Big Tech’s Private Internets: What’s Real, What’s Hidden

private internet - 15

By Tom Morgan

The short version

  • The “shadow internet” is real, but it is not secret. It is a set of private backbones, subsea cables, and routing systems that Google, Microsoft, Meta, and Amazon run for their own traffic. Most of it is publicly documented. What outsiders cannot get is the capacity itself and the traffic data.
  • The scale is large. TeleGeography says content and cloud networks used 75% of all international bandwidth in 2025. In 2010, traditional telecom carriers held roughly that share.
  • It is still accelerating. TeleGeography forecasts content provider bandwidth will grow about nine-fold between 2025 and 2035.
  • Private does not mean invincible. The September 2025 Red Sea cable cuts and the 2026 Persian Gulf work stoppage on Meta’s 2Africa Pearls extension show that the richest networks still depend on the same ships, straits and seabed as everyone else.
  • Closed does not have to mean unverifiable. Apple opened its Private Cloud Compute servers to outside inspection in 2024, which proves the two can be separated.

The Saturday morning the cloud slowed down

At 05:45 UTC on September 6, 2025, Microsoft posted a short note on its Azure status page. Traffic crossing the Middle East, it said, might see higher latency because of undersea fibre cuts in the Red Sea.

Internet monitor NetBlocks traced the damage to two cable systems, SMW4 and IMEWE, near Jeddah in Saudi Arabia. Other systems, including FALCON GCX, were also affected, and Kentik’s Doug Madory later added Europe India Gateway to the list. In India, Pakistan and the Gulf, ordinary users felt it as slow pages and patchy connections.

Microsoft’s customers had a different morning. The company rerouted traffic over longer paths, warned of extra latency, and said traffic that did not touch the Middle East was unaffected.

That gap, between “the internet got worse for millions of people” and “one company’s network absorbed the blow”, is what this article is about. Over two decades, the largest technology companies have built a second layer of plumbing beneath the internet most of us picture. Some people call it a shadow internet. The phrase is catchy, so we tested it against the evidence.

A correction to the headline you may have seen

“Shadow internet” suggests a hidden network built for hidden purposes. After going through the research, we don’t think that framing survives. What the documents show is more ordinary and, in some ways, more consequential.

Claim you may have seenWhat the evidence supports
“Big Tech is building a secret internet.”Cables, papers and filings are public. Google publishes peer-reviewed network research, Meta announces its cables, and TeleGeography maps them.
“They are replacing the public internet.”No. Content still reaches you through ISPs and the public internet. What moved private is the traffic between the companies’ own data centers.
“Most internet traffic is now private.”Careful. The 75% figure is a share of used international bandwidth, not of all internet traffic.
“You can’t access it.”Mostly true for capacity and traffic data. Enterprises can buy private connections to some clouds, and 2Africa is described as open access. Individuals mostly cannot.
“It makes the internet faster for everyone.”Not directly. A private backbone moves a company’s data between its own sites.

So the accurate version of the title is this: Big Tech runs private internets alongside the public one, and the public has very little visibility into them. Six cases show how.

Case 1: The numbers, from the people who count cables

TeleGeography has tracked international bandwidth and submarine cables for decades. Its June 16, 2026 analysis by Research Director Alan Mauldin is the cleanest primary source on this question.

Its category of content providers includes the networks of the major hyperscalers, cloud providers, AI platform companies, and so-called neoclouds. In 2025, that group accounted for 75% of all used international capacity. The category’s bandwidth is forecast to grow nine-fold from 2025 to 2035.

How fast did we get here?

YearContent provider share of used international bandwidthSource
2010Traditional telecom carriers used about 75%; content providers were negligibleTeleGeography, IEEE ComSoc
2017Content providers pass internet backbone providers as the largest usersTeleGeography
202169%TeleGeography, State of the Network 2023
2024“Almost three quarters”TeleGeography, State of the Network 2025
202575%TeleGeography, June 2026

The takeover is uneven. TeleGeography reported that content providers took 92% of used capacity on the trans-Atlantic route in 2021, but only 21% on the Europe to East Asia route. By year-end 2024, content networks accounted for at least 80% of demand on the trans-Atlantic, trans-Pacific and intra-Asia routes, and they had become the main source of used bandwidth on US to Latin America links.

The logic is simple. These companies mostly want to connect their own data centers and the big interconnection hubs. They take huge capacity on the busiest corridors and pay far less attention to secondary long-haul routes.

On the demand side, AppLogic Networks (formerly Sandvine) publishes an annual Global Internet Phenomena Report built from anonymized operator data. Its 2025 edition says Alphabet, Meta, Microsoft, Amazon and Apple, along with Netflix, Disney and TikTok, account for more than 50% of web traffic. Treat that as indicative: the sample comes from the vendor’s operator customers and is not a census of the internet.

What this proves: an independent research firm and an operator-data vendor, measuring different things, point the same way. What it does not prove: that the traffic is secret, or that concentration harms anyone by itself. It shows concentration and opacity. Harms and benefits come later.

The measurement gap, stated plainly

Almost every number in this section traces back to one commercial research firm and to the companies’ own statements. That is a weakness, not a footnote. No regulator, university, or public monitor publishes an independent count of the traffic that actually crosses these private networks. TeleGeography measures used capacity by buyer, which is not the same thing as traffic, and its detailed data sits behind a paywall. Outside observers can see cable routes, outages (through monitors such as NetBlocks and Kentik), and what the companies choose to publish. They cannot see load, failover behavior, or how much headroom is left.

Independent analysts say the same. A 2024 conference slide citing Omdia puts Google and Meta behind almost all content-provider subsea capacity, which means the picture rests on very few operators. CSIS fellow Masha Kotkin, quoted in the Financial Times coverage of the Gulf threats, warns that the region’s cloud and AI build-out depends on cables most users never think about. If you are a carrier, a regulator or an insurer, that is a planning problem: you are pricing and modelling risk on a network whose real state you cannot observe.

Case 2: Google, which told us in 2018

Google has been unusually open. In 2018, its engineers presented a paper at SIGCOMM, a top networking conference, on B4, the private software-defined network linking its data centers. The abstract says B4 is larger and growing faster than Google’s connectivity to the public internet.

Two details stand out. Demand on B4 grew about 100x over five years, roughly a doubling every nine months. And traffic is sorted into reliability classes, with search ads, DNS, and web serving in the strictest, at 99.99% availability.

A year earlier, Google described Espresso, its system for steering traffic out to the public internet at peering points, which already handled more than 20% of Google’s internet-facing traffic.

Google’s cable footprint has grown since. Its stake in the transatlantic Grace Hopper and in Equiano, which links Europe to southern Africa, is well documented. On February 18, 2026, Google announced America-India Connect, which adds a subsea gateway at Visakhapatnam and new paths linking India to Singapore, South Africa and Australia. Its own announcement describes a data corridor from the US East Coast “through the Red Sea” to Mumbai. Keep that in mind for Case 5.

Caveat: the private-versus-public claim dates from 2018. We found no newer Google statement giving a current split, so treat it as a documented milestone, not today’s ratio.

Case 3: Microsoft says its traffic never touches the internet

Microsoft says so in plain words on its Azure engineering blog. Traffic between its data centers, including all traffic between Microsoft services anywhere in the world, stays on its own network and does not flow over the internet. It invests in private fibre for metro, terrestrial and submarine paths and manages the whole thing with its own software-defined networking system, SWAN.

With Facebook, it co-built MAREA, a 6,600 km cable from Virginia Beach to Bilbao with eight fibre pairs and an initial design capacity of 160 Tbps.

More recent figures come from trade-press coverage of Microsoft presentations. At Hot Interconnects 2025, Microsoft reportedly cited about 275,000 miles of fibre, more than 2 Pbps of WAN capacity and 200 Tbps of peering. Those are company claims relayed by industry press. We could not check them against independent measurements, and that is part of the story.

Case 4: Meta, from cable tenant to cable owner

For years, Meta joined consortia. It says it has invested in more than 20 subsea cable projects. Now it is taking the driver’s seat.

ProjectWhat it isStatus (November 7, 2026)
2AfricaAbout 45,000 km system ringing Africa and linking Europe, the Middle East and South Asia. Backed by a consortium including Meta, China Mobile International, Orange, Vodafone, Telecom Egypt, stc’s center3 and othersCore ring declared complete. The Pearls extension (Gulf, Pakistan, India) paused in March 2026, when contractor ASN reportedly declared force majeure. We found no report that work has resumed
Waterworth50,000+ km planned across five continents, 24 fibre pairs. Meta has not announced it as open accessPlanned, no ready-for-service date. Australian environmental filings lodged by Vocus in June 2026 name Darwin, Townsville and Karumba as proposed landings, and a stakeholder register links the work to Waterworth. Vocus declined to confirm the partnership
PetalAbout 7,000 km (4,300 miles) between the US and France. Meta says it will be the first transoceanic cable with petabit capacity (1 Pbps) and the first at scale with two-core fibreAnnounced September 21, 2026. Due in service in 2029. Built with NEC and Sumitomo Electric, with Orange supporting the French landing. Meta has not named the US landing site or disclosed the cost

Two cautions so you don’t overread this. “50,000 km” describes the Waterworth project as a whole, not necessarily one uninterrupted cable. And none of this speeds up your home broadband: it moves Meta’s data between Meta’s sites. Meta has not disclosed Waterworth’s budget, so the widely repeated “$10 billion” is a press estimate.

Case 5: AI labs are becoming network operators

This is the newest development, and it comes from TeleGeography, not a press release. In its June 2026 analysis, TeleGeography says OpenAI and Anthropic rely on the major hyperscalers’ networks for global reach and compute, but are also building their own private networks to interconnect critical data centers. In its words, they are increasingly acting like network operators in their own right.

It adds a second observation: operators such as Google dynamically route computing workloads across data centers worldwide to chase power and compute availability. That only works with a high-capacity mesh between those sites, acting as a global load balancer.

TeleGeography is careful not to overclaim, and so should we. AI-specific demand cannot be separated from the rest. Cloud, video, social media, and a long tail of other applications are still enormous and growing. AI is speeding up a trend that was already running.

Case 6: Chokepoints, where private networks meet very public water

The Red Sea, September 2025

Cuts on SMW4 and IMEWE near Jeddah degraded connectivity in several countries. The Associated Press noted that SMW4 is operated by Tata Communications. Early analysis, including from the International Cable Protection Committee as reported by Tom’s Hardware, pointed to a commercial ship dragging its anchor as the prevailing theory, not sabotage. We found no confirmed final attribution.

Three lessons:

  1. Control helps. Microsoft steered around the damage within hours. Operators without comparable private paths likely felt it more, which is our inference, not Microsoft’s statement.
  2. Control does not remove physical dependence. The cables that broke were shared systems. Even Google’s own corridor, as Google describes it, crosses the Red Sea.
  3. Redundancy still matters most. Kentik’s analysis noted that Europe-Asia traffic kept flowing because other systems on the route, including SEA-ME-WE 5, AAE-1 and PEACE, stayed intact.

For scale, Mauldin told Reuters that capacity routed through the Red Sea accounts for about 3% of global international demand and around 10% of traffic between regions.

The Persian Gulf, 2026

The risk has moved rather than shrunk. In mid-March 2026, Bloomberg reported that Alcatel Submarine Networks (ASN) had issued force majeure notices because it could no longer work safely in the Persian Gulf amid the Iran conflict. The installation ship Ile de Batz was stranded near Dammam, Saudi Arabia. Much of the 2Africa Pearls cable was already on the seabed but not connected to all its landing stations. TeleGeography’s Mauldin told Bloomberg that other Gulf projects, including SEA-ME-WE 6 and Ooredoo’s Fibre in Gulf, had also reportedly paused.

The story has darkened since. On October 23, 2026, Reuters reported that two senior Iranian security officials said Tehran has drawn up plans to target Gulf undersea cables if the US and Israel attack again. On November 6, the Financial Times reported, citing maritime-intelligence firm Windward, that at least four Iran-linked vessels had loitered over cable routes in the Gulf of Oman, some with tracking switched off. A RUSI analyst quoted in the coverage said the movements could be coincidental, or a signal of capability. None of this is confirmed sabotage, and no cable damage was reported in those accounts.

TeleGeography counts six cables passing through the Strait of Hormuz. Iraq, Kuwait, Bahrain and Qatar depend heavily on links running beneath it. A cable can be financed by the richest company on earth and still wait for a ship to be allowed back to sea.

This is also why Meta’s Waterworth route is notable. Its alignment avoids several of today’s most volatile chokepoints and places more of the infrastructure under direct control.

The counter-example: a closed system anyone can inspect

Not every locked door is a hiding place. In October 2024, Apple published a security guide, source code for key components, and a Virtual Research Environment for Private Cloud Compute, the servers behind its heavier Apple Intelligence requests. It also set up bounties from $50,000 to $1 million, with the top prize for code execution that breaks the system’s guarantees.

The servers remain Apple’s, but outsiders can test the claims. It shows that “private” and “unverifiable” are separate choices. We found no independent audit results since launch, so this proves the model is possible, not that it works flawlessly.

What you can and cannot actually see

LayerWho controls itCan outsiders see it?Why it matters
Cable routes and landing pointsConsortia and single ownersLargely yes (submarinecablemap.com)Physical risk and geopolitics are visible
Who owns what capacityOwners and customersPartly, mostly via paid TeleGeography dataPricing, planning, competition
Traffic volumes on private fibreThe companiesRarelyRegulators and carriers plan with gaps
Routing logic (B4 and Espresso style)The companiesOnly what they publishA performance edge nobody else can match
Security of private AI cloudsThe companiesOnly when they open it up, as Apple didTrust depends on audits, not promises

The IEEE ComSoc technology blog, summarizing industry analysis, makes the planning problem concrete: hyperscalers don’t share traffic data, and traffic on fully private dark-fibre segments is invisible to carrier networks and traditional traffic models.

Why they do it: economics, not a plot

  • Cost. At this volume, owning capacity can beat renting it.
  • Latency and control. Application-aware routing and fewer middlemen mean faster, steadier services.
  • Resilience. Owned fibre in several corridors reduces dependence on someone else’s repair schedule.
  • AI. Training and serving models create large, predictable flows between data centers.
  • Necessity. A peer-reviewed study of content provider investment found that private cable building is partly driven by where data centers get built, often where public cable connectivity is thin. There, owning the cable is less a power move than the only way to connect the building.

Many of these systems also bring capacity to places traditional carriers were not investing in, and several include local telecoms as consortium partners.

The real risks, ranked

  1. Opacity, and it is the one that matters most. Three-quarters of used international bandwidth sits with a handful of companies that publish almost no traffic data. Regulators cannot test claims about resilience, carriers cannot size their own networks against a competitor they cannot see, and insurers price cable risk without knowing how much of a region’s traffic a single corridor carries. When a cable fails, the first accurate account of the damage often comes from the company that owns the network, and nobody can check it.
  2. Geopolitical exposure. The Red Sea and Persian Gulf episodes show how fast conflict zones become network risks, and how a repair ship can matter more than a cable.
  3. Concentration. A few firms decide where cables land, and therefore which regions get high-capacity links first.
  4. Dependency. If your business sits on one cloud’s backbone, its incident becomes yours.
  5. Governance by default. Landing a cable requires permits in every country, as the Australian filings for Waterworth illustrate. We found nothing that sets a global limit on how much capacity a few firms can hold.

What to do about it

If you run a website or online business

  1. Don’t treat one cloud’s backbone as a safety net. Put a CDN in front of the site and consider a second CDN or a second cloud region on a different provider.
  2. Subscribe to status pages for your cloud and CDN. Microsoft’s Red Sea notice appeared the same morning as the cuts.
  3. Measure from where your readers are. Use synthetic monitoring from Asia, the Middle East, Africa, and South America, not only Europe and North America. Public tools such as Cloudflare Radar, RIPE Atlas and Kentik’s blog help you tell a regional cable event from your own bug.
  4. Fail over by region, not just by server. Write down which region pair takes over and test it quarterly.

If you work with enterprise cloud

  1. Ask each provider which regions share physical corridors, such as the Red Sea or the Strait of Hormuz, and what the documented reroute plan is.
  2. For latency-sensitive or regulated workloads, price dedicated cloud connections against internet transit.
  3. For Gulf or South Asia deployments, ask specifically how a Hormuz disruption would affect your service level.

If you are a researcher, journalist or policy reader

  1. Start at submarinecablemap.com, then check each system’s owner list. “Owner” can mean full ownership, a minority stake, selected fibre pairs, or leased capacity.
  2. Read the SIGCOMM papers and TeleGeography’s public analyses. They are among the few primary documents.
  3. Distrust any single big percentage. Ask: measured how, by whom, in which year, and of what?

If you are simply curious

Nothing here threatens your ability to browse. The internet is no longer one thing. It is a public layer and several private ones, and the private ones are growing faster.

Frequently asked questions

What is a “shadow internet”?

A popular label, not a technical term. Here it means the private networks large tech companies run alongside the public internet: backbones between their data centers, owned or leased subsea cables, and private routing systems. They are documented, but their capacity and traffic data are closed.

Is Big Tech replacing the public internet?

No. These companies still deliver content to you over ISPs and the public internet. What moved to private networks is the traffic between their own data centers.

Does 75% of internet traffic run on private networks?

No, and this is the most common misreading. TeleGeography’s 75% is the share of used international bandwidth taken by content and cloud networks in 2025. It is not a share of all internet traffic.

Can ordinary people or small companies use these networks?

Not directly. Businesses can buy private connections to some clouds, and 2Africa is described as open access. Most individuals reach these networks only through the services that run on them.

Is this legal and regulated?

It is lawful and visible in public records. Cables need landing permits and environmental approvals in each country. We found no global rule or coordinating body that looks at how much capacity a few firms hold overall.

Are these cables targets in the Iran conflict?

Iranian officials have threatened Gulf cables, according to Reuters, and the Financial Times reported unusual vessel behavior near cable routes in October 2026. As of November 7, we found no reported cable damage tied to those threats. Treat the reports as warnings, not confirmed attacks.

Will this get bigger?

TeleGeography forecasts used international bandwidth growing about 24% a year from 2025 to 2035, a doubling roughly every three years, with content provider bandwidth growing about nine-fold. Treat those as a forecast, not a measurement. A ten-year projection depends on AI demand, cable supply, and geopolitics holding roughly as assumed, and TeleGeography itself says AI-specific demand cannot be separated from the rest.

How we researched this, and what we could not verify

We built this from primary documents wherever possible: TeleGeography’s publications, Google’s peer-reviewed papers, Microsoft’s Azure engineering blog, Meta’s announcements, AppLogic’s report and an academic study. On November 7, 2026, we re-checked the TeleGeography June 2026 figures, Meta’s Petal announcement, the 2Africa Pearls reporting, the Waterworth filings, and the October to November 2026 Gulf reports.

What we could not confirm:

  • A current private-versus-public traffic split for Google. The latest public statement we found dates to 2018.
  • Microsoft’s current WAN figures, which come from trade-press accounts of conference talks.
  • Waterworth’s and Petal’s budgets. Meta has not disclosed them.
  • The cause of the September 2025 Red Sea cuts. Early analysis favored a dragged anchor, with no confirmed final attribution.
  • Whether Pearl’s construction has resumed. We found no report that it has.
  • Whether the Gulf of Oman vessel movements involved any intent to damage cables.
  • Independent audit results for Apple’s Private Cloud Compute since 2024.

If you spot something outdated or wrong, tell us, and we will correct it and log the change here.

Correction log: November 7, 2026, first publication.

Sources and further reading

Primary and first-party

Industry and press

Red Sea and Persian Gulf

Leave a Comment