top of page

Google Just Bet Its Cloud on EPYC. The Generation That Started It Costs $48.

Writer: Brian Lathe
Brian Lathe
Nov 9
6 min read

Google Cloud's newest machines — the C4D, the N4D, the Z4D, even its GPU boxes — all run on AMD EPYC processors now. The current generation is called Turin, and it is a sweep: instance family after instance family, all on the same socket family. Wall Street noticed — AMD's stock jumped about 13% in a single day on the news. I noticed for a different reason. The platform Google just bet its cloud on is the same SP3 socket family sitting at the top of my own build sheet. And the generation that started the whole story — EPYC Rome, from 2019 — now sells used for less than a decent dinner out. The entry chip is $48.


This is the story of how that happened, what that 2019 generation actually costs on the used shelf right now, and the one trap that catches almost everybody who goes shopping for it.


The bet Google just made

Turin is AMD's fifth-generation EPYC, and Google has rolled it out across the fleet:

  • C4D — general compute, up to 384 vCPUs and 3 TB of DDR5 memory. Google reports about 30% higher performance than the generation before it on SPECrate and 55% more MySQL queries — and C4D is Google's first AMD-based bare-metal instance family.

  • N4D — the cost-optimized workhorse: up to 96 vCPUs and 768 GB of memory, with Google claiming up to 50% better price-performance than its predecessor and 3.5x the web-serving throughput.

  • Z4D — storage-optimized, with up to 84,000 GiB of local SSD, 15.6 million random-read IOPS, and roughly 40% gains over the Z3 — and the first AMD instance to support Nutanix Cloud Clusters on bare metal.

  • G4 — even Google's GPU machines, built around NVIDIA's RTX PRO 6000 Blackwell cards, sit on EPYC Turin hosts.

That is not a vendor experiment. That is a fleet decision — the kind a company makes when it expects to buy a platform by the tens of thousands and run it for years.


It started with Rome

Turin is the fifth generation. The break happened in the second. In August 2019 AMD launched EPYC 7002 — Rome — straight at Intel's Cascade Lake Xeon line, and it wasn't a fair fight:

  • Cores: Rome went to 64 cores and 128 threads in a single socket. Cascade Lake topped out at 28.

  • Memory: Rome fed those cores with 8 channels of DDR4-3200, about 205 GB/s of bandwidth. Intel had 6 channels, about 141 GB/s.

  • I/O: Rome offered 128 lanes of PCIe Gen4. Intel offered 48 lanes of Gen3.

  • Design: Rome was 7nm chiplets arranged around an I/O die. Cascade Lake was a 14nm monolith.

Intel's one real edge was AVX-512, a wider vector instruction set — and even that came with clock-throttling caveats when it actually engaged. Everything else ran AMD's way. Under Rome's pressure, Intel was forced into early SKU discontinuations and price cuts on Cascade Lake — Tom's Hardware documented the retreat at the time. Amazon shipped Rome instances. The cloud defection pattern that ends at Google's Turin fleet started here: Rome didn't just win benchmarks, it broke a pricing regime that had held for a decade.


What that generation costs now

Here is the part that matters to a home lab. The chips that broke Intel's pricing in 2019 have now depreciated into pocket money. These are used-market readings from November 2026 — this market moves, so treat them as a dated snapshot, not a quote:

  • EPYC 7452 (32 cores): launched at $2,025. It sells used today around $179–$212 — about a tenth of launch.

  • EPYC 7702P (64 cores): the 7702 launched at $6,450. The single-socket P version goes for roughly $699–$876 unlocked.

  • EPYC 7302P (16 cores): $825 at launch. One Minnesota seller has it at $62.95.

  • EPYC 7252 (8 cores): $47.95, in stock, refurbished with a one-year warranty.

  • EPYC 7502 (32 cores, not vendor-locked): $159 at that same seller, with 91 on the shelf — while a different seller asks about $463 for the same chip. The spread is the market. Shop carefully.

The chips are the cheap part of this platform. The board is the grown-up line item: a Supermicro H11SSL-i runs about $390, and Gigabyte’s MZ32-AR0 — the board on my own build sheet — sits on the same used shelf. Budget for the platform, not just the processor.


That seller is Core 4 Solutions, in Eden Prairie, Minnesota — about 45 minutes from my desk. Local color worth knowing: Core 4 is also the seller behind the viral $59.7M listing — 300 racks holding 9,600 A100 GPUs and 4,200 64-core Rome chips out of a decommissioned hyperscaler cluster. The Rome firehose isn't abstract. It's physically sitting in Eden Prairie.


One caution from my own notes: that same store lists the 7252's part number at two different prices — $47.95 and $295. Ask which is which before money moves. That's not a knock on the seller; it's how this whole market works. Verify everything.


The trap: half-price chips that won't boot for you

Now the reader-service part — the thing that burns first-time buyers. Many used EPYC chips are vendor-locked: a feature called Platform Secure Boot fuses the chip to one manufacturer, so a Dell-locked chip boots only in Dell boards, an HP-locked chip only in HP boards, and so on. Locked chips are the cheap listings. A Dell-locked EPYC 7402P can go as low as $40, while the unlocked version of the same chip runs $78–$110.


That gap is not a bargain. It's the market pricing in the lock. Drop a locked chip into a Supermicro or Gigabyte board and you don't have a server — you have a $40 paperweight. The rule is simple: buy listings that say unlocked or not vendor-locked, verify the part number, and treat any listing that stays silent about lock status as locked until proven otherwise.


Why this socket is on my build sheet

Full disclosure, because this blog runs on it: I don't own one of these machines yet. This is a planned build — sheets drawn, buy rules written, prices watched daily by my own systems. I write about it as a plan, not as a machine I run, and I'll say so again when it's built and measured.


The build on my sheet is a single-socket SP3 machine: a Gigabyte MZ32-AR0 board (Supermicro's H11SSL-i runs around $390 — and mind the board revision, Rome support needs rev 2.x), 256 GB of DDR4-3200 as eight 32 GB sticks — one DIMM per channel, all eight channels fed — dual Tesla V100 32GB cards, and fast NVMe storage.


Why that shape? The workload it's being built for is Colibri, the streaming inference engine at the heart of my local AI work. Colibri doesn't park a giant model in video memory; it streams model experts from fast storage through system memory, token by token. That makes the machine a funnel — and for a funnel, the spec that matters is channels and lanes, not clock speed. Rome is the widest, cheapest funnel the server world ever mass-produced: 8 memory channels, 128 PCIe lanes, and a used market practically giving the platform away. My fleet already runs Colibri work today on other hardware; this build is the engine getting a home shaped exactly like its appetite.


To be fair to the other side: Cascade Lake Xeons are cheap used too, and AVX-512 genuinely helps a few workloads. If yours is one of them, shop that shelf. But channel for channel and lane for lane, the throughput math isn't close.


Next in this series

This piece was the story of the socket. The next one is the shape: why a streaming inference engine dictates a very specific machine — single socket on purpose, memory channels maxed, drive slots counted before cores — and what my exact build costs, line by line, against a modern desktop trying to do the same job. That one's next. If you go shopping the Rome shelf in the meantime, reread the trap section before you reread the prices.


Read more from Brian Lathe on Substack


This blog is one room in a bigger house. My main writing lives on Substack - longer pieces, new posts first, always free to read: Brian Lathe on Substack.

 
 
 

Recent Posts

See All
Do You Need an Integrator If You're a Team of One?

Do You Need an Integrator If You're a Team of One?If you've read anything about EOS — Gino Wickman's Entrepreneurial Operating System, from Traction — you've met the famous pair. The Visionary is the

 
 
 

Comments


Talk to Brian

on WhatsApp

Appointment booking opens soon.

Questions about this article — or about SuccessBrian OS? Text me here.

Texting is best — please don't call. Calls are by confirmed appointment only.

Talk to Brian

Questions, or want a system like this working for your business? Message me directly.

What is SuccessBrian OS?

One system that runs a solo business while its owner lives his life: a checked second brain, a crew of AI agents, and one calendar that never forgets. The whole system is free and open.

Get new posts by email

I build this in public and post the progress. Follow along on Substack.

bottom of page