AI Navigate
アップデート最新記事AI大全カオスマップ
2026 · 06 · 25 · Thu

Updates for 6/25

OpenAI moved inference in-house with Jalapeño as Anthropic pushed Claude into Slack. Chips and enterprise SaaS are both finding new power centers this week.

A · Theme of the day

Custom inference chips are taking over the AI serving stack

Companies not known for chips are all designing their own ASICs this week.

OpenAI moves inference in-house with custom Jalapeño chip

GPT (OpenAI)GPT (OpenAI)
Compared to before

OpenAI's inference has run on NVIDIA GPUs, making cost cuts dependent on buying more hardware — a constraint every major lab was working around.

What changed

OpenAI and Broadcom launched 'Jalapeño,' a custom LLM inference ASIC, moving OpenAI's serving stack beyond pure GPU dependence (OpenAI blog).

Why it matters

No API price cut yet, but structural pressure is real. Expect gradual reductions over 6–12 months; at personal-scale usage, it's still mostly noise.

Source: openai.com　·　→ View on chaos map
Read in depth →

Qualcomm buys AI inference startup Modular for ~$4B

AI Semiconductor / GPU Economics
Compared to before

The AI inference chip space had been a near-duopoly — NVIDIA and Google TPU — with Qualcomm firmly in the mobile SoC lane until today.

What changed

Qualcomm acquired AI inference startup Modular for ~$4B (Wired), entering the inference chip race alongside NVIDIA and Google.

Why it matters

More competition means more pricing pressure on inference infra. Watch whether Modular's software stays open or gets closed into Qualcomm's ecosystem.

Source: localhost　·　→ View in AI Encyclopedia
Read in depth →
B · Theme of the day

Enterprise AI rewrites the rules on which data it touches

Claude moved into Slack as Copilot flipped to hiding confidential docs.

Claude Tag lets Slack conversations kick off AI tasks

Claude (Anthropic)Claude (Anthropic)
Compared to before

Getting Claude into a Slack workflow required a separate window or custom build. Teams + Copilot had a native option; Claude didn't — until now.

What changed

Anthropic launched 'Claude Tag' for Slack: task handoffs, summaries, code generation from chat. Claude writes 65% of Anthropic's own code (THE DECODER).

Why it matters

A direct competitor to Teams + Copilot now exists natively in Slack. For Claude-first teams, the approval threshold for enterprise adoption just dropped.

Source: anthropic.com　·　→ View on chaos map
Read in depth →

Microsoft Purview now hides sensitive docs from Copilot by default

Microsoft CopilotMicrosoft Copilot
Compared to before

For 18+ months, Copilot Cowork accessed all corporate data by default. IT teams worried about sensitive docs leaking into AI outputs.

What changed

Microsoft Purview is making 'isolate sensitive-labeled docs from Copilot' the default, reversing the long-standing opt-in setting (Innovatopia).

Why it matters

Governance hurdle for enterprise Copilot adoption just got lower. If you're in the 'let Copilot read everything' camp, expect extra config steps.

Source: copilot.microsoft.com　·　→ View on chaos map
Read in depth →
C · Theme of the day

The line between licensed and unlicensed AI training data hardens

Two moves clarified who controls what goes into AI training data.

Cloudflare lets site owners block AI scrapers from the dashboard

Copyright and AI: Legal Issues in Generated Outputs and Training Data
Compared to before

Blocking AI crawlers meant writing robots.txt entries or building WAF rules — technical work that locked out non-developer site owners.

What changed

Cloudflare launched a dashboard toggle to allow or block AI scrapers site-wide (GIGAZINE). Same day OpenAI signed a content licensing deal with Getty Images.

Why it matters

The consent infrastructure for AI training data is hardening at the network layer. AI companies clearing licensing hurdles will access better data pools.

Source: localhost　·　→ View in AI Encyclopedia
Read in depth →
D · Theme of the day

China's AI accelerates top-to-bottom despite export controls

Chips doubled in price and still sold; models cut prices to chase US demand.

NVIDIA chip prices in China double despite export controls

US-China AI Geopolitics and Sovereign AI
Compared to before

A year into export controls, buyers split: some switched to local chips, others paid premium for restricted NVIDIA parts. Both camps are now active.

What changed

NVIDIA AI chip street prices in China doubled (Tom's Hardware). China Telecom bought 40,000 servers for ~$1.7B, mostly flowing into Huawei's ecosystem (SCMP).

Why it matters

Controls limit supply but not demand. Huawei keeps absorbing server procurement — a self-sufficient China AI stack grows more credible each quarter.

Source: localhost　·　→ View in AI Encyclopedia
Read in depth →

Alibaba cuts Qwen prices; Baidu open-sources long-doc OCR

China's Moves: DeepSeek / Qwen / Doubao
Compared to before

Until last month, Qwen API pricing was in the same range as GPT-4o — no major cost advantage for Chinese models over US alternatives.

What changed

Alibaba cut Qwen prices on Qoder targeting US-hour traffic (SCMP). Baidu released 'Unlimited OCR' for long-document AI processing as open source (GIGAZINE).

Why it matters

Teams with heavy document loads should try Baidu's OSS release. Qwen's drop means it's time to rerun any cost comparisons from 3+ months ago.

Source: localhost　·　→ View in AI Encyclopedia
Read in depth →

China's 'LineShine' tops the Top500 supercomputer list

AI Semiconductor / GPU Economics
Compared to before

The Top500 #1 slot has been held by US or Japan systems — Frontier, Fugaku — for several years. A Chinese system hadn't topped the list until now.

What changed

China's 'LineShine' took the #1 spot on the Top500 supercomputer ranking for the first time (GIGAZINE) — the first Chinese system to ever lead the list.

Why it matters

AI training scale comes down to raw compute. Even without export-controlled GPUs, China is building enough in-house computing to close the gap.

Source: localhost　·　→ View in AI Encyclopedia
Read in depth →
E · Theme of the day

Big AI money and political risk land on the same day

A $30B Japan pledge and Anthropic's DC friction landed on the same morning.

Blackstone commits ~$30B to Japan's AI infrastructure

The Big Picture of the AI Economy
Compared to before

Japan has lagged the US and China in AI infra build-out, with land and power constraints slowing data center expansion for the past year.

What changed

US PE giant Blackstone announced ~$30B in Japan AI infra investment — data centers, power, cooling, and networking (AI Business).

Why it matters

Japan's AI infra catch-up phase is now officially open. The service-layer competition — who profits above the hardware — follows once the infra is live.

Source: localhost　·　→ View in AI Encyclopedia
Read in depth →

Trump White House sours on Anthropic CEO Dario Amodei

Claude (Anthropic)Claude (Anthropic)
Compared to before

Anthropic published 'Policy on the AI Exponential' on June 11, staking out a pro-regulation stance that was already diverging from the White House.

What changed

Wired reports the Trump administration has soured on Anthropic CEO Dario Amodei, as Anthropic's vocal regulatory advocacy clashes with the White House (Wired).

Why it matters

Political alignment risk is now material for Anthropic. Anyone evaluating Anthropic for US government procurement or regulated industries should factor this in.

Source: anthropic.com　·　→ View on chaos map
Read in depth →
Archive

Past updates

A daily archive of changes actually applied to the site.