Dec 31, 2025 · 24-Hour AI Briefing: ByteDance’s $14.2B GPU Lock-In, Intel 14A’s High-NA Bet, Gemini-3-Pro Takes the VLM Crown
In the past 24 hours, three signals stood out: ByteDance is treating top-end GPUs as strategic inventory, Intel is tying its 14A narrative to High-NA EUV and “manufacturable reality,” and Google’s Gemini-3-Pro is widening the gap in a measurable multimodal benchmark.
1. ByteDance plans to spend about $14.2B in 2026 on NVIDIA chips amid surging AI demand
Commentary:
Doubao, ByteDance’s AI assistant, reportedly grew daily token throughput from 4T at the end of 2024 to 50T by the end of 2025—12.5× growth in roughly half a year.
If the entire $14.2B were allocated to H200s, that implies roughly 71,000 GPUs—less like “capacity topping-up,” more like pre-locking 1–2 years of premium compute for training, multimodal generation, and real-time video AI.
For ByteDance, chips are not routine IT procurement; they are strategic supply. Stable access to higher-spec, larger-scale GPUs directly translates into faster iteration, lower inference cost, and tighter product shipping cadence.
That said, cross-border delivery of high-end GPUs remains constrained by export controls and compliance reviews—budget does not automatically equal deliverable volume or SKU mix.
2. Intel targets 2026 mass production for the Intel 14A process node
Commentary:
Intel 14A is positioned as the first volume node built around High-NA EUV lithography. High-NA EUV can improve resolution by about 50%, helping push below the physical limits that bite hard under 3nm.
If Intel truly ramps 14A in 2026, the message is bigger than one node: advanced manufacturing would no longer be a single-lane road dominated by TSMC. For large buyers, a second viable leading-edge source is both negotiation leverage and supply-chain resilience.
But the real test is not the announcement—it is yield, cost, and the ramp curve. If 14A lands, Intel Foundry’s credibility rises meaningfully; if it slips or yields disappoint, the market will quickly file it under “roadmap promise” rather than “shippable product,” which is toxic for foundry customer acquisition.
Will Intel actually deliver in 2026?
3. SuperCLUE-VLM (December): Google’s Gemini-3-Pro leads with 83.64
Commentary:
In VLMs, a big margin usually reflects one of two things: stronger vision encoding/alignment, or superior engineering optimization on the measurable task set. Either way, Google just demonstrated first-tier “quantifiable multimodal capability.”
The #2–#5 positions are also notable: SenseNova V6.5 Pro (75.35) is second; ByteDance’s Doubao (73.15) is third and scores 82.70 on the basic cognition subtask; Baidu ERNIE-5.0-Preview and Alibaba Qwen3-VL round out the top five, with Qwen3-VL becoming the first open-source model to cross 70 overall.
Beyond Google, the rest of the top five are Chinese models. Notably, Anthropic Claude-opus-4-5 and OpenAI GPT-5.2 (high) did not make the top five.
Gemini-3-Pro’s lead looks like the compounding result of years of multimodal investment—and this time, the distance is visible on the scoreboard.
Top AI stories from the past 72 hours: