Yes, those figures are per rack. I was highlighting the data per GPU, but I agree that the cost-per-token improvement is on a rack basis. You are correct. I will correct the write up.
I just want to understand one thing from the community.
Is Technology Obsolescence a big looming threat for this entire industry as new technologies comes up every year or few years.
How is that mitigated? Can someone explain me this.
1 Like
Performance comparison across generations is not as straightforward and varies with applications. Token throughput is a function of multiple factors (precision, model parameters, latency, concurrency). Blackwellâs key advantage is FP4 support, higher memory and bandwidth. While this helps with frontier models, the performance leap is not that high for mid-tier SLM models with fewer parameters (which donât require as much memory). Most of current enterprise use cases are for application specific workloads for which such SLMs are sufficient. For these, Hopper offers comparable price-performance as well as more mature software ecosystem and hence will continue to remain relevant.
Nvidia sold USD 2 bn of Hopper in Q3FY26 (https://finance.yahoo.com/quote/NVDA/earnings/NVDA-Q3-2026-earnings_call-379484.html?utm_source=chatgpt.com), four years after its launch. Recently, Coreweave also commented that it re-contracted its H100 capacity within 5% of original price, despite Blackwell being widely available. Of late H100 rental pricing has also been rising (https://www.silicondata.com/products/silicon-index), indicating continuing demand.
Even A100 (released in 2020) is seeing steady pricing and demand. The leap from Ampere (7 nm) to Hopper (4 nm) was much larger than Hopper to Blackwell (4 nm). Despite that A100 continue to remain relevant today. This indicates that old GPUs continue to remain relevant even after 2-3 newer generations are released. So, while obsolescence is an ongoing debate, at least past experience does not suggest that GPUs become obsolete so soon.
14 Likes
How will E2E stand to gain from tax holiday announcement? If global hyper scalars are going to come to India to serve international customers (that too likely will be in countries that do not have strict data sovereignty policies which is smaller pie) they will build out their own DCs as they always do instead of using players like E2E. Is this policy targeted more towards non CSPs like salesforce/SAP to use india to serve the world?
Looks like this only means more competition for E2E in competing for international customers as now foreign companies will also come, setup shop and compete from India (while also getting tax benefits which E2E will not get)
There is a possibility that neo-cloud players may choose to partner with them. Many neo-cloud businesses do not have sufficient cash to deploy assets in India, so partnering with players like E2E allows them to leverage Indiaâs tax benefits without heavy capital investment.
Hyperscalers, in contrast, are likely to bring the entire ecosystem into India. While hyperscalers were already present earlier, even with tax benefits they may still find it difficult to compete on certain economics and agility parameters. At the same time, Indian cloud players are also expected to explore partnerships with foreign players to strengthen their offerings.
Disclaimer: Invested and biased.
5 Likes
upto dec25 1100cr gross block
1> loss 7cr
2>utilization 60%
3> dep 47cr fully provided quarterly
assuming MRR goes up to 35 to 40cr and utilization lvl at 85% going forward ,the bottom line will be big positive.
jan26 onwards gross block 1700cr
1> blackwell gpus rents are much more than previous generation (270 to 350 rs/hr vs 180rs/hr) . assuming variable cost will not go up as the utilization goes up. so break even utilization lvl can come down drastically going forward.
2> watching massive demand for AI at least in media, upcoming AI SUMMIT and all gvt push , the tailwind is going to stay for some time.
btw in last two weeks L&T launches 40MW dc in navi mumbai . in BHUMI PUJAN Tarun dua was present. In chennai dc more than 15 mw gpus can be added and now another 40mw.
who will fill the racks? is it e2e or some body else? invested recently.
4 Likes
Amazon Web Services (AWS) CEO Matt Garman has said that the company is still running six-year old Nvidia A100 servers, in part due to customer demand for GPU capacity continuing to outstrip supply.
"Because there is so much more demand than supply, there typically still is demand for the older chips⌠we actually are completely sold out and have never retired an A100 server.â Matt Garman
6 Likes
Here you go the same AWS CEO saying too
Just adding some context, most users donât really care about the name of the GPU being used.
What matters is the throughput/memory required for the specific application, and the lowest cost hardware that satisfies this.
2 Likes