E2E Networks Ltd Q3 FY2025 Concall Summary

Key Takeaways

• Revenue dropped 12.6% QoQ as concentrated training workload completions and customer churn hit a small-scale, concentrated revenue base hard in a single quarter.
• Sales cycles are lengthening, not demand disappearing: the pipeline remains strong but immediate closures have vanished, a shift management admits it probably did not anticipate.
• H100 and H200 GPUs are underutilized right now, and the company will not disclose the utilization rate, calling it a fluctuating number that is hard to measure point-in-time.
• The L&T partnership is weeks old with no revenue yet; it is the company's main bet to shift toward larger customers and longer-cycle business over the medium term.
• What to watch next quarter: Whether the Chennai data center operationalizes on time, any sign of ARPU stabilization after the training churn, and the IndiaAI Mission financial bid outcome.

The Q3 Revenue Shock

• Training concentration caused the drop: Multiple large training deployments completed in Q3, and because the company is small with customer concentration, the de-provisioning effect was highly visible. If the churn had been among smaller customers, the impact would not have registered.
• December seasonality amplified it: Training teams take vacations in December, and training activity is not a 24x7, 365-day operation but more like nine months out of twelve. Training teams may end one cycle and return the next year to redo work.
• ARPU declined as larger training customers churned or downscaled, and management cannot put a sharp timeline on when improvement starts to show up in the numbers.
• Inference grows slowly: Initial inference volumes are always small and scale up over time as enterprise AI adoption grows, so inference cannot yet offset training lumpiness.
• Management stance on the surprise: When asked if Q3 played out as anticipated or contained a negative surprise, management deflected, stating "nothing surprises us basically" and that the company operates on the principle of having the ability to react to anything.

Tarun Dua: "The demand environment continues to remain strong. So, the pipeline is there obviously for both H100 as well as H200. What we are potentially missing is immediate closures that we were seeing in the past. So, I think the sales cycles have grown. I don't think the demand has gone away. I think the sales cycles have become longer."

GPU Utilization & CAPEX Strategy

• H100 and H200 underutilization is acknowledged but not quantified: management calls utilization a fluctuating number that is very hard to measure point-in-time and declined to give a percentage.
• CAPEX is reactive, not predictive: The company will not commit to a spend timeline because GPU versions get superseded quickly, and building inventory of soon-to-be-obsolete equipment is a risk. Deployment will be just-in-time based on pipeline evaluation over the next several quarters.
• Blackwell GPUs are in conversation but not ordered: The company intends to build significant Blackwell capacity when the range becomes available, but no orders have been placed yet.
• Reactive Hopper CAPEX remains possible depending on the outlook from ongoing sales conversations.
• Depreciation jumped because the company uses straight-line over six years: once additions happen in a quarter, depreciation continues on that schedule regardless of utilization.

Tarun Dua: "Very very hard to commit to kind of like spend that money in a market where the infrastructure, as in like the high-tech equipment like GPUs and CPUs, they get like upgraded version. So, it is very, very important not to kind of like build a lot of inventory of what would potentially kind of like will get superseded by a newer version."

Customer & Revenue Mix Shift

• Training dominates revenue and will continue to do so for the foreseeable future. The company does not track training vs inference separately because the line between them is "nebulous."
• Medium-term normalization expected: Over time, the ratio should move toward 50:50 or 60:40 in favor of training, and eventually 40:60 with inference dominant, but no timeline is given.
• Shift to larger customers is a medium-term goal: The company plans to start working with larger customers in conjunction with partners including L&T, pursuing long-cycle business opportunities. This is described as a natural change, not an immediate pivot.
• Geographic concentration remains: All data centers are in India (Delhi NCR, with Chennai coming), and customers are primarily in India. International expansion is a medium-to-long-term aspiration with no specific timeline.
• US GPU regulation impact should be assessed on a roughly two-year timeframe, not immediately. No immediate impact is expected.

L&T Partnership & IndiaAI Opportunity

• The L&T partnership is weeks old: Teams have started interlocking and jointly exploring customer opportunities, but it is too early for revenue contribution. The logic is that both businesses bring their own strengths to the table.
• IndiaAI Mission technical evaluation is cleared: The company has qualified the technical evaluation for the IndiaAI tender. Financial bids have been opened but L1s have not been declared yet. Once declared, L1s would need to be matched by players seeking empanelment.
• IndiaAI is a net positive for the industry: Regardless of who becomes the biggest beneficiary, the AI Mission and its budgetary support will expand the overall market in India.

Tarun Dua: "It's quite early days. I think it's been practically less than a few weeks. So, basically, where we have started interlocking, started working together, and jointly exploring opportunities with the customers, so in the long run, I believe that the interlock in terms of jointly exploring opportunities would obviously help both the businesses who bring their own strengths to the table."

Infrastructure Expansion

• Data center capacity is doubling: From 4.2 MW to nearly 10.2 MW, supported by the recent fundraise.
• Chennai data center adds seismic and connectivity diversity: The southern location gives access to a different seismic zone and landing stations that are significantly different from those in Delhi and Mumbai.
• Chennai timeline is aggressive but uncertain: The company is trying to operationalize the facility during the next quarter, but whether all services can be deployed in the first quarter remains to be seen.
• The TIR platform continues development, designed to streamline AI/ML workloads including training, inference, and model endpoint deployment. The platform has been in continuous operation for a decade, which management says has made it battle-tested.

Competitive & Market Context

• Competition has intensified: Since AI became mainstream around 2019-20, competition in compute and technology has become far more intense, and management acknowledges this is the reality they operate in.
• Edge computing is additive, not zero-sum: Management does not see edge AI taking away the need for data centers; whatever runs on the edge, there will be still further workloads on the cloud.
• Sales cycle lengthening was unanticipated: Management probably did not anticipate the increase in sales cycles, but over time the effect gets mitigated by doing more sales effort and expanding the number of conversations.
• Medium-to-long-term outlook remains growth-oriented: AI is a very large market with many players, and the company sees continued growth over the medium to long term.

Tarun Dua: "As AI has become very, very mainstream since 2019-20, since we have been working with AI, the competition has become far more intense today. And competition in the compute and technology world has always been intense."

Guidance & Commitments

Key Metrics