Industries / Technology / Software and SaaS / Cloud and data centers
Servers and networking
This is the work of choosing, building and connecting the computers in a data center: the servers, the racks that hold them and the network that links them.
6
Milestones achieved, from 1999 to 2022.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1999
Search on cheap PC parts
Google ran its early search engine on racks of ordinary personal computer parts instead of branded server machines. The Computer History Museum holds a 1999 Google rack whose trays hold 80 PCs and two HP Ethernet routers, a gift of Google. London's Science Museum Group holds a 1999 Google "cork board" server and says the young company "made do with inexpensive materials including off the shelf personal computers and thin sheets of corkboard for insulation".
Credit
Google, founded by Larry Page and Sergey Brin; the Science Museum Group names King Star Computer Inc. of California as the maker of its server
2011
Server designs given away
Facebook started the Open Compute Project in 2011 to publish the designs of the servers, racks, power supplies and cooling it had built for its data center in Prineville, Oregon. By Facebook's own figures that building was 38% more energy efficient and 24% cheaper to run than its earlier ones. The project is now run by an independent foundation with members across the industry.
Credit
Facebook, now Meta
2015
Networks from cheap switches
Google reported at SIGCOMM 2015, a peer-reviewed networking conference, that five generations of its data center networks had been built as Clos networks: layers of small switches made from off-the-shelf switch chips, wired so that any server can reach any other by many paths. Google said the design grew its network capacity 100 times in ten years, to more than 1 petabit per second (a million gigabits per second) across the middle of the network.
Credit
Arjun Singh, Amin Vahdat and others at Google
2017
Cards take over the chores
Amazon Web Services launched its Nitro System in 2017. Separate cards in each server handle networking, storage and security, and a small hypervisor (the software that divides one machine among customers) allocates memory and processors, so that, AWS says, its own staff have no way into customer data. In 2023 NCC Group, a security firm AWS hired for an independent architecture review, "found no gaps in the Nitro System that would compromise" AWS's security claims.
Credit
Amazon Web Services; review by NCC Group
2018
Shared processors leak secrets
On 3 January 2018 Google's Project Zero described flaws, later named Spectre and Meltdown, in the way modern processors work ahead of time, which let one program read memory it should never see. It had reported them to Intel, AMD and ARM on 1 June 2017. The Meltdown researchers warned that on some cloud set-ups "data from other customers" could be stolen.
LessonCustomers who share one physical server are only as separate as its processor keeps them.
Credit
Jann Horn of Google Project Zero, with researchers at Graz University of Technology, Cyberus Technology, the University of Pennsylvania, the University of Maryland, Rambus and the University of Adelaide
2022
Mirrors replace patch panels
Google reported at SIGCOMM 2022 that its Jupiter networks now run through optical circuit switches, which steer light from fiber to fiber with tiny movable mirrors, so the network can be rewired by software instead of by hand at a patch panel. Google credits a decade of such changes with 5 times the speed and capacity, 30% lower capital cost and 41% lower power use, and says the network can now be reconfigured 3 times faster.
Credit
Leon Poutievski, Amin Vahdat and others at Google
Watch for
Old memory in new servers
Meta says Vistara, a chip it designed, lets its newest AMD servers, which take only DDR5 memory, also use DDR4 memory sticks pulled from retired servers. The chip connects the old memory over CXL, a link that lets a processor use memory on a separate card as if it were its own. Tom's Hardware reports that each Meta MemServer pairs 768 GB of new DDR5 with 256 GB of recovered DDR4, and EE Times Japan reports that Meta cut the number of servers it needed by 25%.
Achieved whenMeta or an outside party publishes error and failure rates for the recovered memory after a year in service, or a second operator runs the same scheme.
Claimed by
Meta, June 2026
Ethernet rebuilt for AI clusters
The Ultra Ethernet Consortium, a group of chip, network and cloud companies, published Specification 1.0 on 11 June 2025 and version 1.1 on 12 October 2026: a version of Ethernet for AI and supercomputer clusters that covers network cards, switches, optics and cables. Makers certify their own products; Network World reports that "Vendors claim compliance by self-attestation".
Achieved whenan independent test lab publishes results showing network cards and switches from at least two makers working together at full speed under the specification.
Claimed by
Ultra Ethernet Consortium, June 2025 and October 2026
Two AI chips per package
Google says TPU 8i Max, two of its TPU 8i AI chips in one package, delivers "double the compute, double the SRAM" (the fast memory on the chip) and "fits twice the silicon within the existing rack footprint". Google has published no test of it, and the announcement gives no date when customers can rent it (see the Note "Double the compute").
Achieved whencustomers can rent it and an outside benchmark, such as an MLPerf result, shows the gain per rack.
Claimed by
Amin Vahdat of Google, November 2026
Open
01
Test second-hand server memory
A public test of recovered DDR4 server memory sticks (registered DIMMs) bought on the used market. Each stick runs memory stress tests for days, and its corrected and uncorrected error counts are published by age, maker and size. A subset is then run again behind a CXL memory card, the way Meta reuses old memory.
The briefCredit · unclaimed02
Screen used processors for errors
Run open processor test suites on a batch of second-hand server processors bought on the used market, and publish how many give wrong answers without any error signal, by model and age.
The briefCredit · unclaimed
Open 01 · the brief
Test second-hand server memory
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of recovered DDR4 server memory sticks (registered DIMMs) bought on the used market. Each stick runs memory stress tests for days, and its corrected and uncorrected error counts are published by age, maker and size. A subset is then run again behind a CXL memory card, the way Meta reuses old memory.
Why it is still open
Meta's Vistara chip is "available exclusively to Meta", by Tom's Hardware's account, and Meta has published no error rates for the old memory. A Marvell executive told EE Times Japan in August 2026 that one hyperscaler believes failure rates of reusable DDR4 are not high, but that exact figures have to come from each company. No published error rate for second-hand server memory turned up.
Built with what exists
Used DDR4 registered DIMMs and a used two-socket server from resellers; Google's open-source stressapptest or MemTest86; the server's error-correction logs read with the Linux tool rasdaemon; a CXL memory card, where one can be bought.
First step
Buy 20 used 32 GB DDR4 registered DIMMs, run stressapptest on each for 24 hours, and publish the logs.
Done looks like
At least 200 sticks from at least three makers, each run for 72 hours, with per-stick error counts and the test settings published as an open data set.
Open 02 · the brief
Screen used processors for errors
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Run open processor test suites on a batch of second-hand server processors bought on the used market, and publish how many give wrong answers without any error signal, by model and age.
Why it is still open
Hyperscalers have reported "about one CPU per thousand" in their fleets with silent data corruption, but researchers from Meta and the University of Athens wrote in 2025 that "there is no public data on the rate of SDC", and gave their own fleet figures only in relative terms. Used processors, which go to smaller buyers, have no published rate.
Built with what exists
OpenDCDiag, an open-source test suite for finding defects in processors (Apache 2.0 licence, github.com/opendcdiag/opendcdiag); Intel's DCDiag tool for Xeon chips; used two-socket servers; a script to log and compare results.
First step
Run OpenDCDiag for 24 hours on ten used Xeon processors in one server and publish the logs.
Done looks like
At least 300 processors of at least five models, each run through the full suite for 24 hours, with every failure confirmed in a second machine and failure counts per model published.