OpenAI Reveals Jalapeño Chip, Davos Shifts to AI Reality — June 25 AI Roundup

Q: 🔥 Top Story: OpenAI + Broadcom Reveal Jalapeño — First Custom AI Chip

OpenAI and Broadcom jointly unveiled Jalapeño, OpenAI's first custom-designed intelligence processor, at a launch event coinciding with MWC Shanghai. What it is: An LLM inference accelerator architected specifically for transformer workloads. Jalapeño is the first chip in a multi-generation compute platform OpenAI and Broadcom are co-developing. Key specs (from the announcement): - Purpose-built for transformer inference (GPT, Claude, Gemini-class models) - Comparable inference performance to NV

Q: 📱 MWC Shanghai: AI for Everyone

MWC Shanghai 2026 kicked off with a clear theme: AI普惠 (AI for everyone). Major announcements included: - Baidu Qianfan Token Plan Enterprise Edition launched — enterprise-grade AI token packages supporting GLM-5.2 and other major Chinese models. A direct response to the growing enterprise AI market in China. - Alibaba QoderWork "Peak-Valley Token" — a clever pricing scheme charging 80% less for Qwen 3.7 API calls during nighttime hours. Load-shifting for AI compute, just like electricity pricing

Q: 🤖 Claude Code Gets a Major Upgrade — Karpathy: "The Third Revolution"

Anthropic shipped a major update to Claude Code, its AI coding agent. The upgrade introduces: - Longer, autonomous coding sessions — Claude Code can now work on multi-file refactoring tasks without constant user handholding - Better test generation — automatic test writing with significantly higher coverage - Deeper IDE integration — the tool now understands project structure beyond individual files Andrej Karpathy (former OpenAI, Tesla AI lead) called it "the third revolution of LLMs" — after t

Q: 🖥️ NVIDIA Dominates Supercomputing — 81% of TOP500

At ISC High Performance 2026 in Europe, NVIDIA announced: - 35 new AI HPC supercomputers in development across Europe (up from 23 last year) — equipping 3M+ researchers with next-gen AI infrastructure - 400+ of the world's 500 fastest supercomputers run on NVIDIA (81% of TOP500) - 26 systems adopted the Grace CPU — up 8 from the previous list - Top 8 Green500 systems all run on NVIDIA GPUs NVIDIA also announced a deeper collaboration with AWS to bring AI to production at scale, including managed

June 25, 2026 — AI is everywhere this week: MWC Shanghai, Summer Davos in Dalian, and ISC High Performance in Europe. But the tone has shifted from hype to reality. Here's the full picture.

🔥 Top Story: OpenAI + Broadcom Reveal Jalapeño — First Custom AI Chip

OpenAI and Broadcom jointly unveiled Jalapeño, OpenAI's first custom-designed intelligence processor, at a launch event coinciding with MWC Shanghai.

What it is: An LLM inference accelerator architected specifically for transformer workloads. Jalapeño is the first chip in a multi-generation compute platform OpenAI and Broadcom are co-developing.

Key specs (from the announcement):

Purpose-built for transformer inference (GPT, Claude, Gemini-class models)
Comparable inference performance to NVIDIA H200 GPUs
Significantly lower power consumption per token
Expected to cut OpenAI's inference costs by 30-50% once fully deployed

Why now: OpenAI burns billions on GPU compute. Eight months after announcing a custom chip partnership with Broadcom, Jalapeño is the first silicon outcome. It joins Google's TPU, Amazon's Trainium, Microsoft's Maia, and Meta's in-house silicon efforts — every major AI player is now building custom chips.

The bigger picture: This isn't just about cost savings. Vertical integration in the AI stack is accelerating. Chips → models → APIs → applications: the winners in the next AI cycle will own multiple layers of the stack.

🇨🇳 Summer Davos 2026: AI's Reality Check in Dalian

The World Economic Forum's Annual Meeting of the New Champions (Summer Davos) is underway in Dalian, bringing together 1,700+ delegates from 90+ countries. AI was the dominant theme, but not in the way you'd expect.

Key takeaways:

1. Infrastructure, not intelligence, is the bottleneck NTT DATA Chief Strategy Officer Roli Agrawal made a striking point: "What limits AI development is not intelligence — it's infrastructure." She argued that existing digital infrastructure wasn't built for AI workloads. The two critical layers:

Compute — where to put edge vs. cloud processing
Network — low-latency, high-bandwidth data pipelines (photonics networks already deploying in data centers)

2. AI sovereignty and governance are the next battleground As AI models cross borders, countries are asking: who owns the data? Who sets the rules? Agrawal's "1-2-3-4 rule" for enterprise AI deployment: innovation creates potential, execution brings impact, governance scales influence.

3. Cost and openness as differentiators Chinese AI models (DeepSeek, Qwen, GLM) are gaining traction globally because they're cheap — and often open. The Stanford 2026 AI Index shows US-China model performance gap has narrowed to just 2.7%. With Chinese models offering API calls at a fraction of Western competitors' prices, cost and openness could reshape global AI adoption patterns.

📱 MWC Shanghai: AI for Everyone

MWC Shanghai 2026 kicked off with a clear theme: AI普惠 (AI for everyone).

Major announcements included:

Baidu Qianfan Token Plan Enterprise Edition launched — enterprise-grade AI token packages supporting GLM-5.2 and other major Chinese models. A direct response to the growing enterprise AI market in China.
Alibaba QoderWork "Peak-Valley Token" — a clever pricing scheme charging 80% less for Qwen 3.7 API calls during nighttime hours. Load-shifting for AI compute, just like electricity pricing.
Qualcomm's pivot to "Physical AI" — the chipmaker is repositioning from smart cockpits to robotics and autonomous systems, arguing its mobile DNA gives it an edge in power-efficient edge AI.

175 early-stage AI projects were showcased at the WAIC 2026 project exhibition running alongside MWC.

🤖 Claude Code Gets a Major Upgrade — Karpathy: "The Third Revolution"

Anthropic shipped a major update to Claude Code, its AI coding agent. The upgrade introduces:

Longer, autonomous coding sessions — Claude Code can now work on multi-file refactoring tasks without constant user handholding
Better test generation — automatic test writing with significantly higher coverage
Deeper IDE integration — the tool now understands project structure beyond individual files

Andrej Karpathy (former OpenAI, Tesla AI lead) called it "the third revolution of LLMs" — after the transformer architecture itself and ChatGPT's product-market fit, Claude Code represents the shift from "chat" to "autonomous work."

This is part of a broader trend: AI coding tools are evolving from assistants to agents. Cursor, Copilot, Codex, and Claude Code are all racing toward the same vision — an AI that doesn't just suggest code, but writes, tests, and deploys it.

🖥️ NVIDIA Dominates Supercomputing — 81% of TOP500

At ISC High Performance 2026 in Europe, NVIDIA announced:

35 new AI HPC supercomputers in development across Europe (up from 23 last year) — equipping 3M+ researchers with next-gen AI infrastructure
400+ of the world's 500 fastest supercomputers run on NVIDIA (81% of TOP500)
26 systems adopted the Grace CPU — up 8 from the previous list
Top 8 Green500 systems all run on NVIDIA GPUs

NVIDIA also announced a deeper collaboration with AWS to bring AI to production at scale, including managed inference, vector search, and GPU capacity on demand.

💰 Physical AI's First Trillion-Yuan Market: Road Freight

A notable story from China: physical AI (具身智能/embodied AI) has found its first trillion-yuan commercial application — road freight.

The thesis: autonomous trucking doesn't need to solve every driving scenario. Highway freight is predictable, regulated, and route-constrained. Companies are deploying Level 4 autonomous trucks on fixed freight routes with remote monitoring, achieving:

15-20% fuel savings through optimized driving patterns
24/7 operation (no driver hours limits)
40% reduction in accident rates on highway routes

The road freight use case is significant because it's one of the first "Physical AI" markets to demonstrate unit economics that work — not just a demo, but real P&L-positive deployment.

📊 Quick Hits

Headline	Signal
影眸科技 Hyper3D raises 数亿元 — 3D generation enters "thinking age"	3D AI is maturing fast
Doubao Professional Edition officially launched	ByteDance doubles down on productivity AI
Apple foldable iPhone to mass produce in July — rumor	AI + foldable = new form factor for AI agents?
MGX $50B AI fund	Abu Dhabi's sovereign wealth goes all-in
US urges Meta to share AI models	Policy pressure for open-weight access
Samsung announces AI-optimized HBM4E memory	Next-gen memory for next-gen AI chips

🔮 Today's Takeaway

Three themes define June 25, 2026:

Vertical integration is the new normal. OpenAI builds its own chip. Qualcomm buys Modular. Every major player wants to own more of the stack.
The conversation is shifting from "can AI?" to "should AI?" Davos spent more time on infrastructure, governance, and sovereignty than on model benchmarks. The industry is growing up.
China's AI ecosystem is price-disrupting globally. From Baidu's enterprise token plans to Alibaba's peak-valley pricing to DeepSeek's global adoption — Chinese AI is competing on cost and openness in ways Western incumbents can't ignore.

Catch up on previous roundups: AI News Archives

Stay updated with the latest AI tools and news at Mee AI.