E2E Networks Limited

E2E rents GPUs by the hour and by the month. It buys NVIDIA accelerators, puts them in third-party data centres, wraps them in its own software and sells the compute to anyone in India training or running an AI model. The fleet is about 5,100 GPUs live today across A100, V100, H100, H200 and B200, with another 1,024 B200 waiting to be installed.

The software is the part that is not a commodity. TIR is E2E's training and inference platform, Jarvis Labs is a bought-in layer on top, and the Sovereign AI stack lets a customer run open-weight models inside their own boundary and control what the model sees, keeps and forgets. Management's argument is that open-source models are now close enough to frontier models that an enterprise would rather own the whole stack than rent access to somebody else's API.

The group has three new subsidiaries. SovCloud holds and contracts large-scale infrastructure. A Delaware entity sells outside India. L&T is an arm's length partner: E2E buys data centre capacity from them, they buy cloud from E2E, and the two go to customers together.

Tarun Dua is Managing Director and Nitin Jain is CFO. Promoters hold 39.45 percent as of June 2026, down from 59.71 percent three years ago. The stock split 10 for 1 and listed directly on the BSE mainboard during Q1 FY27.

E2E spent FY26 buying assets that did not earn. Revenue was Rs 246 cr, depreciation was Rs 169 cr, and the company reported a full-year loss of Rs 16 cr while burning Rs 1,140 cr of free cash. Three of those four quarters were loss-making. The H200 fleet arrived before the demand for it did, and the market saw a company whose costs were fixed and whose revenue was not.

That gap closed in one quarter. The 1,024 B200 went live in mid-May and were fully utilised on arrival. June-quarter revenue was Rs 157 cr against Rs 36 cr a year ago. EBITDA margin went from 29 percent to 75 percent. The company made Rs 44 cr of profit after tax against a Rs 3 cr loss in the same quarter last year. Almost none of that came from price. Management said twice on the call that the numbers reflect capacity and utilisation rather than the July price increase, which has not hit the P&L yet.

Then on 31 August the business model changed shape. E2E signed a binding term sheet with an unnamed Indian sovereign AI company worth about Rs 1,000 cr, running to June 2029. It converts the Blackwell capacity deployed in May from pay-as-you-go to a committed contract. That is roughly Rs 29 cr a month of revenue with a term attached to it, against a company doing about Rs 52 cr a month in the June quarter.

So the question in front of the price is no longer whether E2E can fill GPUs. It has filled them, at 75 percent margins, and locked a large slice of that for three years. The question is what it costs to buy the next fleet and who pays for it.

The second B200 lot. Another 1,024 units are expected within a couple of months. The first lot filled on day one and E2E says the existing fleet is at maximal utilisation. Growth from here comes from capacity, not price, which management stated plainly.

The Rs 1,500 cr raise. The board approved a QIP, rights issue or FPO on 31 August, going to shareholders at the 28 September AGM, along with a borrowing limit raised to Rs 10,000 cr. The current loan is Rs 450 cr. This is the funding for the Blackwell and Vera Rubin build-out, and how it is priced sets the share count everything else is divided by.

Contract mix hardening. Customers facing higher prices asked to lock rates and offered advances and one to three year terms to get them. E2E took the trade. Every conversion of on-demand capacity into contracted capacity lowers the risk that a fixed depreciation charge meets a variable revenue line, which is exactly what produced the FY26 losses.

Revenue mix already diversified. India AI Mission was 20 to 21 percent of revenue last quarter, international about 37 percent, the rest domestic private. The Delaware entity is the vehicle for the international share.

MSCI Global Small Cap inclusion. Confirmed in the August review. Passive money buys on the schedule regardless of what it thinks of the multiple.

Vera Rubin. Management has begun preliminary planning for B300 and Vera Rubin. No numbers, no dates. Worth watching, not worth paying for.

Assuming the second B200 lot lands in Q3 FY27 and fills the way the first did, EBITDA margin settles near 72 percent rather than holding the 75.2 percent of a single strong quarter, depreciation runs on management's stated six-year GPU life, the Rs 1,500 cr raise closes in H2 FY27 at roughly the current price, and tax at 25 percent. FY28 and FY29 EPS is on 23.1 cr shares after that dilution; FY27 is on the current 20.57 cr.

On gross debt of Rs 450 cr the enterprise value is about Rs 12,800 cr, which is 20 times FY27 EBITDA and 13 times FY28.

The P/E line is not the right lens here and neither is the TTM number, which averages a fleet that was idle with a fleet that is full. This is a business that buys an asset, depreciates it over six years and rents it out. What matters is the cash a GPU returns against what it cost. The June quarter shows a fleet earning Rs 118 cr of EBITDA in three months on a gross block of roughly Rs 1,570 cr including work in progress, which is the whole reason the stock moved.

At Rs 602 the price asks E2E to roughly triple revenue over three years and hold margins near 70 percent while doing it. The first leg of that is already contracted. The Rs 1,000 cr term sheet, the India AI Mission work and the US order together put a floor under a meaningful part of FY28 revenue before a single new GPU is switched on. The rest depends on capital, and the company has just asked for Rs 1,500 cr of equity and headroom for Rs 10,000 cr of debt.

The Rs 1,000 cr contract is a term sheet, not a definitive agreement, and the counterparty is not named. A single customer at that size is concentration risk whichever way it is described, and E2E does not disclose the credit behind it.

That contract also does not add capacity. It re-papers GPUs that were already earning. Treating it as new revenue double counts.

Depreciation is the swing factor. Management assumes a six-year GPU life. If Blackwell prices the older Hopper fleet down faster than that, the charge stays and the revenue does not. FY26 is the demonstration: Rs 169 cr of depreciation against Rs 126 cr of EBITDA, and a loss.

Free cash flow was negative Rs 1,140 cr in FY26. Accounting profit and cash are far apart in this business and will stay apart while the fleet grows.

Funding cuts both ways. Rs 1,500 cr of equity at the current price is about 12 percent dilution, and the Rs 10,000 cr borrowing limit tells you where the balance sheet is heading. CFO Nitin Jain declined to quantify peak debt for the year.

Disclosure is thin. No MRR guidance, no split between training and inference, no customer mix, no capex figure, no contracted versus on-demand percentage. Management's stated position is that the company is too small for these metrics. That leaves an investor underwriting utilisation without being able to see it.

Promoter holding is down to 39.45 percent from 59.71 percent three years ago.

EBITDA margin below 60 percent for two consecutive quarters. The sovereign term sheet not converting into a definitive agreement by the December quarter. The second B200 lot slipping past Q3 FY27. Revenue falling sequentially while depreciation rises. Any extension of the GPU depreciation life beyond six years. Net debt above two times EBITDA.

TenetFour Research. Educational analysis only. Not investment advice or a recommendation to buy, sell, or hold any security. Author may hold a position. Readers are responsible for their own decisions.