AWS is Not a Dumb Pipe

lobste.rs - Fri Jan 14 18:17

In 1997, David Isenberg wrote an article called the Rise of the Stupid Network. Isenberg worked at Bell Labs, which was the R&D lab of AT&T and responsible for some significant breakthroughs during the PC and the early internet era. The telephone companies were struggling with the idea of turning their "dumb pipe" into an "intelligent network" by building over-the-top services on top of calls.

In the article, Isenberg makes a prediction that the internet is going to dis-intermediate the telephone network. Telcos could save themselves by 1. disrupting themselves (unlikely) or 2. reinvent their network value proposition as a "dumb pipe". Later analyses (which ended up being true) showed that the broadband providers could actually make more money if they stopped competing on higher level services and opened their network up.

The same question is often posed against the cloud and the cloud providers: will AWS and Azure become "dumb pipes" of compute, storage, and other fundamental data center primitives? Or will they continue to compete and beat third party providers at high level services?

Fast forward to 2021, when Erik Bernhardsson wrote an article called the Storm in the stratosphere: how the cloud will be reshuffled where he made a prediction about the cloud vendors

Cloud vendors will increasingly focus on the lowest layers in the stack: basically leasing capacity in their data centers through an API.

I hope Bernhardsson is right, but I think it's optimistic thinking. I don't think that the hyperscaler clouds become "dumb pipes" like the AT&T and telcos before them.

Here's  why I think why it's different this time:

  • When there's no product differentiation, distribution wins. In the telco case, products build on top of broadband were 10x improvements over products built into broadband. YouTube, Netflix and other internet content providers could do things that cable and telephone simply couldn't. Now, competitors must differentiate on expertise, community, and developer experience.
  • Hyperscalers have already proven they can move up the stack. Yes, AWS missed the $90B Snowflake opportunity, but won many others. Kinesis vs. Kafka, DocumentDB vs. MongoDB, MemoryDB vs. Redis, OpenSearch vs. ElasticSearch. And it's not just fast-follow, but fundamental new services like AWS Lambda have changed the way we develop software. Google App Engine was a decade before its time.
  • Amazon, Microsoft, and Google get software. Telcos were dealing with a fundamentally different beast. The folks at Bell Labs might have gotten it, but the rest of the company wasn't run by internet-natives. Engineers at AWS and Google Cloud deeply understand the strategy and future of some of these technologies.
  • Businesses don't want more vendors. This plays into the point about distribution, but vendor sprawl is real. Is an in-memory key/value store worth signing a new contract and going through the procurement process?
  • Incentives are aligned between open source and hyperscalers. Open source is the solution to vendor lock-in (not really, but its good enough for most). Startups use open source to build community and get superior distribution over hyperscalers. Companies monetizing open source often need to have proprietary add ons to capture value (services revenue is never enough). Hyperscalers can let development be community driven and fully open source, either becoming a "dumb pipe" or actually capturing more margin or lock-in value (see OpenSearch vs. Elasticsearch).

But there's always another side to it. I'll give a few reasons why it just might work.

  • Cloud infrastructure has better economics than telco infrastructure. Becoming a "dumb pipe" is potentially a much better tradeoff for hyperscalers than it was for telcos. On the other hand, that means that services built over-the-top pay a higher cloud tax. Not sure how this one plays out.
  • There's just too much opportunity. Too many services need to be built, and the customers are demanding them sooner than the hyperscalers can develop in their large, slow-moving organizations. Some of these services will have hidden vendor lock-in or attributes that make them competitive (e.g., counterpositioning like multi-cloud).
  • Bernhardsson makes a good observation that the incentives drive some of the best engineers to join well-funded startups or to go off and start their own instead of continuing to work for an AWS. The amount of money in the private markets also make more startups like Snowflake possible.

The telcos didn't go down without a fight. They successfully got so many regulations passed against VoIP that it served a serious barrier to entry for more than a decade. The hyperscalers have an even better card to play than regulation: open source. By bringing the cost of software down to zero, they can commoditize their complement. If AWS open sourced all higher-level services, they would still be a "dumb pipe", but with fewer competitors.