E2E Networks Ltd

A1. What E2E is, and where it sits

E2E Networks operates GPU cloud infrastructure in India — renting accelerated compute (NVIDIA H100/H200, and since May-2026 B200) to AI startups, enterprises and government-backed programmes. It is a capacity landlord: it buys GPUs with capital, racks them in colocated datacentres, and sells them by the hour or on fixed-term contracts. Revenue quality is therefore a function of two things the company controls — utilisation and contract mix — and one it does not: the price of accelerated compute, which is set globally.

Scale at the report date: ~5,100 GPUs, FY26 audited revenue ₹245.58 Cr, Q1 FY27 revenue ₹156.8 Cr, exit MRR ₹71.8 Cr/month at 30-Jun-2026. Larsen & Toubro holds 37,933,290 shares (~₹2,330 Cr at market) with two board seats. SHP 30-Jun-2026

The single structural fact that shapes everything downstream: E2E discloses no power, bandwidth, datacentre or colocation cost in either FY25 or FY26. Those costs sit bundled inside “Purchase of services and consumables” with no breakdown. FY25’s Annual Report carries a “Power and fuel” line of ₹3.44 lakh — trivial for a GPU cloud, and proof the omission is structural rather than incidental. Any analyst figure for power cost per GPU-hour would be fabricated. GAP

A3. The opportunity — and how much of it is authorised rather than real

India’s IndiaAI Mission authorised ₹10,371.92 Cr, of which ₹4,563.36 Cr targets compute capacity. Cabinet Mar-2024, via Medianama/Elets The empanelled GPU pool moved 18,417 → 34,333 PIB May-2025 and past 38,000 in 2026 MeitY/Tribune, against a stated 100,000 target.

The correction that matters: E2E’s share of the empanelled pool is DERIVED: 2,524 ÷ 38,000 = 6.6%, not the 7.35% carried in the first draft. And the headline authorisation is not a pipeline — only ~₹400 Cr of the ₹10,371.92 Cr had been released as at Apr-2026. The addressable market is an authorisation, and treating it as demand is the most common error made about this company.

₹1,249.5 Cr is the desk figure of record. The valuation section published ₹1,206.42 Cr built on a ₹160.90 Cr undrawn figure that the fundamental section had already withdrawn — the ₹43.1 Cr difference is exactly the lease adjustment. Corrected at the second gate; one figure survives, not two. Separately, ₹893 Cr of issuance headroom exists but is authorisation, not demand. devil gate

A5. Financials — and the one question this quarter cannot answer

Revenue compounds at 35.0% across the available history computed.json, but the shape matters more than the rate: FY26 audited revenue was ₹245.58 Cr and Q1 FY27 alone was ₹156.8 Cr. A single quarter delivered 64% of the prior full year.

Revenue and EBITDA, FY15–TTM — the step is Q1 FY27, not a trend.

Margin trend — expanding, but FY26 audited (51.4%) and TTM (63.9%) are different questions.

A8. Forensic filter — 5 clean, 6 caution, 1 flag

The audited statements tie in every direction tested: expense components sum to totals in both years; the balance sheet balances in both; the cash flow’s three sections tie to ₹0.02 lakh; closing cash ties to the balance sheet; the deposit pool reconciles within ₹0.43 Cr; finance costs reconcile P&L-to-cash-flow in both years; and an independently derived FY25 EBITDA reproduces ICRA’s published 7.3x interest cover exactly. This is not a dressed set of books.

The single 🔴 is not an accounting-integrity finding — it is that the audited balance sheet is four months obsolete on its most important line.

Free cash flow against PAT — FY26 FCF was −₹1,140 Cr on a ₹1,262 Cr capex year.

Jarvis Labs — 🔴 on disclosure form, de-fanged by the balance sheet. Accounted as a business acquisition with no price, no PPA, no goodwill and no fair-value table anywhere — searched across the FY26 filing, the Q1 filing and four transcripts. Ind AS 103 requires the fair value of consideration transferred. But the quantum is bounded: other intangibles rose only ₹1.73 Cr net; adding back ₹1.16 Cr amortisation gives gross intangible additions of ~₹2.89 Cr for the whole year, and no goodwill line exists on the balance sheet at all. The component cannot exceed ~₹3 Cr.

A10. Valuation — the base case sits below the market price

A conventional FCF-DCF is not deliverable, and this desk will not fake one. FY26 FCF was −₹1,140 Cr; capex jumped 10× from ₹125.93 Cr to ₹1,262.45 Cr. Terminal value would be ~100% of any DCF built here, which makes it an opinion with decimal places.

Instead, a reverse exit-multiple bridge. Requiring a 12% p.a. equity return ASSUMPTION — no WACC is derivable from the filings over three years gives a required EV of ₹17,934 Cr. At management’s own maintained 70% EBITDA margin target:

The price implies 14–35% p.a. revenue growth from the current run-rate, or 73–106% p.a. from the last audited fiscal year. The gap between those two numbers is the investment debate — and it is entirely a question of whether the run-rate is a level or a peak.

Debt and coverage — ~₹505.88 Cr of total obligations against a depreciation charge that does not pause.

A16. Scenario analysis — 1 year (FY27)

The base case sits BELOW the current price. On the desk’s own filing-derived base, using E2E’s own median historical multiple, the 12-month reference is ₹467–540 against ₹614.15 — after a +44.5% run in thirteen sessions. The market has already paid for the base case and some of the bull.

Probability-weighted 1Y reference: ₹561 — about 9% below the close. DERIVED: 0.30×272 + 0.45×503.5 + 0.25×1,012.5 Arithmetic on the desk’s assumptions, not a model output, and not a target.

Bear weight is 30%, above a conventional 20–25%, for four filed reasons: contracts reaching term inside FY27 with no renewal disclosed; the ₹1,249.5 Cr funding gap; ICRA’s rating being 8.3 months stale with an explicit debt-funded-capex downgrade trigger; and one quarter being a thin basis for a 75.2% margin. Bull stays a full 25% because management has met or beaten every specific commitment it has made — MRR ₹35–40 Cr → ₹37.4 Cr, margin 70% → 75.2%, B200 mid-May → live late-May. The countervailing fact: it makes very few, and gives no forward numbers.

A13. Risks — and the one that kills the thesis

The renewal cliff, both conventions, with the inference labelled

The first step lands ~end-Aug-2026 — roughly three weeks after this report — not months away. GNANI, the single largest contract, was being computed on the commencement basis while the section claimed award dates throughout; corrected, the cliff’s first step is three months earlier and ₹177 Cr larger than the prior draft published (₹265.02 Cr by ~end-Aug-2026, not ₹161.0 Cr by ~Oct-2026). Both conventions are carried above rather than averaged: the award basis is computable from what E2E has actually disclosed; the commencement basis rests on GNANI’s 4.5-month lag as a single observation. The first draft’s “all expiring ~Sep-2026 to ~Jan-2027” was withdrawn — it mixed go-live basis for one contract with award basis for the others inside a single cell and stated a window narrower than the evidence supports. The only empirical anchor for the award-to-commencement lag is GNANI’s observed 4.5 months, an inference from a single observation, and it is applied as a one-sided ±. devil gate

Three order-book figures circulate and all three are correct at different scopes: ₹265 Cr = the IndiaAI pair (ICRA’s stated book) · ₹338.02 Cr = the three fixed-term contracts · ₹346.51 Cr = total disclosed wins including the ₹8.49 Cr SaaS order. Same underlying set. No contradiction.

A15. Ownership — locked at the top, blind through the move

SMART MONEY: NEUTRAL. Control blocks locked and immobile through the filed endpoint (constructive); marginal institutional flow negative into the print (cautionary). No insider distribution; no institutional accumulation into the last leg.

A correction worth surfacing: the first draft read 28-Jul’s 44% delivery as accumulation. Delivery measures transfer, and is direction-agnostic — the direction that day was down: a lower-circuit close at the day’s low, −5.00%, on the second-heaviest volume of the run, with the following session’s delivery at 36% (the run’s low). devil gate

Return ratios — ROCE and ROE are NaN across all 13 periods, and that is a finding, not a rendering bug.

B0. Stage — a Stage 2 advance in its late, accelerated phase

The 20–24 Jul episodic pivot fired on real volume — 4.72× the trailing-30 average on 23-Jul — and the delivery data confirms transfer rather than churn: the 22–23 Jul circuit locks carried 100% and 99% delivery. The first draft claimed “three times the delivery on a third of the volume.” The volume was not a third — it was essentially identical (41,281,150 vs 39,009,879 shares traded), which makes the real finding stronger rather than weaker: 5,625,529 vs 18,104,683 shares delivered — 3.2× the stock changing owner on the same volume (13.6% vs 46.4% blended). devil gate

The blemish, stated rather than buried. Ten of thirteen sessions from 20-Jul to 05-Aug closed at a circuit band, four of them in the last seven. 28-Jul closed at the LOWER band. Band levels are simply close(t−1) × 1.05, so this is fully derivable from closes — the lock signature is demonstrated, not asserted.

B4. Risk:Reward — no valid entry exists

Research reference levels only — not buy or sell recommendations. All arithmetic uses ₹614.15.

The practical obstacle comes before the arithmetic: the next session’s band caps at ₹644.86 and nine of the last thirteen sessions closed at a band. You cannot reliably fill above a locked circuit — the mechanism that drove the advance is the mechanism that prevents participating in it. The June–July base’s measured move (569.75) was already reached on 04-Aug: the primary target is spent.

What would create a valid entry: a 3–6 week consolidation holding above the ₹516.70–542.50 volume shelf, with the mean absolute daily move compressing from 4.42% toward ~2%, volume drying through the base, then expansion through the ceiling. That is weeks of sideways action, not days — and the ±5% band must go inactive during it, because while the band is active fills are not controllable in either direction.

C1. The investment case

E2E Networks has genuinely inflected: FY26 audited revenue of ₹245.58 Cr became a Q1 FY27 quarter of ₹156.8 Cr at a 75.2% EBITDA margin, on a fleet that reached ~5,100 GPUs after a ₹1,262 Cr capex year. The operating story is real and the books behind it are clean — the forensic pass tied the audited statements in every direction it tested and reproduced ICRA’s published 7.3× interest cover exactly.

But the quarter that re-rated the stock is contract-shaped, not platform-shaped. Three named fixed-term contracts worth ₹338.02 Cr carry 37.5–47.7% of the ₹71.8 Cr exit MRR and occupy ~2,524 GPUs — 49.5% of the entire physical fleet — On the convention actually computable from disclosure, the first step is not months away but weeks: ₹265.02 Cr and that entire ~49.5% of the fleet reach term around end-August 2026 — inside Q2 FY27 itself. E2E has disclosed a service-commencement date for only one of the three contracts, so the alternative reading pushes the same set to ~Mar-2027; that upper bound rests on a single observed award-to-go-live lag.

Against that, the price has already paid for the base case. The 12-month reference range is ₹467–540 against a ₹614.15 close, with a probability-weighted reference of ₹561 — about 9% below spot after a +44.5% run in thirteen sessions. The technical structure offers no entry that survives a 3% stop. The result is a good business at a price requiring the bull case to be delivered, with the single largest determinant — whether ~half the fleet gets re-sold at commercial rates — falling due inside two quarters, under a company with no Chief Revenue Officer since 30-Jun-2026.

C4. Alignment — DIVERGING

Fundamentals and technicals point the same direction on the business and opposite directions on the price.

The technicals say this is a momentum leader in a confirmed uptrend with volume and delivery confirmation. The valuation says the market has already paid for the base case and part of the bull. Those are not contradictory readings of the same evidence — they are the classic signature of a late-stage re-rating, where price has moved faster than the disclosure needed to validate it.

What this means for timing: the divergence resolves on a date, not on a view. Q2 FY27 carries three months of B200 power, colo and bandwidth against one in Q1 — it is the first print that can distinguish operating leverage from a timing effect. Until then the technical structure offers no entry surviving a 3% stop and the fundamental case offers no valuation cushion. Buying here requires being right about the bull case and about the fill, simultaneously. The asymmetry is poor in both dimensions at once.

Working capital — the cycle is not where the risk sits; the contract cliff is.

C7. Primaegis research opinion

⚠ SEBI: Internal analyst opinion only. Primaegis is not a SEBI-registered investment adviser or research analyst. This is research framing, not a buy or sell recommendation, and nothing here is personalised investment advice.

Run notes — what this desk could not establish

Recorded, not filled. Infrastructure cost decomposition (structurally undisclosed in both years) · the definition of exit MRR · numeric Q1 FY27 utilisation · Q1 audited cash flow · Jarvis Labs consideration · SovCloud funding structure · peak FY27 debt · FY26 capital commitments, contingent liabilities, CWIP and receivable ageing · customer-concentration exposure · ROCE/ROE (NaN across all 13 periods) · sell-side estimates, consensus and target price (none exist) · service-commencement dates for the GAN AI and US contracts · whether the ₹1.25 Cr High Court deposit was made.

Chart captures: static TradingView screenshots were attempted and failed — the drivable chart tab was a background tab that had never laid out (all canvases at the 300×150 default), so every capture returned blank, and screen recording was not permitted to the capturing process. Per desk policy, captures are enrichment rather than a gate; the six charts above are generated from the desk’s own computed dataset and the live TradingView widget is embedded in Part B. Recorded as a gap rather than worked around silently.

Primaegis Equity Desk · 7 agents · 2 adversarial gates · generated 06-Aug-2026. Every load-bearing number traces to a filed source; where two analysts disagreed on convention the divergence is carried as a range rather than averaged.