Skip to content
Source · Daily Brief

AI Daily Brief — 26 February 2025

Wednesday delivered NVIDIA’s biggest quarter ever and another Chinese open-source kernel drop. DeepSeek opened Day 3 of Open Source Week with DeepGEMM — a clean FP8 GEMM library powering V3 and R1 training/inference, ~300-line core kernel, Just-In-Time compilation, dense and two MoE layouts, hitting up to 1,350+ FP8 TFLOPS on H800. NVIDIA reported Q4 FY25: $39.3B revenue (+78% YoY, +12% QoQ), Data Center at $35.6B (+93% YoY), and Blackwell contributing $11 billion in its first quarter — the fastest product ramp in NVIDIA history. FY25 full-year revenue hit $130.5B (+114%). Microsoft launched Phi-4-multimodal (5.6B, mixture-of-LoRAs unifying speech/vision/text) and Phi-4-mini (3.8B), MIT-licensed on Azure AI Foundry, Hugging Face, and NVIDIA API Catalog — Phi-4-multimodal topped Hugging Face’s OpenASR leaderboard at 6.14% WER, beating Whisper V3. ElevenLabs released Scribe (96.7% English, 98.7% Italian ASR across 99 languages, $0.40/hour). Hume AI launched Octave — first LLM-based TTS, blind study with 180 raters preferred over ElevenLabs 71.6% on audio quality.

Top stories

  • NVIDIA Q4 FY25 record: $39.3B revenue (+78% YoY); Blackwell $11B in first quarter. Data Center $35.6B (+93% YoY). FY25 full-year $130.5B (+114%). Q1 FY26 guide $43B vs $41.78B expected. CFO commentary: hyperscalers >50% of Data Center; analyst combined-capex tracking ~$700B. Stock muted in after-hours (~+1%). via SEC 8-K · via Futurum
  • DeepSeek open-sources DeepGEMM — Open Source Week Day 3. Clean FP8 GEMM library powering V3 and R1. ~300-line core kernel with JIT compilation. Dense and two MoE layouts. Up to 1,350+ FP8 TFLOPS on NVIDIA Hopper (H800). Surpasses expert-tuned kernels on most matrix sizes. via DeepSeek GitHub
  • Microsoft launches Phi-4-multimodal and Phi-4-mini. Phi-4-multimodal (5.6B) is Microsoft’s first model integrating speech, vision, and text in a unified architecture via mixture-of-LoRAs. Tops OpenASR leaderboard at 6.14% WER (beating Whisper V3’s 6.5%). Phi-4-mini is a 3.8B text-only model. Released MIT on Azure AI Foundry, Hugging Face, and NVIDIA API Catalog. via Microsoft Azure
  • ElevenLabs releases Scribe ASR. First standalone speech-to-text model. 96.7% accuracy in English / 98.7% in Italian on FLEURS/Common Voice — beats Whisper Large V3, Gemini 2.0 Flash, and Deepgram Nova-3. 99 languages; word-level timestamps, speaker diarization, audio-event tagging at $0.40 per hour. via VentureBeat
  • Hume AI launches Octave TTS — first LLM-based text-to-speech. Generates voices from text prompts; takes acting-style instructions (sarcasm, whispering). 180-rater blind study: preferred over ElevenLabs Voice Design 71.6% on audio quality, 51.7% naturalness, 57.7% description-match. Priced ~50% below ElevenLabs. via VentureBeat
  • Figure AI publishes Helix logistics update. Implicit stereo vision (3D depth-aware motion), learned visual proprioception (cross-robot transfer), and “Sport mode” delivering faster-than-demonstrator execution speed while preserving success rate. via Figure AI

Who shipped

NVIDIA shipped the year’s biggest earnings print. DeepSeek shipped DeepGEMM. Microsoft shipped the Phi-4 family. ElevenLabs and Hume shipped competing audio products. OpenAI, Anthropic, Google DeepMind, and Meta made no dated launches.

Open-source pulse

Three Apache/MIT-class shipments in 24 hours (DeepGEMM, Phi-4 family, ongoing Wan 2.1). Phi-4-multimodal (5.6B) topping OpenASR proved small models could punch up; Wan 2.1’s 1.3B running on consumer GPUs continued spreading.

By the numbers

  • $39.3B / +78% YoY — NVIDIA Q4 FY25 revenue / growth
  • $11B — Blackwell first-quarter revenue (fastest ramp ever)
  • $130.5B / +114% — NVIDIA FY25 full-year revenue / growth
  • ~$700B — analyst estimate of combined hyperscaler 2025 capex
  • 1,350+ TFLOPS — DeepGEMM FP8 peak on H800
  • 6.14% WER — Phi-4-multimodal OpenASR (#1)
  • Most-mentioned company: NVIDIA

Compiled by AI Feed’s editor from verified web sources for 26 February 2025.