As of November 10, 2026, Level lists 297 live roles that match a search for CUDA. Google announced two new TPU systems on November 9, weeks after it released TorchTPU, a native PyTorch stack for TPUs. This guide shows what that means for engineers who write GPU code today.
What happened
On November 9, 2026, Google announced TPU 8tx and TPU 8i Max. TPU 8tx targets secure and confidential inference. It runs on fifth generation AMD EPYC hosts. Google says it delivers up to 2x lower latency than TPU 8t. TPU 8i Max combines two 8i chips in one package. Google says it has double the compute, SRAM and HBM bandwidth of TPU 8i. The same post says Google introduced four TPU systems in a single year, a first. It also says TPUs have nearly 100 customers.
The software side moved a few weeks earlier. On October 20, 2026, Google Cloud released TorchTPU as open source. Google built it with Meta. Google says TPUs now join GPUs and CPUs as a native option for PyTorch. Google lists support for vLLM, SGLang and Hugging Face. The post adds that Hugging Face Transformers v5 accepts device_map="tpu".
For job seekers, the question is simple. If PyTorch code can run on TPUs, does CUDA still matter for hiring? The Level catalog gives a first measured answer.
Is CUDA still the skill employers ask for?
Today's data shows CUDA roles far outnumber TPU roles. A search for CUDA returns 297 live roles on Level. A search for TPU returns only 8 live roles. Those TPU roles come from Waymo, OpenAI, Anthropic, Thales and Calico Labs. OpenAI lists a Software Engineer, TPU Compilers role. Anthropic lists a TPU Performance Engineer role.
The CUDA roles are deep technical jobs. Of these CUDA roles, 208 are Builds AI on the Level scale. That is 70% of the set. Another 48 are Works on AI. Only 32 are Little AI. You can browse the whole top level on the Builds AI page.
Most of these jobs sit in engineering teams. Software engineering accounts for 199 of the roles. Research accounts for 42 more.
The tools listed next to CUDA point the same way. PyTorch appears in 123 of the CUDA listings. That is 41% of the set. Python appears in 144 listings. vLLM appears in 46 listings, which ties much of this work to inference serving. Our vLLM jobs guide covers that side of the market.
TorchTPU adds a second target for PyTorch code. It does not remove the first one. Among the CUDA results that name PyTorch, 115 also tag CUDA itself as a tool. The CUDA tool page and the PyTorch tool page list every match.
Should I learn JAX or TPU skills now?
A search for JAX returns 581 live roles. Of those, 525 are Builds AI. Amazon lists the most JAX roles, with 86. NVIDIA follows with 77. Waymo has 73. Mistral AI, Thinking Machines Lab and xAI also hire for JAX. The JAX tool page keeps the full list.
JAX also appears inside CUDA listings. It shows up in 41 of the CUDA roles. NCCL, the NVIDIA library for multi GPU communication, appears in 28. TensorRT appears in 18.
Triton is a third path. The name is used for a GPU kernel language and for the NVIDIA Triton Inference Server, so read each listing. A search for Triton returns 54 live roles. Intel lists 15 of them. NVIDIA lists 13.
JAX returns more roles than CUDA in these searches. Both counts include listings where the tool is only named. Amazon, NVIDIA and Waymo lead JAX hiring. Roles that name TPUs remain rare, with only 8 listings today.
Who is hiring CUDA engineers, and at what level?
One company leads this market. NVIDIA lists 188 of the CUDA roles. That is 63% of the set. SpaceX lists 11 roles. Amazon lists 10. Jump Trading, a trading firm, lists 6. Mistral AI lists 5.
Kernel work shows up in job titles. Exactly 12 CUDA titles contain the word kernel. xAI lists a Sr. Software Engineer, Kernels/CUDA (C++) role. Modular lists an AI Kernel Engineering Manager role. NVIDIA lists a Principal Engineer, CUDA UMD role focused on GPU kernel scheduling.
Seniority leans senior. Senior roles number 170. That is 57% of the set. Lead roles add 21 more. Junior roles are rare, with only 6 listings.
Interns have a real door, though. Internships account for 62 roles. That is 21% of the set. The AI internships page groups them with other student roles.
Location matters for this skill. The United States hosts 138 of the roles. That is 46% of the set. China hosts 64 roles. India hosts 22. Remote work is possible in 43 roles. That is 14% of the set. The remote AI jobs page filters the wider market the same way.
Pay is high where it is published. 109 roles publish a salary range in US dollars. The median midpoint of those ranges is $235,000+. The median minimum is $184,000. For a related market, see our guide to inference engineer jobs.
Employers posted 12 of these CUDA roles in the last seven days.
Open roles right now
The list below comes from live Level listings that mention CUDA.
