Gigaom Structure: Hyperscale Cloud Innovation

Attending Gigaom Structure was an exercise in getting fire-hosed with the leading edge innovation that public cloud providers are bringing to their customers worldwide. These innovations not only will have a profound effect on public cloud computing, but also will ultimately impact data center architectures, costs, and benefits worldwide.

Cloud Vendor Statements at Gigaom Structure

Werner Vogels: CTO –

  • If you care about your data, store it at Amazon
  • AWS gives you control of the security of your data through encryption and other security tools
  • The customer should be in control of the security of their data, not the cloud provider
  • 40% to 60% of corporate IT budgets are related to dev and test, and all of that should be moved to AWS
  • All of the new stuff is hosted up at Amazon
  • The CIA paid AWS $600M over ten years for an AWS private cloud.

Jay Parikh: VP Infrastructure Engineering – Facebook

  • 20% of the world’s population cannot connect to the Internet; investing in high altitude long-endurance planes to beam the Internet into remote areas
  • Facebook was able to cut mobile data consumption by 50% to allow more people to afford to connect to Facebook
  • Need a data center architecture that can handle the diversity of data flows and types across the different applications
  • Facebook looks at problems from a full stack perspective: data center, hardware, software
  • The network has been an agility bottleneck, especially the top-of-rack switch
  • Facebook announced a new switch of its own design that splits the switch hardware into components and splits the software into different components
  • The switch becomes just another server managed by Facebook as just another server
  • The Facebook switch will be put back into open compute at some point
  • The switch is based on an Intel processor and runs the Facebook Linux distribution.
Facebook Top of Rack Switch

Urs Hölzle: SVP Technical Infrastructure – Google

  • Moore’s law is relentless
  • Learned early on how to grow at 50 to 100% a year
  • Android has had an effect on the data center; probably have 2 billion concurrent TCP/IP connections from phones to the data center at any point in time
  • Cloud will force convergence of cloud and data center architectures
  • Has been using Linux containers and not VMs for 10 years
  • Open-sourced a container manager last week
  • Google controls its entire hardware and software stack
  • Automatic management is going to win in the long term
  • “We have a security team that no one else could afford.”

Diane Bryant: SVP Data Center Group – Intel Corporation

  • Moving into custom solutions for hyperscale clouds and applications, which means  cloud customers like Amazon, Google, Facebook, and Microsoft have the option of designing custom variations of Intel chips and having Intel manufacture those custom chips for the cloud customer
  • Shipping an FPGA into a Xeon package
  • Change is the biggest driver of downtime, which is the biggest driver of operational costs
  • Software-defined infrastructure will deliver pools of compute, memory, network, and storage at the rack scale.

Scott Guthrie: EVP Cloud and Enterprise group – Microsoft

  • Azure is tuned for business organizations
  • There are going to be three hyperscale cloud providers: MS, Google, and AWS
  • Hyperscale is adding a million servers a year
  • Differentitate via an enterprise-grade service
  • Integrate with existing on-premises assets and management
  • Going to compete in higher-level services like machine learning and video encoding
  • 57% of the F500 has a deployment on Azure
  • The security FUD is not good for anyone
  • Having an on-premises option makes CIOs value Azure
  • Azure Pack: download to run on Windows Server and System Center.

Vinod Khosla: Principal – Khosla Ventures

  • Has not seen an architectural innovation from IBM, Dell, or HP in 20 years
  • Google, Facebook, AWS, etc. have invented a new computing architecture
  • What has changed is the rate of the change (increase), at the business level and the technical level
  • We have to re-engineer engineering to optimize for the high rate of change, not for a specific performance or cost objective
  • You have to optimize for change and adaptability because you cannot know what you are gong to face next year
  • What is useful about software-defined X is that is allows you to build applications without having to re-engineer the infrastructure
  • Customers do not have a clue as to where their businesses are going to go, so build systems that allow them to react quickly
  • Plan not to plan; plan to be agile
  • SDX changes the way we approach architecting data centers and enables new applications
  • The largest piece of IT budget is cost of ownership
  • Get rid of IT people and replace them with machine learning and autonomic systems
  • It is ridiculous to expect humans to manage the current level of complexity in the data center
  • If you stop linear extrapolation of what we doing you see massive opportunities; running batch Hadoop systems is silly
  • Look at high-margin vendors, and don’t buy anything from them.

Security Insights from the above Comments

Both Amazon and Google made some pretty bold claims about cloud security:

  • If you care about your data, store it at Amazon. Amazon asserts that your data is more secure if you (1) use Amazon’s tools to secure it and (2) rely on Amazon to manage those tools and to ensure they are working correctly and not getting hacked, than if you attempt to buy your own tool set, implement it yourself, and manage it yourself.
  • “We have a security team that no one else could afford,” claims Google.

Taken together, these claims imply that because of the scales of their respective operations, both Amazon and Google can invest more into security than security vendors can invest on behalf of their customers, and can amortize that cost across a huge computing estate. Our security analyst, Edward Haletky, will have to weigh in separately on the technical merits of this argument, but the economic basis of the argument is sound.

Control of the Stack as a Source of Agility and Price/Performance

Several vendor comments addressed how people are innovating either across their entire stack or at certain levels of their stack to attain competitive advantage:

  • It is clear that Microsoft, Amazon, Google, and Facebook all own their entire stacks from the hardware through the system software, to either the application services or the applications themselves.
  • Facebook viewed the top-of-rack switch as an impediment to agility. It therefore built its own switch hardware and developed its own switch software based on its own distribution of Linux. Cisco should now be worried just as much about Facebook switch commoditizing switch hardware as it is about NSX from VMware’s commoditizing Cisco’s investment in IOS.
  • Intel is adding a field-programmable gate array (FPGA) to its core Xeon server chipset, allowing customers to load code directly into the processor for extremely efficient execution. Two use cases were posited for this: situations where the customer wants the server to do different things at different times that require hardware execution speeds, and giving customers the ability to perfect their code before that code is implemented in hardware.
  • There was some discussion of various cloud vendors’ working with Intel to get custom chips for their clouds. Intel said that each cloud vendor is very aware of which instructions it uses most intensely and that options exist for cloud-optimized chipsets in the future. Ultimately, this means that cloud providers are going to be running different and better Intel chips in their servers than will on-premises enterprise customers.

The above statements combine to create a situation in which the cloud vendors are innovating across their entire hardware/software stacks. This will first cause public cloud architectures to diverge from standard data center architectures, and then, as the price/performance difference between clouds and data centers becomes large enough, it will cause workloads to migrate to the more efficient option.

Impacts on the Economics of Cloud and Data Center Services

Several statements provided excellent insights into where the economics of the cloud and the data center are headed:

  • Google stated that “automatic management is going to win in the long term.”
  • Vinod Khosla stated the following:
    • That he “has not seen an architectural innovation from IBM, Dell, or HP in 20 years”
    • Google, Facebook, AWS, etc. have invented a new computing architecture
    • Computing needs to be re-engineered for the ability to cope with rapid change and unplanned-for workload characteristics
    • People should be replaced with machine learning and autonomic sytems
    • Customers should not buy anything from high-margin vendors

The commitments to hyperscale, innovation, automation, and ever-improving price/performance points in the cloud on the part of the cloud vendors and Intel, bolstered by the analysis of one of the deans of the high-tech investing community (Vinod Khosla), have profound implications for the future of the computing industry:

  1. It will become impossible for any enterprise to replicate the price/performance points provided by cloud providers, as cloud providers will use their control over their hardware and software stacks to achieve price/performance points that no other enterprise will be able to achieve by assembling off-the-shelf hardware and software into internal clouds.
  2. The major vendors of expensive on-premises data center hardware, software, and services (IBM, HP, Dell, and their European and Asian equivalents) are in a huge amount of trouble. As the providers of the hardware, some of the software, and much of the services that make their enterprise customers uncompetitive with the cloud providers, these vendors will quickly be identified as the source of the problem, not the source of the solution to the problem.
  3. While cloud vendors today are attacking the expensive hardware, software, and services that enterprises are stuck with, they will soon strike on the management front. If and when the cloud vendors get “autonomic management” in place, they will economically attack all of the people who work in management in IT on-premises as unneeded costs.
  4. Amazon and Google are both betting that the public cloud wins long term at the expense of on-premises computing. Amazon and Google have no on-premises cloud offerings of their own. (Amazon has Eucalyptus as a partner on this front.) Microsoft is in the unique position of having the largest installed base of systems software on the planet (Windows) and a first-class public cloud (Azure). If on-premises computing has a long-term future, then Microsoft is well positioned to be a long-term winner.
  5. If all of this plays out as outlined above, the majority of the new workloads in the world will end up on public clouds. Many of the traditional enterprise workloads that are based on stateful transaction processing will never move to the cloud, so on-premises enterprise computing will continue to exist for twenty or thirty more years. But the long-term trend is clear.


Amazon and Google are going to use innovation across their entire hardware and software stacks to put enterprise computing at an agility and price/performance disadvantage relative to their public cloud offerings, with the goal of completely destroying on-premises enterprise computing. Microsoft will play its on-premises and Azure cloud cards and hope that enough on-premises computing survives to make Microsoft a long-term winner.