Cloud Direct-Peering Maps

Where Google Cloud, AWS and Microsoft Azure hand traffic to other networks, inferred from traceroutes run from inside each cloud.

Cloud Direct-Peering Maps

Where do Google Cloud, Amazon Web Services and Microsoft Azure hand traffic to other networks? Each map places the direct peers a cloud reaches from its own regions, city by city, as seen in traceroutes launched from inside that cloud toward the rest of the Internet.

The neighbor table lists every network seen in October 2026, per provider, with the regions that saw it and the other providers it also connects to.

Google Cloud · IPv4 + IPv6

October 2026 · 42 regions
6,679peer ASNs
2,791cities
230IXPs
42regions

5,195 neighbor ASNs over IPv4 and 4,615 over IPv6, 3,131 of them over both. Probed at 5,000 packets per second, five times September's rate, at which Google rate-limits replies; see what changed below.

Open map →

Amazon Web Services · IPv4 + IPv6

October 2026 · 27 regions
8,176peer ASNs
7,311cities
222IXPs
27regions

AWS's first IPv6 measurement: 6,102 neighbor ASNs over IPv4 and 5,431 over IPv6, 3,357 over both. Re-mapped on 17 October 2026 to count hand-offs behind AWS routers that answer from unrouted addresses; see what we observed in AWS below.

Open map →

Microsoft Azure · IPv4

October 2026 · 39 regions
4,665peer ASNs
3,414cities
216IXPs
39regions

Nearly twice September's regions, with the hot-potato routing preference requested, which only 28 regions honoured. IPv6 was not measured: the probes never left the virtual machines.

Open map →

Neighbor networks · all three clouds

October 2026 · searchable table
9,008IPv4 ASNs
2,370in all three
108regions
3clouds

Every neighbor network per provider, with the regions that saw it, whether it was reached across an exchange, and which other clouds also connect to it. Filter, sort and download as CSV.

Open table →

Google Cloud · IPv4 + IPv6

September 2026 · 43 regions
6,853peer ASNs
2,696cities
232IXPs
43regions

Both address families from the same 43 regions, with a filter to view either one and a side-by-side comparison: 5,365 neighbor ASNs over IPv4 and 4,763 over IPv6, 3,275 of them seen over both. IPv4 used Google's Standard Tier and IPv6 its Premium Tier (see below), so the comparison is also one of routing tiers.

Open map →

Google Cloud · IPv4

August 2026 · 41 regions
5,208peer ASNs
1,803cities
219IXPs
41regions

The previous month's IPv4 campaign, from a slightly different set of regions.

Open map →

Google Cloud · IPv4

March 2026 · 42 regions, 125 zones
5,879peer ASNs
3,351cities
178IXPs
125zones

An earlier campaign with a different design: three zones per region, 44-byte probes, the .1 address of each /24 as target, an unrecorded network tier and older IXP membership lists. Its counts are not directly comparable with the other maps.

Open map →

Amazon Web Services · IPv4

August 2026 · 27 regions
6,280peer ASNs
5,826cities
204IXPs
27regions

AWS commercial regions, measured from one instance each. Re-mapped on 17 October 2026 from the original traceroutes and datasets with the same rule as October (3,434 neighbor ASNs before).

Open map →

Microsoft Azure · IPv4

September 2026 · 20 regions
4,649peer ASNs
3,629cities
215IXPs
20regions

A subset of Azure regions; the map covers those this campaign measured, not all of Azure.

Open map →

October 2026: what changed

In October all three clouds were measured at the same time, against the same freshly generated targets, for the first time. Each comparison with an earlier month below is over the regions both campaigns measured, earlier month first. None of them isolates a change in the Internet: in each cloud, something about the measurement changed as well.

Cloud Against Regions Neighbor ASNs Median per region What else changed
Google Cloud, IPv4 Sep 2026 42 5,357 → 5,195 3,769 → 3,541 Probing rate 1,000 → 5,000 pps
Google Cloud, IPv6 Sep 2026 42 4,754 → 4,615 4,008 → 3,180 Probing rate 1,000 → 5,000 pps
Google Cloud, IPv4 Aug 2026 41 5,208 → 5,191 3,694 → 3,541 Not the rate (both 5,000 pps)
Amazon Web Services, IPv4 Aug 2026 27 6,280 → 6,102 1,662 → 1,470 Both re-mapped with the rule below (both 5,000 pps); there is no September AWS campaign
Microsoft Azure, IPv4 Sep 2026 18 4,637 → 4,472 3,849 → 80 Routing preference: Microsoft network → Internet

Google Cloud. About 87% of neighbors are the same in both months. Per-region counts fall by 6% over IPv4 and 21% over IPv6 against September, and by less against August, which used the same probing rate. That pattern is consistent with the probing rate rather than the network: in a side-by-side test on the same targets, Google's routers answered fewer probes at 5,000 than at 1,000 packets per second, and some neighbors went unseen. Separately, a few hundred networks move in each direction, half of them seen from a single region.

Amazon Web Services. Counted the earlier way, AWS lost 60% of the neighbors it reaches across exchanges between August and October (3,434 → 2,550 neighbors in all), while every other neighbor was unchanged. That drop was not a change in AWS's interconnection; see the next section. Counted as described there, both months agree: 6,280 and 6,102 neighbors, exchange neighbors within 1% of each other, and no region but one losing more than a fifth of them.

Microsoft Azure. October asked for the Internet routing preference, Azure's hot-potato option, in every region. Twenty-eight regions behave that way: 21 to 344 neighbors each, mostly transit providers near the region. Eleven regions, in India, Qatar, the UAE, Korea, South Africa, Australia and Arizona, see 2,700 to 3,900 neighbors each at exchanges worldwide, as Microsoft's own network would carry them. Why the preference took effect in some regions and not others is not yet established. The union of neighbors barely moves, but per-region counts cannot be compared with September's.

Across clouds. Over IPv4, 2,370 networks are neighbors of all three clouds and 9,008 of at least one. AWS and Azure share 3,375; Google Cloud and AWS 3,146; Google Cloud and Azure 2,803. AWS has 1,951 neighbors the other two lack, Google Cloud 1,616 and Azure 857. The AWS sets use the rule described next; the Google Cloud and Azure maps have not yet been re-mapped with it, which on samples adds about a tenth to Google Cloud's neighbors and little to Azure's.

What we observed in AWS

AWS's routers mostly answer traceroute from addresses that are not routed on the Internet: the reserved 240.0.0.0/4 block, the 100.64.0.0/10 shared address space, and public space AWS holds but does not announce in BGP. A neighbor is counted when a hop in the cloud's network is followed by a hop in another network, and a hop with no BGP origin belongs to no network, so a hand-off behind one of these routers was never counted. On 200,000 traceroutes from us-east-1 in each month, the hop just before traffic left AWS was unrouted in about 60% of cases.

The same routers also explain the August-to-October drop. AWS reached as many exchange members in October as in August (376 against 352 in that sample), but the router just before the exchange port answered from AWS's announced space 9% of the time in August and 1% in October, so fewer of those hand-offs could be counted (77 → 21 members).

Both AWS maps were therefore re-mapped from the original traceroutes with one change: walking out from the instance, a hop with no BGP origin counts as AWS's until the first hop that belongs to another network, provided no hop between it and the instance, or the last confirmed AWS hop, went unanswered. A typical path, from us-east-1 (Virginia) to a host at WU Vienna:

Hop Address RTT Whose
1 240.64.221.129 1.8 ms AWS router, reserved 240.0.0.0/4 space: no BGP origin
2 240.3.184.33 1.4 ms AWS router, reserved 240.0.0.0/4 space: no BGP origin
3 149.112.69.20 4.7 ms Hurricane Electric's port at the CIX-NoVA exchange: the hand-off
4–6 184.105.213.174 … 98 ms Hurricane Electric's backbone, Paris then Vienna
7–8 193.171.13.17 … 105 ms ACOnet, then WU Vienna

Neither AWS hop has an ASN, so the hand-off to Hurricane Electric at hop 3 was not counted before; it is now. Neighbor counts roughly double: 3,434 → 6,280 in August and 2,550 → 6,102 in October over IPv4, and 870 → 5,431 over IPv6.

A mistake that slipped through, which we are fixing. Not every address without a BGP origin is AWS's. Neighbors also number routers from public space they hold but do not announce, and when such a router directly follows an AWS one, the rule takes it for AWS's. From ca-central-1 (Montréal), traceroutes toward Columbia University leave AWS for NYSERNet, the New York research network that connects Columbia:

Hop Address Whose Counted as
… 240.x, 100.100.x AWS routers, no BGP origin AWS (right)
next 199.109.114.26 NYSERNet's router (columbia-nyc32-55a1-cdn.nysernet.net), in public space NYSERNet holds but does not announce AWS (wrong)
next 128.59.255.103 Columbia University (AS14) AWS's neighbor (wrong: NYSERNet is)

So Columbia appears as an AWS neighbor from the two Canadian regions, although it reaches AWS only through NYSERNet. On a preliminary check of one region (us-east-1, October, IPv4) this kind of mistake does not change much of the connectivity: 87% of the links through such routers still name the right neighbor network, and about 86 of the region's 1,145 neighbors (7.5%) are affected. The next campaign attributes these addresses to whoever holds them in the regional Internet registries, so a NYSERNet router counts as NYSERNet's. The "Since 2020" and exchange tables further down are generated from the earlier August count (3,434) and have not been regenerated.

How the maps are built

  • From one virtual machine in each cloud region, scamper runs ICMP-echo traceroutes to one address in every routed IPv4 /24 (about 12.2 million targets). For IPv6 (Google Cloud from September 2026, all three clouds in October) it probes one address per announced IPv6 prefix (about 254,000 to 256,000 targets). Probing ran at 5,000 packets per second per vantage point in August and October 2026 and at 1,000 in September.
  • Each hop is assigned to a network: if the address is a member port on an Internet exchange's peering LAN, the member's ASN; otherwise the origin AS of its covering prefix in RouteViews for the measurement date.
  • What the maps count as a peer is, strictly, a neighbor network: the first network a path enters after leaving the cloud's own networks, seen at two consecutive responding hops. It may be a settlement-free peer, a transit provider, or a customer.
  • Each peer interface is placed in a city using IPInfo geolocation and, where its router hostname encodes one, CAIDA HOIHO.

Premium versus standard routing

Where a cloud hands outbound traffic to the rest of the Internet is partly a product choice. Two strategies bracket it: cold-potato routing carries traffic on the provider's own backbone as far as possible and hands it off near the destination, while hot-potato routing hands it off near the cloud region and lets other networks carry it the rest of the way. Which one a measurement used strongly shapes which neighbors and which interconnection cities a traceroute from that region can see.

  • Google Cloud exposes the choice as Network Service Tiers. In the Premium Tier (the default), outbound traffic "typically routes over the Google global network to a point of presence that's as close as possible to the internet user". In the Standard Tier, it "is sent through a peering or transit network in a point of presence near the region". External IPv6 addresses are available only in the Premium Tier.
  • Microsoft Azure exposes it as routing preference on a public IP. The default, the Microsoft global network, is what Microsoft describes as cold potato: egress "exits closest to the user". The Internet option is its hot-potato counterpart: traffic "exits Microsoft network in the same region". It is IPv4-only and cannot be changed after the address is created.
  • Amazon Web Services has no equivalent setting for an instance's Internet traffic, and we found no AWS documentation of where EC2 egress leaves its network. Global Accelerator brings inbound traffic onto the AWS backbone at an edge location, but does not change how an instance reaches the Internet.

What these maps used. Google Cloud IPv4 was measured on the Standard Tier and Google Cloud IPv6 on the Premium Tier, because Google offers external IPv6 only there; every instance's tier was checked and recorded at launch. Azure public IPs used the default routing preference, the Microsoft global network, through September; in October they requested the Internet preference instead (see above). AWS instances used ordinary public addresses, with no option to set. So on the September Google Cloud map the IPv4-versus-IPv6 comparison also compares Standard with Premium routing, and the two cannot be separated from these data.

Our data do not yet show the difference cleanly. IPv6 neighbor sets are more uniform across regions than IPv4 ones (a median overlap between two regions of 0.91 against 0.84; the least similar IPv4 pairs all involve Johannesburg). That is consistent with Premium routing, but the address family and the 50-fold difference in target count could produce it too. Geolocating the first neighbor's interface is too unreliable to say where the hand-off happens: the link is often numbered from the cloud's own addresses, so the first hop we attribute to the neighbor can lie deeper inside it, and transit backbones geolocate poorly. Round-trip times to the first neighbor, which do not depend on geolocation, are the next check. The clean test is the same address family from the same regions on both tiers, measured side by side, which we have not yet run.

Since 2020

In 2020, Arnold et al. (Cloud Provider Connectivity in the Flat Internet, IMC 2020) counted the networks four clouds connect to, using traceroutes from inside each cloud. The table sets their neighbor sets beside these campaigns, counted the same way: distinct neighbor ASNs over all vantage points, and the organizations behind them (CAIDA AS-to-organization tables, with the cloud's own organization excluded).

Cloud 2020 ASNs Campaign ASNs Organizations Vantage points vs 2020
Google Cloud 7,553 2023 9,240 8,765 115 zones 1.22×
Mar 2026 5,879 5,580 125 zones 0.78×
Aug 2026 5,208 4,983 41 regions 0.69×
Sep 2026 5,365 5,130 43 regions 0.71×
Microsoft Azure 3,564 Sep 2026 4,649 4,382 20 regions 1.30×
Amazon Web Services 1,188 Aug 2026 3,434 3,239 27 regions 2.89×
IBM Cloud 2,746 no 2026 campaign

Google Cloud's 2026 counts are below its 2020 figure and its 2023 count is above it; Microsoft's is about a third higher and Amazon's is close to three times larger. These are not like-for-like measurements: vantage points, target lists and probing differ between 2020 and 2026, and between campaigns, so the ratios are not growth or decline rates. For Google Cloud in particular, the BGP view below moves far less over the same years (-9%) than the traceroute count does.

Turnover. Counted as organizations through one reference table, the 2020 and 2026 sets overlap only partly:

Cloud 2020 orgs Still there Gone New Retained
Google Cloud 6,940 2,897 4,043 2,233 41.7%
Microsoft Azure 3,264 2,237 1,027 2,145 68.5%
Amazon Web Services 1,122 483 639 2,740 43.0%

Google Cloud's September 2026 set retains 41.7% of the 2020 organizations, 41.7% of those in a 2023 Google run, and 41.5% of that run's non-IXP set: three reference sets, collected in different years with different tooling, agree within a point on how much survives.

Most of it happens at exchanges

A link counts as an IXP link when either end is a listed member port of an exchange's peering LAN; everything else is counted as private interconnection (PNI). Each neighbor is placed by exchange name where it meets the cloud on a fabric, and by the city encoded in its router hostname where it does not.

Campaign Neighbor ASNs Only at an IXP At an IXP Via PNI Both Placed Placed in several
Google Cloud, Sep 2026 5,365 53.2% 3,465 2,510 610 66.0% 25.9%
Google Cloud, Aug 2026 5,208 55.0% 3,439 2,343 574 67.5% 26.6%
Microsoft Azure, Sep 2026 4,649 57.4% 3,510 1,981 842 76.2% 31.2%
Amazon Web Services, Aug 2026 3,434 54.6% 2,147 1,560 273 64.2% 26.2%
Google Cloud IPv6, Sep 2026 4,763 35.0% 2,186 3,094 517 46.5% 24.1%
Google Cloud, Mar 2026 5,879 29.3% 3,045 4,154 1,320 53.0% 20.9%
Google Cloud, 2023, as recorded 9,240 37.3% 5,903 5,797 2,460 65.2% 25.1%
Google Cloud, 2023, relabelled IXP-first 9,974 41.9% 7,121 5,797 2,944 72.6% 23.4%

Across the three clouds' IPv4 campaigns, between 53.2% and 57.4% of neighbors are reached only across an exchange; on this measure the three clouds are close. These shares are floors, since a peering-LAN address missing from the membership lists counts as private. Two rows are not comparable to the others: the Google Cloud IPv6 set uses a different address family, target list and routing tier, and the March campaign used an older membership list and a different target list. Nor are the two 2023 rows, which come from a 2023 run with its own target list and the June 2023 membership list: one keeps the neighbor that run recorded, the other relabels its links with the 2026 rule (a listed member port's ASN wins), which is as close to the 2026 method as its stored links allow. Microsoft's neighbors are the most often seen in several places: 31.2% of its placed neighbors meet it in more than one place, against about a quarter for the others. Placement leans on exchange membership; private interconnects are placed only when a router hostname names a city, so that side is close to unmeasured.

2023 against 2026. Google Cloud's share of neighbors reached only at an exchange is 37.3% in 2023 as recorded, or 41.9% relabelled, against 53.2% in September 2026. The membership lists, target lists and link selection all differ, so this gap is not by itself evidence of a shift toward exchanges. The share of placed neighbors seen in several places moves much less: 25.1% in 2023 as recorded and 25.9% in September 2026. Neighbors seen both at an exchange and over a private link are 26.6% of the 2023 set but 11.4% in September 2026. Keeping one zone per region barely moves the 2023 figure (26.7%–26.9%), so fewer vantage points do not explain the difference; what does is not established here.

How much do extra vantage points add? The March campaign probed from three zones in each of 42 regions. Keeping one random zone per region keeps 96.6% of the neighbor ASNs (96.0%–97.2% over 20 draws) but only about 34% of the links, and the share placed in several locations barely moves (20.4%–20.8%, against 20.9% with all zones). Neighbor counts survive single-zone probing nearly intact; link-level counts do not. The 2023 run, which probed from 115 zones in 38 regions, agrees independently: one zone per region keeps 96.4%–96.8% of its neighbor ASNs.

What BGP sees

CAIDA's AS relationships give, every month, the neighbors a cloud's ASNs have in public BGP data, independently of any traceroute. Counted the same way for each month (neighbor ASNs of the cloud's own ASNs, links between them excluded):

Cloud Sep 2020 Jun 2023 Mar 2026 Aug 2026 Sep 2026 2020 → 2026
Google Cloud 411 399 360 375 376 -8.5%
Microsoft Azure 318 294 307 307 307 -3.5%
Amazon Web Services 334 345 434 469 466 +39.5%
IBM Cloud (AS36351) 3,027 2,380 274 297 289 -90.5%

BGP sees about 376 neighbors for Google Cloud in September 2026; the traceroutes above see 5,365. Public collectors miss most peering links, because a peer does not pass the cloud's routes on to the networks that feed the collectors, so interconnection has to be measured from inside. IBM's count falls from 3,027 to 289; this page does not investigate why.

Exchange route servers add a second, narrower view: CAIDA also harvests multilateral peering from a small panel of route-server looking glasses (panel size per month, with the number shared with 2020's panel in brackets: Sep 2020 9 (9), Jun 2023 0 (0), Mar 2026 11 (7), Aug 2026 3 (0), Sep 2026 3 (0)). The March 2026 panel keeps seven of 2020's nine; between those two months Google Cloud's route-server neighbors go from 503 to 0 and IBM's from 0 to 589. That says Google is no longer visible on the sampled route servers, not that it left them. The June 2023 snapshot lists no looking glasses, and the August and September 2026 panels share none with 2020, so neither can be read against it.

Every number in these three sections is computed from files by scripts/cloud_peering_trend.py (scamper-analysis 5653f89): the 2026 campaign peer links, Arnold et al.'s 2020 neighbor sets, the 2023 run's neighbor sets and its per-link, per-zone records, and CAIDA AS-relationship and AS-to-organization files for each month.

Reading them carefully

  • A neighbor appears only if some traceroute from the cloud crossed into it, so each map is roughly a lower bound on the provider's interconnection, seen in the cloud-to-Internet direction, though mapping errors can also add networks that are not true neighbors.
  • Locations are where the neighbor's interface geolocates, which can differ from the facility where the interconnection physically happens. Router-interface geolocation is often wrong by hundreds of kilometres or more.
  • The campaigns differ in date and in which regions were measured, and the IPv6 target set is one address per prefix rather than per /24, so compare counts across maps with that in mind.