Moonshot AI, a Chinese lab, launched Kimi K3 on July 16 through its website and its developer interface. It has 2.8 trillion parameters, the internal values a model learns during training, and a context window of about one million tokens (chunks of text), meaning it can hold a very long document in mind at once. It costs $3 per million input tokens and $15 per million output tokens, a sharp rise from the $0.95 and $4 Moonshot charged for the previous version.
Why it matters
Moonshot's claim that K3 beats Claude Opus 4.8 and GPT-5.5 on several coding tasks comes from its own testing, so treat it with care. Even so, a lab outside the US landing this close to the leaders puts downward pressure on what everyone pays to build with AI.
Google pushed back the wide release of Gemini 3.5 Pro, its next major model, after it fell short of internal goals. Coding was the main weak spot, and reports also pointed to problems with how efficiently it uses tokens and how well it handles long tasks that take many steps. The model was originally due in June 2026, and a later target of July 17 slipped as well.
Why it matters
It is rare for a large AI company to publicly admit a model is not good enough to ship. It is a useful reminder that progress is not automatic, and it landed the same week a Chinese rival claimed near-frontier scores.
Representatives of 29 countries signed the founding agreement for the World Artificial Intelligence Cooperation Organization in Shanghai on July 16, and Xi Jinping announced it in a keynote speech the next day. Founding members include Indonesia, Brazil, Malaysia, South Africa, Senegal, Russia and Pakistan, while no G7 economy (the group of seven large, wealthy democracies) and no European Union member state joined. The group is based in Shanghai and says its mission is to promote beneficial, safe and fair AI.
Why it matters
Three weeks ago every UN member state met to discuss governing AI together. This is the opposite move, a smaller bloc with its own rules, and it means what AI is allowed to do may increasingly depend on which country you are in.
Governor Kathy Hochul signed an executive order on July 14 pausing state environmental permits for new hyperscale data centers, the giant facilities that power cloud and AI services, for up to a year. The state will use the pause to write a regulatory framework, including a study of how much energy and water these sites use and how they affect air quality. Hochul cited rising utility bills for residents, strain on the electricity grid, and pressure on energy and water supplies.
Why it matters
The electricity and water that AI runs on now show up on household bills, which turns an abstract cost into local politics. Other states are likely to copy the template.
Microsoft's July security update fixed 570 vulnerabilities (flaws attackers can exploit), its largest ever and more than four times the 137 it fixed in July 2025. Two of them were already being used in real attacks when the fixes shipped. Microsoft said the jump came from an internal AI system that finds vulnerabilities across the Windows code faster than human review did, and warned that the coming months will look similar.
Why it matters
This is a concrete example of AI changing security work now, not a prediction that it will. The practical takeaway if you use Windows: install the update.
Munich-based Helsing announced a $1.8 billion funding round on July 13 at an $18 billion valuation, up from 12 billion euros in June 2025, and said investor demand was greater than the amount it had available. Helsing builds the HX-2 strike drone, battlefield software and a proposed concept for an autonomous fighter aircraft, and works with defense companies Rheinmetall, Kongsberg and Saab.
Why it matters
Defense has gone from roughly 1 percent of European venture funding before 2020 to around 10 percent today. If you are weighing where jobs and money are heading in European tech, this is one of the clearest signals available.
TSMC, the Taiwanese company that manufactures the chips designed by Nvidia, Apple and most other tech firms, reported quarterly revenue of about $40.2 billion on July 16, up 36 percent from a year earlier, with net income (profit) up 77 percent. High-performance computing, the part of the business that covers AI chips, now makes up 66 percent of its wafer revenue. It raised its 2026 capital spending plan, money for factories and equipment, from $52-56 billion to $60-64 billion.
Why it matters
TSMC makes the chips behind essentially every AI product, so its spending plans are the most reliable read on whether the AI buildout is still speeding up. Right now the answer is yes.
The week pulled in two directions at once. On the software side, the frontier looked crowded rather than commanding: a Chinese lab claimed near-leading scores in the same days that the company that invented the transformer said its next model was not ready. Capability is no longer something one or two American labs quietly build up, and the gap between the leaders and everyone else is now measured in months and dollars rather than years.
On the physical side, the bill came due. A record quarter for the world's most important chipmaker and a state pausing data center permits are the same story told from opposite ends: demand for computing power is real enough to reset a chipmaker's spending plans, and heavy enough that a governor is now weighing it against household electricity bills. Add a new governance bloc formed without any Western members, and AI's hard limits are shifting from what models can do to where they are allowed to run and who pays for the power.
What to watch
Watch for Google's next release date for Gemini 3.5 Pro, which has already missed both its original June target and a later July 17 date.
Test yourself
A few quick questions on this week. Just for you, nothing is recorded.
1. Why did Google delay the wide release of Gemini 3.5 Pro?
2. What stands out about the membership of the new World Artificial Intelligence Cooperation Organization?
3. Why did Microsoft's July security update fix a record number of flaws?
4. What reasons did New York give for pausing new hyperscale data centers?
