Thursday, February 21, 2013

Kai Li's FAST Keynote

Kai Li's keynote at the FAST 2013 conference was entitled Disruptive Innovation: Data Domain Experience. Data Domain was the pioneer of deduplication for backups. I was one of the people Sutter Hill asked to look at Data Domain when they were considering a B-round investment in 2003. I was very impressed, not just with their technology, but more with the way it was packaged as an appliance so that it was very easy to sell. The elevator pitch was "It is a box. You plug it into your network. Backups work better."

I loved Kai's talk. Not just because I had a small investment in the B round, so he made me money, but more because just about everything he said matched experiences I had at Sun or nVIDIA. Below the fold I discuss some of the details.

Kai quoted Dr. Geoffrey Nicholson:
Research is the transformation of money into knowledge.
Innovation is the transformation of knowledge into money.
Data Domain is a stellar example of the second. The outlines of the story are simple. They started in late 2001, raised a total of about $41M in 3 rounds, IPO-ed less than 6 years later at a $1B valuation having spent only $27M, and were acquired 2 years after that at a $2.4B valuation. Before their IPO they had more than 60% of the market and more than 70% gross margin. That is an extraordinary performance.

The vision was to replace tape for backup with disk at roughly the same price but much lower space, power, and network costs, and to make restoring from a backup much faster, thereby reducing the operational impact of failures.  The only way to do this was to use deduplication to get a high enough compression factor to swamp the cost per byte difference between disk and tape. Kai illustrated their success by showing a line of 17 full racks each containing an IBM tape library that were replaced by 3 3U Data Domain systems.

The key to implementing this vision was to bet on long-term technology trends. The two that Kai pointed out were that disk had already replaced tape in personal audio (Walkman to iPod) and in TV time-shifting (VHS to Tivo), and that Moore's Law had already shifted from faster CPUs to more cores.

The two major challenges they faced were:
  • They had to sell for no more than a tape system, so their gross margin was directly related to the compression ratio they could achieve.
  • The amount of data to be backed up was doubling every 18 months, but there are only 24 hours in a day, so their throughput needed to at least double every 18 months.
The three founders started the company just after 9/11, at a time when no-one was starting companies. We started nVIDIA in 1993 in one of Silicon Valley's periodic downturns; we were the only semiconductor company to get any funding the quarter of our A round. Starting a company when no-one else is - absolutely the best time to do it.

Kai laid out a list of key precepts, all of which I agree with despite some caveats:
  • Build "must have" products.
  • Customer driven technology. For nVIDIA, this was more difficult, since we had customers (PC and board manufacturers) and end-users (game players).
  • Work with the best VC funds.  The difference between the best and the merely good in VCs is at least as big as the difference between the best and the merely good programmers. At nVIDIA we had two of the very best, Sutter Hill and Sequoia. The result is that, like Kai but unlike many entrepreneurs, we think VCs are enormously helpful.
  • Raise more than we need - give up a lot of equity. One downside of starting a company when the market has gone south is that you have to give up more equity. But you are giving it up to VCs willing to invest when no-one else is, who are the ones you want to work with. And having cash in a downturn gives you the ability to move fast.
  • High standard for early team, even if you miss the hiring plan. After the IPO at Sun came the bozo invasion, but then the company was well-enough established to survive it.
  • No egos - take the best ideas wherever they come from. This is often hard for the best people to handle, and is a real test of the initial management.
Kai's slides plotting revenue and things like lines of code and deduplication throughput on the same  graph were fascinating. They matched closely, with throughput increasing 100-fold in 6 years, and lines of code growing much faster after the IPO than before.