In 1984 an NSF panel suggested a two-stage approach: first to link the NSF supercomputer centers and then to construct a high-speed backbone network for connecting existing and future regional, campus, and other networks. In 1985 six federally funded supercomputer centers were linked in a 56 Kbps network, and in 1987 bids were solicited on the NSFNET backbone. Merit Network, Inc., which had experience linking four universities in Michigan, was selected along with partners MCI, IBM, and the State of Michigan.
By 1988, 13 backbone nodes linked supercomputer centers and regional networks in a 1.544 Mbps network, which was upgraded to 45 Mbps in 1991.
In 1990, the NSF began offering Internet connection grants to US research and education institutions. Schools could apply for $20,000 in connection assistance, typically a router and a link to a regional NSFNet point of presence. NSF also made grants to connect foreign research and education networks to the NSFNet, eventually linking 28 research and education networks in 26 nations.
When decommissioned in April, 1995, NSFNET was the global backbone, linking 28,470 domestic and 22,296 foreign networks. The NSFNet had three characteristics that led to its success and impact.
1. NSFNet was highly leveraged.
The NSFNet was a high-return investment. It became the first global Internet backbone at a cost of less than $100 million to the US taxpayer. The aggregate cost of staff, equipment and installation of university local area networks far exceeded the cost of the NSFNet program. The following table shows the cost of NSFNet and the important federally funded networks that preceded it.
| Project | Cost ($Million) |
| Morse telegraph | .03 |
| ARPANet | 25 |
| CSNET | 5 |
| NSFNet backbone | 57.9 |
| NSF US higher education connections | 30 |
| NSF international connections | 6 |
2. NSFNet and the ARPANet before it were research projects, designed by highly qualified researchers.
Although these networks eventually went into production, they were applied research projects. At the time they were being designed, active debates on packet versus circuit switching, OSI versus TCP/IP, the separation of the network and transport layers, etc. were taking place. Routing algorithms had to be invented and the Domain Name System was being defined.
Highly qualified researchers from leading universities and research labs were brought in to design and oversee the implementation of these networks. The work was not done by career government employees, but by top scientists on temporary assignment who funded research with grants and oversaw deployment after contracts were awarded by competitive bidding.
3. NSFNet was a "dumb," "end-to-end" network.
We have seen that the NSF funded only the Internet backbone, and left the bulk of the funding to connecting networks. Application development was also left to the users. The IP protocol was developed to route packets of data from one computer to another, nothing else. The network was "dumb," since it ignored the content of those packets. They were treated the same if they contained music or pictures, email messages or images from Mars. The data was interpreted and acted upon by application programs running on the computers connected to the network. Applications were invented by users at the network ends, not by the operators of the network. Every user was a potential application inventor and developer. Every user is a content provider.
