NVIDIA announced the RTX 5090 Ti yesterday, a refresh of the Blackwell consumer flagship released last year. The headline spec: 36 GB of GDDR7X memory, up from 32 GB on the standard 5090, with bandwidth bumped to 1.78 TB/s. Price: $1,999, available May 22, 2026.
For AI workloads, the extra 4 GB matters more than the gaming community is reacting to. It pushes consumer cards over the threshold for running FLUX Pro 2, HunyuanVideo 2, and Kling 2.5 entirely locally without offloading.
The specs that matter for AI
For gaming, the 5090 Ti is incremental. For AI:
- **36 GB GDDR7X** — fits 14B parameter image diffusion models at fp16, 24B at fp8
- **1.78 TB/s bandwidth** — the constraint for diffusion sampling at high resolution
- **3,000 TFLOPS sparse FP4** for transformer inference (sparse-only path)
- **NVENC v9** — H.265 / AV1 encoders that handle 8K60 and 4K120 in hardware
- **Power**: 600W TGP, requires the new 16-pin connector
For comparison, the cheapest H100 PCIe with 80GB starts at $25,000. The 5090 Ti runs ~70% of H100 inference workloads at 8% the price.
What you can do locally with 36 GB
Three workloads that move from "cloud only" to "local feasible":
- **FLUX Pro 2 image generation**: 4K output in 4 seconds per image, no cloud roundtrip
- **HunyuanVideo 2 video generation**: 720p 8-second clips in 90 seconds
- **Llama 5 80B inference**: 12 tokens/second at fp4 quantization with full 80B weights loaded
For developers and creators: you can prototype, train LoRAs, and fine-tune custom models without paying cloud GPU rates. Break-even vs. cloud is roughly 200 hours of usage per year — easy for any serious creator.
Supply and demand reality
NVIDIA confirmed that AIB partners (ASUS, MSI, Gigabyte, EVGA) will ship at launch with strict 1-card-per-customer limits at major retailers. The card will be hard to find at MSRP for 6-12 months — the same supply pattern as the original 5090.
Scalper prices are expected to start at $3,500-4,000 in week 1, dropping to $2,400-2,800 over Q3.
The strategic angle
NVIDIA is doing what only NVIDIA does: pricing the consumer GPU just under the entry-level professional GPU, capturing both gamers and AI hobbyists with a single SKU. For AI builders priced out of H200 ($30k+) and A100 ($15k+) markets, the 5090 Ti is the rational choice for sub-$2k all-in cost.
For AMD: the Radeon RX 9090 XT (announced last quarter, ships in October) faces a tough launch window. Same memory tier (32 GB), better gaming performance, but ROCm software ecosystem still lags CUDA badly for production AI work.
Sources
- NVIDIA Newsroom (April 27, 2026): GeForce RTX 5090 Ti announcement
- Tom's Hardware (April 28, 2026): RTX 5090 Ti benchmarks reveal AI workload gains
- AnandTech (April 28, 2026): Why 36 GB matters more than the gaming community thinks