What’s Changed in AI This Week? OpenAI Comes to Google Cloud, Uber Scales ChatGPT and Microsoft Pushes Sovereign AI
What’s Changed in AI This Week? OpenAI Comes to Google Cloud, Uber Scales ChatGPT, and Microsoft Pushes Sovereign AI
OpenAI, Claude and Gemini could soon be competing for your business through the same cloud platform.
OpenAI, Claude and Gemini could soon be competing for your business through the same cloud platform. OpenAI's move into Google Cloud could make it easier for businesses to access and compare leading AI models without changing providers or signing separate agreements.
Meanwhile, Uber is rolling out ChatGPT to more than 30,000 employees, Microsoft is giving organisations greater control over where AI processes their data, and Anthropic is making progress in automated AI security testing.
And there's more this week. ChatGPT has introduced locked conversations, while Google and Meta are working to make AI infrastructure more flexible.
As always, we've picked out this week's biggest developments and what they could mean for businesses so you don’t have to. So, let’s unpack all those headlines and make sense of them, shall we?
This week's AI headlines
OpenAI expands to Google Cloud
Businesses will be able to access OpenAI models through Google Cloud's existing marketplace, with eligible customers using their current cloud commitments.
Uber rolls out ChatGPT to 30,000+ employees
The company is supporting adoption with approved integrations, employee training and an internal AI skills library.
Microsoft expands sovereign AI capabilities
New Azure features will give organisations greater control over data, infrastructure and locally hosted AI models.
ChatGPT introduces locked chats
A new privacy feature lets users protect selected conversations, although enterprise retention and compliance rules still apply.
Anthropic advances automated AI safety testing
New results suggest automated red-teaming could make ongoing model security assessments more effective.
Google and Meta develop TorchTPU
A new open-source project could make it easier to run AI workloads on Google's TPUs, potentially improving infrastructure choice.
OpenAI models are coming to Google Cloud
OpenAI and Google Cloud announced on the 21st of October that OpenAI's frontier models and Codex agentic workflows will become available through Google Cloud's Model Garden and Marketplace.
Google says GPT-Astra, GPT-6.1 Sol and GPT-6 Luna are entering private preview for selected enterprise customers, with wider availability expected in the coming weeks.
For businesses already using Google Cloud, this is great news, as it could (and I’d like to say ‘should’) make adopting OpenAI considerably easier.
Basically, rather than establishing separate contracts and billing arrangements, eligible organisations will be able to purchase OpenAI services through their existing Google Cloud accounts. Usage can also count towards certain cloud spending commitments and Flexible Savings Plans.
It means businesses could potentially evaluate OpenAI alongside Gemini, Claude and other models within an environment they're already using, which makes everything much more uniform, flexible and - perhaps most importantly - secure.
That flexibility angle is what’s particularly useful, because the fact of the matter is that different models perform better for different tasks. One might be more effective at analysing information, while another could offer better performance or lower automation costs.
Businesses should be comparing the quality, speed, cost and security of each option against the work they actually need it to perform.
It's worth remembering that this is still an initial private preview, so don’t get too excited just yet. Wider availability is planned and on the horizon, but organisations shouldn't assume these capabilities are ready for production deployment straight out the gate.
Uber shows what a large-scale ChatGPT rollout looks like
On the 22nd of October, Uber announced plans to expand ChatGPT Work across more than 30,000 employees, following an initial pilot involving 1,000 people.
The rollout includes integrations with Google Drive, Gmail, Google Calendar and Slack, allowing employees to use AI across everyday business processes.
Uber expects the technology to support teams across operations, finance, marketing, and product development, as well as software engineering, through Codex.
What's particularly relevant for other businesses is how Uber is managing adoption.
The company is introducing an internal AI skills library, approved integrations and dedicated employee training, which are important considerations for any organisation introducing AI at scale.
Giving employees access to a tool doesn't automatically mean they'll use it effectively. There's also a significant difference between asking an AI assistant to summarise a document and giving an AI agent access to business systems.
For smaller organisations, the same approach can work without requiring thousands of licences.
Start with specific processes where AI could make a measurable difference. Establish which systems employees and AI agents can access, provide guidance on appropriate use and measure whether the technology improves efficiency or quality.
As it was only announced this week, Uber hasn't yet published measurable productivity or financial outcomes from the wider rollout, obviously. OpenAI and Uber say they intend to share adoption and results data later. We’ll keep an eye out for it, and share when we hear something.
Microsoft makes sovereign and local AI more practical
Microsoft's latest Sovereign Cloud announcements focus on giving organisations greater control over where their systems operate, who can access them and how information is protected.
The updates are particularly relevant to government, critical infrastructure and regulated industries, although businesses handling sensitive information may also benefit.
Microsoft's Sovereign Control Panel is expected to enter public preview at Microsoft Ignite, bringing information about sovereignty controls, security gaps and workload dependencies into a single Azure interface.
The company is also expanding Data Guardian and its tamper-evident Production Touch Ledger across the EU and EFTA, while external key management for Azure Managed HSM is now generally available.
Alongside that, there are some notable developments around AI itself. Microsoft says GitHub Copilot can use locally hosted models for inference, while Foundry Local is expected to become generally available on Azure Local, supporting both connected and disconnected environments.
All of this could, in theory, give organisations more flexibility over where AI processes information, particularly when sensitive data can't leave certain environments.
Most SMEs won't need a fully disconnected sovereign cloud. However, businesses working with regulated customers or handling confidential client information still need to understand where their data is processed and what controls apply.
That can only really start after you’ve identified which workloads can safely use cloud-based AI services and which require additional protection. Of course, the appropriate controls will depend on the sensitivity of the information, the systems involved and any contractual or regulatory requirements.
ChatGPT locked chats improve privacy, but businesses still need proper data controls
OpenAI has started rolling out locked chats, allowing users to protect selected ChatGPT conversations with a four-digit PIN or supported device authentication.
Locked conversations are removed from regular chat history and search, making them less visible when someone is sharing their screen or using a shared device.
For Enterprise and Edu workspaces, the feature is being introduced across web, mobile and desktop applications, which is a particularly useful addition for those regularly working with confidential conversations, such as HR.
However, businesses need to understand exactly what the feature protects, and we can’t stress this part enough.
Locked chats don't change an organisation's data retention settings, workspace policies or legal obligations. They also remain accessible through OpenAI's Compliance Platform for Enterprise and Edu customers.
Employees shouldn't assume that locking a conversation makes it invisible to their employer or removes it from organisational records (so don’t think you can do anything sneaky or that you shouldn’t!).
For businesses using ChatGPT, this is yet another reason to make sure staff understand which information they're allowed to enter into AI tools and what happens to that information afterwards.
Privacy features can help reduce accidental exposure, but they don't replace clear policies, appropriate permissions or effective data governance.
Anthropic's automated AI security testing shows promising results
Anthropic has published an update to its Frontier Safety Roadmap, including new findings around automated red-teaming.
Red-teaming involves deliberately testing a system to identify vulnerabilities, unsafe behaviours or weaknesses that could be exploited.
Anthropic says its automated testing system matched or exceeded the strongest human bug-bounty submissions across all 24 question-and-classifier comparisons, outperforming them in 23.
The company also reported that its automated testing took between 24 and 48 hours, compared with weeks or months for the human-led bounty process. For what it’s worth, these findings haven't yet been independently verified, and Anthropic says a more detailed technical explanation will follow.
Even so, the results suggest that automated testing could play a larger role in assessing AI systems after deployment, which is particularly relevant as businesses begin connecting AI agents to internal applications and sensitive information.
Security testing needs to account for how systems behave over time, including when models, integrations, or permissions change. Anthropic also reported that it had retired a separate oversight objective after finding that its sampling process wasn't identifying additional issues beyond existing reviews.
It's yet another useful reminder that governance processes need to deliver measurable value. Adding more checks doesn't necessarily improve security if those checks don't identify new risks.
For organisations evaluating AI providers, it's worth understanding how models are tested, how vulnerabilities are addressed, and whether the provider publishes independently reviewed evidence of its security practices.
Google and Meta make AI infrastructure more flexible
Google Cloud has open-sourced TorchTPU, a project developed with Meta that makes it easier to run PyTorch workloads on Google's Tensor Processing Units (TPUs).
PyTorch is a widely used framework for developing and running AI models. TPUs are specialised processors designed to accelerate machine learning workloads.
Google says TorchTPU will allow developers to run existing PyTorch workloads on TPUs with minimal code changes, while continuing to use familiar tools such as Hugging Face, vLLM, Ray and TorchTitan.
For most SMEs, this isn't something that requires immediate attention; however, it’s worth keeping in mind. That’s because the infrastructure used to develop and operate AI models has a direct influence on the cost and availability of AI services.
Making it easier to run workloads across different types of hardware could increase competition and give organisations more flexibility when choosing infrastructure providers, which is never a bad thing. But still, the immediate benefits are likely to be greatest for organisations training or operating their own models at scale.
For businesses buying AI services, the longer-term impact could be greater choice and potentially more competitive pricing further down the line.
What does this week's AI news mean for businesses?
This week's developments show how AI adoption is increasingly tied to the systems, processes and controls businesses already have in place.
OpenAI's expansion into Google Cloud could make it easier to evaluate models through existing suppliers. Uber's rollout demonstrates the importance of training and approved integrations, while Microsoft's developments give organisations more options for managing sensitive AI workloads.
There are a few areas worth reviewing as businesses develop their own AI strategies:
Review your existing technology providers
New AI capabilities may become available through cloud platforms and software agreements you're already paying for.
Focus on practical adoption
Identify processes where AI could make a measurable difference and give employees the training and guidance to use it effectively.
Understand where your data goes
Different workloads carry different risks, particularly when AI tools connect to internal systems or process confidential information.
Make sure employees understand the controls
Features such as locked chats can be useful, but they shouldn't create confusion about privacy, retention or organisational access.
As AI becomes more widely available, businesses will have more choice over the tools they use and how they're deployed.
Making those choices effectively will depend on having the right data, security controls and processes in place, alongside a clear understanding of what the technology is supposed to achieve.
This week's AI news FAQs
Will OpenAI models be available through Google Cloud?
Google Cloud has announced plans to make OpenAI models and Codex workflows available through Model Garden and Marketplace. Selected enterprise customers are entering private preview first, with wider availability expected later.
What can businesses learn from Uber's ChatGPT rollout?
Uber's planned expansion to more than 30,000 employees highlights the importance of training, approved integrations, an internal AI skills library and clear business use cases. Smaller businesses can apply the same principles without deploying AI at the same scale.
What is sovereign AI?
Sovereign AI refers to approaches that give organisations greater control over where AI systems run, where data is processed and which parties can access infrastructure and information. Requirements depend on the organisation's security, contractual and regulatory obligations.
Do locked chats in ChatGPT prevent employers from accessing conversations?
No. Locked chats add a privacy layer within the ChatGPT interface, but they don't override workspace retention, compliance or administrative access policies. Enterprise and Edu conversations remain subject to applicable organisational controls.
What is automated AI red-teaming?
Automated AI red-teaming uses testing systems to identify vulnerabilities, unsafe behaviour and weaknesses in AI models or applications. It can complement human security testing, particularly as models and integrations change over time.
What is TorchTPU?
TorchTPU is an open-source project developed by Google Cloud with Meta to help developers run PyTorch workloads on Google's Tensor Processing Units with fewer code changes. It could increase infrastructure flexibility for organisations developing and operating AI models.
What should UK SMEs prioritise when adopting AI?
Businesses should focus on defined use cases, employee training, approved tools, data protection, access permissions and measurable results. Choosing a model is only one part of adoption; the supporting processes and governance determine whether it delivers sustainable value.
Turn AI capability into controlled business value
We can help you identify useful AI workflows, understand the systems and data involved, control access and build governance around what happens next.
Talk to our team
Bring us the workflow, AI tool or business process you’re considering and we’ll help you work through the practical next step.
Explore AI with Fifosys
Understand how AI can fit into your business with the right foundations, data controls and governance in place.