Dakarda AI Newsletter · 20 July 2026
Monday Edition
GPU is the new oil – enough for everyone?
The world's largest open-source model suspends subscriptions because GPU capacity ran out. The European Union is forcing Apple and Google to open their AI assistants to competition. And McKinsey reveals the number that will change how we think about agents: 60% of costs aren't inference — they're fixing answers.
Intro · Alex
This was the week AI hit three walls at once: computational, regulatory, and economic. Kimi K3 (2.8 trillion parameters!) turned out so popular that Moonshot AI had to suspend new subscriptions — they ran out of GPU capacity. At the same time, the EU, instead of just watching the arms race, decided to cut the Gordian knot: it's forcing Google and Apple to let us choose any AI assistant as the default on our phones. In today's edition, I've traced four stories that in my assessment will have the biggest consequences in the coming weeks. From China's compute capacity exhaustion, through the regulatory assault on the Apple-Google duopoly, to McKinsey's financial anatomy of AI agents. As a bonus — Alibaba's Qwen 3.8, which aims to be 'second only to Fable 5' and does it with open weights.
What's worth knowing
Kimi K3 halts subscriptions — demand crushed GPU capacity
Moonshot AI suspended new subscriptions for the Kimi K3 model just 48 hours after launch. The reason? Massive GPU load — demand nearly reached the limit of available compute capacity. The company plans to split membership into Kimi Membership (web/app) and Kimi Code Membership (for developers). This is the first time a Chinese model has achieved parity with Fable 5 and GPT-5.6 Sol, generating such global demand — and the first time GPU access, not technology, has turned out to be the bottleneck.
EU forces Google and Apple to open smartphone AI assistants
The European Commission, citing the DMA (Digital Markets Act), ordered Google and Apple to allow users to choose alternative AI assistants (e.g., ChatGPT, Claude) as defaults on Android and iOS. This is a fundamental shift in AI distribution — instead of platforms controlling assistants, users can select any AI agent as the system default. Reminder: Apple announced in June it would not launch the new Siri AI in the EU precisely because of these regulations.
McKinsey: 60% of AI agent costs go to response refinement — most firms already over budget
A new McKinsey report reveals a game-changing number for anyone building agents: 60% of operational costs go to the response refinement phase, not inference itself. Most enterprise firms have already exceeded their AI agent budgets. The real cost lies in verification, validation, and iteration loops — cheap inference is a myth.
Alibaba releases Qwen 3.8 (2.4T parameters) — 'second only to Fable 5'
Alibaba launched Qwen 3.8-Max-Preview — a model with 2.4 trillion parameters, available on Token Plan, Qoder, and QoderWork. The company announced a forthcoming full open-weight release. This is a direct response to Kimi K3 (2.8T) released just 3 days earlier. Two models over 2 trillion parameters in a single week — the Chinese open-weight model race is accelerating at cosmic speed.
From the tech world
16-year-old Linux KVM flaw lets guest machines escape to host
A critical vulnerability was discovered in KVM (Kernel-based Virtual Machine) for Linux, present in the code since 2009. It allows virtual machines to escape (VM escape) to the host system on Intel and AMD x86 processors. Critical information for cloud administrators, data centers, and DevOps — breaking isolation between VM and host is a worst-case scenario.
US government warns: Russian hackers taking over home and small office routers
The US government warns of mass compromise of home and small office routers by hackers linked to the Russian state. Attackers are creating botnets to mask attacks on sensitive institutions, exploiting poorly configured SNMP. The article includes practical recommendations for securing network equipment.
Tip of the day
How to estimate the real cost of an AI agent before you blow your budget
Most teams calculate agent costs like traditional LLMs: price per input and output token. Yet, as the McKinsey report shows, 60% of spending goes to response refinement — verification, validation, re-query, and correction loops. Before deploying an agent to production, run a cost test: execute 100 simulated tasks, measure the total number of API calls (not just the main prompt), sum up validation costs, and compare against your budget. Practical step: add a cost logger to your agent at every loop stage — inference, validation, correction, re-inference. Only when you see the breakdown will you know where the money is actually going. In 9 out of 10 cases, you'll be surprised at how little the 'thinking' costs and how much the 'checking' costs.
Reading list
60% of Agentic AI Costs Go to Response Refinement — and Most Enterprises Are Already Over Budget
The full McKinsey analysis that should be mandatory reading for every CTO and product owner building agents. It shows where the money really goes and why a 'cheap LLM' doesn't mean a 'cheap agent'.
EU forces Apple and Google to open smartphones to rival AI assistants
A detailed overview of the EC decision that could change the global AI assistant market. If you think this is just a European matter — think again. EU regulatory precedents have a habit of becoming global standards.
This week, AI taught us three things: GPUs are the new oil, regulation can be faster than technology, and the most expensive part of an agent isn't thinking — it's checking whether it thought correctly.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.