The cost of a public cloud depends on a combination of factors:
Once all those costs are known, it is possible to calculate the costs and profitability of a cloud computing company. The following paragraphs will detail the cost factors and assumptions of our business model.
The cost of a public cloud depends on the price of electricity, real estate, internet transit and loans. Those cost can vary a lot from one region of the world to another. The valus provided bellow are based on a real example in France in 2018.
Rapid.Space rented a space for one rack in a data centre close to Paris. Professional Internet transit (1 Gbps) is charged 270€ per month. It could be 40% cheaper if we purchased 10 Gbps). Electricity is charged at 0.13€ per kWh and includes all kinds of value added services based on redundant electrical infrastructure. It could be 30% cheaper (0.085€ per kWh) if it was purchased directly. We pay 483€ per month for the surface (4 m2 or less) where we can place our rack. This price includes real estate as well as value added services (cooling, conditioning, cameras, etc.). Again, this could be much cheaper if we were using a low cost data centre design based on "free cooling" technologies. We have also included in the basic cost factors the current value of interest rates as in France (about 3%).
In order to keep on building our business model, we need to evaluate the density and power usage of a single rack.
A rack contains 36 servers (it could be more, up to 48, at the expense of reliability). The average power usage for 36 serves is about 10 kW. A shared Internet bandwidth of 1 Gbps is sufficient for 36 servers. Servers in the rack are interconnected through a 10 Gbps local area network. Rack contains power supplies connected to tri-phase 380 V power. It also contains batteries that can protect against short outages (1 minute) which represent the vast majority of outages. Two racks can be interconnected so that one rack can acts as a redundant internet transit route for the other rack. Rack requires a facility that can handle a weight of 2 tons per m2.
Let us now study the performance and cost of servers present in a rack.
Each rack server provides 24 Xeon cores, 256 GB RAM, 4 TB SSD. The price of such configuration is about 2500€ per server (including the rack, switch, cables, power, etc.). We have added a line name "cost to import a server" because in certain countries, it can be difficult or costly to purchase low cost servers, either due to tariff or due to a requirement to knock down server into small parts and re-assemble them. This is not the case in our example in France where such costs are zero. We then have to consider a ratio of servers that are lost every year (2%) and a ratio of servers that need to be repaired (3%). For each server that has to be repaired, we consider an average cost of repair of 400€. These ratio are based on our experience in VIFIB with a significant added margin but we expect lower values in real world because OCP servers we purchase have already bees re-certified through strict industrial process.
The cost of managing servers is about 20€ per month for software (operation of SlapOS cloud) and 20€ per month for hardware (sending staff to reboot or re-install). This is again based on VIFIB's experience where 300 servers are managed by a team of two who also does software development.
Last, we consider a 3 years amortisation period. In reality amortisation is longer: servers can be used up to 7 years. But what people are ready to pay for a 7 years old server is much lower than for a recent server. We thus consider that 36 months is a reasonable approximation for a business model to cover the price devaluation of service over time.
The first step in the business model consists of purchasing servers. This leads to cash out at a rate of 2500€ per server.
If the business is so successful that 1000 new customers suddenly order the same month a server, the quantity of cash required to meet their demand is... 2.5 million euros. There is thus only two ways to implement a low cost cloud business model:
We suppose here that cash is being financed by a bank at typical interest rate of 3% with amortisation over 36 months. Based on the previous list of cost factors, operating a server during one month therefore translates into the following list of expenses:
A spreadsheet with details of this calculation can be downloaded here: NXD-Cloud.Business.Model. What is intersting in this table is that the combination of electricity, management software and hardware management costs about the same as the server itself.
Let us suppose that each server is sold online at 195€ per month and that demand is high enough for all servers to be rented in average 36 months at this price. We end up with the following profit analysis:
While the server is being reimbursed, profits represent 27€ (out of 195€ income) which is about 14% of income. Once loan has been reimbursed, profits jump to 52% of income. Price competition can be surpassed by extending the operation of servers beyond the initial amortisation period of 36 months.
If we now consider the business model from a cash point of view, once the server has been paid (2500€), every month of operation generates a net cash income of 97€ after all expenses have been paid.
Initial cash out could thus be paid back in about 26 months. Over the last 6 months of the amortisation period, a total cash out of 980€ is generated. The business model is thus very similar to fashion business:
In our business model, we keep the same price (195€) all over the 36 months. A more sophisticated business model involves price evolution (ex. 145€ after 24 months) and a longer period of operation (ex. 48 months or even longer).
Based on the above figures calculated for one server, we have created a simplified business plan. Supposing that 20 customers subscribe during the first quarter and that with each quarter, 100 new customers subscribe, a total of 6840 customers - and servers - is reached after 3 years.
The total number of clients represents the value at the end of the year. The cash out is based on the number of servers that had to be purchased to server customers. Income is the total of monthly subscriptions at 195€ per server per month.
Details of this calculation can be downloaded here: NXD-Cloud.Business.Model. This spreadsheet shows in particular that, thanks to net cash generated on existing customers, capital requirements for new customers can be reduced by half after one year and by 75% after 3 years.
If we consider a small data center of 200 racks (7,200 servers), the capital needed to fill it is about 7 million euros. Compared to this, the capital needed to build the data center (200.000 to 500.000€ for a free flow data center) is irrelevant. And since a free flow data center costs much less to operate than the cost factors we considered previously (483€ / rack / month), it seems obvious that one should build their own data centres to further save costs.
There is however one exception to this: Edge Computing. With Edge Computing, servers should instead be located all over the word, in the data centres of telecommunication companies. Cost of hosting might be in this case more expensive than 483€ / rack / month.
In countries such as China (and Japan to some extent), the cost factors to provide cloud can be very different. They also differ a lot from one city to another one. From one company to another one.
If network in China is 10 times more expensive than in France, then it represents up 30% of the expenses for a cloud company instead of 5% in France. This has a huge impact on the final price for the consumer. Also, China's import regulations and tariffs add a lot of complexity to the purchasing process of OCP servers. This can increase their price by 10% to 30%. Again, this has a huge impact on the final price for the consumer. Overall, it might be possible to reach in China costs which are similar to those in Europe by negociating lower prices on electricity, loan and real estate and compensate higher prices on Internet transit.
Even with a bit higher price, low cost cloud would still be of great use in China since the current solutions offered to consumers are 4 to 20 times more expensive (see: "Unleashing the Full Economic Potential of Cloud Computing in China").