AI Pretraining Scaling Laws With Compute Are Still Working – XAI and Google Will Pull Away

 

Step by step forward, one chip at a time.  We have been doing all this since the sixties and we are over seventy years in.

Yet understand that this is one channel of tech been progressively optimized because we can.

Way more interesting is what we have left behind, and will one day return to.

AI Pretraining Scaling Laws With Compute Are Still Working – XAI and Google Will Pull Away

November 19, 2025 by Brian Wang

Multi-billionaire fund manager Gavin Baker gives the overview of the state of AI.

https://www.nextbigfuture.com/2025/11/ai-pretraining-scaling-laws-with-compute-are-still-working-xai-and-google-will-pull-away.html

Nextbigfuture identified the AI data center scaling of google and XAI over a year ago.

AI Scaling Laws are Intact.

Four leaders in AI – Google, xAI, OpenAI and Anthropic.

BUT Google and XAI will pull away with faster and more compute scaling.

Nvidia is still dominant.

Best models are still coming mid-2026 using B300 chips.

AI Scaling Laws are Intact

Gemini 3 is the strongest evidence since o1 that pre-training scaling laws still hold.

Expect large jumps from Blackwell-trained models in Q2 2026.

GPT-5 was intentionally a smaller, cheaper-to-infer model (router-based), not a max-performance push → not evidence of scaling slowdown.

Frontier Model Landscape

(Emerging Oligopoly)Google – Currently pulling ahead. Lowest-cost high-quality tokens, massive coherent TPU fabric advantage.

xAI – Close second on cost/quality. Grok 4.1 leverages huge coherent GPU clusters. Google + xAI clearly best models right now.

Anthropic & OpenAI – Both have strong unreleased checkpoints, but currently behind Google/xAI in public model quality (OpenAI in 3rd place for the first time).

Meta – Small chance to stay competitive because Chinese open-source is only ~9 months behind Meta’s Llama series.

China open-source – Falling further behind. DeepSeek hasn’t released a new frontier model in a year, Huawei H20 issues persist. Blackwell gap + US rare-earth ramp will widen this dramatically.

Nvidia Dominance

Nvidia remains overwhelmingly dominant due to Blackwell (B300 variant now ramping smoothly after a rocky start with B200 delays, mask changes, canceled variants, and extreme datacenter complexity).

Blackwell delivers superior coherent FLOPs and tokens-per-watt, the two metrics that matter most going forward.

Power shortages make tokens/watt the key decision driver → favors general-purpose GPUs over custom ASICs; most non-Google ASIC programs likely dead.

Hopper rental prices are still rising and even A100s remain highly profitable → GPU useful life likely over 6 years, financing costs dropping further.

Blackwell will dramatically widen the gap vs Chinese domestic silicon (much bigger lag vs Blackwell than vs Hopper), further entrenching Nvidia’s moat.

Reasoning Flywheel & Rising Barriers

Reasoning models (o1-style, Gemini 3, etc.) create the classic internet flywheel: users → high-quality interaction data → better models → more users.

Pre-reasoning era had only pre-training data scaling. Now post-training/reasoning data is a new, proprietary, compounding moat.

All four U.S. leaders (Google, xAI, Anthropic, OpenAI) have much better internal checkpoints than public releases → very hard for anyone to catch up.

Other Key Themes

Power shortages are bullish: act as a natural governor against overbuild, extend cycle length and smoothness; push decisions toward tokens/watt (Blackwell wins).

Optics (optical interconnect) becoming critical for multi-campus training and moving workloads to cheap/available power. Also ironically helps China offset some compute deficit (at cost of much higher power draw).

Hyperscaler ROIC still higher than pre-AI capex ramp. Early signs of real enterprise productivity gains appearing in S&P 500 earnings.

Overall token demand (driven by customer ROI) is what matters for the sector. Individual leader share battles (OpenAI losing ground) won’t kill the secular growth story.

Bottom line from Gavin Baker

We are still very early. Next decade will be steady progress driven by compute scaling, reasoning flywheels, and power/tokens-per-watt optimization, with Nvidia, Google, and xAI as the biggest structural winners.