ARCHIVE · DAILY SELECTIONS
Selections for September 1, 2026
BitShovel tracked 26 signals that day; start with the 3 focus picks. This page is an immutable record of the choices made then — useful for checking them against what happened next.
Amazon triples its NVIDIA GPU order: 2M more chips headed to AWS through 2028
Announced on NVIDIA's Q2 earnings call: Amazon adds 2 million GPUs (Blackwell Ultra, Rubin, Rubin Ultra) to AWS, delivering in 2027-2028 with deployments starting this quarter — tripling the earlier 1M-chip agreement. NVIDIA says demand "has exceeded those expectations" since the first deal, alongside Vera CPU supply and Q2 data-center revenue of $89B (+117% YoY).
Why it was selectedSupply-side orders are the hardest demand evidence: hyperscalers back AI compute demand with multi-year purchases, and supply — not model capability — is becoming the real bottleneck for the next model generation. A directly verifiable anchor for judging compute pricing and the sustainability of the model arms race.
CaveatThe evidence is official statements from both companies on the earnings call, as reported by TechCrunch — not a transcript; financial terms are undisclosed and the "tens of billions" valuation is a media estimate from GPU pricing. Not an independent review; order volume is not proof that end demand is fully real.
Google ships Gemini 3.5 Transcribe into public preview
Google DeepMind launched Gemini 3.5 Transcribe, its self-described most precise speech-to-text model yet: Artificial Analysis measures average word error at 4.0% streaming and 2.6% non-streaming across 85+ auto-detected languages, and time to final transcription improves about 70% over Chirp 3. It ships as a public preview in the Gemini API and already powers Gboard on Android plus the macOS Gemini app. Two endpoints: streaming gemini-3.5-transcribe-live via the Live API, and gemini-3.5-transcribe for pre-recorded audio with speaker attribution and word-level timestamps.
Why it was selectedThe frontier-model lane also lacked a vertical anchor for Google Gemini; this lands squarely on voice workflows — 85-language coverage plus word-level timestamps is a clear capability jump for transcription, QA, and meeting-notes tooling.
CaveatEvery figure comes from Google DeepMind's own announcement and its cited Artificial Analysis measurements; this is not an independent evaluation, and official claims during a public preview are not proof of production stability or cost behavior.
Qwen open-sources Qwen3.8-Flash-Next as an early Qwen4 architecture preview
Alibaba's Qwen team announced the open-sourcing of Qwen3.8-Flash-Next together with an FP8 variant, positioning it officially as a multimodal MoE early preview built on the next-generation Qwen4 architecture so developers can adapt ahead of the full Qwen4 family. The official ModelScope model page for the model is already live. No concrete specs were disclosed at release, and circulating parameter counts have no official backing yet.
Why it was selectedThe frontier-model comparison lane previously had no vertical anchor for Qwen; this first architectural preview of the Qwen4 family directly shapes version expectations across the open-source ecosystem.
CaveatEvidence anchors on the Qwen blog address and the ModelScope model page going live, discovered through the AIHOT featured feed snapshot; none of this is an independent review, and release-day aggregator or community attention is not proof of any performance or cost claim — rumored figures like 125B total parameters are not adopted here.
26 tracked signals that day