Why Nobody Else Built It
Forty-three carriers. Every asset required. A three-person startup built it instead.
Digital Island was first, and being first is inseparable from why. Forty-three operators were in the market with more capital, more plant, more engineers, and more customers. Every one of them could have built this. None did. They watched us do it.
The reasons are specific, they are documented, and there are eleven of them. Sections twelve and thirteen close the record with what the build produced and the two transactions that prove it.
1. Commerce was prohibited on the backbone until 1995
TCP/IP was twenty-two years old. The Web was six. Neither had carried commerce at global scale, and the reason was not technical.
From 1987 to 1995 the Internet’s backbone was NSFNET, operated under a cooperative agreement NSF awarded in November 1987 to Merit Network with IBM and MCI as partners, and governed by the NSFNET Acceptable Use Policy, which listed “use for for-profit activities” among its Unacceptable Uses. The spine of the Internet was not permitted to carry a sale.
The prohibition had a landlord with a side door. In September 1990 the partners formed Advanced Network and Services, installing IBM’s Allan Weis as president and Merit’s board chairman Douglas Van Houweling as ANS chairman, and in May 1991 ANS created the for-profit ANS CO+RE to sell commercial carriage over the same subsidized backbone the AUP closed to everyone else. The excluded commercial providers, PSINet, UUNET, and CERFnet, formed the Commercial Internet Exchange in 1991 to interconnect around that chokepoint, and ANS initially refused to connect to it. Congress examined the arrangement at the March 12, 1992 Boucher hearing. The NSF Inspector General’s 1993 review, OIG 93-01, found no illegality, and faulted NSF for extending the agreement without competition, for weak oversight, and for never accounting for the public’s share when the position passed into ANS. When ANS was sold to America Online in 1995 for roughly $35 million, the Treasury that funded the buildout received nothing.
The purpose of the structure is stated by its own governance. Mort Rahimi, who chaired Merit’s board, states on his own faculty homepage that Merit built the NSFNET, “which became the Internet when Advanced Network Services was formed to make the NSFNet available for commercial use.” The chairman’s biography says ANS existed to commercialize the backbone. The Acceptable Use Policy said nobody else could. Neither sentence is mine.
So for the entire period during which the protocol architects are credited with creating the Internet, the Internet was structurally prohibited from doing the thing this record is about. When NSF retired NSFNET on April 30, 1995, what it kept was MCI’s vBNS, a research network under a five-year, $50 million cooperative agreement. Still not commerce.
The commercial Internet has a start date. NSFNET decommissioned April 30, 1995. The Telecommunications Act signed February 8, 1996, and for the first time a new entrant could contract capacity, negotiate interconnection, and compete against incumbents rather than ask their permission. Nobody built a commerce-grade global fabric before 1996 because nobody lawfully could. That is why activation is a separate contribution from invention. Not as a rhetorical distinction. As a legal one.
MCI had every advantage going into that window: the backbone heritage, the circuits, the operations, the NSF’s confidence, and the co-author of the protocol carrying Internet Architecture in his title. What MCI sold was access to its own network, and carriage across its own portion of the route. Where MCI’s network ended, so did MCI’s responsibility.
Within 90 days of the Act I produced the global wide-area network diagrams. Nine months after it, I signed Cisco.
2. The market was twenty countries and it rounded to zero
At the start of 1996, only about 20 of the 193 United Nations-recognized countries participated in the Internet in any meaningful way. The measure was simple: where Cisco shipped gear, the Internet was in use. Where it didn’t, it wasn’t. Twenty countries, a handful of metros inside each, and almost nothing between those metros. Regional ISPs peered locally or not at all.
That was the entire addressable market. Any CFO looking at it said no, and they were right to.
3. It was all of it or none of it
Twenty markets is forty half-circuits: a foreign half in each country and the domestic half it lands on. Every carrier on Earth could sell exactly one of those. None could sell the set, because none owned both ends of anything.
The counterparty math was twenty-one: twenty foreign carriers, one per market, plus the United States, where every domestic half landed. Twenty-one counterparties. Twenty-one jurisdictions. Twenty-one regulators. Every dollar committed before one packet crossed and paid. And a fabric that reaches eighteen of twenty is not a fabric.
Inside the United States I split the path on purpose. The local loops were ordered from a separate provider from the interexchange carrier, ILEC or CLEC direct, so the IXC could never call its outage the local carrier’s problem. Every segment of every route had a named owner, a separate invoice, and nowhere to point, and I held insight into the full route path end to end. That was not procurement convenience. That was accountability engineered into the purchase orders, and it is the same discipline that later made the SLA signable: you can only sign for a path when no vendor on it can hide behind another.
There was no pilot. No toe in the water. No route that pays for itself while you grow into the rest. It was the whole thing or nothing, against a market that did not exist on paper yet.
4. Forty-three competitors cannot build a fabric
A global fabric requires competitors to carry each other’s traffic into each other’s markets. What was in it for NTT to light a path that lets France Télécom’s customer reach Tokyo? What was in it for Bezeq to carry a session that lets Telstra sell into Israel? Every one of them would have been building the thing that let the other forty-two into the market they owned.
Competitors compete. That is not a failure of imagination. That is the definition of the job.
And suppose the circuits were funded and every regulator said yes. The fabric still requires each carrier to announce full routes to the others, and a full table is a customer list. Every prefix, every downstream, every account, handed to the rival most motivated to take it. No carrier does that. Peering between competitors was filtered, selective, and grudging for exactly that reason, and it still is.
Filtered announcements produce reachability with holes in it and nobody accountable for the whole path. That is the condition this network was built to end.
5. Being a carrier was the disqualification
International telecom in 1996 ran on half-circuits, bilateral operating agreements, accounting rates, and correspondent relationships. AT&T could not unilaterally control a path into Japan. It had a relationship with a carrier in Japan, governed by an agreement neither party could rewrite alone. Every incumbent was a party to that regime, and the regime is why no incumbent controlled an end-to-end path.
I was not a party to it. I bought half-circuits as a customer, from both ends, and terminated them under my own ASN. Same regime, opposite side of the table. The thing that let me assemble an end-to-end path was precisely that I was not a correspondent carrier. Being one was the disqualification.
Seven of the forty-three could not even try. Pacific Bell, Southwestern Bell, Bell Atlantic, NYNEX, Ameritech, US West, and BellSouth were barred by Section 271 of the Act from in-region interLATA service pending FCC approval. The first approvals came in 1999. In November 1996, no lawful offer could exist.
6. Every incentive on every balance sheet pointed at no
Every incumbent’s enterprise data revenue came from selling the oversubscribed, DE-bit-managed transit this architecture bypasses. To offer what I offered, a carrier would have had to sell clear-channel IPLC capacity terminated under a customer’s independent AS, with routing policy handed to that customer at the demarcation. That meant displacing its own installed base, surrendering the margin oversubscription produced, and giving routing control to a customer it preferred dependent.
I asked them. In person, in their own offices, on six continents. The answers sorted three ways and they sorted the same in every language.
70 percent: “After obtaining and paying for your monthly internet access, you can use the internet at no additional cost. Nobody will pay extra to have their web content distributed globally for a premium user experience. Your business model’s proposition is stupid; nobody will pay for it.”
20 percent: “What makes you think you’re so smart? If it were such a good idea, somebody else would have thought of it and done it already; you’re wasting my time with this.”
10 percent: “I totally get it, it’s a total no-brainer. How much port capacity do you want? How many cabinets do you need? How about DNS? Do you need any IPs? Is 15 amps per rack enough?”
That is the market pricing this proposition at zero, in the words of the people who priced it, and it priced the same in every language. The 90 percent were not wrong about their own incentives. They were wrong about one thing only: whether anyone would pay.
I got the same verdict from my own employer. On August 5, 1996, a Sprint regional manager pulled me into a conference room and said: “Digital Island will never happen; stop working on it immediately.” Ron’s offer letter is dated the same day. I accepted it and resigned. Three months later I signed Cisco.
7. MCI had the best claim on Earth
The verdict was no different in the United States. AT&T and Sprint had the plant, the capital, and the carrier relationships. MCI had all of that, and more Internet than any company alive.
From 1987 to 1995, MCI ran the backbone. It won the NSF contract with IBM in November 1987 and supplied the circuits. It formed Advanced Network and Services with Merit and IBM in 1990 to operate the NSFNET backbone. It built MCI Mail, and in 1989 an NSF-approved experiment connected MCI Mail to the Internet, making MCI’s own product the first commercial email service on the network while the Acceptable Use Policy barred general commerce for every other company in America. The single exception the gate ever granted went to the gate-keeper’s partner. When NSF retired NSFNET on April 30, 1995, it kept MCI: a five-year, $50 million cooperative agreement for the vBNS. And in February 1994 it brought back Vint Cerf as its senior Internet architecture executive; his title by 1996 was Senior Vice President, Internet Architecture and Technology.
Now note what the backbone MCI operated actually was. NSFNET ran under acceptable use policies that constrained commercial traffic. The vBNS that replaced it was a research network. For eight years, the company that ran the Internet’s spine ran a spine that was not allowed to carry a sale. That is not MCI’s failing. That was the policy, and it is the single most important fact in this record.
Then the window opened, and every asset required to build the fabric was already on MCI’s balance sheet: the backbone heritage, the circuits, the operations, the NSF’s confidence, and the co-author of the protocol with Internet Architecture in his title.
MCI had the best claim on Earth to build a commerce-grade global utility, and it did not. It sold transit and access piecemeal. WorldCom-MCI was my largest single-vendor cost center expense globally by international IPLC half-circuits. I know exactly what they sold, because I bought more of it from them than from anyone else, and what they sold stopped at their border.
And then the record closes MCI’s window from the other side. On November 10, 1997, WorldCom and MCI announced their merger. On July 15, 1998, the Department of Justice cleared it on the condition that MCI divest its Internet business entirely, and MCI sold internetMCI to Cable & Wireless for approximately $1.75 billion, the largest divestiture of a company in merger history. Mark the timing against this record: during the exact window in which Digital Island was onboarding E*TRADE and Schwab and approaching its NASDAQ listing, the company with the best claim on Earth was not declining to build the path product. It was under government order to exit the Internet business altogether. The structural incapacity was joined by a legal one, and the asset itself deteriorated after transfer, ending in litigation and a $200 million settlement in March 2000.
8. Why didn’t Cisco just buy it from MCI?
This is the question, and it should be asked in its harshest form.
Cisco Systems built the routers the Internet ran on. Its engineers could evaluate any network claim on the merits. It had unlimited capital and standing relationships with every major carrier on Earth. Picking up the phone to MCI cost nothing. MCI had run the backbone since 1987, held the vBNS under a $50 million NSF agreement, built the first commercial service ever connected to the Internet, and employed Vint Cerf as its senior Internet architecture executive.
Now look at what Cisco was actually buying. Not connectivity. Cisco had connectivity. Cisco sold the boxes that made connectivity. What Cisco needed was a 16MB IOS image arriving intact in Singapore, Moscow, and Tel Aviv, in one session, every time, with someone contractually on the hook when it didn’t.
MCI could sell Cisco a circuit to Tokyo. MCI could not sell Cisco that outcome, because MCI did not control the far end of any path outside its own network. Neither could AT&T, BT, NTT, or anyone else, for the reasons above. What every one of them could offer was a segment, and a segment cannot carry an end-to-end guarantee no matter who signs it or how much they’re paid.
So Cisco’s choice was not between a three-person startup and MCI. It was between the outcome and no outcome. MCI wasn’t a worse version of what Digital Island sold. MCI was in a different business, and the business Cisco needed did not exist as a product anywhere on Earth.
Now measure the counterparty Cisco accepted instead.
Three people. The CEO, Ron Higgins, had been a Director of Sales at Radius, a computer hardware company. Sanne Higgins came from media communications. Neither had ever worked a day in telecom, internetworking, or commercial website operations. There was no network. There was no funding. There was a diagram, an AS number issued in August, and me.
Cisco signed a $300,000 agreement with that, in November 1996, and put cisco.com on it.
That is the measure of the alternative. Not that Cisco was impressed by us. That Cisco, with every carrier on the planet returning its calls, found nothing on any of those calls worth choosing over three people and a promise, because the promise was the only one anybody would make.
9. The physics were never the obstacle. The signature was.
Clear-channel IPLC with no oversubscription and no discard eligibility has a latency floor set by propagation and switching. Any carrier engineer could compute it. That was never scarce.
What was scarce was a party who could sign it. A service level agreement binds only what the signer controls, and no incumbent controlled the far end of the path. MCI could not sign anywhere to anywhere. Neither could AT&T, BT, NTT, or any of the other forty-three.
I could, because I controlled all of it. That is why the number went into an instrument, and it is why Cisco, the company whose routers ran the Internet, with relationships with every carrier on Earth, signed a $300,000 agreement with three people in November 1996.
10. Why me
Not capital. Not fiber. Not engineers. Every one of the forty-three had more of all three.
I had been the merchant. In 1995 I ran PerfectWheels.com out of my garage. I watched checkout sessions die mid-order. I was on the side of the counter that loses money when a session hangs, and that is a different education than any carrier could give you. The carriers were selling connectivity. I was trying to sell products across it. We were solving different problems, and only one of us knew there was a problem.
I had been the carrier. Sprint and Pacific Bell. I knew the tariffs, the DE bits, the oversubscription ratios, the half-circuit structure, the correspondent regime, and how to buy.
That combination is the whole thing. Both sides of the table, in one person, in the ninety days after the Act. That is what was scarce. Any one of the forty-three could have hired it. One of them, Sprint, already had.
And I had nothing to protect. No installed base to cannibalize. No transit margin to defend. No market of my own that another operator could take. Which made me the only party on Earth that forty-three rivals could each say yes to without saying yes to each other.
NTT could not put six cabinets of Cisco.com inside France Télécom’s data center. France Télécom would be hosting NTT’s customer relationship in its own building, cutting NTT’s transit bill, and handing NTT a beachhead in Paris. Never, at any price. But a company that competed with neither of them could carry Cisco to both, and both helped, because both got cheaper and faster and neither armed the other.
11. What I actually did
I did not buy transit. I went to each ISP and asked for one thing: announce me your routes only. Nothing you learned from anyone else. Just your own customers.
That ask changes everything on their side of the table. I am not asking them to haul my traffic anywhere, so there is no transit cost to them and no capacity consumed on their backbone. I cannot use their table to reach their rivals, and I cannot use it to compete for their accounts. Their cost to serve me was a cross-connect and a filtered announcement. Once each of them worked out what I was actually asking for, the price came down. Their own pricing sheet said I was not a threat.
Then I paid for everything else myself. The IPLC was mine, end to end. I took delivery at the cross-connect in their own data center and pulled it straight onto my switch: served locally if the content was already sitting there, or onto the circuit if the session had to reach an origin on another continent. Nothing of mine touched their network past the demarcation. My circuit, my routers, my routing policy under AS6553, my money.
And I brought them something before I asked for anything. I rented a minimum of six cabinets inside each ISP, plus thousands of square feet in Northern California, New York, London, and Hong Kong. Those cabinets held Cisco.com, microsoft.com, Stanford’s journals, Intel. The content their subscribers requested most, sitting in their own building, one cross-connect away.
Before Digital Island, every one of those requests was an international transit charge on their books, hauled across an ocean they were paying someone else for, and hauled badly. I did not just remove that cost. I paid to remove it. My cabinets, my servers, my power, my circuits. They got their heaviest content localized, their trans-Pacific and trans-Atlantic bill cut, and their subscribers’ experience fixed, from a party who then asked to buy something from them.
That was not a competitive concession. That was an expense reduction that also wrote checks.
Every path on that fabric was the same shape: my circuit, their cross-connect, their own routes, their own customer. One AS boundary. No transit. No hot potato. Nothing to congest. Single-hop was not a marketing phrase. It was the literal topology of routes-only plus my own haul, and it is why the number was deterministic enough to sign.
They sold me the parts, one half-circuit at a time, and announced me full routes at every demarcation, because a customer under his own AS is a purchase order and not a threat. Forty-three operators who could not do that for each other all did it for me, separately, and every one of them was just booking revenue and cutting cost.
I assembled the fabric out of their refusal to build it.
And they did not watch from across the market. Many of them watched from their own meet-me rooms. My cabinets stood on their raised floors. My circuits terminated in their patch panels. My traffic rode cable systems they held the IRUs on. Every packet on my fabric touched plant somebody on that list owned, and they invoiced me monthly for the privilege. The operators hosting my equipment had physical custody of the product that was replacing their category, and port-side metrics on its growth. I lit the major metros in 1996 and 1997, then densified through 1998 and 1999 with redundant, diverse paths across multiple incumbents per market: four ISPs in Japan, three in Korea, seven in the United States. Inside each host ISP the installation was standardized, a minimum of one T-1 IPLC and six cabinets of ATM, routing, and CDN equipment, so every host was metering the same known footprint, and the only number moving on their dashboards was the growth, month over month, until the day Cable & Wireless bought the company. But what they were metering was only the edge. The core of the product, the hub data centers running 2,000 to 15,000 square feet where the fabric terminated and the end-to-end control lived, sat in my own facilities that no incumbent ever entered. From inside their buildings, six cabinets and a T-1 looked like a hosting customer. Without the core, the edge reads as a customer, not a competitor, and that is why nobody smoked us. And because rivals do not compare notes, each of them held a fragment of the telemetry and dismissed the fragment. The closest vantage point in the history of competitive intelligence, instrumented in parallel, and not one of them assembled the picture. That is what watched means in this record: visibility, access, custody, telemetry, revenue, and no product.
12. What it produced
Enforceable QoS across the major Internet markets: SSL session completion under 300ms round-trip, anywhere to anywhere, guaranteed in the contract, reaching approximately 99% of internet-accessible users.
That is the Buy Now button. That is the Trade Now button. A dropped checkout is an annoyance. A dropped trade is a filled or missed position and a compliance event. That is where a guaranteed path stops being a convenience and becomes the reason an enterprise clears the Internet for use. It is what let corporate counsel say yes.
Cisco, November 1996. Stanford, January 1997. Visa, Q2 1997. E*TRADE and Schwab, 1998. CERNET and one-fifth of humanity, February 1998. MasterCard, 1999. Microsoft, Intel, and Compaq, 2000. 881 customers in under four years. One per business day, every business day, for four years.
13. The two transactions that close it
April 1999. Telefónica operated Spain’s national network and was expanding across Latin America through privatization-driven acquisitions, with billions in infrastructure capital and engineering staff on multiple continents. It did not build a competing product. It signed a reseller agreement and sold Digital Island’s network to its own enterprise customers.
May 2001. Cable & Wireless held submarine cable assets across more markets than Digital Island ever built, and it did not come to this purchase empty-handed. It already owned MCI’s entire former Internet business, internetMCI, acquired for approximately $1.75 billion under the 1998 divestiture order, the largest in merger history. It did not build a competing product. It paid approximately $340 million to acquire Digital Island. Owning the segment business did not give Cable & Wireless the path business. It had to buy that separately, from Digital Island. One acquirer, both product classes on its books within three years, and two purchase prices, both public record, pricing the difference between reachability and an outcome.
A company builds when building is faster, cheaper, or more capable than buying. Neither of them built.
There is no stronger carrier exclusion proof than a transaction.
The thirteen sections above share one ancestor. The half-circuit regime, the correspondent agreements, the oversubscription economics, the mutual distrust, and the twenty-country market were not natural conditions. They were the residue of an eight-year period in which the backbone’s operators prohibited the commerce that would have forced those structures to evolve, while positioning their own affiliate inside the exemption. Digital Island was not merely first through an open gate. Digital Island was first through a gate that specific, named institutions had held shut, and the dated instruments on this site all fall inside the window that opened when they lost their grip: NSFNET decommissioned April 30, 1995, the Act signed February 8, 1996, AS6553 issued August 29, 1996, Cisco executed November 7, 1996.
Standards define possibility. Infrastructure delivers reality.
MCI ran the Internet’s backbone for eight years and employed the co-author of TCP/IP with Internet Architecture in his title. It did not deliver the reality of eCommerce. It sold me half-circuits, and it was my largest single-vendor expense in the world. In 1998, while still selling me those half-circuits, MCI’s entire Internet business was sold under government order to the company that bought Digital Island three years later.
That reality was a signature under a number, anywhere to anywhere. It was our entire value proposition because it was nobody else’s product.
The complete operator-by-operator exclusion record, with the Section 271 statutory bar, the six deployment records, and the Route Views and CAIDA verification layer, is here: The Digital Island Litmus Test and Carrier Architecture Record, 1996
Corrections supported by documentation are welcomed and will be incorporated with attribution: mark@marknichols.com