Infrastructure as a Service: The Race Won’t Go All the Way to the Bottom
SEE LAST PAGE OF THIS REPORT Paul Sagawa / Artur Pylak
FOR IMPORTANT DISCLOSURES 203.901.1633 /.1634
psagawa@ / email@example.com
August 19, 2014
Infrastructure as a Service: The Race Won’t Go All the Way to the Bottom
The stakes for public cloud operators (IaaS) are huge. Annual global spending on enterprise data centers – hardware, infrastructure software, staffing, facilities, etc. – is more than $1.2T. We expect most of this will migrate to the cloud over the next 2 decades, due to the compelling economic advantages of web-scale distributed data centers, with cloud operators hosting both SaaS and customized applications. The costs and performance of IaaS are highly levered by scale and technical sophistication, factors that will force all but the biggest and best from the market with time. Currently, the 3 best positioned players – AMZN, MSFT, and GOOG – are engaged in a price war, temporarily sapping industry growth and maybe driving all 3 into losses. However, the real casualties of this war will be smaller IaaS operators, who will be uncompetitive without massive consumer cloud franchises to drive scale and sophistication. Meanwhile, we do not expect a “race to the bottom” amongst AMZN, MSFT and GOOG. Rather, as the market concentrates into their hands, we believe natural points of differentiation for each of them will allow lucrative value added services atop commodity processing and storage services. All three should be profitable long-term, with sales growing into $10s of billions before decade end. However, given a slightly richer data center cost structure and the lack of a high margin core business to cover for any losses, we are concerned that AWS will be a burden to AMZN in the near term, as it works to re-establish the confidence of investors.
- Global spending on data center infrastructure is huge. Infrastructure-as-a-Service (IaaS) solutions compete directly for data processing jobs that would otherwise be run in-house. Globally, we estimate that enterprises spend more than $1.2T annually in building and operating their own data centers, including the costs of hardware, infrastructure software, facilities, maintenance, user support, and management. Long term, we believe that all but the most latency sensitive workloads will be addressable by cloud-based IaaS operations, even those currently constrained by regulation and security concerns.
- Modern web-scale distributed data centers have compelling advantages. Public cloud distributed data center platforms have huge cost and performance advantages vs. private data centers based on several factors: 1) Use of commodity components vs. value-added configured systems; 2) Minimal non-productive costs; 3) Much higher utilization; 4) Superior system availability and recovery; 5) Very low personnel costs; 6) Flexibility, scalability, power and convenience; 7) Substantial economies of scale. For example, we estimate that the all-in costs of running computing workloads in the biggest and most sophisticated IaaS data centers may be as little as 10% of those for a typical enterprise IT operation, while offering the advantages of nearly limitless scalability, global access, and world class support.
- AMZN, MSFT and GOOG very well positioned vs. would-be rivals. These three IaaS competitors lever enormous scale, institutional skills, data center R&D and operational experience from their massive consumer cloud franchises into their commercial IaaS platforms. This has given them big head starts, with significant cost and performance advantages over traditional enterprise IT players looking to move into IaaS. Moreover, AMZN, MSFT and GOOG continue to make massive CAPEX and R&D investments in their data center infrastructures, setting fast moving targets far out ahead of potential rivals, such as IBM, RACK, CRM, VMW, HP, ORCL, VZ and others.
- Aggressive pricing will drive IaaS adoption and concentrate share. AMZN, well known for its punishing pricing vs. retail rivals, got a taste of its own medicine earlier this year when GOOG cut its computing and storage prices by roughly 32% and 68%. Both MSFT and AMZN quickly responded with price cuts of their own, with AMZN getting punished by investors after reporting 2Q14 “other” revenues (N.B. largely comprised of AWS) down QoQ and delivering a much wider loss than had been promised. Despite this short term revenue hit, these step function price reductions should greatly accelerate adoption of hosting by enterprises and facilitate more compelling SaaS offerings by 3rd parties running on IaaS infrastructures. We expect that this incremental growth will be captured primarily by AMZN, GOOG and MSFT, and that the ongoing share shifts will be apparent in coming quarters.
- Traditional IT spending will suffer. The aggressive IaaS pricing is hastening enterprise adoption of cloud hosting and removing barriers for SaaS applications vendors. Conversely, we expect it to negatively affect sales of data center hardware and software, which have already been decelerating. This will compound the suffering of traditional vendors, who will see dying sales in their core products even as they struggle to compete in the IaaS market. While many these companies are pressing the idea of a “hybrid” cloud, hoping to preserve lucrative proprietary software solutions and premium priced “value added” hosting, we believe that this is a transitory opportunity that will quickly erode as tools for managing workloads (such as “containers”) improve and as barriers to cloud adoption (regulation, security/privacy concerns, etc.) recede. In the long run, we believe it will put the viability of venerable IT names, such as IBM, HPQ, ORCL, VMW and others, at serious risk.
- No “race for the bottom” – the big 3 will be huge and profitable long term. As the market concentrates and grows, we believe AMZN, MSFT and GOOG will all reach $10’s of billions in annual IaaS sales by 2020, with $100’s of billions realistic within 20 years. While we anticipate much of the volumes to be in commodity processing and storage services, we expect the price rivalry to ease somewhat as the market matures and weak hands exit. Moreover, we believe that there will be substantial demand for differentiated, value-added services atop simple hosting that will afford each of the big 3 opportunities for attractive margins playing to their disparate strengths.
- AMZN will be challenged in the near term. MSFT and GOOG have high margin core businesses that can absorb the costs of losses in their small but fast growing IaaS operations. AMZN will be much more challenged. Not only is AWS considerably larger than its rivals, but AMZN’s core businesses compete with razor thin margins while the company invests heavily in their future growth. Losses in AWS are more than noticeable on the bottom line and will serve to undermine the company as it seeks to re-establish the confidence of its investor base, since we do not expect CEO Jeff Bezos to cede IaaS market share to his rivals without a fight.
IaaS: Many Firms Enter, Three Firms Will Leave
Investors are worrying about Amazon Web Services (AWS). When AMZN posted 2Q14 results with a much wider than expected loss, a QoQ decline in “other revenues” (where AWS resides) and unrelenting capex, the AWS narrative jarringly shifted from unstoppable juggernaut to vulnerable leader in an unwinnable race to the bottom overnight. We see this as an overreaction. First, the prize is huge – AMZN, GOOG and MSFT are playing for leadership in a business that we believe will subsume more than $1T in enterprise data center spending over the next 10-15 years. Second, the price war that has erupted amongst the three may last a few years and will pressure margins for sure, but it will also hasten the transition of enterprise IT to the cloud while thinning the herd, as smaller and less sophisticated wannabe competitors fall out. Third, the big three will be positioned to capture lucrative value-added services that will ride atop commoditized computing, storage and networking infrastructure. AMZN, MSFT and GOOG each brings a distinctive array of strengths that will allow them to earn premiums. Those differentiated services will keep IaaS from entirely unraveling into a profitless commodity slog.
We believe that the large majority of enterprise computing will migrate to public cloud platforms by 2025, either directly or via SaaS applications, which will also migrate to large public infrastructures. With total global IT spending, including personnel and facilities, running at nearly $4T per year, that is a massive addressable market. It will migrate because modern web-scale data center operations can run at 10% the all-in per unit cost of the typical in-house enterprise data center, with superior scalability and performance – a differential that is widening. The regulatory, security and privacy concerns that skeptics raise are not endemic – they can and are being addressed – with latency, immutable by the laws of physics, a limiting factor for only the very most time sensitive uses. The migration will be inexorable.
AMZN, GOOG and MSFT will win. Their would-be rivals – IBM, VMW, HP, and RACK, etc. – lack the gigantic consumer franchises that give the big three scale advantage. They also do not have the same institutional skill set, leaving them far enough from the leading edge to leave them at a perhaps insurmountable design disadvantage as well. The price war amongst GOOG, MSFT and AMZN will have its casualties amongst the also-rans and market share will concentrate to the top.
Eventually, the triumvirate will begin to compete on factors other than price. Enterprises will have other needs that will open opportunities to add value, atop commodity computing and storage services. MSFT, AMZN and GOOG are not identical in their approaches and skill sets, leaving further room for differentiation. MSFT has a strong position in SaaS application software, a well established facility for infrastructure software tools and a deep understanding of enterprise IT requirements. AMZN has a big lead in serving web-based businesses, a large, well established ecosystem, and a demonstrated flexibility in accommodating the heterogenous needs of its commercial customers. GOOG has the low cost position and world leadership in analytics, data management and other key areas of computer science. By 2020, we expect each of these companies to have established a more than 20% share of a $100B+ IaaS market that will be growing at a better than 25% annual pace, with other potential rivals dead or fading. We also believe that they will have found a way to deliver profits as well.
The Old Way
The early ‘90’s saw a sea change in the architecture of enterprise data centers to an approach often called “Client Server”. Servers, built on clusters of the same Intel x86 processors that powered PCs, began to replace the proprietary and expensive minicomputers and mainframes that had ruled in the decades before (Exhibit 1). While the change played out incrementally and incompletely (N.B. IBM still sold nearly $5B of its 50 year old mainframe technology in 2013), it was as devastating to the old guard of IT vendors (DEC, Data General, Prime, Unisys, etc.) as it was nourishing to a new breed of competitors. Most of today’s biggest enterprise players – Microsoft, Intel, Dell, EMC, HP, Oracle, SAP, Cisco, and others – made their bones during this paradigm shift and rewarded their investors handsomely in the process (Exhibit 2). The Y2K communal freak-out hastened the process, pushing laggards to step up and replace whole systems ahead of the millennium change to avoid the perceived risk.
Exh 1: Major eras in computing
Exh 2: Worldwide Mainframe Market, 2011-18
Today, the very large majority of enterprise data centers are built this way (Exhibit 3). Configured servers with clusters of x86 processor cores, storage systems built on RAID (random array of independent disks) architecture, and networking based on IP routers. Atop this hardware, largely sold as fully configured boxes with proprietary control software and vendor supplied value-added features, the servers run some form of vendor supported Linux or Microsoft Windows Server, with a structured relational data base (most often Oracle, but also Microsoft, IBM and others) to manage the data, and a virtualization “hypervisor” to divvy up access to resources efficiently and separate user workloads from one another. These data centers then run applications to support the workings of the enterprise – enterprise resource planning (ERP) systems, business intelligence/data analytics tools, communications/collaboration solutions, etc. – which, traditionally, have been either licensed from vendors or written as custom software specifically for that organization. All of this technology resides in specially configured rooms, either within the enterprise’s own facilities or in space leased from a third party. Support for the data center and to individual users is provided via maintenance and support contracts with the various vendors, and by IT organization personnel. While this is a vastly simplified schematic, ignoring smaller product categories and homogenizing the variances across different organization, it is fairly representative of standard practice for enterprise data centers over the last decade.
Exh 3: Datacenter architecture of the past decade and today
The New Way
As the transition to Client Server ramped in the late ‘90’s the seeds of the next revolution were already germinating. The progenitors of Google, Larry Page and Sergey Brin were collaborating on a new approach to index and search the already huge and growing World-Wide Web. Early on, Page and Brin realized that the data management architecture of the moment, the relational data base systems championed by Oracle and others, would not be able to scale to the requirements of their ambitions. Thus, the newly formed Google embarked on an institutional quest to redesign the data center, offering much of their development work to the open source community. Thus was born the techniques of big data – Hadoop, unstructured data bases and web-scale distributed data center design (Exhibit 4).
Exh 4: MapReduce/Hadoop parallel computing flow
Modern cloud data bases are built on the cheapest possible hardware. No expensive configured servers with built in security and other bells and whistles or RAID storage systems with integrated back-up and “value-added” storage management software. Instead, bare bones CPU chips, memory and disk drives bought in bulk, directly from the manufacturer at volume wholesale prices, and installed onto low cost system boards of the data center operator’s design by contract manufacturers. Networking is accomplished with cheap commodity data switches. No money wasted on heat sinks, cooling fans, enclosures, or other extraneous hardware – all of those boards are mounted into open racks in warehouses designed to promote good airflow with shared primary and back-up power supplies. The racks are mounted into modular containers that can be moved in and out with ease. Disaster recovery is handled through the operating software and is based on redundancy. Data is duplicated 2 or 3 times over in disparate locations. If a part goes bad, the process can shift to a good part and use duplicate data without missing a beat – the bad board is pulled out of service and replaced. The modern data center is designed to use as little extraneous electric power as possible and to efficiently utilize real estate – ideally located where land is cheap and fiber telecommunications bandwidth is plentiful. It is highly automated in its operating requirements and can be monitored remotely with only a skeleton staff on site.
All of this cheap commodity hardware is controlled by sophisticated proprietary software. In the old paradigm, it was assumed that structured data, conforming to pre-determined and rigid formats, was necessary for systems that adhered to the ACID (Atomicity, Consistency, Isolation and Durability) requirements of computer transactions (Exhibit 5). Moreover, it was also assumed that fixed data formats were necessary to allow records to be sorted quickly and accurately enough to deliver specific information in a timely manner. Over the past decade, computer scientists have surmounted these assumptions, defining unstructured data systems that conform to ACID and that can be searched quickly and accurately at nearly limitless scale.
Exh 5: ACID Database Transaction Properties
Key to this is an underlying system architecture that breaks processing jobs into many simple steps that can be performed in parallel on many different CPU cores and reassembled. Additional important software innovations include data storage systems that rely on added disk drive redundancy rather than expensive dedicated logic to deliver reliability and disaster recovery, “container” technology that separates workloads into discrete and secure units that can be assigned to the most appropriate processors, software that manages the flow of data across communications networks built of inexpensive hardware switches rather than routers, and many others. All of these innovations were implemented first as proprietary solutions in their progenitors’ (mostly Google) own data centers before later being contributed to the group knowledge of the open source community. As the new data center paradigm continues to evolve, much of the leading edge development is happening at a small set of consumer cloud leaders (Google, Microsoft, Amazon, and Facebook) who have the resources to maintain serious computer science research efforts and massive cloud platforms. This is a formidable advantage for these companies.
Better? Stronger, Faster …
Web-scale distributed IaaS cloud infrastructure built on the new paradigm have substantial cost and performance advantages, which we codify into seven factors: 1) Use of commodity components vs. value-added configured systems; 2) Minimal non-productive costs; 3) Much higher utilization; 4) Superior system availability and recovery; 5) Very low personnel costs; 6) Flexibility, scalability, power and convenience; 7) Substantial economies of scale. These factors enable as much as 90% lower all-in costs relative to the typical enterprise data center based on virtualized client-server architecture, and as much as a 50% advantages over less sophisticated operations applying the new paradigm on a smaller scale.
Exh 6: The 7 Advantages of Cloud Infrastructure
Exh 7: Server / component cost comparison
Use of commodity components. By buying chips and drives directly from component manufacturers and having them installed onto circuit boards of their own design by low cost contract manufacturers, the top cloud operators eliminate costly hardware bells and whistles, unnecessary software and system vendor margins from their equipment investment (Exhibit 7). Moreover, other elements of hardware can be eliminated or replaced with cheaper alternatives – for example, open racks allow for minimal systems cooling hardware and commodity data switches can replace high cost routers. All together, we believe the top IaaS hosts can have hardware costs as much as 50% lower than even large scale data centers built out with configured systems. We also note that both Google and Microsoft own tens of thousands of miles of fiber optic cable connecting their data center locations, greatly reducing costs in comparison to leasing such lines from network operators or using commercial telecommunications services.
Exh 8: Calculating Power Usage Effectiveness (PUE)
Minimal non-productive costs. Traditional data centers are designed to avoid failure. Cooling systems maintain components within temperatures optimal for performance, while power management systems work to ensure the constant availability of electricity. This is expensive – the costs of this environmental infrastructure are as much as 25% of the non-personnel costs of a data center. Modern data center design dispenses with as much of this cost as possible, allowing data centers to operate outside of “optimal” parameters for temperature and power consistency. While this yields greater failure of components, cloud data centers are designed to accommodate it with storage redundancy and the fast reassignment of workloads to working processors. The failed boards are then simply replaced. Not only does this dramatically reduce the investment in environmental infrastructure, but it also cuts the power draw of the data center as well. Electricity is more than 15% of non-personnel costs for the average data center, with power for environmental control, power conditioning, lighting and other non-productive uses amounting to more than 80% of that needed for computing, storage and networking. This ratio, known as Power Usage Effectiveness (PUE) and expressed as the power drawn relative to the amount needed for productive uses, shows top cloud operators at a distinct advantage (Exhibit 8). For example, Google’s mean self-reported PUE for its data centers is 1.12 vs. the 1.80 enterprise average, and includes a broader range of non-computing power use categories than is normally reported. Using the less rigorous metrics applied by most other data center operators, Google notes that its PUE could be as low as 1.06 and has been trending down steadily over the 6 years that it has reported the number. Beyond PUE, it is also likely true that the top cloud operations are much more efficient in the use of power for productive purposes as well. Overall, it adds to as much as a 50% advantage in electricity costs, a 25% advantage in non-personnel operating costs and a ~10% advantage in the all-in costs of computing.
Exh 9: Server Utilization Rates
Higher Utilization. Many of the costs of a data center are fixed – hardware and software investment, facilities costs, some personnel, etc. The more that the data center is utilized, the more broadly these costs can be amortized against productive work (Exhibit 9). The larger scale cloud operators are able to run effectively at higher levels of utilization, as their size and the broad mix of applications that they serve statistically tighten the degree of slack capacity necessary to handle peak demand. Moreover, the flexible nature of modern distributed data center design (workloads can be easily shifted from location to location to balance utilization), optimized hardware configurations (the commercial systems used in typical data centers may “waste” capacity if the CPU, memory, storage, etc. do not ideally fit the applications being run), the base load demand from their huge consumer cloud franchises, and their global reach, also help the top IaaS players maintain higher average utilization. Google has reported that its data centers run at an average 30% capacity utilization, nearly 3 times the 12% enterprise data center average that has been calculated by Gartner, amongst others. This factor alone could give Google as much as 60% cost advantage for computing and storage workloads.
Superior availability and recovery. Data center capacity occasionally goes out of service, both for planned reasons and for unanticipated outages. Downtime is costly, requiring work to be shifted to alternative capacity or delayed while expensive systems engineers execute the necessary steps to bring the systems back into service. Cloud architecture is designed to handle system upgrades and maintenance with less planned downtime, while having fewer unplanned outages and recovering from them more quickly than enterprise data centers. The modern cloud data center has the advantage of geographic diversity, with data typically stored redundantly at multiple locations. Processing loads and network communications can automatically shift to alternative, available data centers should any one site suffer failure. All of this is handled in the system software infrastructure, saving the cost and complexity of functionally specific back-up systems, such as RAID storage. With these techniques, cloud operators can offer lower systemic risk to customer data than within traditional data centers at a lower overall cost. The overall cost impact of this impact, including the staffing and consulting support needed to respond to outages, is very difficult to quantify but it is a real advantage for IaaS (Exhibit 10).
Exh 10: Cloud provider Service Level Guarantees
Very low personnel costs. Modern cloud data centers are highly automated and monitored remotely. According to Microsoft, web scale operations typically run with less than one employee for every 1,000 servers. Anecdotal reports suggest that Google may be operating data centers of 100,000 servers or more with IT staffs of fewer than 20, implying an even more impressive 1/5,000 ratio. In contrast, even in a well-run in-house enterprise data center automation is minimal and human error causes considerable error. The IT staff to server ratio typically runs at an expensive 1/100. For such organizations, personnel can make up 30-50% of total data center operating costs. By comparison, data center operations personnel costs for any of the top IaaS players is less than 5% of the total (Exhibit 11).
Exh 11: Web-scale data center cost distribution
Flexibility, scalability, power and convenience. Beyond integrated backup, the web-scale approach of cloud architecture also offers valuable benefits to commercial customers – no practical limits on application scale, ability to accommodate enormous short term usage spikes, predictable all-in pay-as-you-go pricing, global scope, automatic system upgrades, world-class IT support and maintenance, and many others. To illustrate one of these advantages, Google’s Senior Vice President of Infrastructure, Urs Hotzel, demoed the capabilities of the company’s Compute Engine hosting service during the 2012 I/O developers conference, employing more than 750,000 server cores in parallel to complete an analysis of a human genome in less than a minute, vs. the 15 hours that it usually took on the customer’s own computing infrastructure. The combination of enormous power, complete usage flexibility and all inclusive pricing, available on a global basis with little previous notice is a powerful selling proposition. We note that the top cloud operators (including Facebook, which has not shown interest in applying its infrastructure to the commercial IaaS market) have been the employers of choice for the most talented computer scientists over the last decade, and that the R&D excellence that this talent represents will only push the proprietary implementations of cloud architecture further ahead in the future.
Exh 12: Quarterly CAPEX of Select Cloud Operators
Economies of scale. The infrastructures of the top three IaaS hosts, with massive consumer franchises to drive volume growth, are hundreds of times bigger than the typical large enterprise IT operation. In 2013, Google spent $7.3B on CAPEX, bringing its PP&E to $16.5B. At the same time, Microsoft spent $4.3B in CAPEX reaching $10B in PP&E. Amazon spent $3.4B and had total PP&E of $10.9B. Not all of this is IT spending – Amazon, in particular, also spends heavily on its fulfillment infrastructure – but the majority, and in some cases a large majority, is spent building out distributed cloud data centers. Meanwhile, the old line IT players, like IBM, HP, VMWare and others, have been scaling back their CAPEX (most of it spent on other things than cloud infrastructure), allowing depreciation to take down the book value of their PP&E, while new breed competitors, like Salesforce.com and RackSpace, operate with CAPEX and accumulated infrastructure investment a magnitude smaller than the leaders (Exhibit 12-13).
Exh 13: Quarterly Net PPE of Select Cloud Operators
For data systems costs, bigger is definitely better. Technology is purchased more cheaply in bulk. Personnel and other development costs can be levered over a much larger base. Scale enables greater geographic dispersion and justifies dedicated fiber connections, both of which reduce telecom costs. Large operations can buy their own real estate, better suited to the needs of a data center and cheaper to boot. The best in IT talent is attracted to the challenges at the cutting edge, and can be leveraged across a massive base of IT driven business. Custom hardware and software development makes sense when applied to lucrative consumer cloud franchises and can be utilized in the hosting business as well.
It Adds Up
In 2012, Forrester Research estimated that buying storage as a service from a major cloud provider was 75% cheaper than the all-in costs of buying, implementing and operating storage in an in-house data center for most enterprises. In the intervening two years, the prices for cloud storage have dropped by more than 80% and the prices for CPU usage have fallen 60%, while the relevant costs for the in-house solution, the bulk of which are personnel and facilities, have likely risen (Exhibit 14). Given some indication that the recent price cuts have, at least temporarily, pushed Amazon Web Services into mild losses, it would seem that its costs of service are now 90% lower than the enterprise benchmark. We suspect that Microsoft and, particularly, Google may be in an even more advantaged cost position, given evidence of lower PUE ratios, more proprietary hardware and data center designs, and more computationally demanding consumer cloud franchises.
Exh 14: Costs of 100TB of Storage, Internal versus Cloud, 2012 vs 2014
We note that these advantages apply not only against in-house enterprise data centers, but also against smaller cloud operators and so-called “private” clouds. Even when smaller competitors have implemented modern distributed data center design – and many of them (e.g. Salesforce, Oracle, and others) haven’t – they still suffer from significant scale disadvantages and typically run at a significant efficiency deficit vs. the big three. For example, IBM reports that that the majority of its data center locations are 10-30 years old, and that the average PUE rating for its facilities is 1.73, with its most recent sites running at 1.4 to 1.6. With this infrastructure, it will compete with Google, which has no data center older than 10 years and operates with an average PUE rating of 1.12. A recent academic study suggests data centers built with commercial hardware systems may be limited to a floor of 1.15-1.2. This also implies significantly higher spending for the cooling and power conditioning infrastructure that is drawing so much of the non-productive power.
Capacity utilization may be the biggest difference between the big guys and the hoi polloi. With massive volumes stemming from consumer cloud franchises, global operations across many time zones, flexible system designs that allow workloads to be shifted to idle sites, and the benefits described by queing theory (large systems can handle demand spikes with a lower percentage margin of slack capacity than smaller systems), we believe that the servers at Google, Amazon and Microsoft are far better utilized than those at other would-be rivals. We also suspect that few of the smaller operators have been able to implement the same degree of automation that has been built in to the top IaaS data centers, likely adding significant personnel costs as well.
Begun the Cloud War Has
In March, Google announced massive price cuts across all of its commercial cloud service offerings, dropping its rates for computing by an average of 32%, for storage by 68% and for use of its Big Query data base platform by 85%. These cuts reverberated across the market, prompting responses in kind. The next day, Amazon announced cuts of its own, ranging from 36 to 66% across the extensive AWS product line. A week later, Microsoft followed with a 35% cut in computing and a 65% cut in storage. We expect this era of hyper-aggressive pricing to persist for a few years, with a number of important effects (Exhibit 15-16).
Exh 15: Cloud Price Reductions by Vendor, 2013
Exh 16: Sample Basic Cloud Services Pricing from Major Providers
Smaller IaaS players will suffer badly. The situation presents a long term existential dilemma for smaller cloud operators. Cut prices to match the big boys and hemorrhage cash OR hold the line on price and watch customers slowly take their business elsewhere. The size and breadth of the price cuts will be impossible for most enterprises to ignore, particularly since there is no indication that further cuts will not be forthcoming in the future. These operators will tout their value-added software and services as justification for premium pricing, but these products are bundled with an underlying commodity that is badly disadvantaged. For example, relatively pure cloud play Rackspace, which bundles IaaS hosting services with IT consulting, has been earning operating margins of less than 8%. If it were to drastically cut price or begin to lose customers, the results could be catastrophic. The same drama will play out within the IaaS business units of the IT dinosaurs that have been turning to the cloud as an attempt to remain relevant to their customers. We expect that the herd will be effectively culled.
Amazon will take damage. In the aftermath of these price cuts, Amazon reported a sequential decline in its “other revenue” category that contains AWS and a much wider than expected loss for 2Q14, shortfalls that have been widely attributed to the effects of those price cuts (Exhibit 17). The impact on Google and Microsoft, with much bigger margin cushions from their core businesses, was far less apparent. While the top three operators have serious advantage vs. the smaller IaaS players, it is also likely that Google has serious cost advantage on the other two, and that Microsoft has a modest cost advantage vs. AWS. Given Amazon’s vulnerability to shortfalls in operating contribution from its cloud business that would be rounding errors for the other two, and its dominant share leadership in the commercial IaaS market, it is in the unusual position of having to play defense against aggressive price competitors with lower cost positions. It will be interesting to see Jeff Bezos play from the opposite side of the table.
Exh 17: Amazon “Other” Revenue
The shift to the cloud will accelerate. CIO surveys for the past several years have seen cloud hosting and SaaS applications rise to the top of the priority list. SaaS start-ups built on public cloud IaaS platforms have proliferated to challenge both traditional packaged software and the older SaaS players competing off of their own subscale and old paradigm data center operations. Amazon reported that usage on AWS was up 90% YoY, obscured somewhat from the lurid impact of the price cuts on revenue. Microsoft reported that its Azure cloud REVENUE had more than doubled YoY in the most recent quarter, despite the impact of the price cuts. Over the next decade, we believe that the cost and performance advantages of the modern web scale data center will compel most enterprises to shift the large majority of their IT use to the cloud. The current pricing environment will only ease and hasten the decision.
Turtles, All the Way Down
While we expect the aggressive pricing environment to continue for a few years as the market expands and the market share concentrates, we don’t believe that it is a true “race for the bottom” for two main reasons. First, cloud hosting is now a very small business for top players, particularly for Microsoft and Google, but it is growing very rapidly. As it becomes a bigger piece of each of the company’s sales, the pressure to generate profits will undoubtedly soften their resolve to use pricing as weapon to gain market share from AWS. Second, we expect that there will be considerable opportunities for all three players to bundle premium value-added services on top of the bundled commodity. While Microsoft, Amazon and Google share the same basic data center architecture and offer similar web-scale size and breadth, there are differences in the institutional skill sets of the three companies that suggest differentiated opportunities for each within the developing IaaS market.
Microsoft’s relative strengths are obvious, as it is the only one of three that begins with a substantial enterprise IT franchise, with the sales relationships and deep understanding of enterprise customer needs that come with it. Azure is a clear choice for the legion of enterprises that already operate on Microsoft infrastructure software – Windows Server, Microsoft SQL server, etc. – and a good choice for those with a strong commitment to Microsoft SaaS applications. The combination of a leading edge web-scale cloud product and a strong legacy IT position is unique and exploitable toward new SaaS applications and software infrastructure tools to manage a cloud future that will likely be heterogenous, combining in-house resources with SaaS and IaaS, perhaps from multiple vendors. All of this can mean premium margins for Microsoft.
Google is newer to the IaaS game than either of its rivals and has been seen as inflexible by some potential customers, but brings formidable strengths. Google is by far the biggest and most sophisticated consumer cloud operator and is able to harness more computing power than any company on the planet. It invented many of today’s state of the art cloud technologies including MapReduce and Hadoop that allow for parallel computing, and can summon hundreds of thousands if not millions of server cores on demand to simultaneously crunch problems in parallel, making it the obvious choice for use cases requiring the periodic application of maximum computing firepower. Google leads the bleeding edge of data center design, and almost certainly has dramatic cost advantages over all other cloud rivals stemming from its huge scale, its extraordinary efficiency, and exceptional utilization. It also has invented groundbreaking techniques for “big data” analysis and ACID compliant high performance unstructured data bases that could make it a unique platform for high value business intelligence applications.
Amazon has the first mover advantage, with the experience and market share accumulated over 6 years of offering commercial hosting services. With this, Amazon carries a well-earned reputation for flexibility with its customers – Microsoft has been slow to support non-Microsoft software, while Google has been even more resistant to software outside of its preferred standards. As a result, AWS boasts an impressive ecosystem of 3rd party software vendors working closely with the Amazon cloud APIs and infrastructure software implementations. This should stand AWS in good stead as it defends its customer base in the midst of the budding price war, and should allow it to develop integrative premium services to address the heterogeneity of its customers’ computing needs.
Exh 18: Winners and Losers
Winners and Losers
In the long run, we believe that Microsoft, Amazon and Google will all be winners in IaaS, perhaps in that order, while, essentially, all other would be rivals to this market are likely to turn out as losers. This would include old paradigm IT vendors like IBM, HP, VMWare, CSC, Oracle, and others, telecom carriers like Verizon and Century Link, and sub-scale new breed hosts like Salesforce, Rackspace, and Virtustream. Traditional IT gets the double whammy, as the growth of public cloud IaaS will commoditize data center hardware and erode demand for value-added infrastructure software. New consumer cloud and SaaS enterprise application vendors may also be winners, given the opportunity to use low cost hosted platforms to move past once formidable barriers to entry.
In the more immediate term, the current price war has its own implications for the three IaaS leaders. Microsoft is well positioned to take market share and grow revenues with Azure, with the margin hit from the IaaS price war covered by the ample profitability of its Office 365 franchise and other businesses. From a stock performance perspective, strong growth in Azure should help push the changing perception of the company from stodgy old paradigm loser to cloud era champion.
Google hasn’t made selling its IaaS services a priority, but with the massive cuts announced in March of this year and its provocative statement that cloud prices haven’t kept up with Moore’s Law, Google may finally be getting serious about it. While it has made many high profile moonshot investments, with a series of acquisitions in robotics and satellite companies in the past year as well as its own internal efforts like the self driving car and CALICO, Google’s underlying technology is fundamentally driven by its datacenters. Monetizing IaaS will give the company an ancillary revenue stream beyond advertising to buttress growth and give investors greater faith in its ability to deliver long term growth. With the industry’s best cost structure and the exceptional profitability of its advertising businesses, Google has been the instigator of the aggressive pricing environment and is best positioned to absorb its effects.
For Amazon, the company that invented IaaS with the launch of AWS in 2006, the immediate situation appears challenging. Despite its leading position in IaaS, Amazon seems to have higher hardware expenses than its peers given it equips its servers with premium Intel chips and more expense solid state memory. The company also likely has higher relative labor costs than its peers given it hired “thousands” of new employees in the past year and is on track to hire more sales reps, engineers, developers, and solutions architects. With the recent AWS price cuts and more likely coming, given signals from cloud rivals and new entrants, Jeff Bezos may have to make some tradeoffs, as Amazon is in precarious position spending heavily on all fronts against razor thin margins in its core business. Playing defense against cost advantaged rivals will be a new experience for Amazon, which has earned a fearsome reputation for its relentless attacks on high cost incumbents in all of its other businesses. Certainly it will soak up many degrees of freedom for Bezos as he looks to reassure his shell-shocked investors, and we fear, increase the volatility of the already unpredictable results of the company.