Part 4.1 of Planetary-Scale Computation: An industry primer on the hyperscale CSP oligopoly (AWS/Azure/GCP):
- Let’s Get Physical, (Cyber)Physical!: Flows of Atoms, Flows of Electrons
- A Cloudy History: Four Histories of Cloud Computing
- Primer on the Economics of Cloud Computing
- Three-Body: Competitive Dynamics in the Hyperscale Oligopoly
- Initial Positions and Laws of [Competitive] Motion
- Mass and the Law of [Economic] Gravitation
- Velocity and the n-body problem
- The Telos of Planetary-Scale Computation: Ongoing and Future Developments
Table of Contents for Initial Positions and Laws of [Competitive] Motion:
- Initial Positions: Where are the Three Bodies?
- Initial Positions and Laws of Motion: Why Three Bodies?
- Laws of [Competitive] Motion: The Law of Conservation of Attractive Profits
Initial Positions: Where are the Three Bodies?
Let t(0) = Jan ’22. What are the initial positions of the three bodies?
As Covid-19 impacts every aspect of our work and life, we have seen two years' worth of digital transformation in two months.
— Satya Nadella
An oversimplified, high-level overview about the relative strategic positioning of the Big Three cloud hyperscalers from early 2020 (say, pre-NYC lockdown) would have gone something like this: AWS is Goliath, Azure might catch up to AWS because incremental market share gains in cloud penetration are likelier to be from larger, legacy (non-“tech-first”) enterprises where Microsoft has an advantage due to near ubiquitous penetration of software products (Windows, Office, VSCode, Github, etc.) and a clearly articulated hybrid cloud strategy, and GCP’s new CEO has to prove that they’ve sufficiently internalized the need to build out their enterprise sales muscle so that they can quickly get to sufficient scale to reach the promised land of mid-20% operating margins.
In other words, intra-industry competitive positioning hasn’t dramatically changed between then and now — plenty of analysts could have (and did) pretty much give the same synopsis circa 2019. At the risk of being only slighty less reductionist than I’ve already been, the even higher-level story goes along the lines of:
- Docker and Kubernetes enabled widespread developer adoption of containerization and container orchestration, a development that was driven by and, in turn, accelerated the mainstreaming of microservice architectures to replace traditional, monolithic architectures
- An ethos of vendor-agnostic modularity, driven by concerns of vendor lock-in and empowered by the adoption of containerization, normalized the rhetorical and actual adoption of hybrid cloud and multi-cloud architectures, possibly allowing for the long forewarned phenomenon of cloud commoditization to grow deeper roots
- The decade-long process of commoditization of basic compute, network, and storage services from cloud providers accelerated hyperscaler focus on higher margin, differentiated services like industry/vertical-specific solutions (effectively going up the stack) and AI/ML, as well as cement their infrastructure dominance by integrating backwards into chip design
In relative terms, GCP had/has more to gain from positive trends in containerization adoption and resultant commodification of lower-level offerings because of their third place position, Azure had/has more to gain from hybrid cloud adoption where they have a clearly articulated strategy [ctrl+F “hybrid” in each hyperscaler’s respective 10-K’s — you hit 0 results for AMZN], AWS has more to lose from multi-cloud adoption [a category that Google considers to subsume hybrid cloud and claims as a trend it is positioned to gain from], both GCP and Azure claim horses in the race for dominance in AI with respect to both developer mindshare [where Google’s TensorFlow platform competes with Facebook’s/Meta’s Pytorch platform that AWS partners with them on to combat TensorFlow] and exploration of industry use cases [where Google’s DeepMind battles with the Azure x OpenAI partnership], and Azure has a relative advantage in providing industry-specific solutions due to strong enterprise relationships and simply by virtue of not being Amazon.
Despite macro uncertainties justifying IT budget cuts, reduced spend on non-critical cloud workloads throughout 2020, and lockdown-induced delays in executing hybrid cloud implementations, Cloud companies were clear beneficiaries of an unprecedented catalyst for accelerated digital transformation in the form of COVID-19.
That cloud companies were beneficiaries of what Satya Nadella dubbed “two years’ worth of digital transformation in two months” and the broader narratives continuing to develop around remote working/learning is clearly reflected in IaaS run rate numbers and, despite numerous to-be-expected sector rotations out of tech since 2020, the aggregate market capitalization of the cloud sector.
While the acceleration of digital transformation forced on companies by a global pandemic led to strategic shifts for most industries, the accelerated digital future brought about by COVID was more a change in speed than direction for an industry that already sought to abstract away materiality, physicality, and distance as part its core value proposition. In other words, the relative competitive positioning of the hyperscale cloud players hasn’t dramatically changed from what it had been before the global pandemic.
Even this synopsis from 2017 about the Cloud industry written by ...
... is still relevant.
Ongoing developments have, for the most part, been anticipated by those closely following the industry.
Initial Positions and Laws of Motion: Why Three bodies?
Why three bodies? Why not one? Why not one hundred? How do they interact with each other?
That the public cloud infrastructure market consolidated into a three-party oligopoly (ex-China) is usually taken for granted in discussions about the industry. While economies of scale logics provide easy explanations for why cloud infrastructure isn’t a fragmented market with dozens of similarly-sized cloud service providers, why cloud infrastructure didn’t become a monopoly is less well articulated. One explanation is that a select few companies, out of a limited number of potential entrants, chose to capitalize on the market opportunity while barriers to entry were still surmountable. A more overlooked explanation is that the market for public cloud infrastructure services demanded more entrants. With respect to the former explanation I covered how the economics of leasing excess internal capacity favored consumer-facing tech companies in another section, but the latter explanation for why the industry tended away from a monopoly structure can be found in Porter’s case study of the corn wet milling oligopoly.
This bottom-up explanation (assuming [“bottom-up” = emergent, market-based demand] and [”top-down” = hierarchical, calculation-based decision making]) provides a complement to existing narratives about big tech incumbents engaging in opportunistic backwards integration into providing cloud services — the market opportunity had to be created through the interaction of multiple players such that there was sufficient capacity and choice for enterprise customers (this argument is less relevant to start-ups who weren’t starting with legacy IT tech debt) to consider re-architecting their technology stack. Porter’s idea is that supply and demand are actually interdependent phenomena which bootstrap each other in a market.
How big the Cloud sector would be today if Microsoft and Google did not exist and AWS had a monopoly on cloud infrastructure is a moot question because other eligible entrants would have eventually recognized excess industry returns and potential customers would have been incentivized to help these competitors develop (which isn’t to say AWS could not have been a monopoly under different path-dependent circumstances). Businesses are afraid of vendor lock-in for the same reason why twenty-something single New Yorkers might date more than one person at a time — they want optionality and they’re afraid of being tied down.
Like the final three contestants in The Bachelor or Flavor of Love [by far the worst show “concept” category in existence, imo], the presence of other competitors creates a favorable competitive dynamic for would-be customers. Leverage and negotiation posture exist along a spectrum: digital Switzerland cloud companies like Snowflake wouldn’t be able to exist without the ability to play infrastructure providers off on each other and even wholly committed, long-time single-CSP customers like Netflix (NFLX completed their cloud migration to AWS in 2016 but use OpenConnect, their internal CDN, to stream their content to you) can tell their cloud provider what basically amounts to “You better treat me right, there are two other fish in the sea ...” but in our modern-day equivalent MBA/corpo-speak mixed with legalese. Sure I’m talking to other people, aren’t you?
In fact, the empirical evidence on the cloud computing industry supports the existence of this “more competitors is better for customers”-dynamic.
[Note: For those with access to Goldman’s research portal, Goldman’s Cloud Quarterly series gives detailed summaries of LTM price cuts between the three hyperscalers; for a certain subset of employees at the Big Three, you probably already know where to find the relevant internal dashboard]
The data clearly shows the emergence of competition within the public cloud infrastructure from 2013 to 2015 in the form of accelerated price cutting from the market incumbent, AWS. While, in classic textbook fashion, Microsoft sought to stabilize price cutting in 2013 [1][2][3] for what Microsoft referred to as “commodity services”, in equally classic textbook fashion, game theoretic logics [Porter analogizes the competitive situation in an oligopoly to the Prisoner’s Dilemma and specifically references Thomas Schelling’s work on game theory] regarding competition over market share dictated that détente would be unsustainable: in 2014, with the helpful push of Google, price wars resumed [4][5][6][7].
Incumbent initiates price cuts to combat market entry → runner-up commits to retaliation on price cuts to reduce competitive uncertainty → third-place underdog destabilizes the competitive dynamic to gain market share. Variations of this dynamic continue to play themselves out in the cloud oligopoly to this day, although the basis of competition is not necessarily always price. As we’ll explore in the next section, Google’s (”The small firm ... may have much to gain and little to lose by initiating a move ...”) open-sourcing of their container orchestration software, Kubernetes, and their internal AI/ML framework, TensorFlow, both represent examples of the competitive pursuit of corporate self-interest to the detriment of competitors (from an analysis of only first-order effects, that is) and to the benefit of the broader market (WOO, go capitalism!!!!).
Whereas the simulation of a three body system under Newton’s Laws deterministically tend towards unpredictable chaos, the hyperscale oligipoly’s near complete information condition and inherent competitive logics of this three body/cloud system tend towards predictably intense competition.
[Although Cloudflare is an Nth-body, it’s still worth mentioning here how their recent elimination of egress fees from their R2 object store immediately pressured AWS into expanding free data transfer limits, but more on this later]
Laws of [Competitive] Motion: The Law of Conservation of Attractive Profits
On [modularization vs interdependence], [commoditization vs differentiation], and dynamically shifing bases of competition within the cloud computing value chain. On the competitive laws governing the motion and positioning of bodies within the system.
In addition to the cloud computing industry’s unique grounding in economic and financial theory discussed here in Primer on the Economics of Cloud Computing and its distinct status as an oligopoly with near complete information, the Cloud’s value chain is arguably one of the most appropriate industries to analyze through the lens of Clayton Christensen’s interdependence-modularity framework that he develops in The Innovator’s Solution.
Analyzing the cloud computing industry could be considered to be a natural continuation of Christensen’s own analysis of the mainframe computer and PC industries that he undertakes in his book, with obvious connections that have not been lost on other analysts like scuttleblurb ...
... and from Ben Thompson indirectly here and here, more directly here, and explicitly here and here (and doubtless a bunch of other times as well).
What’s more is that Christensen’s framework interdependence-modularity framework explicitly builds on Porter’s value chain concepts from the well-known five forces and differentiation vs low cost stategy frameworks featured in Competitive Strategy.
Applying Porter’s base framework and Christensen’s dynamic overlay to a competitive analysis of the Cloud industry’s value chain provides us with a highly explanatory lens through which the strategic positioning and tactical moves of the players within the industry can be understood. Each player within the Cloud ecosystem is aiming to commoditize adjacent subsystems while positioning to be in that area of the value chain where differential profits accrue to after disintegrated subsystems [temporarily] recongeal into new, interdependent architectures.
For Google, the decision to open source Kubernetes can be interpreted as an attempt to commoditize basic compute instances and lower switching costs for customers to try out Google’s cloud offerings, in which Google positions its AI/ML solutions as the newly integrated, interdependent, differentiated solution through optimizing their TPU chips for the TensorFlow framework which they open sourced [a tactic known as commoditizing the complement].
Microsoft has clearly articulated their strategy to be one of providing a full-stack offering that’s greater than the sum of its parts. Instead of a dogmatic insistence on the Windows OS, Microsoft has assented to the operating system to becoming a modular consideration in its datacenters (Azure offers Linux-based instances alongside their Windows VM instances) and is, in fact, seeking to commoditizing/modularizing not just the OS but everything and claiming that it, alone, can provide the best interdependent, full-stack solution for enterprises. This is not just my conjecture but Microsoft’s express strategy.
Satya wants to “commoditize digital tech” while positioning Microsoft to provide a differentiated approach in the form of an “interdependent whole ... in these platforms and clouds”. Recognizing that Satya is explicitly following a Christensen-esque approach provides a coherent explanation Microsoft’s approach towards integrating VSCode and GitHub and Teams and Dynamics and Loop and HoloLens, etc. into a tidy whole of which Azure is the centerpiece and its cloud services (IaaS, PaaS, and SaaS ... and Metaverse(?)) justify recurring subscription revenue.
From the perspective of an Independent Software Vendor (ISV), Snowflake’s strategy is predicated on modularizing (and therefore commoditizing) the infrastructure layer provided by hyperscale cloud providers and capturing value through a superior database service that comes procedurally from pure product focus and structurally from vendor-agnosticism. Their data marketplace initiative is meant to create a platform that constitutes a “performance-defining subsystem” that lies on top of an increasingly commoditized substrate.
With respect to the application of this lens to buzzwordy deployment models, the movement towards hybrid and multi-cloud architectures can be interpreted as a customer-driven impetus to modularize lower-level compute/network/storage products and further preempt/unwind vendor lock-in, as well as preserve the negotiating leverage and optionality that comes with an architecture that has minimal switching costs between infrastructure providers — you’ll be nicer to me come contract renewal season if switching my workloads to Azure can be done at a literal push of a button. The existence of other capable competitors within the oligopoly makes the threat credible, which motivates all three hyperscalers to accommodate customers lest they lose market share to an even more accommodating competitor. Further out, edge architectures and the hypothetical fully-IoT-connected world holds the promise of reintegrating the subsystems to a decentralized/distributed edge that services latency-sensitive workloads which can’t wait for a response from centralized datacenters.
The hyperscalers’ attempts at vertical integration can be understood as a strategic response to the continual disintegration and modularization of those subsystems in the middle of the value chain. The middle is being hollowed out, with IaaS and PaaS increasingly seen as an integrated I+PaaS offering in which the infrastructure layer delivers the value and the platform layer is table stakes because there exist open-source alternatives that can be easily plugged in (and if one hyperscaler doesn’t make it easy to plug in, then the other two will, thereby making composability the default strategy for all three). This vertical integration by cloud hyperscalers is taking the form of backwards integration through chip design as well as what is effectively forwards integration through an increased industry-specific focus for customers.
Through this lens, the Cloud ecosystem looks like one big game of different players trying to commoditize competitors and adjacent value chains while simultaneously integrating core subsystems along the value chain and making tactical moves to catalyze a reformation of the value chain to their advantage.
[More on hyperscaler strategies around non-public deployment models and vertical integration later.]
Porter’s observation that products have a tendency to become commoditized as buyers accumulate knowledge about them over time, thereby shifting the basis of competition towards price as the product becomes less differentiated is borne out in the data.
Not only is the rate of technological improvement of products in growing industries superlinear (and certainly in tech-oriented industries like those within adjacent to cloud computing), customers exhibit diminishing marginal utility (a supralinear function) for more of the same thing — the intersection of superlinear innovation and supralinear satiation leads to a situation of previously valuable interdependent systems overshooting customer needs.
Companies need to make sure they surf the right S-curve and manage their dis/advantages and competencies along the value chain as previously differentiable products and services become “good enough” and modularizable, the industry’s basis of competition shifts, and new areas for reintegration of subsystems appear at adjacent stages of the value chain.
More on the “reciprocal process of commoditization and decommoditization” within the Cloud’s value chain as we continue to flesh out this three body system.