Thinking Machines Corporation

In May 1985, Thinking Machines became the third company to register a.com domain name (think.com). The company became profitable in 1989, in part because of its contracts from the Defense Advanced Research Projects Agency (DARPA). The next year, it sold US$65 million worth of hardware and software, making it the market leader in parallel supercomputers. Thinking Machines' primary supercomputer competitor was Cray Research. Other parallel computing competitors included nCUBE, nearby Kendall Square Research, and MasPar, which made a computer similar to the CM-2, and Meiko Scientific, whose CS-2 was similar to the CM-5. In 1991, DARPA and the United States Department of Energy reduced their purchases amid criticism they were unfairly favoring Thinking Machines at the expense of Cray, nCUBE, and MasPar. Tightening export laws also prevented the most powerful Connection Machines from being exported. By 1992, the company was losing money, and CEO Sheryl Handler was forced out.

In August 1994, Thinking Machines filed for Chapter 11 bankruptcy. The hardware portion of the company was purchased by Sun Microsystems, and TMC re-emerged as a small software company specializing in parallel software tools for commodity clusters and data mining software for its installed base and former competitors' parallel supercomputers. In December 1996, the parallel software development section was also acquired by Sun Microsystems.

Thinking Machines continued as a pure data mining company until it was acquired in 1999 by Oracle Corporation. Oracle later acquired Sun Microsystems, thus re-uniting much of Thinking Machines' intellectual property.

The program wide area information server (WAIS), developed at Thinking Machines by Brewster Kahle, would later be influential in starting the Internet Archive and associated projects, including the Rosetta Project as part of Danny Hillis' Clock of the Long Now.

Architect Greg Papadopoulos later became Sun Microsystems's chief technology officer (CTO).

On the hardware side, Thinking Machines produced several Connection Machine models (in chronological order): the CM-1, CM-2, CM-200, CM-5, and CM-5E. The CM-1 and 2 came first in models with 64K (65,536) bit-serial processors (16 processors per chip) and later, the smaller 16K and 4K configurations. The Connection Machine was programmed in a variety of specialized programming languages, including *Lisp and CM Lisp (derived from Common Lisp), C* (derived by Thinking Machines from C), and CM Fortran. These languages used proprietary compilers to translate code into the parallel instruction set of the Connection Machine. The CM-1 through CM-200 were examples of single instruction, multiple data (SIMD) architecture using a hypercube interconnect to reduce hop-count and latency, while the later CM-5 and CM-5E were multiple instruction, multiple data (MIMD) that combined commodity SPARC processors and proprietary vector processors in a fat tree computer network.

All Connection Machine models required a serial front-end processor, which was most often a Sun Microsystems workstation, but on early models could also be a Digital Equipment Corporation (DEC) VAX minicomputer or Symbolics Lisp machine.

Thinking Machines also introduced an early commercial redundant array of independent disks (RAID) 2 disk array, the DataVault, circa 1988.