Insights into the Cloud Business Model

Cloud computing is often related to a lease vs buy decision of an asset like real estate. There are scenarios in which you would lease a hotel room, take out an annual lease for an apartment, or buy a house. If you only need it for a short time, then you don’t mind paying for a likely higher daily rate for a room. If you are going to be using that room for 30 years, then owning it likely makes sense.

Because cloud computing can be leased on demand, it’s closest analogy is the hotel room. I need extra capacity for the holiday season? Great, let’s lease some servers for that period.

Unlike real estate, cloud providers have economies of scale that enable them to squeeze out every efficiency to make the price so cheap that it’s hard to justify owning and operating your own server infrastructure.

Thus, cloud computing provides you with both price and flexibility. While true in comparison to on-premise infrastructure, the analogous buy decision, the standard cloud computing product is still priced for flexibility. If you don’t need the flexibility, you are likely overpaying for cloud resources.

Cost of on-demand availability

As I mentioned in my previous post about planning Microsoft’s cloud, one of the core scale challenges was handling demand uncertainty when dealing with long-lead time physical assets. Across all the cloud assets, providers are spending over $20B annually. This seems astronomically high, however it’s likely conservative when you start looking at the numbers.

Let’s consider the typical datacenter we planned. They cost around $500M as they needed to be massive to achieve the scale efficiencies. Given the size of the construction project, let’s say these datacenters took 2 years to build on average. Which means I need to have enough inventory in the pipeline to cover the next 2 years worth of demand. It was pretty common to fill up these datacenters in 6 months in our major regions. So in a very rudimentary analysis, I would need to position $2B worth of datacenters in a single region to anticipate demand. Blend in the other assets such as servers, fiber, land, water, and power, multiplied by the number of regions you operate and you get a pretty large bill.

Now this estimate would be true if the demand was perfectly predictable. Let’s factor in forecast uncertainty. The eventual answer for the example above may be somewhere between $1.5B (lower than expected demand) or $3B (higher than expected demand). Furthermore, the industry’s performance on multi-year forecasts have had higher error than this example range.

While the additional spend to cover uncertainty may seem tough to swallow, not having cloud capacity for your customers can be a catastrophic on the business. It’s common for customers to go to your competitors.

A huge area of exploration we did for years was finding ways to reduce the safety stock. In theory, the scale that we operated should allow us to use the large portfolio to dampen demand volatility.

Learnings about cloud demand

Turns out that there are a variety of demand artifacts that prevent us from pooling uncertainty given the broad portfolio we operated.

1. Demand is local. Bandwidth costs prevent load balancing workloads between regions. Latency requirements often required that different cloud components sit together in the same region. Furthermore, data sovereignty and privacy created hard requirements for customers to stay within certain geographies. While the cloud is a global portfolio, it operates on a local level.

2. Demand is seasonal. Suppose each customer’s demand volatility was independent and random. Law of large numbers would suggest that as you add more samples, it would start to converge towards the average. However, the underlying demand drivers create seasonal elements across many customers. An easy one to point out is day vs night. You need to have enough capacity to support daytime peak hours, despite there being extra resources during the evening.

3. Demand is relatively price inelastic. We experimented with different pricing across regions to see how much it could influence future demand. Specifically, some of our highest demand regions were considerably marked up to incentivize choosing more available regions. It wasn’t a strong motivator to shape demand as cloud requirements mattered more than price.

4. Business model encourages volatility. When you give people the flexibility to spin-up and down on demand, it creates a lot of jitter in the data. Furthermore, we’re seeing more cost control companies pop up that furthers this behavior.

To reduce uncertainty, we met with countless enterprise customers to have a conversation around their future needs. The thinking was that we can improve our forecast with additional data points from our customers. For some accounts this work well, but the overall effectiveness was mediocre as there wasn’t an incentive for customers to provide us high quality data.

Commitments as the economic solution

Commitments aren’t a new concept. In the Software-as-a-Service (SaaS) world, it’s pretty common to offer discounts for annual and multi-year commitments. The difference is the economic value and the size of the discount. For Infrastructure-as-a-Service (IaaS), the lowest margin cloud product, you’ll see discounts upwards to 70% for a three year commitment. Coincidentally, the depreciation life of a server is also 3 years. Basically, you would buy the usable life of the server for 30% of what it would cost you if you ended up using it for 3 years at the no-commitment rate.

Furthermore, as the cloud grows, the uncertainty problem for the cloud grows with it. With cloud cost control companies exacerbating volatility, we’ll see further bifurcation of on-demand cloud products and commitment-priced products. There is a strong economic pressure on the cloud provider to add more predictability to its supply chain.

Given the business mechanics of the cloud, it is highly likely many companies will have lackluster long-term performance of cloud cost control initiatives as it assumes the status quo on cloud pricing. When you look at the broader ecosystem, enterprises have an opportunity to create market value through cloud commitments, stabilizing the cloud supply chain.