# AI capex — X 热门讨论 (2026-09-27 19:06 UTC)
## @MMatters22596 (The Analyst) · 09-27 13:24 · ♥38 ↻0 💬21 Everyone is talking about $NVDA’s “cheap” forward P/E.
But the real question isn’t valuation today.
It’s the durability of tomorrow’s earnings. $NVDA doesn’t have a traditional multi-year customer backlog.
At the same time, AI is entering a new phase.
The market no longer just wants bigger models. It wants to see a return on trillions of dollars of AI infrastructure spending.
If monetization fails to justify that spending, CapEx growth eventually slows.
$NVDA would still be an incredible company. But an incredible company can trade at a very different valuation. https://x.com/MMatters22596/status/2104200383937540405
## @mezza9_Equity (Mezzanine Equity Pvt Ltd) · 09-27 16:32 · ♥39 ↻6 💬5 ESDS FY27–FY29: My Estimate
First the summary.
FY27E Revenue: 1,420–1,750 cr PAT: 265–378 cr
FY28E Revenue: ~7,230–7,280 cr PAT: ~1,165–1,525 cr Average revenue-generating GPUs: ~15,000 FY28-end capacity assumption: ~50,000 GPUs
FY29E Revenue: ~22,500–23,000 cr PAT: ~3,500–4,650 cr Full-year benefit of ~50,000 GPUs
No additional large GPU deal beyond 50k is included.
These are my estimates, not management guidance.
Why these numbers?
I see 4 separate revenue engines with different economics.
1. Existing business
FY26 had a sizeable one-off technical/design contribution which also carried good margins.
I am not assuming any similar one-off in FY27.
So headline YoY growth may not show the actual underlying growth. After normalising the FY26 base, I estimate underlying FY27 growth around 50–70%.
FY27 revenue: 520–550 cr PAT margin: ~25% PAT: 130–138 cr
2. Sharon AI
ESDS has contracted 8,208 NVIDIA B300 GPUs from Sharon AI for 60 months.
GreenSquare has separately announced Sharon's planned deployment of approximately 8,200 B300 GPUs at its SYD1 facility, with initial operations targeted in Q4 CY26.
The numbers are almost identical. I still want management confirmation that this is the exact ESDS cluster, but it gives an important execution timeline to track.
My estimated full-year ESDS revenue from 8,208 GPUs is ~3,600 cr.
That is roughly 300 cr per month.
Management commentary indicates contribution can begin around late Q3 FY27.
I am keeping execution buffer and counting only 900–1,200 cr in FY27.
PAT margin assumption: 15–20%
FY27 Sharon PAT: 135–240 cr
From FY28, I assume a full-year contribution.
Revenue: ~3,600 cr PAT: ~540–720 cr
3. ESDS own AI factory
Management has discussed around 1,500 cr of own AI infrastructure capex.
After considering the complete AI factory cost including GPU servers, networking, storage, interconnects, racks, cooling, power and supporting hardware, my estimate is roughly 2,500–3,000 B300-class GPUs.
Management expects commissioning around Q4 FY27.
I am adding execution buffer.
My assumption is commissioning in Q1 FY28 and billing from Q2 FY28.
So FY27 contribution is zero and FY28 gets only around 9 months.
PAT margin assumption: ~20%.
4. Journey towards 50k GPUs
Management has spoken about reaching around 50,000 GPUs within roughly a year.
I am not assuming 50k GPUs become operational together.
AI infrastructure gets installed in tranches. GPU delivery, power, liquid cooling, networking, storage, rack installation, testing and customer acceptance all take time.
So I assume ESDS reaches approximately 50k GPUs only by FY28-end.
More importantly, I count only ~15k average revenue-generating GPUs during FY28.
That 15k already includes Sharon's 8,208 GPUs, partial-year contribution from ESDS-owned 2,500–3,000 GPUs and additional capacity progressively installed during FY28.
For additional partner/third-party GPU capacity, I assume 15–20% PAT margin.
FY27E
Existing business: 520–550 cr Sharon: 900–1,200 cr
Revenue: 1,420–1,750 cr PAT: 265–378 cr
FY28E
Existing business: 650–700 cr Sharon: ~3,600 cr Own GPU capacity: ~820–990 cr Other GPU capacity: ~1,990–2,160 cr
Revenue: ~7,230–7,280 cr PAT: ~1,165–1,525 cr
FY29E
This is the first year where I assume the full ~50k GPU capacity contributes for the year.
Existing business: 800–900 cr Sharon: ~3,600 cr Own GPU capacity: ~1,100–1,300 cr Other GPU capacity: ~17,000–17,200 cr
Revenue: ~22,500–23,000 cr PAT: ~3,500–4,650 cr
I stop the calculation at 50k GPUs.
No 60k, 75k or 100k assumption.
No additional large AI deal is counted.
FY27 is about proving Sharon commissioning and billing.
FY28 is about scaling towards 50k GPUs while I count only ~15k average revenue-generating GPUs.
FY29 is the first year where the full economics of ~50k GPUs can become visible.
The numbers look extraordinary, which is exactly why execution matters more than projections from here.
#ESDS #AI https://x.com/mezza9_Equity/status/2104247922573652362
## @Microinteracti1 (Gandalv) · 09-27 15:54 · ♥32 ↻7 💬1 Oracle's five-year credit default swaps hit roughly 222 basis points on Thursday, the highest in a series stretching back to 2008, and its 6.7 percent bond due 2056 crossed 8 percent yield for the first time.
The trigger, according to Bloomberg, was a force majeure notice Oracle sent to Blue Owl over Project Jupiter, the $165 billion Stargate campus in New Mexico whose gas pipeline has slipped to 2027 after repeated permit denials. Oracle wants up to three years' relief on rent if the site misses 2028. The project's own debt now trades below 90 cents on the dollar.
None of this is a collapsing business. Revenue is growing 30 percent and the contracted backlog stands at $664 billion. The trouble is sequencing: $90 to 95 billion of capex this year, negative free cash flow, and a BBB- rating one notch above junk while 30-year Treasuries yield above 5.4 percent.
Nor is Oracle alone. Nvidia, Alphabet and SpaceX all set CDS records this month, AI-linked bonds trade 115 basis points over Treasuries against 78 for the wider market, and hyperscalers are expected to borrow $420 billion next year. Equity markets price the dream. Bond markets ask who pays the interest. https://x.com/Microinteracti1/status/2104238143176065168