When to be Open, When to be Closed

Having reviewed business models that support open-source software development, we can now approach the general question of when it makes economic sense to be open-source and when to be closed-source. First, we must be clear what the payoffs are from each strategy.

The closed-source approach allows you to collect rent from your secret bits; on the other hand, it forecloses the possibility of truly independent peer review. The open-source approach sets up conditions for independent peer review, but you don't get rent from your secret bits.

The payoff from having secret bits is well understood; traditionally, software business models have been constructed around it. Until recently, the payoff from independent peer review was not well understood. The Linux operating system, however, drives home a lesson that we should probably have learned years ago from the history of the Internet's core software and other branches of engineering—that open-source peer review is the only scalable method for achieving high reliability and quality.

In a competitive market, therefore, customers seeking high reliability and quality will reward software producers who go open-source and discover how to maintain a revenue stream in the service, value-add, and ancilliary markets associated with software. This phenomenon is what's behind the astonishing success of Linux, which came from nowhere in 1996 to be the second-most-popular operating system in the business server market by mid-2000 (and some surveys actually showed it passing Microsoft's share in late 2000). In early 1999 IDC projected that Linux would grow faster than all other operating systems combined through 2003; this projection has held true so far.

An almost equally important payoff of open source is its utility as a way to propagate open standards and build markets around them. The dramatic growth of the Internet owes much to the fact that nobody owns TCP/IP; nobody has a proprietary lock on the core Internet protocols.

The network effects behind TCP/IP's and Linux's success are fairly clear and reduce ultimately to issues of trust and symmetry—potential parties to a shared infrastructure can rationally trust it more if they can see how it works all the way down, and will prefer an infrastructure in which all parties have symmetrical rights to one in which a single party is in a privileged position to extract rents or exert control.

It is not, however, actually necessary to assume network effects in order for symmetry issues to be important to software consumers. No software consumer will rationally choose to lock itself into a supplier-controlled monopoly by becoming dependent on closed source if any open-source alternative of acceptable quality is available. This argument gains force as the software becomes more critical to the software consumer's business—the more vital it is, the less the consumer can tolerate having it controlled by an outside party.

There's a flip side to this. Economists know that, in general, asymmetric information makes markets work poorly. Higher-quality goods get driven out when it's more lucrative to collect rent on privileged information than it is to invest in producing better products. In general, not just in software, secrecy is the enemy of quality.

Finally, an important customer payoff of open-source software related to the trust issue is that it's future-proof. If sources are open, the customer has some recourse if the vendor goes belly-up. This may be particularly important for widget frosting, since hardware tends to have short life cycles, but the effect is more general and translates into increased value for all kinds of open-source software.

When the rent from secret bits is higher than the return from open source, it makes economic sense to be closed-source. When the return from open source is higher than the rent from secret bits, it makes sense to go open source.

In itself, this is a trivial observation. It becomes nontrivial when we notice that the payoff from open source is harder to measure and predict than the rent from secret bits—and that said payoff is grossly underestimated much more often than it is overestimated. Indeed, until the mainstream business world began to rethink its premises following the Mozilla source release in early 1998, the open-source payoff was incorrectly but very generally assumed to be zero.

So how can we evaluate the payoff from open source? It's a difficult question in general, but we can approach it as we would any other predictive problem. We can start from observed cases where the open-source approach has succeeded or failed. We can try to generalize to a model that gives at least a qualitative feel for the contexts in which open source is a net win for the investor or business trying to maximize returns. We can then go back to the data and try to refine the model.

From the analysis presented in The Cathedral and the Bazaar [CatB], we can expect that open source has a high payoff where (a) reliability/stability/scalability are critical, and (b) correctness of design and implementation is not readily verified by means other than independent peer review. (The second criterion is met in practice by most non-trivial programs.)

A consumer's rational desire to avoid being locked into a monopoly supplier will increase its interest in open source (and, hence, the competitive-market value for suppliers of going open) as the software becomes more critical to that consumer. Thus, another criterion (c) pushes towards open source when the software is a business-critical capital good (as, for example, in many corporate MIS departments).

As for application area, we observed above that open-source infrastructure creates trust and symmetry effects that, over time, will tend to attract more customers and to outcompete closed-source infrastructure; and it is often better to have a smaller piece of such a rapidly-expanding market than a bigger piece of a closed and stagnant one. Accordingly, for infrastructure software, an open-source play for ubiquity is quite likely to have a higher long-term payoff than a closed-source play for rent from intellectual property.

In fact, the ability of potential customers to reason about the future consequences of vendor strategies and their reluctance to accept a supplier monopoly implies a stronger constraint; without already having overwhelming market power, you can choose either an open-source ubiquity play or a direct-revenue-from-closed-source play—but not both. (Analogues of this principle are visible elsewhere—e.g., in electronics markets where customers often refuse to buy sole-source designs.) The case can be put less negatively: where network effects (positive network externalities) dominate, open source is likely to be the right thing.

We may sum up this logic by observing that open source seems to be most successful in generating greater returns than is closed source in software that (d) establishes or enables a common computing and communications infrastructure.

Finally, we may note that purveyors of unique or just highly differentiated services have more incentive to fear the copying of their methods by competitors than do vendors of services for which the critical algorithms and knowledge bases are well understood. Accordingly, open source is more likely to dominate when (e) key methods (or functional equivalents) are part of common engineering knowledge.

The Internet core software, Apache, and Linux's implementation of the standard Unix API are prime exemplars of all five criteria. The path towards open source in the evolution of such markets are well-illustrated by the reconvergence of data networking on TCP/IP in the mid-1990s following fifteen years of failed empire-building attempts with closed protocols such as DECNET, XNS, IPX, and the like.

On the other hand, open source seems to make the least sense for companies that have unique possession of a value-generating software technology (strongly fulfilling criterion (e)) which is (a) relatively insensitive to failure, which can (b) readily be verified by means other than independent peer review, which is not (c) business-critical, and which would not have its value substantially increased by (d) network effects or ubiquity.

As an example of this extreme case, in early 1999 I was asked "Should we go open source?" by a company that writes software to calculate cutting patterns for sawmills that want to extract the maximum yardage of planks from logs. My conclusion was ``No.'' The only criterion this comes even close to fulfilling is (c); but at a pinch, an experienced operator could generate cut patterns by hand.

Note that my answer night have been very different if the cut-pattern calculator had been written by a sawmill-equipment manufacturer. In that case, opening the code would have increased the value of the associated hardware they were selling. Also note that if some open-source cut-pattern calculator already existed (perhaps the one written by the sawmill-equipment manufacturer) the closed-source product would have trouble competing with it—not so much for reasons of price, but because customers would perceive on open-source advantage in customizability and other traits.

An important point is that where a particular product or technology sits on these scales may change over time, as we'll see in the following case study.

In summary, the following discriminators push towards open source:


Reliability/stability/scalability are critical.


Correctness of design and implementation cannot readily be verified by means other than independent peer review.


The software is critical to the user's control of his/her business.


The software establishes or enables a common computing and communications infrastructure.


Key methods (or functional equivalents of them) are part of common engineering knowledge.

The history of id software's best-selling game Doom illustrates ways in which market pressure and product evolution can critically change the payoff magnitudes for closed versus open source.

When Doom was first released in late 1993, its first-person, real-time animation made it utterly unique (the antithesis of criterion (e)). Not only was the visual impact of the techniques stunning (far exceeding the flat-world animation in its predecessor Wolfenstein 3D), but for many months nobody could figure out how it had been achieved on the underpowered microprocessors of that time. These secret bits were worth some very serious rent. In addition, the potential payoff from open source was low. As a solo game, the software (a) incurred tolerably low costs on failure, (b) was not tremendously hard to verify, (c) was not business-critical for any consumer, (d) did not benefit from network effects. It was economically rational for Doom to be closed source.

However, the market around Doom did not stand still. Would-be competitors invented functional equivalents of its animation techniques, and other ``first-person shooter'' games like Duke Nukem began to appear. As these games ate into Doom's market share, the value of the rent from secret bits went down.

On the other hand, efforts to expand that share brought on new technical challenges—better reliability, more game features, a larger user base, and multiple platforms. With the advent of multiplayer `deathmatch' play and Doom gaming services, the market began to display substantial network effects. All this was demanding programmer-hours that id would have preferred to spend on the next game.

From the time the game was first released, id had looked benignly on the publication of technical specs that helped people to create data objects for the game, and occasionally cooperated directly with hackers by answering specific questions or publishing an existing specs document of their own. They also encouraged the Internet distribution of new Doom data.

The technical and market trends raised the payoff from opening the source; Doom's opening of specifications and encouragement of third-party add-ons increased the perceived value of the game and created a secondary market for them to exploit. At some point the payoff curves crossed over and it became economically rational for id to shift to making money in that secondary market (with products such as game-scenario anthologies) and then open up the Doom source. Sometime after this point, it actually happened. The full source for Doom was released in late 1997.

Doom makes an interesting case study because it is neither an operating system nor communications/networking software; it is thus far removed from the usual and obvious examples of open-source success. Indeed, Doom's life cycle, complete with crossover point, may be coming to typify that of applications software in today's code ecology—one in which communications and distributed computation both create serious robustness/reliability/scalability problems only addressible by peer review, and frequently cross boundaries both between technical environments and between competing actors (with all the trust and symmetry issues that implies).

Doom evolved from solo to deathmatch play. Increasingly, the network effect is the computation. Similar trends are visible even in the heaviest business applications, such as ERP systems, as businesses network ever more intensively with suppliers and customers—and, of course, they are implicit in the whole architecture of the World Wide Web. It follows that almost everywhere, the open-source payoff is steadily increasing.

If present trends continue, the central challenge of software technology and product management in the next century will be knowing when to let go—when to allow closed code to pass into the open-source infrastructure in order to exploit the peer-review effect and capture higher returns in service and other secondary markets.

There are obvious revenue incentives not to miss the crossover point too far in either direction. Beyond that, there's a serious opportunity risk in waiting too long—you could get scooped by a competitor going open-source in the same market niche.

The reason this is a serious issue is that both the pool of users and the pool of talent available to be recruited into open-source cooperation for any given product category is limited, and recruitment tends to stick. If two producers are the first and second to open-source competing code of roughly equal function, the first is likely to attract the most users and the most and best-motivated co-developers; the second will have to take leavings. Recruitment tends to stick, as users gain familiarity and developers sink time investments in the code itself.