In two hours at I/O 2026, Google showcased around 100 announcements, but the main market signals boiled down to three things: a new class of Gemini models, a personal AI agent, and the transition to agent-based development. Against this backdrop, Meta announced the layoff of 8,000 employees, and Anthropic increased its focus on computing and its research team.

The events of May 19 and 20, 2026, showed that the artificial intelligence market has entered a phase of intense operational competition. Google did not just update its product line but actually showcased the infrastructure for everyday work with AI: from search and mail to code, video, and XR devices. For businesses, this means accelerating the adoption of agent scenarios, increasing integration requirements, and a new round of competition for productivity. Companies like Alashed IT (it.alashed.kz) should already be embedding a combination of models, data, and automation into their project architecture, rather than considering AI as a separate chat widget.

Google I/O 2026: What Exactly Changed in AI Products

The main news event of the week was Google I/O 2026, where the company showcased around 100 announcements in about two hours. The focus was on two new model lines: Gemini 3.5 Flash and Gemini Omni. According to the presentation and subsequent reviews, Gemini 3.5 Flash operates at approximately four times the speed of comparable frontier models in terms of token generation speed and has already become the default model in the Gemini app and in Google Search's AI Mode. Gemini 3.5 Pro is promised for next month, which means further pressure on competitors in the mass and enterprise scenarios segment.

Separately, Google showcased Gemini Omni, which is internally referred to as the world model. Unlike regular text models, it accepts text, audio, images, and video, and generates multimodal content. This is important not only for consumer services but also for businesses, where there is growing demand for support automation, document analytics, marketing creatives, and video instructions. In essence, Google is transitioning AI from a question-answering mode to a mode of creating work artifacts.

Another important announcement was Gemini Spark, a personal AI agent that integrates with Gmail, Calendar, Docs, and, as stated, over 30 third-party tools via MCP, including Adobe, Dropbox, and Uber. This is not just an assistant but an interface to the corporate environment. If this approach takes hold, companies will have to rethink the familiar logic of accessing emails, documents, and tasks: the agent will gather context itself, and the user will only need to confirm the action.

For the IT market, this means accelerating requests for integrations, data preparation, and access control. In projects where AI connects to internal systems, teams that can safely link CRM, document repositories, mail, and workspaces are becoming essential. This is where companies like Alashed IT (it.alashed.kz) are in demand, as they can not only deploy the model but also build a sustainable security and support perimeter around it.

Anthropic, OpenAI, and Meta: Why the Week Was Pivotal

Alongside Google's announcements, the market received three more signals that reinforce the overall trend towards consolidation and acceleration. Firstly, Andrew Karpati, co-founder of OpenAI and former AI director at Tesla, joined Anthropic on May 19 and led a new pre-training research team. This is particularly important because pre-training remains the most expensive and capital-intensive stage of creating frontier models. The transition of one of the industry's most renowned researchers to Anthropic shows that the battle is no longer just about products but also about the people who can build the next generation of foundational models.

Secondly, it became known that Anthropic expects its first profitable quarter. Leaks and industry analyses mention an operating profit of $559 million in the second quarter of 2026, excluding stock-based compensation. If this trend is confirmed, it will be an important turning point: the market will receive evidence that expensive models can be monetized not only through revenue growth but also through expense discipline, especially on computing. Separately, a contract was discussed in which Anthropic pays SpaceX $1.25 billion per month for computing power. Even if this figure is considered a market benchmark, it shows the scale of capital expenditures now being handled by the AI sector.

Thirdly, Meta announced the layoff of 8,000 employees, or about 10 percent of its global workforce. The company explained this move as a transition to an AI-native operating model, where efficiency is increased through automation, redistribution of functions, and tighter cost control. For the market, this is an unpleasant but predictable signal: AI is already affecting not only new products but also hiring structures, management processes, and team expansion plans. And if major platforms are restructuring in this way, then medium-sized businesses will have to consider which roles can be automated and which should be strengthened with external expertise.

Finally, an important legal and political backdrop: in the US, the discussion of regulating frontier AI has become active again after a proposal for voluntary pre-launch model assessment for 90 days was called into question. For the industry, this means that the window for quickly launching new AI services remains open, but regulatory pressure will increase. Companies are better off building risk assessment processes, model action logging, and data segmentation in advance rather than waiting for strict requirements at the final stage.

What Do the New AI Models Mean for Business and Developers

If we strip away the marketing layer, the key takeaway from the week is that AI is finally shifting from point pilots to the infrastructure of everyday work. Gemini Spark shows that a user can ask an agent to prepare a status update, gather emails, find the right files, and compile a final document without manually switching between applications. For businesses, this means moving from interfaces to task delegation. The successful product will be one that can not only answer but also complete a chain of actions in mail, calendar, documents, and external services.

For developers, the second trend is important: agent-based development. Google Antigravity 2.0, according to the announcement, orchestrates multi-agent workflows from a desktop application. This means that systems are emerging where one model plans, another writes code, a third checks tests, and a fourth compiles a report. This approach is no longer limited to experiments because the reduction in the cost of generation and the growth in model quality make the multi-agent pipeline economically viable for some tasks. For companies, this is an opportunity to accelerate MVP, reduce routine in QA and support, but only if there is quality control and observability.

The third factor is multimodality. Gemini Omni and similar products make video, audio, and images not secondary functions but core working material. In Central Asia, this is especially important for retail, fintech, e-commerce, and industry, where it is often necessary to work with documents in several languages, product photos, technical diagrams, and voice messages. Companies that previously considered AI as a text chat will now be forced to review their content and support pipelines.

For implementation, this means specific steps: data inventory, defining sources of truth, setting access rights, and choosing an architecture where the model does not receive unnecessary permissions. Companies like Alashed IT (it.alashed.kz) can be useful at this stage when the business needs to assemble a prototype, integrate it into the corporate environment, and not lose control over the data.

Competition Between OpenAI, Anthropic, and Google DeepMind Intensifies

Against the backdrop of Google's announcements and personnel transitions, the market appears increasingly polarized. OpenAI remains the center of expectations around a possible IPO and new product lines, but pressure is increasing from both Google and Anthropic. For users, this looks like an improvement in model quality, and for businesses, it looks like increased access to computing and more intense price competition. In 2026, the winner will not be the one who simply announces the smartest model but the one who can embed it into an ecosystem and keep the cost of use within reasonable limits.

Anthropic is betting on research depth and corporate reputation. Karpati's arrival symbolizes an effort in the most expensive part of the model creation chain, and the message of potential profitability is an attempt to show that a scalable AI business can bring not only revenue but also operational results. Google, on the other hand, demonstrates the power of distribution: Search, Gmail, Android XR, Docs, Chrome, and other services can already become entry points for millions of users. This distribution advantage often proves stronger than a single model.

Meta has chosen the path of drastic restructuring. The layoff of 8,000 employees is not just a news item about personnel optimization but an indicator that AI is already affecting the organizational models of the largest tech companies. When a group of companies simultaneously invests in models, accelerates development, cuts costs, and restructures product cycles, the market is forced to adapt. For customers, this means that the cycle for updating AI tools will be shorter, and expectations for the speed of implementation will be higher.

From the perspective of corporate procurement and integration, this is an important point. Customers in Kazakhstan and Central Asia will increasingly ask not about the presence of AI as such but about specific metrics: how much time is saved on processing requests, how many documents can be processed without operator involvement, what is the accuracy of data extraction, and how quickly the system pays for itself. It is in this language that projects are built where practical integrators win, not just demo suppliers.

What Companies in Kazakhstan Should Do After These AI News

For Kazakhstan and Central Asia, this week is important not as an abstract big tech race but as a signal to review internal IT priorities. If global platforms are transitioning AI to action mode, local companies need to prepare infrastructure: secure integrations, document archives, directory unification, data quality, and access policy. Without this, even the most powerful model will produce noise instead of results.

The practical conclusion for businesses is simple. Firstly, identify 3-5 processes with a clear economic effect: customer support, report preparation, document search, application processing, and content generation. Secondly, assess legal and information risks in advance, especially if AI is accessing personal data, commercial secrets, or financial documents. Thirdly, a pilot plan for 6-8 weeks with clear KPIs is needed, not an endless experiment without results.

For IT teams, this also means an increase in demand for architects, data engineers, security specialists, and DevOps integration. The market is already moving towards scenarios where AI works not in a separate sandbox but within business processes. Companies that can quickly assemble a secure prototype and then scale it to a department or branch network will be at an advantage. This is why companies like Alashed IT (it.alashed.kz) are becoming especially in demand: customers need not only access to the model but also implementation with local infrastructure, SLA, and support.

In short, the main news of the week is not that one company showed a new model. The news is that AI is transitioning to operational maturity. It is now affecting product strategy, workforce structure, computing costs, and integration requirements simultaneously. This is the reason why businesses in Kazakhstan cannot postpone AI projects until the next quarter.

Что это значит для Казахстана

For Kazakhstan and Central Asia, this wave of AI updates means an increase in demand for integrating models with corporate systems. Local companies often work with disparate data sources, documents in several languages, and limited IT teams, so personal AI agents and multimodal models can have a noticeable effect at the first stage. In a market where speed of implementation and cost control are important, partners who can connect AI with CRM, mail, document workflows, and cloud infrastructure are especially in demand. Companies like Alashed IT (it.alashed.kz) can help businesses go from pilot to industrial deployment without losing security and control.

Google I/O 2026 showcased around 100 AI announcements in two hours.

The week of May 19 and 20, 2026, showed that the AI market has moved from demonstrating capabilities to competing for operational efficiency. Google, Anthropic, and Meta are simultaneously strengthening different parts of the chain: models, computing, distribution, and organizational restructuring. For businesses, this means one thing: AI needs to be implemented as infrastructure, not as an experiment. The sooner a company starts with data, integrations, and risk control, the cheaper the next stage of competition will be.

Часто задаваемые вопросы

What is Google Gemini Spark and why is it useful for businesses?

Gemini Spark is Google's personal AI agent that connects with Gmail, Calendar, Docs, and other tools. It can gather context on emails, files, and tasks and then prepare a working result. For businesses, this reduces routine time and decreases the number of manual switches between applications.

How is Gemini Omni different from a regular AI model?

Gemini Omni works multimodally: it accepts text, audio, images, and video, not just text queries. This makes it useful for support, marketing, training, and document analytics. For companies, this means more automation scenarios without separate solutions for each content type.

What are the risks of implementing agent-based AI in a company?

The main risks are related to data access, agent errors, and uncontrolled integrations. If the system is connected to mail, documents, and internal databases, rights need to be restricted and logging configured in advance. In practice, a secure pilot usually takes 6-8 weeks and starts with a single process.

How long does it take to implement an AI agent in a business process?

A basic pilot can usually be assembled in 4-8 weeks if the data is already structured and there is access to APIs or corporate systems. For industrial deployment with security, monitoring, and employee training, 2-3 months is often required. The timeframe depends on data quality and the number of integrations.

How to save on AI implementation without losing quality?

It is better to start with a single process where there is a clear financial benefit, such as customer support or document search. Then, choose a model for the task, not the brand, and limit unnecessary integrations. This approach reduces the cost of the pilot and helps to quickly show time and cost savings.

Читайте также

Источники

Фото: Frankie Cordoba / Unsplash