Modders have unlocked hidden AI compute performance on NVIDIA’s RTX 20, 30 and 40-series GPUs. The modifications remove software-level limits on certain Tensor Core operations, without changing the hardware or flashing a new VBIOS. - RTX 4090: 165.6 → 331.6 TFLOPS FP16 Tensor Core throughput -RTX 3090 Ti: 87.5 → 157.4 TFLOPS BF16 performance - RTX 2080 Ti: 59.9 → 111 TFLOPS FP16→FP32 performance - RTX 3090: 73.0 → 122.8 TFLOPS in a PyTorch GEMM test The gains are mainly relevant to AI and compute workloads, not gaming.

JPMorgan just dropped an ABSOLUTELY INSANE forecast for AI compute. According to JPMorgan, global AI compute demand is expected to reach roughly 530 GW by 2031, while available supply is projected to reach just 250 GW. That leaves us with a staggering 280 GW shortfall, even after the industry spends hundreds of billions of dollars expanding capacity. We're talking about global AI compute supply potentially QUADRUPLING between 2025 and 2031, yet the shortage is expected to get even WORSE. In fact, JPMorgan estimates the gap between supply and demand could expand from roughly 50 GW today to nearly 280 GW by 2031. And here's why this matters so much. When demand consistently outpaces supply, the companies controlling scarce compute capacity gain enormous pricing power. This is especially important for neoclouds like $NBIS and $CRWV, which are building out GPU infrastructure to serve customers that simply cannot secure enough compute capacity through traditional hyperscalers. If JPMorgan's projections are even remotely accurate, GPU rental pricing could remain stronger for much longer than investors currently expect, supporting higher utilization rates, long term contracts, and potentially stronger returns on deployed infrastructure. But the opportunity extends FAR beyond neoclouds. Companies like $NVDA and $AMD benefit from the continued need for accelerators, memory manufacturers like $MU benefit from increasing demand for high-bandwidth memory, and networking, storage, power, and cooling providers all become increasingly important as the industry attempts to close this gap. And this is where I think the market is making a BIG mistake. Investors are so focused on whether AI companies are spending TOO MUCH money that they're overlooking the possibility that the industry STILL isn't spending enough to satisfy future demand. The real constraint isn't necessarily a lack of customers willing to pay for compute. It's the ability to manufacture enough chips, secure enough electricity, build data centers fast enough, and actually bring that capacity online. And that is EXACTLY why I continue to believe the biggest winners in AI won't necessarily be the companies building the models, but the companies supplying the infrastructure required to run them.

LATEST: 🤖 Bill Gates told CNN that AI “changes the world more than all previous technologies,” calling the PC and the internet “nothing compared to this.”

This post has 200K+ of views, and this is how local AI gets a bad name. "Run a 70B model on a 4GB GPU. 405B on 8GB." Yes, AirLLM can do that. It loads 1 layer from disk, computes it, drops it and loads the next. For every single token. Do the math. A 70B model is about 140GB of weights. To make 1 token you read all of it from disk. On a fast NVMe thats most of a minute per token. A 405B is about 800GB per token. Go make tea. Go make dinner. "Runs" and "usable" are different words. Here is usable on a small card today: 27B, 12GB of VRAM, 262K context, 70 tok/s. I measured it on my own 3090 held to 12GB. Weights and engine are public. Stop telling beginners this works. They try it, wait 10 minutes for 1 sentence and decide local AI is a joke.