In contrast to the U.S., most AI product development and consulting is done by the JCMs rather than small firms or specialized consulting firms. (This is, of course, the usual case in Japan, but is even more extreme in the case of KB systems. The large software-only houses have very little presence.) Table 7.1 shows the number of shells shipped since KB systems were first commercialized, approximately 10 years ago.
If we assume that the average workstation costs $40,000 (actually ranging from $16,000 to $80,000) and that a shell costs between $30,000 and $35,000, then the total cost to a customer is $75,000. Multiplying by the number of shells shipped gives an overall revenue estimate for hardware and software.
Total KBS Shells Shipped
Revenues from consulting or system building, which were not available to us, are not included here and may possibly be the major revenue component. The estimates given in Table 7.2 for hardware and software revenues are probably too high, because a shell might be bought to run on an existing workstation. Also, the shell numbers include internal shipments (one-third of the total in Toshiba's case). Overall, the business appears to be between $3 and $20 million per annum per JCM (again, not counting revenues from consulting or system building for external clients).
Estimated Revenue From KBS Shell Sales
($ in millions)
($ in millions)
We observed that there are many internal applications, that is, applications built by the JCMs for themselves. The return on investment (ROI) was not made available, but there was a clear need and usage. Areas of application included intelligent aids for engineering, production planning, configuration, and marketing.
The main platforms are engineering workstations and PCs, using the client/server model. Symbolic processors are not being used at all, and mainframes are in declining usage. As to parallel processing, no one we spoke with sees a commercial role for PIM (from ICOT); they believe other approaches, such as those taken by Connection Machine and Intel Paragon, are more promising. However, the disinterest in the PIM may relate more to the language issue (PROLOG) than to the computer architecture. Task-specific shells for diagnosis and planning are just coming out, and the hope is that they will have a wide market since they are intended to be used by end-users, not system engineers. Sales are small compared to the expectations of several years ago. Presently, sales growth is flat and growth in the number of applications is also flat (exception: Hitachi). No Japanese company the JTEC panel visited had written an application that is being resold. This parallels U.S. experience, where generic products for the financial and insurance industry, offered by several start-up companies, did not succeed. However, some PC-based products such as Tax Cut are selling in the United States. In Japan, application products that rely heavily on AI are not selling well. For instance, only 20 copies of ICAD and 10 copies of Concept Modeler have been sold; in the United States hundreds of ICAD systems have been sold. Poor sales of these systems may be due to a language problem (no kanji version), and/or to the fact that there is no version that originated in Japan, and/or to the lack of an effective marketing channel.