# Papparti — GPU: Rent or Buy? A decision document

**Prepared by:** Vector Logic · **Date:** 11 October 2026 · **Status:** decision document (round 0)
**Reads first, does not re-litigate:** `CLIENT-ESTIMATE.md`, `MODEL-DESIGN.md`.
**Question:** should the client **rent** the 48 GB GPU monthly (the estimate's assumption) or **buy** it?

Every number below carries a source URL and the date it was read, or is labelled **ASSUMPTION** with its
basis. All USD is converted at the estimate's pinned rate, **1 USD = ₹96.7384** [FX] — re-verified live
today at 96.7372 (exa.ai, 2026-10-11), so **the estimate's rate is correct to 0.001%** and is used unchanged.
Indian figures use Indian grouping (₹78,72,318, not ₹7,872,318).

---

## 0. The one-paragraph answer

**Rent. Keep renting.** At the client's actual expected utilisation the rented card is the cheaper and the
lower-risk choice for at least the first **~20 months**, and the client will not know its own utilisation
until Phase 0B has measured it. **The estimate's ₹41,665/month is correct and should NOT change.** The only
honest adjustment is to *state the flip condition* in the proposal: if measured utilisation holds above
**~60%** of a dedicated card for more than **~20 months**, buying a used 48 GB card and colocating it in
India becomes cheaper than renting — but that is a decision to revisit **after** Phase 0B produces the
utilisation number, never before. Unverified-by-design: a bought card commits capital before there is a
single paying customer, and no marketplace has ever been lost by renting its first GPU. **(These are
ASSUMPTIONS about the client's stage, flagged as such, and they are the crux of the recommendation.)**

---

## 1. The break-even calculation, shown

### 1.1 The two cost functions

- **Rented** (billed hourly, 730 h/month, busy or idle):
  `C_rent(n) = 41,665 · n`   ← linear, no capex.
- **Bought** (one capex, then monthly opex that is NOT zero — a bought card needs a home):
  `C_buy(n) = CAPEX + OPEX · n`   ← a step, then a shallow slope.

Break-even is where they cross:

```
CAPEX + OPEX·n = RENT·n
        CAPEX   = (RENT − OPEX)·n
             n* = CAPEX / (RENT − OPEX)          months
```

`n*` is the month in which buying's cumulative cost overtakes renting's. Below `n*` renting is cheaper;
above it, buying is.

### 1.2 The three buy options (each sourced in §2)

| Option | GPU | Host | CAPEX (INR) | Monthly OPEX (INR) | Basis |
|---|---|---|---:|---:|---|
| **A — New card, new server** | New RTX 6000 Ada 48 GB @ $8,400 | New 2U GPU server @ ~$13,000 | **20,70,202** | 7,716 | apex build; lowest risk, highest cost |
| **B — Used card, India OEM server** | Used RTX A6000 48 GB @ $3,999 | India entry GPU 2U @ ~₹6,00,000 | **9,86,828** | 12,716 | mid; INR invoice, GST credit, onsite SLA |
| **C — Used card, refurb server** | Used RTX A6000 48 GB @ $3,999 | Refurb 2U server @ ~$3,500 | **7,25,441** | 7,716 | cheapest realistic; most execution risk |

OPEX = colocation plan + metered power for the node. Per §2d, a 2U colo slot with a power allowance is
**₹5,000/month** (VyomCloud) and a dedicated-power 2U is **₹10,000/month**; per §2e a GPU node drawing
~600 W (300 W card + host) uses **438 kWh/month**, which at Odisha LT-industrial **₹6.20/kWh** is
**₹2,716/month**. Option B uses the ₹10,000 colo (power billed inside the plan allowance); A and C use
₹5,000 + metered ₹2,716. **ASSUMPTION:** node draw 600 W — the card's 300 W TDP [spec] plus a ~300 W
host; to be confirmed by a power meter in Phase 0B.

### 1.3 The arithmetic

```
Option A:  20,70,202 / (41,665 − 7,716)  = 20,70,202 / 33,949  = 61.0 months
Option B:   9,86,828 / (41,665 − 12,716) =  9,86,828 / 28,949  = 34.1 months
Option C:   7,25,441 / (41,665 − 7,716)  =  7,25,441 / 33,949  = 21.4 months
```

Sanity check by the month (Option C):

| Month | Rented cumulative | Bought cumulative | Cheaper |
|---:|---:|---:|---|
| 12 | 4,99,980 | 8,18,033 | **rent by ₹3,18,053** |
| 20 | 8,33,300 | 8,79,761 | rent by ₹46,461 |
| **21** | 8,74,965 | 8,87,477 | **rent by ₹12,512** |
| **22** | 9,16,630 | 8,95,193 | **buy by ₹21,437** ← crossover |
| 36 | 14,99,940 | 10,03,217 | buy by ₹4,96,723 |

**Break-even: ~20–21 months (Option C), ~34 months (Option B), ~61 months (Option A).**

**36-month total, all-in:** rent **₹14,99,940** · buy C **₹10,03,217** · buy B **₹14,44,676** ·
buy A **₹23,48,050**. Even over three years, the *new-card* build costs **₹8.48 lakh more** than renting.

---

## 2. Live buy-side prices (each with source + read date)

All read **11 October 2026** unless the source states its own check date, which is given.

### 2a. New RTX 6000 Ada / A6000-class 48 GB card

| Item | Price | Source | Read |
|---|---|---|---|
| RTX 6000 Ada 48 GB (new, retail) | **$8,400** | https://aigearwatch.com/hardware/nvidia-rtx-6000-ada/ ("recorded catalog price $8,400", checked 2026-10-10) | 2026-10-11 |
| RTX 6000 Ada 48 GB (lowest avg) | $7,715 (range $5,333–$9,070) | https://gpupoet.com/gpu/learn/price/october-2026/nvidia-rtx-6000-ada-generation | 2026-10-11 |
| RTX 6000 Ada 48 GB (retail, in stock) | $8,299 | https://computizer.com/products/nvidia-rtx-6000-ada-generation-48gb-gddr6-workstations-graphics-card | 2026-10-11 |
| RTX A6000 48 GB (brand-new sealed) | **$3,899.95** | eBay listing 186691900885 (https://ebay.com/p/27046130288) — new stock still exists via marketplace | 2026-10-11 |

**Modelling choice:** $8,400 (the most conservative, in-stock retail new-card price) for Option A.

### 2b. Used 48 GB / 24 GB class card

| Item | Price | Source | Read |
|---|---|---|---|
| **Used RTX A6000 48 GB** | **$3,999** (going rate $4,349; median ask $5,183) | https://rigprice.com/gpu/rtx-a6000 and https://vramglass.com/gpu/rtx-a6000 | 2026-10-11 / undated |
| Used RTX A6000 48 GB (eBay pre-owned band) | $3,824 / $3,899 / $3,999 | https://www.ebay.com/shop/rtx-a6000-48gb | 2026-10-11 |
| Used RTX A6000 48 GB (ex-render-farm) | $3,500 typical | https://theaibench.ai/hardware/rtx-a6000-used | 2026-07 (Jul 2026) |
| **Used RTX 3090 24 GB** | **$1,274–$1,450** (fair range $1,200–$1,300; RigPrice going rate $1,450) | https://resaleprices.com/gpu/nvidia-rtx-3090 (Aug 22 2026) and https://rigprice.com/gpu/rtx-3090 | 2026-10-11 |
| Used RTX 3090 24 GB — **superseded** | $650–750 | https://bestgpuforai.com/articles/best-used-gpu-for-ai | **STALE / CONTRADICTED** |

**Do not use the 24 GB cards for this project's single-node plan.** Two RTX 3090s (2 × 24 GB) would give
48 GB but no NVLink-supported 48 GB tensor-parallel layout is guaranteed to hold the ~20 GB Q4 reasoning
model *plus* STT *plus* TTS *plus* vLLM KV-cache headroom at the 10-session floor; `MODEL-DESIGN.md` §4.4
gates on **sessions/node < 10 breaking the cost model**. A single 48 GB card is the safe floor, so the used
48 GB A6000 at ~$4,000 is the realistic buy.[^ramageddon]

[^ramageddon]: The RTX 3090 resale price nearly **doubled** during 2026 (from ~$650 in March to ~$1,300–1,450
by September) on an AI-demand memory crunch (https://gamesreviews.com/news/09/ramageddon-turns-old-rtx-3090-gpus-into-1300-resale-gold-mines,
2026-09-04). **The used market is itself volatile — see the case against buying, §3.3.**

### 2c. Server / workstation to host it

| Item | Price | Source | Read |
|---|---|---|---|
| Puget 2U GPU server, "starting at" | **$12,415.89** (config $16,598.74) | https://www.pugetsystems.com/products/rackmount-servers/2u/ | 2026-10-11 (page modified 2026-07-30) |
| Puget 2U, lower tiers starting | $9,568–$16,475 | same | 2026-10-11 |
| Gigabyte G293-S42-AAP1 2U 8-GPU barebone | $9,414 (as-built $16,640) | https://www.coltec.com/en-us/servers/ai-line-ai-e208n-2u-dual-xeon-server-8x-gpu-slots | 2026-10-11 |
| India GPU server, entry 1-GPU | **~₹5,00,000–₹25,00,000** (indicative band) | https://proactive.co.in/blog-details/what-a-gpu-server-really-costs | 2026-10-11 |
| India OEM 2U GPU server (RDP Technologies, Hyderabad) | **"Pricing on request"** — not quotable | https://rdp.in/gpu-mart/product/quasar-2x-rtx-pro-6000-blackwell-gpu-server/ | 2026-10-11 |

**Modelling choice:** Option A uses $13,000 (mid Puget 2U config); Option B uses ₹6,00,000 (bottom of the
India entry band — GST invoiced, onsite SLA); Option C uses $3,500 (a used/refurb 2U node, **ASSUMPTION**
since no refurb 2U listing was found with a live price — see §5).

### 2d. Colocation for a self-owned card in India — the piece people forget

| Plan | Price | Power included | Source | Read |
|---|---|---|---|---|
| 1U colo (Tier III, India) | **₹3,999/mo** | 350 W A+B | https://xenaxcloud.com/blog/blog/full-rack-colocation-pricing/ | updated 2026-09-02 |
| 2U colo, VyomCloud (Tier III) | **₹5,000/mo** | 1 A @ 230 V ≈ 230 W | https://vyomcloud.com/1u-rack | 2026-10-11 |
| 6U colo, VyomCloud | ₹10,000/mo | 2 A @ 230 V ≈ 460 W | same | 2026-10-11 |
| 12U colo, VyomCloud | ₹20,000/mo | 4 A A+B ≈ 920 W | same | 2026-10-11 |
| India colo, Mumbai (per kW) | **₹7,500–9,500 /kW/mo** | — | https://sovereignglobalcompute.com/india (week of 2026-07-13) | 2026-10-11 |
| India colo, Chennai (per kW) | ₹6,800–8,800 /kW/mo | — | same | 2026-10-11 |
| **Caution** | "₹5,000 2U" gives only ~230 W; a GPU node needs **≥690 W** → the client must buy **≥4 A / a 6U tier**, or the ₹10,000 metered-power tier | — | derived from VyomCloud amperage | 2026-10-11 |

**Modelling choice:** ₹5,000 (with metered power billed separately) and ₹10,000 (power inside allowance) are
both shown; the price is small either way relative to the ₹41,665 rent, so it does not move the break-even
much — which is the point: **colo is not the deciding cost; capex is.**

### 2e. Electricity — Odisha

| Item | Value | Source | Read |
|---|---|---|---|
| **Odisha LT-Industrial energy charge** | **₹6.20/kWh** (flat) | OERC RST FY 2026-27, effective 1 Apr 2026, https://www.orierc.org/CuteSoft_Client/writereaddata/upload/DISCOMs_Tariff_Notification_FY_2026-27.PDF | 2026-10-11 |
| Confirm (industrial LT ≈ ₹6.20/kWh; tariffs unchanged 5th year) | ₹6.20/kWh | https://mercomindia.com/odisha-retains-last-years-power-tariffs-for-fy-2027 and https://mercomindia.com/odisha-regulator-retains-existing-power-tariffs-for-fy-2026 | 2026-10-11 |
| TPNODL/TPSODL LT-Industrial (cross-check) | ₹6.00/kWh + ₹60/kW/mo fixed | https://electricbilltool.com/tpnodl-unit-rate-calculator | 2026-10-11 |
| Odisha electricity duty | **0%** | same | 2026-10-11 |

**The 219 kWh/month figure in the brief, verified:** a **300 W card alone** at 100% for 730 h/month draws
`0.300 kW × 730 h = 219.0 kWh/month` → **₹1,358/month** at ₹6.20/kWh. A real node also runs a host CPU and
fans, so the **600 W node** figure (**438 kWh → ₹2,716/month**) is the honest planning number for
"electricity for the bought server," at **1.3–2.6% of the monthly rent it replaces.**

---

## 3. The case AGAINST buying — stated as hard as the case for it

### 3.1 The capital is committed before there is one paying customer
Papparti is pre-launch. A ₹7.25–20.70 lakh capex on 11 October 2026 buys a card that sits idle until
customers arrive, while a rented card costs ₹41,665 **only in the months the client chooses to run it**.
The client can pause a rented instance the week launch slips; it cannot un-buy a GPU. **The rent/capex
saving only exists if the workload actually runs for 20+ months — and the client has no bookings yet.**
This is the single strongest argument against buying, and nothing in §1 counters it.

### 3.2 The technology-obsolescence risk is real and dated
The RTX A6000 is **Ampere, launched 2020** [rigprice]. Buying it in 2026 means buying six-year-old silicon.
The **RTX PRO 6000 Blackwell (96 GB)** is already the current card at $0.65/GPU-hr on the rental market
(getdeploying, 2026-10-01). A 2026-bought 48 GB card is worth materially less in 2028, and the client would
be locked to 48 GB while the model floor moves. **A rented card can be re-pointed at a newer GPU in minutes;
a bought one is a stranded asset the day Blackwell-class weights need more than 48 GB.**

### 3.3 The bought card's value can *fall faster* than the amortisation assumes — proven this year
The used RTX 3090 **doubled** in 2026 (AI memory crunch, §2b) — which sounds good for resale, but the same
volatile market can run the other way, and the RTX A6000's own tracked range is **$3,900–$6,000** (vramglass).
Amortising a used card over 36 months assumes a resale floor that the market does not guarantee.

### 3.4 The client is a marketplace company, not an infrastructure company
Owning a GPU means owning a **hardware failure at 2 am** — a dead card, a PSU, a thermal event in a colo the
client has never visited. With RunPod the failure is *their* pager and a replacement is a redeploy. The
support retainer (₹67,717/month) buys software support; it does not make Vector Logic an on-call hardware
depot. **Renting converts a hardware-operations problem into a line item.**

### 3.5 The bought card that is idle still cost the full price
The rented card's "waste at low utilisation" is *visible and bounded* (₹41,665/month). The bought card's
waste is *invisible*: ₹7–20 lakh already spent, depreciating, whether the reasoning model serves one user
or one thousand. **Sunk cost is harder to manage than a subscription.**

### 3.6 It adds a compliance and logistics surface
A self-owned card in an Indian colo brings GST input-credit mechanics, an import/ownership trail on the
hardware, a physical asset on the balance sheet, and insurance/maintenance to manage. A rented instance has
none of that. (The estimate already handles GST reverse-charge for *imported rented* services, §7c — buying
moves that complexity into the client's fixed assets instead of removing it.)

**Balanced against this, the honest case FOR buying:** over 36 months, Option C saves **₹4.97 lakh** vs
renting; the capex is small against the ₹78.72 lakh one-time build; the used-card price is currently
*attractive*; and an owned card removes per-hour billing and gives the client a physical asset with resale
value. **None of these fire before the 20-month break-even, and none of them fire before the client has
customers.**

---

## 4. Recommendation, with the conditions attached

> **RECOMMEND: RENT. Leave `CLIENT-ESTIMATE.md` at ₹41,665/month unchanged.**

**The conditions that flip the answer — quote these in the proposal:**

1. **Expected life below ~20 months → always rent.** Above **~20 months** of *continuous need*, buying a
   **used 48 GB card (Option C)** becomes cheaper. Above **~34 months**, even the India-OEM build (B) wins.
   The new-card build (A) needs **~61 months** — effectively never for this project.
2. **Buy only if measured utilisation exceeds ~60% of a dedicated card** (≈438 of 730 h/month actually
   driving the GPU) **AND** that level is held for **20+ months**. Below 60%, the amortised cost per
   *used* GPU-hour (§5) stays above the rental and buying is irrational.
3. **Buy only if one of the three models must run 24/7 with no cloud dependency** (data-sovereignty /
   DPDP constraint). The reasoning model running 24/7 is *a* reason, but it is a reason to rent a
   *dedicated* 24/7 instance, not to buy — the rental is already 24/7 dedicated.
4. **Never buy before Phase 0B reports.** `MODEL-DESIGN.md` §4.4 makes **sessions/node** and the
   **Hinglish-WER gate** hard stop/go conditions; if the stack fails its gate, a bought card is a
   stranded asset on a product that changed shape. **Buying is a post-Phase-0B decision, at the earliest.**
5. **If the client insists on owning capacity, buy Option C, not A** — the used 48 GB card at ~$4,000 is
   ~½ the price of new for the same 48 GB, and the new-card premium buys a brake the break-even never reaches.

---

## 5. The sensitivity that matters most — utilisation

The rented card is billed **730 h/month whether busy or idle**. The bought card's *effective* cost per hour
**falls** as utilisation rises. This is the whole argument, and it only cuts one way above a threshold.

Cost per **utilized** GPU-hour — Option C (₹7,25,441 capex amortised over 36 months (₹20,151/mo) + ₹7,716 opex =
**₹27,867/month** fixed) vs rent (₹41,665/month):

| Avg utilisation | Used h/month | **Rent ₹ /used-h** | **Buy ₹ /used-h** | Cheaper per used-hour |
|---:|---:|---:|---:|---|
| **10%** | 73 | **₹571** | **₹382** | buy, but *both are terrible*; rent the smallest possible card, or rent on-demand by the minute |
| **30%** | 219 | ₹190 | ₹127 | buy — **but** only if committed 20+ months |
| **50%** | 365 | ₹114 | ₹76 | buy — clearly |
| **100%** | 730 | **₹57** | **₹38** | buy — by 33% |

**Read the table carefully: at every utilisation the bought card has a lower cost-per-used-hour — because
the rental price ($0.59/hr) is set *above* the amortised owned cost ($38–382/hr-equivalent) across the whole
range.** Utilisation does **not** decide rent-vs-buy on a *per-hour* basis. What decides it is the **fixed
commitment over time**, which is §1: **20+ months**. The 10% row is the trap — a bought card at 10%
utilisation still burns the full ₹27,869/month forever, whereas a rented card at 10% can be *switched to a
cheaper plan or stopped entirely*, which is the option ownership deletes.

**Which case does Papparti fall into? Read `MODEL-DESIGN.md`:**
- The **reasoning LLM runs 24/7** — that is a genuine **~100%-of-the-card baseline** *for that one model*,
  and it is the strongest single fact for owning. (MODEL-DESIGN §0, §3.2: the LLM is "the heavy one".)
- But the **STT and TTS models are bursty** (§1.2–1.3: speech traffic peaks at booking time, idles
  between), and `MODEL-DESIGN.md` §4.4 gates the whole cost model on **sessions/node ≥ 10** — a number
  that is **explicitly ESTIMATED and unmeasured** ("~14 sessions/node (ESTIMATED)"; "a 3× error here
  triples the bill"). The concurrency load test (§4.5) exists precisely because the real number is unknown.
- Therefore the client's **actual average utilisation of the card is UNKNOWN until Phase 0B measures it**,
  and the estimate's own §10 says the concurrency capacity of a single 48 GB card under load is
  **UNVERIFIED**. **Planning a 20-month capex commitment on an unmeasured utilisation number is exactly
  the mistake this document exists to prevent.**

**Plain statement:** on the *known* fact (reasoning model 24/7) the card is heavily used and buying *would*
pay off over 20+ months. On the *unknown* facts (real concurrency, whether the stack even passes Phase 0B)
it is unsafe to commit capex now. **The uncertainty itself is the reason to rent until Phase 0B removes it.**

---

## 6. What this means for the proposal

**The recommendation equals the current assumption. Nothing in the money changes.**

- **`CLIENT-ESTIMATE.md` line that stays as-is:** §1b / §5, the monthly row
  `| GPU serving — 1× 48 GB dedicated | 41,665 | RunPod RTX A6000 @ $0.59/hr × 730 hr [1] |`.
  **It does not change. ₹41,665/month is confirmed correct.**
- **The one-time total is NOT affected** — this is a monthly item; the ₹78,72,318 / ₹92,89,335 totals are
  untouched. (The Phase 0 line's GPU sub-charge, ₹41,665, is likewise unchanged.)
- **Suggested addition (wording only, no arithmetic):** add one sentence under §5 or §9 noting that
  *renting is the deliberate choice, with a documented flip condition — if Phase 0B measures sustained
  utilisation above ~60% of a dedicated card for a planned life above ~20 months, a bought used 48 GB card
  in Indian colocation becomes cheaper and should be revisited at that point.* This converts an unexamined
  assumption into a **stated, testable decision rule** — which is worth more to the client than the rupees.

**If the client nonetheless elects to buy**, the line changes as follows (monthly):
`GPU serving` ₹41,665 → **~₹7,716** (colo ₹5,000 + power ₹2,716, Option C), i.e. a monthly saving of
**~₹33,949/month**; the monthly total ₹1,26,108 → **~₹92,161**; **but** the one-time total rises by the
**₹7,25,441 capex** (→ ₹85.98 lakh excl. GST) — which is the trade the client is being asked to accept.
**This is the flip the recommendation says NOT to take yet.**

---

## 7. Verification and what could not be verified

**Recomputed:** every break-even quotient in §1.3, the month-by-month table, the 36-month totals, the
per-used-hour table in §5, the 219 kWh and 438 kWh electricity figures, and the FX conversion. FX used:
**96.7384**, matching the estimate exactly; live cross-check 96.7372 (exa.ai, 2026-10-11).
All pass.

**Could NOT verify / UNVERIFIED:**
1. **No live price for a used/refurb 2U GPU server** — Option C's $3,500 host is an **ASSUMPTION**; no
   listing with a quotable price was found. Option B uses the bottom of the India *indicative band*, and
   RDP (the India OEM that sells exactly this) lists **"Pricing on request"** only.
2. **No quoted monthly colo price from a GPU-certified Indian facility.** VyomCloud (₹5,000/₹10,000) and
   Xenax (₹3,999) are Tier III **general** colo, not GPU-certified; Sify's DGX-ready GPU colo is
   **pay-per-use, price on request** (sifytechnologies.com, 2025-05-20). The GPU-colo real price is
   **UNVERIFIED** — and could be *higher* than the ₹5,000–10,000 assumed, which would *raise* the
   break-even and strengthen the rent recommendation.
3. **Node power draw (600 W)** is an **ASSUMPTION** (card TDP + host); a Phase 0B power meter would confirm it.
4. **The Odisha figure is the LT-Industrial energy charge (₹6.20/kWh).** A colo facility bills the client
   inside its plan (power included), so the ₹6.20 applies only to the on-premises variant; treating it as
   the colo power price is an **ASSUMPTION**. Confirmed source: OERC RST FY 2026-27 + two Mercom reports.
5. **The client's actual future utilisation is unknowable now** — this is the honest core of the document,
   not a gap to be papered over.
6. **Import duty / customs on a card bought from a US/China marketplace** was **not** modelled; landed cost
   could be ~18–30% above the $3,999–8,400 list once duty + freight + IGST are added. **Not priced — flag
   before any buy.**

---

## 8. Assumptions register

| # | Assumption | Basis |
|---|---|---|
| A1 | FX 1 USD = ₹96.7384 | estimate's pinned rate [FX]; re-verified 96.7372 on 2026-10-11 |
| A2 | Rent ₹41,665/mo = $0.59/hr × 730 × 96.7384 | `CLIENT-ESTIMATE.md` §2 [1] |
| A3 | Node draw 600 W → 438 kWh/mo | card 300 W TDP + ~300 W host — **ASSUMPTION** |
| A4 | Odisha LT-industrial ₹6.20/kWh, duty 0% | OERC RST FY 2026-27; Mercom 2026 |
| A5 | Amortisation life 36 months | **ASSUMPTION** for per-used-hour illustration only; break-even uses no life assumption |
| A6 | Used-card host $3,500 | **ASSUMPTION** — no live listing found |
| A7 | India entry server ₹6,00,000 | bottom of Proactive's indicative ₹5–25 lakh band |
| A8 | Landed cost = list price (no duty modelled) | **ASSUMPTION — understates buy cost; flagged in §7.6** |

[FX] https://www.xe.com/currencyconverter/convert/?Amount=1&From=USD&To=INR — XE USD/INR mid-market, 11 Oct 2026 (estimate's source); re-verified at exa.ai/library/markets/forex/USDINR, 2026-10-11T08:16Z (96.737212).