AI Weekly: weekly roundup (13–19 July 2026)

Last week showed that competition in AI is now unfolding on several levels at once. Model providers are fighting over quality, pricing and a place in everyday work. Technology companies are moving into hardware, cybersecurity and robotics. Governments are building their own regulatory frameworks, while investors are directing an ever-larger share of available capital into the sector.

The most prominent story was the open conflict between Apple and OpenAI, but the less spectacular signals may prove just as important: the growth of open-weight models, record results from a leading chip manufacturer, delays at Google and the first major labour dispute over deploying humanoid robots in factories.

We have collected the most important developments in one place, so they can be reviewed over a cup of coffee and quickly separated into interesting headlines and changes that may have a real impact on businesses using AI.

The format remains simple: first, a quick overview of the week, followed by short explanations of the most important themes.

In brief

Apple and OpenAI enter an open conflict

Apple sued OpenAI and people connected with its hardware division, accusing them of using trade secrets taken by former employees of the iPhone maker. The complaint includes allegations that confidential materials were transferred and that physical Apple components were brought to recruitment interviews.

Movement of specialists between technology companies is not unusual in itself. The difference is that OpenAI is no longer developing only models and applications. It is building its own hardware division and preparing devices that may compete for the same user attention as the iPhone, headphones or a smartwatch.

The lawsuit is therefore about more than the conduct of a few employees. It could delay OpenAI’s hardware projects, make further recruitment from Apple more difficult and increase legal risk ahead of a planned stock-market debut. Even if the proceedings take years, they are already affecting how partners and investors assess the company’s credibility.

At the same time, Apple is not stepping back from AI. During the same week, Apple Intelligence received approval to enter China, where some functions are expected to use technology from Alibaba and Baidu. The company also returned to first place among the world’s largest companies by market capitalisation, moving ahead of Nvidia.

This is only an apparent paradox. Apple can defend its intellectual property from OpenAI, use Chinese models in a local market and still convince investors that its advantage does not depend on building the most powerful model. It may be enough to become the most important distribution channel for AI across billions of devices.

For businesses, the practical lesson is straightforward: agreements with AI providers do not replace rules for protecting know-how. When recruiting specialists, integrating with a partner or building joint products, companies need to distinguish clearly between an employee’s experience and the materials, data and solutions owned by a previous employer.

GPT-5.6 brings structure to OpenAI’s portfolio

OpenAI completed the rollout of the GPT-5.6 family in three variants: Sol, Terra and Luna. Sol is the flagship model for the most demanding tasks, Terra is intended to balance capability and cost, and Luna is designed for fast, lower-cost operations at scale.

The most important change is therefore not simply better benchmark performance. OpenAI is beginning to sell models in a way that resembles cloud infrastructure: customers are not choosing one “best AI”, but a level of quality and price appropriate to a specific process.

This matters for teams running thousands or millions of queries. It makes little economic sense to use the most powerful model to classify simple requests, extract fields from a document or prepare the first version of a summary. The more expensive option can be reserved for difficult reasoning, work involving sensitive data and operations where the cost of an error is high.

In parallel, OpenAI is developing ChatGPT Work and the GPT-Live interface. The first moves the model towards producing finished documents and carrying out office work. The second enables a more natural voice conversation in which the system can listen, speak and respond to interruptions at the same time.

These functions point in the same direction: AI is becoming a less visible layer inside a process. Users do not have to open a separate chatbot, copy data and manually transfer the result. The model begins to operate within a conversation, document or development tool.

Greater convenience also makes operational controls more important. Widely discussed cases of agents behaving unexpectedly around user files are a reminder that access to the operating system, email or repositories should be granted gradually. Start with read-only access and proposed actions, move to approved changes, and only then allow limited autonomy in processes that are easy to reverse.

AI safety becomes a political issue

In mid-July, reports resurfaced about talks between OpenAI and the US administration over a possible transfer of approximately 5% of the company’s equity to the government. This does not mean that an agreement has been reached. The idea itself would nevertheless be unprecedented, because it would link the government’s financial interest to the success of one of the largest AI providers.

For OpenAI, such an arrangement could make it easier to transform its ownership structure, negotiate with regulators and expand internationally. For the state, it would provide direct participation in the value created by a company regarded as strategically important. The difficulty is that a regulator that is also a shareholder may struggle to maintain complete distance.

Social pressure is moving in the opposite direction. Shortly before the week covered by this roundup, around 200 people marched in San Francisco between the offices of OpenAI, Anthropic and Google DeepMind, calling for a temporary halt to training new frontier models and greater investment in safety. The protest was not large, but it showed that criticism of the AI race is beginning to move beyond conferences and expert reports.

A day later, the Future of Life Institute published another edition of the AI Safety Index. Anthropic received the highest score at C+, OpenAI and Google DeepMind received C, Meta received D+, while xAI, DeepSeek and Mistral failed. This is not a direct ranking of how safe the models themselves are. The index primarily assesses risk management, transparency, safety frameworks and whether company practices match their public commitments.

The most telling point is that C+ remains the best result in the entire industry. The leader may use that position in enterprise sales and government discussions, but it is difficult to treat it as proof of sector-wide maturity. The report suggests instead that even the most advanced companies have not yet built mechanisms proportionate to the capabilities they claim for their models.

For organisations purchasing AI, this means that a provider’s brand cannot replace an internal risk assessment. Contracts should address audits, incident reporting, rules for data use, the ability to disable functions and the response procedure when a model or agent behaves unexpectedly.

Open-weight models return to the game

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, introduced Inkling, its first open-weight model. The model can be downloaded, run in an organisation’s own environment and adapted to its needs rather than accessed only through a provider’s closed API.

Inkling does not have to beat the most powerful models from OpenAI, Anthropic or Google in every benchmark immediately. Its advantage may lie elsewhere: companies gain greater control over deployment, data, fine-tuning and cost. This is particularly important in sectors where a process cannot depend on sending confidential information to an external service.

A day later, Moonshot AI presented Kimi K3, a mixture-of-experts model with 2.8 trillion parameters and a context window of up to one million tokens. Full weights are scheduled for release on 27 July. The announcement alone attracted considerable interest because the model is expected to compete with closed systems in agentic tasks, programming and long-context work.

It is worth remembering that parameter count is not a simple measure of quality. In a sparse MoE architecture, only part of the model is activated for a given task. Even so, the scale of Kimi K3 shows that open weights are no longer limited to small models and community experiments.

This is also part of geopolitics. US companies most often export AI through controlled APIs, access limits and their own cloud platforms. Chinese providers are increasingly building influence through models that can be downloaded and run locally. For governments and companies seeking to reduce dependence on US platforms, this is an attractive alternative.

Companies should therefore prepare for a multi-model environment. Closed APIs may still provide the best quality and easiest deployment, but open-weight models will increasingly make economic sense where control, customisation or predictable cost at scale matter more.

China builds its own AI order

In Shanghai, 29 countries signed an agreement establishing the World Artificial Intelligence Cooperation Organization, or WAICO. The organisation is to be headquartered in China and operate as an intergovernmental forum for cooperation on the development and governance of AI.

The founding countries include Indonesia, Pakistan, Russia, Kazakhstan and Laos. UN Secretary-General António Guterres was also present. At the same time, no Western democracy joined the initiative, giving the organisation a clear geopolitical dimension from the outset.

Western AI frameworks usually emphasise human rights, transparency, safety and provider accountability. The Chinese proposal places more emphasis on state sovereignty, economic development and equal access to technology. For many countries in the Global South, this language may be more attractive than adopting rules written by the G7, OECD or European Union.

WAICO does not have to become a United Nations equivalent for artificial intelligence overnight. It may, however, serve as a coordination centre for countries using Chinese models, infrastructure and financing. Combined with open-weight systems such as Kimi K3, this creates a more complete offer: technology, cooperation rules and an institution capable of promoting them.

For international businesses, this means regulatory requirements may continue to diverge. A product compliant with the European AI Act will not automatically meet local rules in countries aligned with WAICO. Companies will have to manage not only technical differences, but also distinct requirements concerning data, moderation, audits and government access to systems.

Google slips while TSMC grows

Google postponed the launch of Gemini 3.5 Pro again. According to reports, the model requires further work following problems found in areas including tool use and programming tasks. The news caused Alphabet’s share price to fall by more than 4%, as investors feared that the company was losing momentum relative to OpenAI, Anthropic and new models from China.

A delay alone does not determine the quality of a product. For frontier models, additional testing may be a better decision than releasing an unfinished system. Google’s problem is that the market assesses not only a final benchmark, but also the ability to deliver models regularly, integrate them into products and turn research into revenue.

At the same time, TSMC reported another record quarter. Revenue reached approximately $40.2 billion and net profit around $22 billion. High-performance computing, which includes chips for AI and data centres, already accounted for 66% of the company’s revenue.

The contrast between these two stories is revealing. A model provider may delay a launch, lose a benchmark or change its product strategy. A manufacturer of advanced chips earns money regardless of whether a particular accelerator is ultimately used by Google, Nvidia, Apple or another AI laboratory.

This does not mean that infrastructure is free of risk. Such a high concentration of revenue in HPC makes TSMC increasingly dependent on sustained investment in data centres. If demand slows or models become significantly more computationally efficient, today’s advantage could turn into excess capacity.

For now, however, TSMC’s results provide hard evidence that the expansion of AI infrastructure is not merely a promise made in investor presentations. Companies are buying chips and reserving manufacturing capacity on a massive scale.

Robots enter the dispute over jobs

Hyundai trade unions in South Korea began a partial strike in a dispute that included plans to deploy Atlas humanoid robots on production lines. Workers are demanding that robots should not enter factories without a prior agreement with the unions.

The situation is unusual because Hyundai owns Boston Dynamics, the company that makes Atlas. It can therefore develop the robot, test it in its own facilities and directly measure the savings generated by automation. From a business perspective, this is close to an ideal deployment model.

From the workers’ perspective, the same closed loop creates a real risk. The discussion is no longer about the abstract possibility that robots may “one day” replace some tasks. It is about deployment schedules, the number of jobs, reskilling, responsibility for safety and workers’ share in the benefits of automation.

Factories are a natural environment for the first humanoids. The setting is controlled, tasks can be described and repeated, and results can be measured through cycle time, error rates and labour costs. This is precisely why disputes similar to the one in Korea may quickly appear in European and US industry.

For companies, the most important conclusion is not about choosing a particular robot. Automation introduced without a plan for people can lead to delays, protests and a loss of trust that offsets part of the expected savings. Discussions about roles, training and the distribution of benefits should begin before the first machine is installed.

Microsoft bets on multiple models

According to reports, Microsoft is preparing Project Perception, a tool intended to use AI to detect vulnerabilities in enterprise systems and propose ways to remediate them. The project is expected to combine models from Anthropic and OpenAI with Microsoft’s own systems.

This is an important signal because Microsoft is not trying to prove that one model should perform every task. The system is expected to route individual operations to different models depending on cost, quality and the nature of the problem. A more expensive model can analyse a complex case, while a cheaper one handles routine classifications and large-scale scanning.

The strategy resembles the development of cloud computing. Microsoft does not have to produce every best application if it controls the platform on which customers run third-party applications. In AI, Azure, security tools, identity management, data and routing between models are becoming the platform layer.

For enterprise customers, this may be attractive because it reduces dependence on one model provider and allows costs to be optimised. At the same time, it makes auditing more complex. An organisation must know which model is processing its data at a given moment, where the data is being sent and what retention rules apply to each subcontractor.

Project Perception has not yet been publicly confirmed as a finished product, so its reported features should not be treated as final. The direction is nevertheless clear: the advantage may go not to the company with one best model, but to the platform capable of connecting several models safely within one business process.

Capital concentrates in AI

Weekly market summaries returned to data from PitchBook and the National Venture Capital Association for the first half of 2026. US startups reportedly raised approximately $412.7 billion in total, of which 86%, or around $355.9 billion, went to AI-related companies.

This does not mean that hundreds of small AI startups received equally large amounts of funding. Most of the capital is concentrated in the biggest rounds and flows to a limited number of companies building models, data centres, chips and infrastructure tools. The concentration exists both at the sector level and within individual transactions.

Some of the money comes from traditional venture funds, but strategic investors also play a major role: technology companies, chip manufacturers and businesses building their own cloud ecosystems. In many cases, the funding is not merely a bet on an increase in a startup’s valuation. It also secures access to models, customers, computing capacity or a future distribution channel.

For the AI industry, this is an enormous acceleration. Companies can train larger models, recruit specialists and finance expensive enterprise deployments. For the rest of the venture market, however, a crowding-out effect appears. A biotechnology, energy or logistics startup competes for investor attention with companies that can add “AI” to their strategy and promise much faster growth.

Such concentration also increases systemic risk. If AI revenue fails to keep pace with infrastructure costs, or if the valuations of several major companies are corrected, the effects will not be limited to their shareholders. Capital links now extend across cloud providers, chip manufacturers, energy, data centres and a significant part of the startup market.

---

What it all means

Several parallel changes emerge from the events of last week.

First, the AI market is ceasing to be a race between individual models. OpenAI is dividing its portfolio into pricing tiers, Microsoft is building a multi-model solution, and open-weight providers are offering alternatives to closed APIs. In practice, organisations should prepare an architecture in which the model can be changed according to the task, risk and cost.

Second, the layer around the model is becoming increasingly important: hardware, infrastructure, distribution, access to data and integration with work processes. TSMC earns money regardless of who leads the benchmarks, while Apple can build a strong position in AI through devices and its relationship with users, even without a flagship model of its own.

Third, safety is becoming a technical, political and ownership issue at the same time. Weak scores in the AI Safety Index, talks about a government stake in OpenAI and Apple’s trade-secrets dispute show that declarations of responsibility are not enough. Audits, clear permissions, separation of interests and incident-response procedures are needed.

Fourth, AI geopolitics is beginning to acquire its own institutions and products. WAICO and large Chinese open-weight models are creating an alternative ecosystem for governments and companies that do not want to rely exclusively on US providers. This means more choice, but also wider regulatory differences and more difficult compliance management.

Finally, automation is entering a stage of social negotiation. The Hyundai strike shows that deploying humanoids will not be only a decision for the technology department. It will also concern working conditions, responsibility and how the benefits are divided between the company and its employees.

For businesses, this does not mean reacting to every new headline. It does mean taking care of several fundamentals: avoid making processes dependent on one model, grant agent permissions gradually, control data flows between providers, plan for automation’s effect on employees and calculate the full cost of deployment — including infrastructure, security and legal risk.