The Anchor-and-Expand Principle

§ 01The idea in plain terms, before the theory

Before diving into academic frameworks, it helps to state the idea the way you might explain it to a friend over coffee. A new technology arrives. Nobody yet agrees on how its pieces should fit together — which plug goes where, which format wins, which chip talks to which software.

In that messy, undecided window, whoever controls one small, well-chosen piece of the puzzle can set the standard the rest of the industry has to build around. Gillette gave away razors and made its money on blades. HP priced printers low and profited from ink. Nespresso sold machines cheaply to sell capsules for years afterward. None of these companies wanted to be in the razor, printer, or machine business forever — the physical object was never the point. It was the entry ticket to a much bigger, much stickier business built on top of it.

The same logic shows up in far more technical settings — mainframe computers, PCs, smartphones, cloud servers, and now edge AI cameras and sensors — except the "razor" is a chip, a reference device, or a piece of infrastructure, and the "blades" are the software, subscriptions, or workflows that keep customers coming back. This article calls the razor the anchor and the blades the expansion layer, and it argues that the time a company gets to enjoy that advantage before competitors catch up — what we call the anchor's half-life — keeps shrinking, generation after generation.

§ 02What earlier thinkers already told us

Several respected strands of management theory already explain parts of this pattern, even though none of them, on their own, tell the whole story.

Transaction cost economicsWhy own anything at all?

Ronald Coase (1937) asked a deceptively simple question: if markets are efficient, why do firms exist rather than everyone just contracting everything out? Oliver Williamson (1975, 1985) answered that firms step in and own things when a relationship is too specific and too easy to be exploited if left to an outside contract — think of a supplier who builds a custom part just for you, and then worries you will walk away once they have made the investment. In plain terms: you buy the printer factory when you cannot trust an outside printer-maker to keep making exactly the printer your ink cartridges need.

Complementary assetsWhy does owning it protect your profits?

David Teece's (1986) work on "profiting from innovation" made a related but different point: the clever idea itself is rarely what makes money, because ideas get copied. What protects profit is control of the complementary assets — the factory, the distribution network, or, in a software business, the platform — especially when the core idea is easy to imitate. In plain terms: anyone can write a similar app; not everyone can own the hardware standard the app depends on.

Resource-based viewWhy is it hard to copy?

Jay Barney's (1991) resource-based view says a real, lasting advantage needs a resource that is valuable, rare, hard to imitate, and hard to substitute. Richard Rumelt (1984) called the mechanisms that make copying slow "isolating mechanisms" — often just the accumulated, unglamorous know-how of having deployed something in the field hundreds of times, which a fast-following competitor cannot download or buy.

Platform envelopmentHow does it spread rather than trap you?

Thomas Eisenmann, Geoffrey Parker, and Marshall Van Alstyne (2011) described "platform envelopment" — entering a neighbouring market by bundling your product with an incumbent's, riding their existing customer relationship instead of fighting for a new one from scratch. Jean-Charles Rochet and Jean Tirole's (2003) work on two-sided markets explains why a firm sitting between two dependent groups — say, hardware makers on one side and enterprise buyers on the other — can profit by keeping both sides calibrated to each other.

Ecosystems and keystonesHow do you become indispensable without owning everything?

Marco Iansiti and Roy Levien (2004), building on James Moore's (1993) earlier idea of the "business ecosystem," described keystone firms — companies whose importance comes not from owning the most, but from occupying the node the whole surrounding network depends on to function, much like a keystone species in an ecosystem. Annabelle Gawer and Michael Cusumano (2002) reached a similar conclusion studying Intel and Microsoft, calling it "platform leadership."

Each of these ideas is genuinely useful, and each is also, on its own, incomplete: they explain why owning a small anchor can work, but almost none of them ask how long it keeps working, or why that window seems to be shrinking every time a new technology cycle begins. That is the gap this article tries to close.

§ 03A simple new model: Anchor-and-Expand

Here is the core claim, stated as plainly as possible: when an industry's technical building blocks have not yet settled into an agreed standard, a company that owns a small, deliberately chosen anchor — the one component whose design quietly sets the performance ceiling for everything built on top of it — gains outsized influence, but only for a limited window.

The anchor is never meant to be the business itself. It exists to make a much larger expansion layer of software, workflows, or services credible, including to customers and partners who never touch the anchor directly.

Four simple ideas make the model concrete:

A quick, current example. An Indian edge-AI company called IndoAI builds a reference camera and compute unit called EdgeBox, alongside a software and deployment layer (its installable AI app platform) and a workflow tool for designing camera networks (its planning and adviser tools). EdgeBox is the anchor — a fixed, known-good hardware baseline used to prove out latency, accuracy, and security claims in the field. The software layer is the expansion — and, notably, it is also designed to run on third-party camera and sensor hardware from established manufacturers, so that IndoAI's software footprint can grow even on devices it never built or sold. That is the whole model in miniature, and we will return to it throughout the historical review below to show it is not a one-off case but a repeating pattern.

Three testable propositions follow, offered as a starting point for further research rather than settled fact:

§ 04Seeing it on a timeline

Two simple charts make this easier to picture than words alone.

Figure 1 lines up six broad technology eras and shows, roughly, how long each era's hardware anchor held its advantage. Mainframes held their edge for about 25 years; PCs, somewhat less; internet infrastructure, mobile hardware, and cloud infrastructure each show a shorter window than the one before; and today's edge-AI hardware — the category EdgeBox sits in — appears, on current evidence, to be the shortest window yet, as open-source AI models and cheap accelerator chips make cloning faster than ever.

Figure 2 shows the same idea from a single company's point of view over time: the anchor's value rises quickly, peaks, and then fades as the category standardizes and commoditizes, while the expansion layer's value builds slowly at first — because it depends on the anchor for credibility — and then compounds once it proves it can run independently of the anchor, on other people's hardware. The point where the two lines cross is the moment every company running this playbook should be planning for from the very first day, not discovering by surprise a few years in.

§ 05The pattern in real life: six stories, six decades

Theory is easier to trust once you can see it play out in real companies, with real dates. Here are six technology transitions, spanning more than sixty years, that all follow some version of the same anchor-and-expand arc — with brief notes throughout on how the current edge-AI wave, including IndoAI, echoes each one.

5.1 · MainframesIBM's System/360 (1964)

IBM's System/360, announced in 1964, is one of the clearest early anchor stories: a single, compatible hardware architecture across a whole family of machines, paired with the OS/360 operating system built to run across that family. The hardware was the anchor; the software, services, and applications built on top became the far bigger and stickier business. Because the wider computing industry took decades to agree on open, interoperable standards, IBM's anchor held its advantage for roughly twenty years — one of the longest windows in this entire review, and a useful reminder that P1 can produce very long windows when an entire industry's architecture is still up for grabs.

IndoAI's EdgeBox, working in a much younger and far less settled industry — edge AI hardware for surveillance and industrial sensing — is betting on nothing like a twenty-year window; the whole point of designing its software to also run on third-party hardware from day one is to not repeat IBM's later mistake of assuming the anchor's advantage would last indefinitely.

5.2 · Personal computersThe IBM PC and the rise of Wintel (1981–1985)

IBM's original PC, launched in 1981, was meant to be an anchor in the System/360 tradition. It did not hold for long. Columbia Data Products shipped a legally reverse-engineered, "clean room" clone of the PC's BIOS within about a year, in June 1982, and Compaq followed later that same year with its own million-dollar clean-room effort, shipping a fully IBM-compatible Portable in 1983.

Clean-room design — one team studying how the original behaves and writing down only what it does, a second team who never sees the original code implementing the same behaviour from that description — proved IBM's hardware could be legally copied within about eighteen months, one of the shortest anchor windows in this whole set. What survived the collapse of IBM's hardware advantage was the layer one level up: Microsoft's operating system and applications, designed from the outset to run on IBM's machines and its clones alike, which went on to dominate the industry for the next two decades.

It is a clean illustration of P2: the hardware edge vanished almost overnight, but a software layer built for portability from day one kept compounding long after. It is also the clearest historical parallel to what IndoAI is trying to do with its own software stack — build the expansion layer to survive the anchor's inevitable commoditization, rather than after the fact.

5.3 · SmartphonesTwo different bets (2007–2010)

Apple's iPhone, launched in 2007, followed a year later by the App Store, is a tightly closed anchor-and-expand play: proprietary hardware and software, never licensed to other phone makers, used to underwrite a hugely profitable app ecosystem. Google took the opposite bet with Android, launched commercially in 2008 on the HTC Dream: rather than owning hardware, Google made its operating system a free, horizontally licensed expansion layer that any phone maker could adopt — a software-side version of the "platform envelopment" move described by Eisenmann, Parker, and Van Alstyne.

Within three to four years, Android's licensed approach had reached more devices, across more manufacturers, than Apple's single integrated hardware line ever could alone. The lesson for edge AI is direct: a software layer designed for portability across other companies' hardware from the start can spread faster than owning the best hardware ever could on its own — which is exactly the bet IndoAI is making by licensing its software to run on established third-party camera and sensor hardware rather than insisting every customer buy EdgeBox.

5.4 · Cloud computingAmazon's Nitro and Graviton (2015–2018)

Amazon Web Services acquired the chip design firm Annapurna Labs in 2015 and used it to build the Nitro hypervisor and card system, formally unveiled at AWS re:Invent in November 2017, followed by the first generation of the Graviton processor in 2018. Neither Nitro nor Graviton was ever sold as a standalone product; both exist purely to give AWS's own cloud computing layer — its true expansion layer — a performance, cost, and security baseline that off-the-shelf hardware of the time could not match.

Unlike the PC-era anchor, cloned within about a year, AWS's cloud anchor has lasted far longer, because chip design is simply much harder and more expensive to copy than a BIOS was, and because AWS kept investing in new generations of the chip rather than treating it as a one-time move.

5.5 · A software anchorNVIDIA's CUDA (2006–present)

Not every anchor is a piece of hardware. NVIDIA released CUDA in 2006 as a way to let its own graphics chips run general-purpose computing workloads, not just graphics. Two decades later, CUDA's programming interfaces, not the chips themselves, are widely regarded as NVIDIA's real moat — an enormous body of software written by researchers and companies over nearly twenty years that would need to be rewritten to run efficiently on any other company's chips.

CUDA shows that an anchor does not have to be a physical object at all; it can be a software interface that becomes so deeply embedded in everyone else's tooling that switching away is more expensive than staying. It is also, alongside the AWS case, a reminder that anchor half-life can sometimes stretch on for a very long time when the switching costs for everyone else are high enough — which is exactly the kind of durable, interface-level advantage a much younger company like IndoAI cannot assume it will get for free, and has to actively design for instead.

5.6 · Edge AI todayA faster, louder version of the same story

The current wave — on-device AI models, camera-based sensing, and autonomous decision-making, all evolving at once — shows the same underlying pattern playing out on a visibly compressed timescale compared with every case above. Open-source AI models, cheap accelerator chips, and fast, well-documented hardware cloning mean a hardware anchor launched today faces a shorter road to being copied than any of the anchors described in this article.

IndoAI's own stack is a useful, current-generation illustration of how a smaller company navigates this compressed window in practice: EdgeBox sets the performance and security baseline; the AI camera platform and its installable apps carry that credibility into a software layer built to run on other companies' cameras and sensors too; and the adviser and planning tools turn the whole thing into a workflow customers use regardless of whose hardware they are standing in front of. Consistent with P1, this is exactly the high-uncertainty condition in which an anchor earns its keep; consistent with P3, the company building the expansion layer for outside hardware from day one — rather than years later, once competitors have already caught up — is the one best placed to keep the advantage once the hardware itself stops being special.

The anchor buys you time. The expansion layer is what you are supposed to build with it.

§ 06Where the idea runs out

No model applies everywhere, and this one should not be stretched past its limits. Once an industry's interfaces are already standardized, owning an anchor tends to destroy value rather than create it — Clayton Christensen's (1997) well-known work on disruptive innovation, and his later work with Richard Rosenbloom (1995) on architectural versus component-level knowledge, both explain why integrated players lose their edge once everyone understands how the pieces fit together.

Utterback and Abernathy's (1975) classic idea of the "dominant design" is really just another name for what this article calls the diffusion threshold — the moment competition stops being about architecture and starts being about cost and incremental improvement. Nor does the model suggest every company should build its own anchor: a smaller firm without the capital to sustain hardware through its whole half-life is often better off licensing or contracting for anchor access from a specialist, exactly as Teece's own framework recommends when appropriability conditions allow it.

Finally, the claim that anchor windows keep shrinking (P3) should be read as a general historical tendency, illustrated by the cases above, not a proven law — the AWS and NVIDIA cases both show that a well-defended anchor can last far longer than the trend line would suggest, and testing this properly across a much larger set of industries is the obvious next step for future research.

§ 07Conclusion

None of the older theories discussed here are being replaced. Transaction cost economics still explains why a company chooses to own something in the first place; complementary-asset theory still explains when that ownership protects profit; the resource-based view still explains why it is hard to copy; platform and ecosystem theory still explain how it spreads. What the Anchor-and-Expand model adds is simply the piece of the puzzle those theories leave open: time.

Every company navigating a period of technical uncertainty — whether it is IBM in 1964, Compaq in 1982, Google in 2008, Amazon in 2017, or a much smaller edge-AI company like IndoAI navigating its own hardware and software stack today — is really answering the same three questions, just on a faster and faster clock each time: how small should the anchor be, how quickly should the expansion layer be built to survive without it, and how much time is really left before the whole industry agrees on a standard and the anchor stops being special.

§ 08What this means for IndoAI — and for anyone buying AI video in India

Read against this framework, IndoAI's structure is not an accident of engineering; it is a deliberate answer to P3. EdgeBox is the anchor, and it is meant to be a small one. Its job is to make otherwise unfalsifiable claims checkable in the field: that inference happens on site rather than in someone else's cloud, that alert latency is measured in hundreds of milliseconds rather than seconds, that footage and biometric templates never leave the premises, and that a known, fixed compute baseline underpins every accuracy figure quoted in a proposal. Those are the claims a CISO or a facility head can actually test on a Tuesday afternoon, and no amount of software marketing substitutes for a box that demonstrably does them. But EdgeBox was never designed to be the thing customers keep buying forever — in a category where open models and commodity accelerators shorten the copy cycle every year, a hardware-only moat would be the shortest-lived asset on the balance sheet.

The expansion layer is where the compounding happens, and it was built hardware-independent from the start rather than retrofitted later. IndoAI's installable AI apps and analytics run against RTSP streams from cameras and NVRs the company never manufactured — existing Indian estates, established global brands, mixed-vintage installations that no one is going to rip out. That choice is the Wintel lesson and the Android lesson applied deliberately: the software footprint can grow on other people's boxes, which is the only mechanism that lets a smaller company outrun its own anchor's half-life. The free planning and adviser tools extend the same logic one level further out — they make IndoAI useful at the design stage, before any hardware decision has been taken at all, on estates that may end up only partly IndoAI. A workflow tool that a system integrator opens on every project is a far stickier asset than an appliance they specify on some of them.

For buyers, the framework yields one practical diagnostic that cuts through most vendor comparisons: ask whether the vendor's analytics run on hardware the vendor does not sell. If the answer is no, you are not buying intelligence — you are buying that vendor's anchor, and you are absorbing the commoditization risk of their hardware on a CCTV refresh cycle that typically runs seven to ten years, far longer than any anchor half-life in Figure 1 except IBM's. If the answer is yes, the vendor has already made the bet that their software has to earn its place on merit, which is exactly the incentive alignment you want from a supplier you will still be depending on in 2033. That is the same question this article has been asking of IBM, Compaq, Apple, Google, Amazon and NVIDIA for sixty years, only now it is on your procurement checklist. For the concrete architecture version of this argument, see our comparison of AI edge boxes versus NVRs in Indian CCTV and our guide to integrating AI analytics with any existing CCTV system.

§ 09Frequently asked questions