Peter Wildeford: I’d expect the 10GW OpenAI cluster becomes operational around 2027-2028.
There are no 10 GW OpenAI clusters, there is a conditional investment by Nvidia for every additional 1 GW in total across all their datacenters, which won’t be forming a single training system. Inference and smaller training experiments need a lot of compute, and OpenAI is now building their own even for inference, so most of what they are building is not for frontier model training.
Zvi Mowshowitz: these announcements plus the original cite in Abilene, Texas cover over $400 billion and 7-gigawatts over three years
Depending on OpenAI growth, this is more of a soft upper bound on what gets built. This is evidence that 5 GW training systems likely aren’t going to be built by (end of) 2028. So there’s going to be a slowdown compared to the trend of 12x in compute (and 6x in power) every 2 years for the largest frontier AI training systems, which held in 2022-2026[1]. In 2027-2028, the largest training systems are likely going to be merely 2 GW instead of the on-trend 5 GW. Though for the 2024 systems, FP8 is likely relevant, and for 2026 systems maybe even FP4, which turns the 12x in compute every 2 years in 2022-2026 into 24x in compute every 2 years (5x per year in pretraining-relevant raw compute for a single training system).
This 24x every 2 years is even less plausible to remain on-trend in 2028, so 2027+ is going to be the time of scaling slowdown, at least for pretraining. Though the AIs trained on 2026 compute might only come out in 2027-2028, judging by how 2024 training compute is still held back by inference capabilities, and some AIs enabled by 2024 levels of compute might only come out in 2026. So the slowdown in the scale of frontier AI training systems after 2026 might only start being observable in scaling of deployed AIs starting in 2028-2029.
Perhaps some of these sites will be connected with sufficient bandwidth, and training with RLVR at multiple sites doesn’t need a lot of bandwidth. Actual plans for larger training runs should urge them to build larger individual sites, as this ensures optionality for unusual training processes. So the fact that this isn’t happening suggests that there are no such plans for now (except perhaps for RLVR-like things specifically).
24K A100s in 2022 (7e18 BF16 FLOP/s, 22 MW), 100K H100s in 2024 (1e20 BF16 FLOP/s, 150 MW), 400K chips in GB200/GB300 NVL72 racks in 2026 (1e21 BF16 FLOP/s, 900 MW). The power estimates are all-in, for the whole datacenter site.
I want to withdraw my prediction “I’d expect the 10GW OpenAI cluster becomes operational around 2027-2028.” I spoke too quickly in my Twitter thread and this was based on a confusion on my part.
I will have a more detailed article soon to give more thoughtful updated predictions. I apologize for this error.
Forgive if a naive question, but what about distributed training runs — any view on whether progress on that front will result in training runs larger than what you describe here?
PS thank you for the extremely useful analysis as always — I would 100% subscribe if you had a newsletter or something (with a strictly positive willingness-to-pay even!).
There doesn’t necessarily need to be algorithmic progress to get there, sufficient bandwidth enables traditional pretraining across multiple sites. But it might be difficult to ensure it’s available across the geographically distributed sites on short notice, if you aren’t already a well-established hyperscaler building near your older datacenter sites.
In 2028, targeting inference on Rubin Ultra NVL576 (150 TB of HBM in a scale-up world) might want a MoE model with 80 TB of total params (80T params if in FP8, 160T in FP4). If training uses the same precision for gradients, that’s also 80 TB of gradients to exchange. If averaged gradients use more precision, this could be 2x-8x more data.
If training is done using 2 GW of some kind of Rubin GPUs, that’s about 2e22-3e22 FP4 FLOP/s, and at 30% utilization for 4 months it produces 8e28 FP4 FLOPs. At 120 tokens/param (anchoring to 40 tokens/param for the dense Llama 3 405B and adjusting 3x for 1:8 sparsity), this system might want about 10T active params (so we get 1:16 sparsity, with 160T total FP4 params, or about 1:8 for FP8). This needs 1,200T tokens, maybe 250T unique, which is a problem, but not yet orders of magnitude beyond the pale, so probably something can still be done without needing bigger models.
With large scale-up worlds, processing sequences of 32K tokens with non-CPX Rubin NVL144 at 30% utilization would take just 2.7 seconds (for pretraining). A 2 GW system has 9K racks, so that’s a batch of 300M tokens, which is already a lot (Llama 3 405B used 16M token batches in the main phase of pretraining), so that should be the target characteristic time for exchanging gradients.
Moving 80 TB in 2.7 seconds needs 240 Tbps, or 500-2,000 Tbps if averaged gradients use 2x-8x more precision bits (even more if not all-to-all, which is likely with more than 2 sites), and this already loses half of utilization or asks for even larger batches. A DWDM system might transmit 30-70 Tbps over a fiber optic pair, so this is 4-70 fiber optic pairs, which seems in principle feasible to secure for overland fiber cables (which hold hundreds of pairs), especially towards the lower end of the estimate.
Depending on OpenAI growth, this is more of a soft upper bound on what gets built.
I’m confused: the announcement indicates that the $400B has been committed, and is not dependent on OpenAI’s growth (although perhaps you’re implying that there’s no way they actually spend the $400B unless OpenAI revenue continues to rapidly grow)?
Also, why would this $400B / 7GW be an upper bound? A recent WSJ article suggests they are planning to surpass that, although details are super light.
There are no 10 GW OpenAI clusters, there is a conditional investment by Nvidia for every additional 1 GW in total across all their datacenters, which won’t be forming a single training system. Inference and smaller training experiments need a lot of compute, and OpenAI is now building their own even for inference, so most of what they are building is not for frontier model training.
Depending on OpenAI growth, this is more of a soft upper bound on what gets built. This is evidence that 5 GW training systems likely aren’t going to be built by (end of) 2028. So there’s going to be a slowdown compared to the trend of 12x in compute (and 6x in power) every 2 years for the largest frontier AI training systems, which held in 2022-2026[1]. In 2027-2028, the largest training systems are likely going to be merely 2 GW instead of the on-trend 5 GW. Though for the 2024 systems, FP8 is likely relevant, and for 2026 systems maybe even FP4, which turns the 12x in compute every 2 years in 2022-2026 into 24x in compute every 2 years (5x per year in pretraining-relevant raw compute for a single training system).
This 24x every 2 years is even less plausible to remain on-trend in 2028, so 2027+ is going to be the time of scaling slowdown, at least for pretraining. Though the AIs trained on 2026 compute might only come out in 2027-2028, judging by how 2024 training compute is still held back by inference capabilities, and some AIs enabled by 2024 levels of compute might only come out in 2026. So the slowdown in the scale of frontier AI training systems after 2026 might only start being observable in scaling of deployed AIs starting in 2028-2029.
Perhaps some of these sites will be connected with sufficient bandwidth, and training with RLVR at multiple sites doesn’t need a lot of bandwidth. Actual plans for larger training runs should urge them to build larger individual sites, as this ensures optionality for unusual training processes. So the fact that this isn’t happening suggests that there are no such plans for now (except perhaps for RLVR-like things specifically).
24K A100s in 2022 (7e18 BF16 FLOP/s, 22 MW), 100K H100s in 2024 (1e20 BF16 FLOP/s, 150 MW), 400K chips in GB200/GB300 NVL72 racks in 2026 (1e21 BF16 FLOP/s, 900 MW). The power estimates are all-in, for the whole datacenter site.
Thanks for your thoughtful pushback.
I want to withdraw my prediction “I’d expect the 10GW OpenAI cluster becomes operational around 2027-2028.” I spoke too quickly in my Twitter thread and this was based on a confusion on my part.
I will have a more detailed article soon to give more thoughtful updated predictions. I apologize for this error.
Forgive if a naive question, but what about distributed training runs — any view on whether progress on that front will result in training runs larger than what you describe here?
PS thank you for the extremely useful analysis as always — I would 100% subscribe if you had a newsletter or something (with a strictly positive willingness-to-pay even!).
There doesn’t necessarily need to be algorithmic progress to get there, sufficient bandwidth enables traditional pretraining across multiple sites. But it might be difficult to ensure it’s available across the geographically distributed sites on short notice, if you aren’t already a well-established hyperscaler building near your older datacenter sites.
In 2028, targeting inference on Rubin Ultra NVL576 (150 TB of HBM in a scale-up world) might want a MoE model with 80 TB of total params (80T params if in FP8, 160T in FP4). If training uses the same precision for gradients, that’s also 80 TB of gradients to exchange. If averaged gradients use more precision, this could be 2x-8x more data.
If training is done using 2 GW of some kind of Rubin GPUs, that’s about 2e22-3e22 FP4 FLOP/s, and at 30% utilization for 4 months it produces 8e28 FP4 FLOPs. At 120 tokens/param (anchoring to 40 tokens/param for the dense Llama 3 405B and adjusting 3x for 1:8 sparsity), this system might want about 10T active params (so we get 1:16 sparsity, with 160T total FP4 params, or about 1:8 for FP8). This needs 1,200T tokens, maybe 250T unique, which is a problem, but not yet orders of magnitude beyond the pale, so probably something can still be done without needing bigger models.
With large scale-up worlds, processing sequences of 32K tokens with non-CPX Rubin NVL144 at 30% utilization would take just 2.7 seconds (for pretraining). A 2 GW system has 9K racks, so that’s a batch of 300M tokens, which is already a lot (Llama 3 405B used 16M token batches in the main phase of pretraining), so that should be the target characteristic time for exchanging gradients.
Moving 80 TB in 2.7 seconds needs 240 Tbps, or 500-2,000 Tbps if averaged gradients use 2x-8x more precision bits (even more if not all-to-all, which is likely with more than 2 sites), and this already loses half of utilization or asks for even larger batches. A DWDM system might transmit 30-70 Tbps over a fiber optic pair, so this is 4-70 fiber optic pairs, which seems in principle feasible to secure for overland fiber cables (which hold hundreds of pairs), especially towards the lower end of the estimate.
As a naive follow-up: let’s say GPT-6 could be trained in 3 months on a 3GW cluster. Could I instead train it in 9 months on a 1GW cluster?
I’m confused: the announcement indicates that the $400B has been committed, and is not dependent on OpenAI’s growth (although perhaps you’re implying that there’s no way they actually spend the $400B unless OpenAI revenue continues to rapidly grow)?
Also, why would this $400B / 7GW be an upper bound? A recent WSJ article suggests they are planning to surpass that, although details are super light.
Do you think that pretraining is currently bottlenecked by compute, or clean text data? How do you forecast this will change?