Google, Nvidia, OpenAI, Europe: four AI commitments to track now (2026)

Four events around August 24-25 put AI decisions on a calendar. Google Cloud and Verizon announced a multi-year enterprise deployment built around Gemini Enterprise. Reuters reported that Nvidia was discussing an investment in Perplexity at a valuation above $30 billion. OpenAI says o3 will leave ChatGPT on August 26, 2026. The European Commission says major AI Act enforcement powers have applied since August 2. 1 2 3 4

These four items have different levels of certainty. Verizon's deployment is planned and partly underway. The Nvidia-Perplexity financing remains a reported discussion. OpenAI's retirement date is confirmed. The EU enforcement framework is already operating. That distinction tells a reader what to verify before buying, deploying, or relying on an AI workflow.

Development	What changed	Action window
Google Cloud and Verizon	Verizon says Gemini Enterprise for Customer Experience handles the majority of its inbound consumer calls and chats each month; the release gives no volume or resolution-rate figures. 1	Before a rollout: define the automated share, data sources, handoff rules, and success measures.
Nvidia and Perplexity	Reuters says Nvidia is discussing an investment in a funding round valuing Perplexity above $30 billion; neither company confirmed the terms. 2	Before treating the deal as settled: verify the signed terms, cloud dependency, and economics of the agent product.
OpenAI o3	OpenAI says o3 will be retired from ChatGPT on August 26, 2026 after a 90-day sunset period; the announcement says the API is unchanged. 3	Before August 26: inventory ChatGPT workflows that depend on o3 and rerun their checks on the replacement.
European Union AI Act	The Commission says enforcement powers for prohibited practices, general-purpose AI obligations, and some transparency rules apply from August 2, 2026. 4	Now: assign an owner for model documentation, transparency, safety evidence, and regulator requests.

Verizon is turning Gemini Enterprise into an operating layer

The Google Cloud-Verizon announcement covers more than a chatbot for customer service. Google Cloud says Verizon will use its full AI stack, including advanced data infrastructure, Gemini models, and custom business agents, across customer experience, contact centers, network intelligence, marketing, cloud security, agent orchestration, and employee productivity. 1

The customer-service piece is the most concrete. Google Cloud says Gemini Enterprise for Customer Experience handles the majority of Verizon's inbound consumer calls and chats each month. The release also says the platform has produced measurable improvements in customer satisfaction and automated resolution rates, while giving no call volume, percentage, or score. 1

The data foundation matters as much as the model. Verizon is consolidating legacy data lakes into Google's Agentic Data Cloud over multiple years, according to the release. Google uses "agentic" for software that can use context, take actions, and coordinate tasks across business systems. The data move is therefore part of the deployment, rather than a separate housekeeping project. 1

A buyer should read the announcement as a scope statement and a set of vendor-reported outcomes. The next questions are operational:

Which customer-service tasks does the AI complete, and which tasks still go to a representative?
Which databases, documents, and network signals can the agents read?
What happens when the model gives an incorrect answer or starts the wrong task?
Which measures will Verizon publish so a claimed improvement can be checked?

The answers determine whether the deployment is a controlled workflow or a broad permission grant with a friendly interface.

Nvidia's Perplexity talks are also a bet on agents

Reuters reported on August 23 that Nvidia was in talks to invest in Perplexity as part of an equity round that would value the AI startup above $30 billion. The report cited The Information and said the proposed valuation would be more than 50% above Perplexity's financing a year earlier. Perplexity declined to comment, and Nvidia did not immediately respond. The size, timing, and final valuation remained unconfirmed. 2

The product detail explains why this financing story belongs in an AI briefing. Reuters said Perplexity's annualized revenue had risen above $750 million from below $250 million at the start of the year, with part of the growth driven by Perplexity Computer. Reuters described Perplexity Computer as a cloud-based AI agent that professionals use to automate computer-based tasks. 2

The reported investment would therefore connect three parts of the AI economy: Nvidia's compute business, Perplexity's agent product, and the capital needed to expand an AI service that performs work for users. The connection is a reading of the reported facts, while the financing itself remains unsettled.

That boundary matters for anyone evaluating an AI supplier. A funding headline can describe a company's direction without proving that a contract has been signed or that the product's economics work at scale. Ask for the executed terms, the cloud and compute commitments, the tasks that generate revenue, and the costs a customer carries when an agent runs many steps to complete one request.

OpenAI's o3 deadline makes model life cycles a user problem

OpenAI's Model Release Notes say that o3 will be retired from ChatGPT on August 26, 2026 after a 90-day sunset period. The same entry says GPT-4.5 left ChatGPT on June 27 after a 30-day sunset period. Both changes apply to ChatGPT; OpenAI says the API has no changes. The models were available to paid users through model settings. 3

The date turns model selection into an operating dependency. A person can build a reliable habit around a model's speed, tone, tool behavior, or formatting, then lose that behavior when the product removes the model. A team can also have saved prompts, evaluation examples, and internal instructions that were tuned against o3. Those are workflow facts a team can inspect even when the model retirement notice gives no migration recipe.

The practical work before August 26 is small and specific. List the ChatGPT tasks that use o3. Save a few representative inputs and expected outputs. Run those checks on the replacement model. Mark the tasks that need a human review, a prompt change, or a different product surface. API users can separate that exercise from the retirement notice because OpenAI's entry explicitly limits the change to ChatGPT. 3

A model catalogue is therefore part of operational risk management. The important field is not only which model performs best today. The field is also how much notice a user gets, what changes when a model leaves, and which tests prove that a replacement still fits the work.

Europe's AI Act has moved from text to enforcement

The European Commission's enforcement page was updated on August 24, 2026. The page says the AI Office enforces rules for providers of general-purpose AI models, including advanced models that may pose systemic risks. National competent authorities enforce rules for other AI systems, while the European Data Protection Supervisor handles AI systems used by EU institutions. 4

Here, "general-purpose AI" means a model that can perform many kinds of tasks and can be integrated into different AI systems. The Commission says the enforcement powers that took effect on August 2 cover prohibited AI practices, obligations for general-purpose AI providers, and transparency requirements for certain systems. Those requirements include informing people when they are interacting with a chatbot, labelling deepfakes, and embedding machine-readable marks in synthetic content. 4

The AI Office can request information, evaluate models, request access to models, require measures that can restrict public availability, interview people, and inspect providers' premises. For prohibited AI practices, the Commission lists maximum penalties of €35 million or 7% of worldwide annual turnover, whichever is higher. Other breaches, including general-purpose AI obligations, can bring penalties of up to €15 million or 3% of worldwide annual turnover. 4

The calendar has more than one date. The Commission says prohibitions related to non-consensual intimate material and child sexual abuse material apply from December 2, 2026. Rules for high-risk systems listed in Annex III apply from December 2, 2027, while high-risk AI embedded in regulated products follows on August 2, 2028. 4

A company using a general-purpose model should therefore identify its role before it writes a generic AI policy. The company may be a provider, a downstream developer, a deployer, or an operator of a service that integrates AI. Each role can change the records, notices, safety evidence, and response path the company needs. The first practical question is which legal entity owns each obligation and can answer an information request quickly.

The bottom line

Four questions separate an AI demo from an AI commitment:

Scope: Which data can the workflow read, which actions can it take, and where does a human take over?
Dependency: Which model, cloud provider, hardware supplier, and funding assumption support the service?
Life cycle: What happens when the selected model changes or retires, and which tests verify the replacement?
Accountability: Which person and legal entity keep the logs, supply the evidence, and respond when a regulator or customer asks what happened?

The week's developments put those questions in four different places: a planned enterprise rollout, a reported financing discussion, a confirmed model deadline, and a live enforcement regime. The common task is simple to state and expensive to skip: verify what the AI can do, what it depends on, when it can change, and who answers for it.