Bit Market · Sourced milestones
Racks and Rows
SI clusters, as a record of progress: how many accelerators one cluster holds, the flash and storage behind them, and the gigawatt campuses still on paper. 17 milestones from 2020 to 2026, each linked to the release it comes from.
Chart of accelerators, storage, memory and compute in one SI cluster by year
Accelerators in one cluster went from 10,000 in 2020 to over 220,000 in 2026
Storage in one cluster went from 231,000TB in 2022 to over 500,000TB in 2025
- Record: the most in one running cluster by then
- Hollow: a plan, not yet running
5 records, oldest first. Each cluster ran more accelerators than any running cluster before it.
5/2020
10,000
Microsoft supercomputer for OpenAI
- Who
- Microsoft
- Stage
- Operating
5/2023
16,000
- Who
- Meta
- Stage
- Operating
3/2024
24,576
- Who
- Meta
- Stage
- Operating
2/2025
200,000
- Who
- xAI
- Stage
- Operating
5/2026
over 220,000
- Who
- SpaceX (xAI)
- Stage
- Operating
2 records, oldest first. Each cluster ran more storage than any running cluster before it.
1/2022
231,000TB
- Who
- Meta
- Stage
- Operating
2/2025
over 500,000TB
- Who
- xAI
- Stage
- Operating
How to read it
Each row is a milestone: a cluster its owner says is running, a planned campus, or a hardware or supply deal. The page holds 17 milestones from 17 public sources, 8 of them plans. A row sits at the date its own source prints, to the month at least.
The Accelerators chart counts the accelerators in one cluster, on a log scale where each gridline step multiplies the value. Most are NVIDIA GPUs; AWS’s Project Rainier counts Trainium2 chips, which its source gives as nearly half a million. A cluster sets a record only once its owner says it runs. A planned campus, such as a Stargate site, is a plan for power and chips, not a running cluster, so it is drawn hollow and never sets a record; its gigawatts are in the table.
Storage is the figure the owner prints for the cluster, in terabytes: Meta’s 2022 Research SuperCluster added up 231 PB across three tiers, and xAI gives at least 500 PB for Colossus. Sources are the companies themselves; news sites and trade press can point to a milestone, but the row cites the primary page.
Every milestone and its source
Newest first. The Lane chips above filter this table and the chart together.
- means the source does not state this figure.
- D Derived: worked out from what the source states (a unit change, or one figure divided by another), not printed by it.
Tap any column heading to sort the table by it. Tap it again to reverse the order.
| Milestone | Who | Lane | Stage | Date | GPUs | Storage | Compute | Power | Source | Note |
|---|---|---|---|---|---|---|---|---|---|---|
| New Compute Partnership with Anthropic | SpaceX (xAI) | SI clusters | Operating | 5/2026 | ≥220,000 | not published | not published | not published | SpaceXSI 2026 release | Colossus 1 has over 220,000 NVIDIA GPUs. Anthropic signed for access to it and said it is interested in developing multiple gigawatts of orbital compute with SpaceX. |
| AWS activates Project Rainier: One of the world’s largest AI compute clusters comes online | AWS | SI clusters | Operating | 10/2025 | ≤500,000 | not published | not published | not published | Amazon 2025 release | Nearly half a million Trainium2 accelerators in use less than a year after the project was announced, with Anthropic running workloads. |
| AMD and OpenAI Announce Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs | AMD and OpenAI | Hardware and supply deals | Forecast | 10/2025 | not published | not published | not published | 6GW | AMD 2025 release | The first 1GW of AMD Instinct MI450 GPUs starts in the second half of 2026. |
| Meta’s Infrastructure Evolution and the Advent of AI | Meta | SI clusters | Forecast | 9/2025 | not published | not published | not published | 5GW | Meta 2025 release | Expected to come online from 2028 and scale to 5 gigawatts. A plan, drawn faint. |
| Meta’s Infrastructure Evolution and the Advent of AI | Meta | SI clusters | Forecast | 9/2025 | not published | not published | not published | 1GW | Meta 2025 release | A 1 gigawatt cluster spanning several data center buildings, under construction when Meta wrote. A plan, drawn faint. |
| OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites | OpenAI, Oracle and SoftBank | SI clusters | Forecast | 9/2025 | not published | not published | not published | ≤7GW | OpenAI 2025 release | The new sites bring Stargate to nearly 7GW of planned capacity, counting Abilene and the CoreWeave projects. |
| OpenAI and NVIDIA Announce Strategic Partnership to Deploy 10 Gigawatts of NVIDIA Systems | NVIDIA and OpenAI | Hardware and supply deals | Forecast | 9/2025 | not published | not published | not published | ≥10GW | NVIDIA 2025 release | At least 10GW; the first gigawatt is on Vera Rubin in the second half of 2026, and NVIDIA intends to invest up to $100 billion. |
| Inside the world’s most powerful AI datacenter | Microsoft | SI clusters | Operating | 9/2025 | not published | not published | not published | 3.3GW | Microsoft 2025 releaseEpoch AI 2025 | Fairwater in Wisconsin: one cluster of NVIDIA GB200 servers with exabytes of storage, on a 315 acre site. Microsoft prints no power figure; the 3.3GW is Epoch AI’s projection for late 2027. |
| Introducing Stargate Norway | OpenAI | SI clusters | Forecast | 7/2025 | 100,000 | not published | not published | 0.23GW | OpenAI 2025 release | 230MW planned, with ambitions for 290MW more; the target is 100,000 NVIDIA GPUs by the end of 2026. |
| Stargate Advances with 4.5 GW Partnership with Oracle | OpenAI and Oracle | SI clusters | Forecast | 7/2025 | not published | not published | not published | 4.5GW | OpenAI 2025 release | |
| Introducing Stargate UAE | OpenAI | SI clusters | Forecast | 5/2025 | not published | not published | not published | 1GW | OpenAI 2025 release | A 1GW cluster in Abu Dhabi, with 200MW expected to go live in 2026. |
| Doubled to 200K GPUs | xAI | SI clusters | Operating | 2/2025 | 200,000 | ≥500,000TB | not published | not published | xAI 2026 release | The Colossus page dates full scale to February 2025 and lists more than 0.5EB of storage for training data and checkpoints. |
| NVIDIA Blackwell Platform Arrives to Power a New Era of Computing | NVIDIA | Hardware and supply deals | Announced | 3/2024 | 72 | not published | not published | not published | NVIDIA 2024 release | One GB200 NVL72 rack joins 72 Blackwell GPUs and 36 Grace CPUs over fifth-generation NVLink. |
| Building Meta’s GenAI Infrastructure | Meta | SI clusters | Operating | 3/2024 | 24,576 | not published | not published | not published | Meta 2024 release | Two clusters of 24,576 H100 GPUs each, used to train Llama 3, with storage on high-capacity E1.S SSDs. |
| Pursuing groundbreaking scale and accelerating research using Meta’s Research SuperCluster | Meta | SI clusters | Operating | 5/2023 | 16,000 | not published | ≤5,000,000TFLOPS | not published | Meta 2023 release | The second phase complete: 2,000 DGX A100 systems, 16,000 A100 GPUs and almost 5 exaflops. |
| Introducing the AI Research SuperCluster | Meta | SI clusters | Operating | 1/2022 | 6,080 | 231,000TB, derived | not published | not published | Meta 2022 release | Phase one: 6,080 A100 GPUs, with a storage tier of 175 PB of FlashArray from Pure Storage (renamed Everpure on February 23, 2026), 46 PB of cache storage and 10 PB of FlashBlade. |
| Microsoft announces new supercomputer, lays out vision for future AI work | Microsoft | SI clusters | Operating | 5/2020 | 10,000 | not published | not published | not published | Microsoft 2020 release | One system with more than 285,000 CPU cores, 10,000 GPUs and 400 Gb/s of network per GPU server, among the top five publicly disclosed supercomputers. |
- SI clusters. Clusters their owners say are running, with the accelerators and storage they print, and the planned campuses that are not yet built.
- Hardware and supply deals. Accelerator platforms and the supply deals behind the build-out. Memory per accelerator is on the HBM timeline.
What it will not tell you
This is a record of clusters with a dated public source, not a list of every data center. Many owners never print how many accelerators a cluster holds, and Microsoft gives no count for Fairwater, so it sits in the table without a point. A gigawatt in an announcement is a plan for capacity, not power drawn today, and an investment pledge with no capacity behind it is left off.
The memory on each accelerator is on the HBM Timeline, and what one gigawatt takes in wafers and EUV scanners is in What a Gigawatt Takes. For memory and compute in space, see Signals; for optics for memory, Photon Fabric; and for the makers’ results each quarter, the Earnings Desk.