Player wiki
Hardware roadmap
Keynotes, pre-orders, lead times and allocation — knowing what ships next decides when you buy, build and sell.
Silicon is announced at keynotes months before it ships. The game fires keynotes on their real calendar dates — GTC (spring), Computex (May/June), SC (November), OCP, and AMD's Advancing AI events — and each one pauses the clock with a keynote card explaining the technology, what it changes in your hall, and the one thing to remember. The Roadmap tab in Procurement lists every past keynote with a Read the keynote button and shows the next one coming.
Pre-orders and allocation
After a keynote, announced parts become pre-orderable:
- Pre-orders open at the keynote, never earlier than 12 months before the ship date
- You pay a launch premium and queue behind the hyperscalers; delivery lands after the ship date
- Cancelling refunds most of what you paid — a pre-order is a cheap option on supply
- Even regular orders of hot parts carry per-quarter allocation caps; vendor relationships and frameworks raise your allocation (see Money and operations)
If you wait instead, you pay the going rate and stand in the queue with everyone else — hot generations sell out before they ship.
The shape of the campaign
A few beats worth planning around (dates as they land in-game):
- GTC 2021 — Grace. A roadmap keynote: an Arm CPU built to sit next to GPUs over NVLink-C2C. Nothing to buy for years; note the direction.
- GTC 2022 — Hopper H100. FP8 Transformer Engine, NVLink 4 at 900 GB/s between GPUs, ~3.35 TB/s HBM3, 700 W per SXM GPU — an 8-GPU node is ~10 kW. A GPU generation is a power and cooling project first and a chip purchase second.
- 2023 — GH200, Spectrum-X, MI300X, H200. Memory becomes the story: MI300X brings 192 GB of HBM3, and H200 shows that when a workload is memory-bound, faster memory is worth more than faster math. Ethernet becomes a real alternative to InfiniBand for scale-out.
- GTC 2024 — Blackwell. GB200 NVL72 links 72 GPUs into one liquid-cooled ~120 kW rack (NVLink 5 at 1.8 TB/s per GPU). Air-cooled halls cannot host it at all; the floor needs a CDU loop.
- Computex 2024 — the yearly cadence. All vendors move to annual generations. Everything you buy depreciates faster; resale timing becomes part of the business model.
- GTC 2025 — Blackwell Ultra, Vera Rubin roadmap. GB300 drops into GB200-class racks: same ~120 kW, so the hall you built keeps working. Rubin racks will draw considerably more — plan bigger feeds and CDUs early.
- Oct 2025 — Vera Rubin orders open. Longer queues than any prior generation; earlier-generation prices soften as buyers wait. Sell older racks while residual value holds.
- GTC 2026 — Rubin Ultra, 102.4 Tb/s switches. 576-GPU NVLink domains; bigger switch silicon means fewer tiers and hops per GPU.
Budget cards: RTX and other PCIe GPUs
Not every rack needs a flagship. PCIe cards fit ordinary servers, run on air and arrive in a month or two.
| Card | Ships | Memory | Notes |
|---|---|---|---|
| GeForce RTX 3090 | late 2020 | 24 GB | cheapest compute of its year |
| RTX A6000 | late 2020 | 48 GB | workstation card: ECC memory, blower cooler |
| GeForce RTX 4090 | late 2022 | 24 GB | close to L40S compute at a fifth of the price |
| RTX 6000 Ada | late 2022 | 48 GB | an L40S a year early |
| L40S | late 2023 | 48 GB | the standard budget server card |
| GeForce RTX 5090 | early 2025 | 32 GB | 575 W each; supports FP4 |
| RTX PRO 6000 Blackwell | 2025 | 96 GB | most memory per dollar |
The GeForce cards are consumer parts: no ECC memory and coolers built for a desktop case, so in a packed server they deliver less than the spec sheet says. The game models that as lower efficiency, not as a ban. With 24 to 32 GB each they also cannot hold large models, and with no fast link between cards a model split across several of them runs slowly. They suit small inference contracts. Training contracts need GPUs linked in groups of eight or more, which no PCIe card can do.
When to buy
- Plan the building two generations ahead. Racks change yearly; power and cooling plants last decades. Buy transformers and CDU capacity for the rack that ships next, not the one shipping today.
- Pre-order at the keynote when the part changes your market — e.g. a rack format that unlocks contracts you can't serve — and treat the cancellation loss as the price of a place in the queue.
- Sell into strength. When a generation lands, its predecessor becomes a discounted workhorse: reassign it to inference or spot, or sell it while residual value is real.
- Watch the calendar beats. The 2023 demand shock slashes allocation and slips in-transit GPU orders; later, vendors reward operators with demonstrated deployments — a campus that ships on schedule gets first call on the next generation.
The Codex's accelerator timeline view shows every generation with memory, bandwidth, power and scale-up — the fastest way to compare what to buy.
Related: Hardware lifecycle · Servers and racks · Power and cooling.
Found a mistake or something missing? Edit this page on GitHub or ask in Discussions.