<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0"><channel><title>The Raw Logs — #compute</title><link>https://therawlogs.com/tag/compute</link><description>Top takes tagged #compute.</description><language>en</language><lastBuildDate>Tue, 08 Sep 2026 16:37:33 GMT</lastBuildDate>    <item><title>#compute — OpenAI, Thailand launch national AI compute for research labs</title><link>https://therawlogs.com/take/14-5</link><guid>https://therawlogs.com/take/14-5</guid><pubDate>Sun, 30 Aug 2026 16:15:26 GMT</pubDate><description><![CDATA[OpenAI, Thailand launch national AI compute for research labs: OpenAI partnered with Thailand to provide frontier AI compute directly to national labs and universities. Academic labs get priced out by commercial API token costs. Structured national compute tiers give research teams access to models for biology and materials discovery. [#compute]]]></description></item>    <item><title>#compute — Epoch AI: frontier datacenters to reach 5M-GPU capacity</title><link>https://therawlogs.com/take/12-4</link><guid>https://therawlogs.com/take/12-4</guid><pubDate>Fri, 28 Aug 2026 14:58:33 GMT</pubDate><description><![CDATA[Epoch AI: frontier datacenters to reach 5M-GPU capacity: Epoch AI projected frontier datacenters to reach 5 million GPU-equivalents by 2028. Single facilities run out of power grid capacity. Scaling requires distributing model training across separate regional campuses linked by dedicated high-speed fiber backbones. [#compute]]]></description></item>    <item><title>#compute — OpenAI releases benchmark data for Jalapeño custom inference silicon.</title><link>https://therawlogs.com/take/9-1</link><guid>https://therawlogs.com/take/9-1</guid><pubDate>Tue, 25 Aug 2026 16:40:45 GMT</pubDate><description><![CDATA[OpenAI releases benchmark data for Jalapeño custom inference silicon.: OpenAI built Jalapeño, a custom inference chip for transformer models. Standard graphics cards waste power on unused 64-bit units. Purpose-built silicon cuts electricity use by 3x per word, lowering daily data center costs for high-volume serving. [#compute]]]></description></item>    <item><title>#compute — Economic value changes from simply creating tokens to pricing power and results.</title><link>https://therawlogs.com/take/8-4</link><guid>https://therawlogs.com/take/8-4</guid><pubDate>Mon, 24 Aug 2026 16:49:40 GMT</pubDate><description><![CDATA[Economic value changes from simply creating tokens to pricing power and results.: Token creation costs drop tenfold each year, moving profit to firms that own power contracts and deliver full tasks. Raw token sales lose value fast. Companies earn by charging for finished results, not per token. Price software by proven outcomes. [#compute]]]></description></item>    <item><title>#compute — Special processors let machines see in under a millisecond.</title><link>https://therawlogs.com/take/6-5</link><guid>https://therawlogs.com/take/6-5</guid><pubDate>Sat, 22 Aug 2026 16:49:15 GMT</pubDate><description><![CDATA[Special processors let machines see in under a millisecond.: AI chips in cars and phones process camera video locally in under 1 ms without the internet. This eliminates the 200–500 ms delay that can cause failures when the connection drops. Install these chips to keep the system running, reduce cloud costs, and improve safety. [#compute]]]></description></item>    <item><title>#compute — The government is moving to protect AI data centers as critical infrastructure.</title><link>https://therawlogs.com/take/4-2</link><guid>https://therawlogs.com/take/4-2</guid><pubDate>Thu, 20 Aug 2026 17:33:41 GMT</pubDate><description><![CDATA[The government is moving to protect AI data centers as critical infrastructure.: AI and data centers are now national critical centers. Wow! There can be no shutdowns or single points of failures according to the government. Is it a win for AI, or is it a battle for supremacy of who wins the AI war? [#compute]]]></description></item>    <item><title>#compute — Humanoid Robotics &amp; Neural Forecasting</title><link>https://therawlogs.com/take/3-5</link><guid>https://therawlogs.com/take/3-5</guid><pubDate>Thu, 20 Aug 2026 08:00:00 GMT</pubDate><description><![CDATA[Humanoid Robotics & Neural Forecasting: In a short time, 60,000 humanoid robots were deployed. They can be reconfigured for any task as they are general-purpose. The robot age starts now? Google DeepMind has figured out the neural forecasting models for weather. I wonder if we will get more accurate forecasts. [#compute]]]></description></item>    <item><title>#compute — Automatic Model Routing Economics</title><link>https://therawlogs.com/take/2-2</link><guid>https://therawlogs.com/take/2-2</guid><pubDate>Wed, 19 Aug 2026 08:00:00 GMT</pubDate><description><![CDATA[Automatic Model Routing Economics: Sending simple text questions to giant models wastes money. Snowflake added automatic routing to send 80% of routine work to cheap, small tools. If your software sends every basic request to an expensive model, your system architecture is broken. [#compute]]]></description></item>    <item><title>#compute — GeoPT &amp; Physics Simulation</title><link>https://therawlogs.com/take/1-2</link><guid>https://therawlogs.com/take/1-2</guid><pubDate>Tue, 18 Aug 2026 08:00:00 GMT</pubDate><description><![CDATA[GeoPT & Physics Simulation: Using billions of random tokens to simulate physics is inefficient. MIT's GeoPT shows that adding geometric rules directly into the model reduces the required training data by 60%. Structured design outperforms random trial. [#compute]]]></description></item></channel></rss>