NVIDIA Vera Rubin vs Grace Blackwell: relative performance

Hey there!

Welcome back to The Pulse, where we dive into interesting AI stories and trends backed by data, all presented through simple visuals.

> unveiled Vera Rubin, a rack-scale AI compute platform - in full production with deployment expected in H2 2026

> end-to-end co-design across Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet Switch

> up to 10× lower inference token cost & 4× fewer GPUs for MoE training vs Blackwell

> Huang claims “more bandwidth than the entire internet”

> launch follows record data-center revenue (+66% YoY) driven by Blackwell; Rubin expected to accelerate ramp

> Huang (Oct ’25): $3-4T AI infra spend over next 5 years; McKinsey: $7T global data-center investment by 2030

> H200s chip export to China to resume by mid-Feb following Trump’s conditional ban reversal

> H200 - only NVIDIA GPU with China export clearance - now 3rd most powerful after Rubin release

> ByteDance overall AI spend to be ~$23B in 2026 (FT)

> already runs a 1,000-member internal chip design unit - at final stage on a chip comparable to H20

> their chatbot, Doubao reportedly processes >50T tokens/day (Dec ‘25), up from 4T YoY (12.5× increase)

> Huang says Chinese demand for H200s - now back in production - “very high”; to be worth $50B in sales every year

> ChatGPT slips below 70% share, Gemini nears ~20%, Grok keeps gaining

> ChatGPT monthly active users at ~900M: in Nov '25, Gemini MAUs rose 30% to 346M while ChatGPT MAUs up 5% to 810M (Sensor Tower)

> Avg daily time spent in Nov: ChatGPT 17 min, Gemini 11 min (up from ~5 min/day in March)

> Gemini surge partly driven by Nano Banana 3, but recently overtaken by GPT Image 1.5 in Dec '25, potentially affecting current traffic share

> among developers, Anthropic shows fastest relative growth: OpenAI-to-Anthropic SDK download ratio = 47:1 to 4.2:1 from early 2024 to 2025 end (Greptile)