Dakarda AI Newsletter · 26 June 2026
Friday Edition
Government says: stop, OpenAI makes a chip, SpaceX pays 60 bn
History accelerates. In two days, the US controls access to an AI model at the customer level for the first time, OpenAI shows its own silicon, and SpaceX buys a coding tool for 60 billion dollars. These are not rumors — they are precedents.
Intro · Alex
Today I woke up feeling like I'm looking at a different world than yesterday. The US government asked OpenAI to limit GPT-5.6 to a small group of partners — the first time an administration decides who can use a model. At the same time, SpaceX breaks the bank with the acquisition of Cursor, and OpenAI shows its own ASIC. These are not separate stories — they are one tale of AI ceasing to be software and becoming infrastructure. In this issue: the Jalapeño chip, which OpenAI designed in 9 months, Claude as a team member in Slack, the final deadline for the EU AI Act, and an attack showing that even security tools can be the weakest link. Ready?
What's worth knowing
US Government limits GPT-5.6 — access 'customer by customer'
The White House asked OpenAI for a staggered release of GPT-5.6 — the model will only go to a handful of approved partners. This is a historic precedent: for the first time, the government directly controls the distribution of an AI model at the customer level. Sam Altman is not thrilled, but he agrees. For the industry, it means one thing: every subsequent frontier model may require certification before anyone touches it.
OpenAI and Broadcom unveil Jalapeño — the first custom AI chip
The OpenAI and Broadcom team completed the project from concept to tape-out in 9 months — using their own AI models to accelerate design. Jalapeño is a dedicated ASIC for LLM inference that will go into Microsoft data centers by the end of the year. This is a first step toward independence from Nvidia and drastically reducing inference costs for GPT-5.3 and Codex.
Anthropic turns Slack into an agent hub — Claude Tag launched
@Claude now works in Slack as a team member — it has access to selected channels, tools, and code repositories. You can delegate tasks from product metrics to debugging. The old integration will expire on August 3, but the new feature turns Slack from a messenger into a platform where AI is a coworker with the context of the entire workspace.
SpaceX acquires Cursor (Anysphere) for 60 billion USD — the largest startup M&A
SpaceX paid 60 billion dollars for Anysphere, the creator of the AI coding tool Cursor. This is the largest startup acquisition in history — for comparison, Facebook bought WhatsApp for 19 billion. The signal is clear: the biggest players treat AI-assisted coding as strategic infrastructure, and Cursor has grown from a developer tool to a milestone.
EU AI Act: delays and new timeline — full enforcement in August 2026
The European Commission postpones deadlines: full enforcement of the AI Act will come into effect by August 2, 2026. The Digital Omnibus amendments give companies some breathing room, but penalties for prohibited practices (Article 5) still reach up to 35 million euros or 7% of annual turnover. For anyone building AI products in Europe, August is the final deadline — don't miss it.
From the tech world
Massive attack on Fortinet devices — nearly 74,000 firewalls breached
The FortiBleed campaign shows that complicated exploits are not necessary — vulnerabilities that the vendor patched, but administrators didn't apply the fixes, were enough. Irony: the devices that were supposed to provide security were themselves taken over. A lesson on patch management that you will remember for a long time.
Read this before you vibe-code another app
The Verge warns: vibe-coding (creating apps through AI without programming knowledge) leads to serious security vulnerabilities. The article describes a case where an AI-generated application left open access to data. Ease of creation does not go hand in hand with ease of securing — a practical caution for anyone using AI to generate code.
Tip of the day
Don't trust AI that designed itself — but learn from it
OpenAI used its own models to design the Jalapeño chip. It's a tempting pattern, but remember: AI excels at optimizing known design spaces, but may overlook fundamental assumptions that a human would take for granted. Practical step: if you use AI for design — whether code, architecture, or processes — always add a human review at the level of assumptions, not just the result. Ask: 'Does this solution make sense, even if AI says it does?'.
Reading list
EU AI Act for US Companies
If you think the AI Act only concerns Europe — read this. US companies also need to comply, and penalties are harsh. A short but substantial compliance guide.
The Verge — vibe-coding security risks
A concrete case study showing how easy it is to generate an app with a security hole. Required reading for anyone who 'programs' with the help of AI.
In a week when the government says 'stop' and SpaceX pays 60 billion for a coding tool, one thing is certain: the game is changing faster than anyone was ready for.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.