What to Do With Old Phones: Google Research Says Build a Cloud Out of Them

What to do with old phones: Google Research says build a cloud out of them

5.3 billion phones fell out of use in 2022, and most of them went into a drawer. In June, Google Research published its answer to what to do with old phones: don’t shred them, don’t hoard them. Rack them, and run cloud workloads on them.

 

The post, co-authored by Turing Award winner David Patterson, describes a platform being built at UC San Diego with Google’s backing: a datacenter of 2,000 retired Pixel phones, scheduled to go live in Fall 2026. It will serve coding notebooks and course backends for hundreds of researchers and students: the kind of workloads that normally rent a small server from Amazon Web Services.

 

It is the same reasoning Acurast has been running in production for years, and the numbers are better than most people expect.

The science: your phone’s carbon is already spent

A computer’s climate footprint splits in two. Operational carbon is the electricity it burns while running. Embodied carbon is everything emitted while making it: mining, fabs, assembly, shipping.

 

For servers in a modern datacenter, embodied carbon is the part that dominates as grids get cleaner. For phones, it is not even close: by Apple’s own environmental report, around 80% of an iPhone’s lifetime CO2 is emitted before the box is opened. The phone pays its climate cost up front. Then it gets retired after roughly a quarter of its functional lifespan, since the average phone is replaced within a few years while its processor works for ten or more.

 

The academic result underneath Google’s post is a 2023 paper called Junkyard Computing, from the UCSD group now building the platform. It won a Distinguished Paper Award at ASPLOS, one of the top venues in computer architecture, and became the most-downloaded paper in the conference’s 28-year history. The team bought ten used Pixel 3As on eBay, clustered them, and ran the same web-service benchmarks a cloud server would run:

 

Carbon

Per query over three years, the phone cluster came out 9.8x to 18.9x more carbon efficient than a comparable Amazon cloud server, because the phones’ embodied carbon was already spent, and a Pixel 3A draws about 1.5 watts against a server’s 300.

 

Cost

The ten-phone cluster cost $1,027 to buy and power for three years. The equivalent Amazon server rental cost $40,404.

 

Performance

The cluster held 4,000 requests per second with even the slowest requests answered inside 100 milliseconds, on par with the server it replaced. Google’s newer benchmarks go further: single-threaded, a recent Pixel’s performance cores match or beat baseline server cores on most of SPEC CPU2017, an industry-standard set of processor benchmarks.

 

“Junkyard computers grow global computing capacity by extending device lifetimes, which supplants the manufacture of new devices,” the paper concludes. The greenest computer is the one that already exists.

The timing: a GPU shortage and the worst memory crunch in decades

The carbon argument has been true for years. What changed is that the economic argument became impossible to ignore.

 

AI datacenter demand has swallowed the hardware supply chain. Memory contract prices jumped 90 to 95% in a single quarter in early 2026, according to market researcher TrendForce. SK Hynix sold out its entire 2026 memory production in October 2025. Samsung and SK Hynix have signed letters of intent to supply OpenAI’s Stargate datacenter project with up to 40% of the world’s output of DRAM, the working memory in every computer and phone. Nvidia’s cloud GPUs are, in its CEO’s words, sold out.

 

The bill has reached consumers. Apple raised MacBook and iPad prices by up to $300 in June. “We have never seen a component price increase this much, this quickly,” the company said. Xbox prices rose $100 to $150 in August. Xiaomi warned that phones get more expensive in 2026 for the same reason. Micron’s CEO expects supply to improve gradually in 2028, not before.

 

Even the hyperscalers are stuck. Microsoft’s CEO said late last year that his constraint is no longer chips but electricity: GPUs sitting in inventory with no powered building to plug them into. The big four cloud providers plan roughly $700 billion of capital spending in 2026, much of it on buildings and power.

 

Now look back at the drawer. A retired phone is a computer with 8 to 12 GB of memory, the component at the center of a global shortage. It is already manufactured, already paid for, already attached to a battery and a network connection. There are billions of them. During a hardware famine, the world’s largest stock of idle silicon is sitting in kitchen drawers.

 

From a 2,000-phone lab to a 287,000-phone network

Google’s platform is a lab deployment: motherboards extracted from the phones, batteries and screens removed, boards wired into racks, Android swapped for Linux, Kubernetes on top. It launches in Fall 2026 with 2,000 phones, roughly 50 server-equivalents.

 

Acurast reached the same conclusion from the other direction, and it has been live for years. The network runs the phone whole, and at a different scale:

Keeping the phone intact is not laziness. It is the point. Three things survive that the rack approach throws away:

 

The security chip

Modern phones ship a Trusted Execution Environment, the hardware-isolated part of the chip that banking apps and biometrics rely on. Acurast runs workloads inside it and uses key attestation: the silicon cryptographically proving what it is and what it runs. Strip the phone down to a bare board with a fresh Linux image and that verifiable, confidential execution story is exactly what you give up.

 

The battery and the location

A phone carries its own uninterruptible power supply, and it is already distributed to wherever people live, which is why the network spans more than 175 countries without a single building being constructed. No land, no shell, no cooling plant, no queue for grid power.

 

The owner

Nobody has to collect, disassemble and rack millions of phones. The owner installs the Processor app, scans a QR code, and is onboarded in about three minutes. Phones with broken screens work fine. The phone earns instead of depreciating.

 

Developers on the other side deploy Node.js workloads, Linux containers, and LLM inference: models in GGUF format (the standard file format for running open-source models) behind OpenAI-compatible endpoints. How the network matches work to phones is its own story.

 

The honest limits

A phone is not a rack of AI accelerators. Nobody trains a frontier model on smartphones, and the Junkyard Computing authors are clear that phones favour bursty workloads over sustained full-throttle compute.

 

But the bulk of the cloud is not frontier training. It is exactly what UCSD is targeting: notebooks, APIs, bots, scheduled jobs, small inference. The workloads of the everyday internet fit in a phone. That is not a compromise; it is a match.

 

Why this matters

The Global E-waste Monitor counted 62 million tonnes of e-waste in 2022 and rising, with $91 billion of metals inside it and less than a quarter formally recycled. Meanwhile the industry plans hundreds of billions of dollars of new datacenters it struggles to power, built from components it struggles to buy.

 

Google Research looked at both facts and proposed a low-carbon cloud built from retired phones. The proposal validates a bet Acurast made years ago: the most sustainable, most affordable compute buildout is the one that requires no building at all.

Sources: WEEE Forum, 2022; Junkyard Computing, ASPLOS 2023 (10% of five years of decommissioned phones); Acurast network, August 2026.

 

Second-life compute is no longer a research curiosity. It is a live network: 287,000 phones in more than 175 countries, secured by the hardware in your pocket, powered by silicon whose carbon was spent long ago.

 

If you have a spare phone, you can be part of the infrastructure shift and get rewarded for your contribution. For the decentralized compute network, every phone matters! Become a Compute Provider.

 

Building on Acurast? If you’re running your own models, agent backends, or other services on attested devices, come talk to us. Join the Discord.

Related Blogs