumped 334% YoY and 64% QoQ, driven almost entirely by capacity addition and utilization as the first Blackwell cluster went live mid-quarter, not by pricing.
• Margins hit a new plateau: EBITDA margin expanded 1,450 bps to 75.2%, a level management calls sustainable over the medium-to-long term given the mix of contracted, on-demand, and spot revenue.
• The AI super cycle is just starting, per management: Leadership frames the moment as "day zero of AI," with a decadal runway for infrastructure build-out, and is already laying preliminary groundwork for B300 and Vera Rubin architectures.
• Sovereign AI is the strategic differentiator: The company is building a full-stack sovereign AI platform on open-weight models, with a new infrastructure subsidiary (SovCloud) and a U.S. entity to chase international demand, though both are in very early days.
• What to watch next quarter: Whether the next lot of 1,024 B200 GPUs arrives and deploys on schedule, any disclosure on SovCloud funding, and whether the 75% margin level holds as electricity costs rise with utilization.

The Blackwell Inflection

• B200 cluster went live mid-May and is running at "quite maximal utilization," making it the primary driver of the quarter's step-change in revenue and margins.
• Revenue expansion is a volume story, not a price story: Management was explicit that the numbers reflect increased capacity and operating leverage, with only a very moderate pricing impact.
• GPU pricing is renegotiated post-contract based on current market conditions, and customers are generally happy to renegotiate, suggesting pricing power exists but is exercised judiciously.
• CPU pricing rose due to input cost increases (memory costs hit CPUs harder than GPUs in percentage terms), with customers offered the option to lock current prices via longer contracts with advance payment.
• Another 1,024 B200 GPUs are expected "soon over the next couple of months," and the company plans to deploy more Blackwell including non-flagship variants, plus expand into Vera Rubin over time.
• GPU life cycle is a minimum 6 years, with older generations continuing to run strong as each generation finds its sweet spot of price and performance aided by open-source model efficiency improvements.

Dua: "We've always maintained that at least at the minimum a 6-year life cycle for all the GPU generations. And we have seen even our older generations continue to run very strong with our customers."

Sovereign AI & Platform Strategy

• The sovereign AI platform is built on open-weight models running on E2E's AI factory, with software layers including TIR and Jarvis Labs that give customers control over model deployment, fine-tuning, data access, and state retention or forgetting.
• The pitch is risk elimination: Open-source models on a sovereign AI factory remove the risk of a foreign entity "judging you for how you're going to use your AI," a framing aimed at governments and regulated enterprises.
• SovCloud was established as an infrastructure subsidiary focused on large-scale GPU deployment and enabling funding arrangements to accelerate AI infrastructure, though management says it is "very early days" and details will be announced later.
• A Delaware (U.S.) entity was established to sell into international markets, establish presence outside India, handle market excellence and alliance management, and service non-India customers.
• The L&T partnership is arm's length: E2E buys L&T data center capacity, L&T buys E2E cloud infrastructure, and they jointly go to customers on platform; if L&T builds compute capacity, E2E may offtake some.

Dua: "Open-source based model on Sovereign AI factory like us eliminates that risk of someone sitting a couple of thousand miles away and judging you for how you're going to use your AI."

Workload Dynamics

• Training still dominates volumes, but workloads are fungible: customers run training during working hours and inference at night on the same clusters, making a precise split impossible to pin down on a small GPU fleet.
• Agentic workloads also run on the same infrastructure, further blurring the line between training and inference.
• At a global level, inference will have preponderance over training, but on a subset of roughly 5,000 GPUs, management says it is "very hard to say" what the local mix looks like.
• Management declined to share revenue contribution by customer type (enterprise, AI startups, etc.), arguing the company is too small at ~5,000 GPUs for such tracking metrics to be relevant.
• India AI revenue was about 20-21% in the previous quarter, international roughly 37%, and the rest domestic, compared to a 60-40 domestic-foreign split and India AI mission at ~40% last quarter, per an analyst's math.

Margins & Cost Structure

• The 75.2% EBITDA margin is a combination of increased utilization, robust demand, and operating leverage; management believes it is "quite sustainable over the medium term and potentially the long term."
• Sustainability rests on a judicious mix of contracted, on-demand, and spot revenue, plus exploration of AI services (inference, agentic) based on open-source models.
• Interest costs spiked due to the loan taken to fund the first lot of B200 GPUs, and with another lot coming, the loan will increase in the near term.
• Data center costs will fluctuate as usage increases, with the electricity portion rising; for existing capacity, most cost is already baked in this quarter, and coming-quarter increases will not be a massive percentage.
• Management declined to quantify peak debt for the year or provide a breakdown of the ₹881.1 crores capex utilization.

Funding & Capacity Expansion

• The expansion philosophy is "both aggressive as well as judicious": management intends to keep balance between funding and building capacity, avoiding being too aggressive while also not missing the AI bus.
• No specific GPU count target for two years out was given; management only reiterated the "judicious as well as aggressive" framing.
• Sovereign Cloud funding details were deflected: "Let us build and execute the plans and then put them out there."
• On future equity raises, management said only: "Obviously, we would let everyone know if and when that happens."
• Some customers are signing 1-3 year contracts to lock pricing, and E2E has executed some and will continue.

Dua: "It is still day zero of AI. We are still in the midst of largest build-outs in the history of IT itself where the whole world is building the AI super cycle. And I think we are at the beginning of that AI super cycle."

What Management Declined to Answer

• Exit MRR guidance for the full year: "We don't provide guidance on MRR. We always advise everyone to look at what we have done in the past rather than predict the future."
• Training vs. inference revenue split: Deflected as too hard to measure on a small GPU subset with fungible workloads.
• Revenue mix by customer type: Declined, citing the company's small size at ~5,000 GPUs.
• SovCloud funding and strategic details: "Very early days, as we do more, we'll obviously announce that."
• Peak debt quantification for the year: "We will not be quantifying that amount."
• Capex utilization breakdown for ₹881.1 crores: No specific breakdown provided.

Guidance & Commitments

Key Metrics