GPUs are expensive and setting up the infrastructure to make GPUs work for you properly is complex, making experimentation on cutting-edge models challenging for researchers and ML practitioners. Providing high quality research tooling is one of the most effective ways to improve research productivity of the wider community and Tinker API is one step towards our mission there. Tinker API is built on top of our experimental results on fine-tuning with LoRA: Beta starts and you can join the waitlist today:

96 GPUs in under 24 hours with zero waiting. That's what it took for @wonderaai to go from compute-constrained to shipping. 64 H100s and 32 H200s rapidly provisioned on @ionet. Not queued for months on a waitlist. The results: - 200,000 users in 4 months, across 171 countries - 50% month-over-month growth - Launched 3 months ahead of schedule This is what happens when your compute keeps up with how fast you can build.

Idle GPUs waiting on hard drives are burning capital, says @wd_corporation SVP Tim Rausch. At AI Infra Summit 2026, @MattKimball_MIS spoke w/ Rausch about why storage may determine how far AI can scale. They are targeting: -60-terabyte drives by 2028 -100-terabyte drives by 2030 -Read/write technology delivering up to 8x today’s throughput $WDC

Free GPUs 2x NVIDIA T4s & run Qwen 3.8 27B at 14 t/s with 120k context - 120k context on FREE Kaggle T4s. - FP16 KV - 14 t/s - No credit card. - No expensive GPU. - Total VRAM used: only 26.6 GB across both cards. - Under 5 minutes setup. - Kaggle gives you 30 hours/week of dual T4 for free. Free kaggle notebook from Alok : -

*consumer* gpus have basically become full computers some of these things push 1kw. these days handle long standing computational loads by themselves we used to have a computer, with components, working together. a tight coupling. now we have a few computers, the general one and the parallel one. they kind of communicate and then apus come in and blur the line even more. we truly live in wild times