Together AI rents India's largest GPU cluster rather than building it

Larsen & Toubro's data center arm said this week that its Vyoma.AI unit, through the AI infrastructure subsidiary LTN Compute, will host an Nvidia B300 "AI Factory" for Together AI, with 10,000 GPUs at Vyoma's Chennai campus and a first phase offering 250MW ^[1][2][3]. The two companies describe it as India's largest single-cluster AI infrastructure, and the notable part is who owns what: Together AI arrives as a tenant of a construction conglomerate's sovereign cloud business, not as a developer ^[4][8].

That is consistent with how Together AI operates. Founded in 2022, it leases chips from other cloud providers and re-leases them to developers, while also buying its own servers to rent out from its own facilities ^[6]. Its documentation places those facilities in North America, and details are otherwise sparse ^[7]. The company has recently said it will deploy GPU capacity via IBM Cloud, has a multi-year cloud capacity agreement with Rumble Inc, and has signed with Hypertec and 5C for European capacity ^[11][12]. It has also announced a dedicated cluster for Y Combinator portfolio startups ^[13]. Chennai is another line in a procurement sheet, not a change of strategy.

The deal value is reported at around $1bn to $1.75bn ^[5]. Across 10,000 GPUs that implies $100,000 to $175,000 per GPU, which is more than the silicon alone would plausibly account for, so the figure is doing more than describing a hardware order ^[4]. The report does not break out term, power pricing, or who holds title to the servers ^[16].

The scale is the part buyers should sit with. Vyoma's 2025 annual review lists two live data centers, 2MW in Navi Mumbai and 30MW in Chennai, plus two more in development, against a stated target of building over 200MW across India ^[10]. That is 32MW live today ^[1]. A 250MW first phase is roughly eight times the operating footprint and, on its own, exceeds the company-wide construction target ^[2][3]. Vyoma itself is barely a year old under that name, rebranded from Cloudfiniti with a cloud platform added in February ^[9]. The chairman, S N Subrahmanyan, framed the deployment as a milestone in what L&T calls its Gigawatt AI Infrastructure Mission ^[14].

One ratio is worth checking against any vendor's capacity language: 250MW divided by 10,000 GPUs is 25kW per GPU ^[5]. No GPU cluster draws that, so the 250MW is site capacity for a campus phase rather than the load of the announced hardware. Buyers reading "250MW" as a proxy for this cluster's size will be off by a large multiple.

The other structural point is the word sovereign. Vyoma sells sovereign cloud platforms, GPU-as-a-Service and hyperscale colocation ^[8], and the anchor tenant for its largest publicly announced block is a foreign AI cloud reselling to developers. Together AI's co-founder Vipul Ved Prakash said the partnership was meant "to bring the scale, resilience and engineering excellence this moment demands to India" ^[15].

What to watch: no deployment schedule appears in the report ^[16], so the first thing to confirm is whether the 250MW is energised, contracted, or designed, and whether the two facilities under development are the vehicle ^[10]. Then watch whether Together's next capacity announcements continue to be other people's buildings.