**The most surprising finding is that the funding totals are dominated by a handful of deals, and the largest one looks like a data error rather than a real deal.** - **Concentration:** 2,065 deals have a usable amount, and together they total about **$38.06B**. The median deal is only **$1.7M**, yet the **73 deals of $100M or more make up 63.6% of all dollars** (steps 1 and 10). Fewer than 4% of the valued deals therefore account for almost two-thirds of the money. - **The top deal:** the largest single record is **Rapido Bike Taxi, Series B, 27/08/2019, $3.9B**. That one row is about **10.2% of all funding dollars** in the dataset (steps 0 and 1). It is bigger than Flipkart's $2.5B round in 2017 and Paytm's $1.4B round, which fits the known Flipkart and Paytm mega-rounds. A $3.9B Series B for a bike-taxi startup is very implausible. I can't confirm from this dataset that it is a unit or typo error, but it is a strong candidate. If it is wrong, it inflates every total, average and trend. - **Missing amounts:** **978 of 3,044 rows (32%)** have no usable amount (steps 7 and 10). Any "total funding" figure covers only about two-thirds of the deals. - **A poisoned row:** the **Drums Food** deal (21/07/2016) has an amount that parses to a non-finite number (`nan`). One row like this turns every SUM and AVG into `nan` unless it is filtered out. This is why several of my earlier aggregate attempts returned `nan` totals (steps 2 to 9). **Takeaway:** the typical Indian startup deal is small, about $1.7M, but the totals are shaped by a few extreme rows and by one suspicious record. Any ranking by city, sector or year based on summed dollars needs those outliers, the missing amounts and the corrupt row handled first. Medians or counts are safer. **Evidence:** SQL results in steps 0, 1, 8 and 10 against `raw.startup_funding`, parsing `Amount in USD` from text and excluding the non-finite row.