BIG TECHAWAYS

China’s AI Chips Get Pricier

• ​Reuters reports that Chinese AI-chip makers are raising prices as supplies of high-bandwidth memory, or HBM, tighten. Huawei’s Ascend 950DT was quoted above 250,000 yuan, 20% to 50% higher than two months earlier.
• For a team choosing infrastructure, that turns HBM from a line in a specification sheet into a procurement risk that can hit unit economics directly. A faster accelerator does not necessarily lower costs if its accompanying memory is rationed or has to be bought at a premium.
• Reuters is describing a specific price increase and supply squeeze, not concluding that China’s entire market will face a permanent chip shortage. For buyers, the practical question is whether memory supply needs to become a capacity-planning variable from the start.

DeepSeek’s 4x KV-Cache Claim

• ​DeepSeek says V4.1-Flash is a 552 billion-parameter MoE model, with 8 billion parameters active for input and 16 billion active for output. The company also says the new version needs one-quarter as much KV-cache HBM as its predecessor.
• KV-cache retains context the model has already read so it does not have to recompute it on every interaction. If DeepSeek’s claim holds up in independent measurements, the same hardware budget could serve more long-running sessions, or leave more memory available for other workflows.
• This is a claim from DeepSeek, not an independent measurement. The final test is whether the HBM saving translates into better throughput, concurrency, or serving costs in real deployments.

PRESENTED BY THE POUR OVER

News without the Noise

If you’re tired of shouting matches dominating your morning news scroll, you’ll love The Pour Over. It’s brief, politically neutral news for Christians of all denominations. Instead of rage-bait, you get clear reporting paired with biblical reminders to keep you focused on the eternal—not the temporary. The Pour Over helps 1.5 million believers navigate a noisy world. Plus, it’s easier to unsubscribe than it is to sign up.

​Sign up for The Pour Over today.​

BIG THINK

Anthropic’s AI Diffusion Map

• ​Anthropic’s Economics team has published an interactive model, Economic Scenarios for Transformative AI, for exploring how AI could affect US growth, jobs, and wages through 2030. It divides the economy into task bundles that remain unchanged, are augmented by AI, are automated, or are newly created.
• In a survey of 10,980 people, the typical response was close to the “substantial” scenario, where 2030 GDP is 8.3% above a no-AI baseline. That is a model output, not a forecast.
• The model makes the assumptions behind each result visible. AI capability becomes economic growth only when businesses adopt it, tasks are redesigned, and workers can move into new roles. Two scenarios with the same technology but different adoption speeds can produce very different outcomes for GDP and labor’s share of income.
• Anthropic calls this a scenario explorer, not a forecast. The approach shifts the question from “Will AI create a boom or mass unemployment?” to “What has to change in real work for a new capability to reach the economy?”

AROUND THE AI WORLD

1. ​Payment networks build a trust layer for AI agents: Reuters reports that Visa, Mastercard, and Ant International have launched an initiative to identify and verify AI agents making transactions. As software starts buying on a user’s behalf, identity and authorization become product infrastructure. This is an announced framework, not evidence that agent commerce is broadly deployed.
2. ​Google puts more AI features into everyday work surfaces: Google announced new AI features for Gmail, Docs, Keep, and Sheets, including voice, image, canvas, and Gemini Spark. AI distribution is moving through surfaces people already open every day, while plan tiers and regional limits become part of the product strategy.
3. ​OpenAI makes a public case for capability-based safety rules: OpenAI says it supports national AI safety requirements based on capability thresholds and names four California bills. This is OpenAI’s policy position, not a prediction of legislative outcomes. For frontier labs, auditable definitions of capability are becoming part of the competitive environment.
4. ​Unity brings native skills into Claude Code: Unity announced an official Claude Code plugin with 29 native Unity skills. When AI tools can operate directly inside a professional environment, their usefulness comes partly from access to the product’s native operations.
5. ​ChatGPT Voice gets more capable models for harder tasks: OpenAI says ChatGPT Voice can use GPT-5.6 or GPT-6 Astra for harder search and reasoning requests, with limits that vary by plan. More capable models are reaching everyday interfaces, but actual usage still depends on plan and availability.
6. ​OpenAI describes a continuous security loop: OpenAI says Defense Factory has mobilized more than 250 people across hundreds of systems, using agents to find vulnerabilities, test them, and verify fixes. The notable pattern is a loop of discovery, testing, and fix verification rather than a one-time promise of automated security. These details describe an OpenAI deployment.