GPU payback may be as short as 2 months in the US, analyst argues amid Ascend-vs-Blackwell debate
teortaxesTex · x · 2026-10-20
teortaxesTex argues that serving the public on Blackwells makes sense given Flash's speed advantage, and if labs figure out training on Ascends while locking a significant allocation, per-GPU throughput becomes a quaint concern.
In the quoted tweet, he pushes back on a markup-based payback estimate as unsupported by current data: payback might be 6 months or less at current revenue productivity, roughly 9-10 months at half that assumption, and possibly as little as 2 months in the US.
Related event: DeepSeek Reportedly Hits $1B ARR, Needs Only ~3,000 GPUs(4 posts)→
More from Infra
- Inference framework economics: payback period reportedly shrinks from 10 months to 4 — oran_ge · 2026-10-20
- New DeepSeek inference stack released: how will it compare to vLLM, Modal, Together? — finbarrtimbers · 2026-10-20
- COUNT Paper: N-gram Counts on Host Memory Equal 57% More Training Compute — pmttyji · 2026-10-20
- Daytona, building an 'agentic cloud,' raises $60M at $1B valuation — jeremyakahn · 2026-10-20
- ObsessionDB launches realtime data warehouse for agents, claiming 100X faster and 20X cheaper than Snowflake — ycombinator · 2026-10-20
- Polars 2.0 makes streaming the default, speeding queries up to 2.8x with zero code changes — KhuyenTran16 · 2026-10-20