Summary Bullets:

• Data center cooling has become an increasingly insurmountable challenge because AI accelerators consume massive amounts of power.

• Liquid cooling adoption is progressively evolving from experimental to mainstream starting with AI labs and hyperscalers, then moving into the colocation space and later enterprises.

As Generative AI (GenAI) takes an ever-stronger hold in our lives, the demands on data centers continue to grow. The heat generated by the high-density computing required to run AI applications that are more resource-intensive than ever is pushing companies to adopt ever more innovative cooling techniques. As a result, liquid cooling, which used to be a fairly experimental technique, is becoming more mainstream.

Eye-watering amounts of money continue to pour into data center investment to run AI workloads. Heat management has become top of mind due to the high rack densities deployed in data centers. GlobalData forecasts that AI revenue worldwide will reach $165 billion in 2025, marking an annual growth of 26% over the previous year. The growth rate will accelerate from 2026 at 34%, and in subsequent years; in fact, the CAGR in the period 2004-2025 will reach 37%.

Source: GlobalData

The powerful hardware designed for AI workloads is growing in density. Although average density racks are usually below 10 kW, it is feasible to think of AI training clusters of 200 kW per rack in the not-too-distant future. Of course, the average number of kW per rack varies a lot, depending on the application, with traditional IT workloads for mainstream business applications requiring far fewer kW-per-rack than frontier AI workloads.

Liquid cooling is a heat management technique that uses liquid to remove heat from computing components in data centers. Liquid has a much higher thermal conductivity than air as it can absorb and transfer heat more effectively. By bringing a liquid coolant into direct contact with heat-generating components like CPUs and GPUs, liquid cooling systems can remove heat at its source, maintaining stable operating temperatures.

Although there are many diverse types of liquid cooling techniques, direct to chip is the most popular cooling method, also known as “cold plate,” accounting for approximately half of the liquid cooling market. This technique uses a cold plate directly mounted on the chip inside the server, enabling efficient heat dissipation. This direct contact enhances the heat transfer efficiency. This method allows high-end, specialized servers to be installed in standard IT cabinets, similar to legacy air-cooled equipment.

There are innovative variations on the cold plate technique that are currently under experimentation. Microsoft is currently prototyping a new method that takes the direct to chip technique one step further by bringing liquid coolant directly inside the silicon where the heat is generated. The method entails applying microfluidics via tiny channels etched into the silicon chip, creating grooves that allow cooling liquid to flow directly onto the chip and more efficiently remove heat.

Swiss startup Corintis is behind the novel technique, which blends the electronics and the heat management system that have been historically designed and made separately, creating unnecessary obstacles when heat has to propagate through multiple materials. Corintis created a design that blends the electronics and the cooling together from the beginning so the microchannels are right underneath the transistor.

Summary Bullets:

• Mistral AI’s valuation went up to EUR11.7 billion after a funding round of EUR1.7 billion spearheaded by Netherlands-based ASML.

• The French company has the edge in open source and is well positioned to capitalize on the sovereign AI trend sweeping Europe right now.

Semiconductor equipment manufacturer ASML and Mistral AI announced a partnership to explore the use of AI models across ASML’s product portfolio to enhance its holistic lithography systems. In addition, ASML was the lead investor in the latest funding round in the AI startup and now holds 11% share on a fully diluted basis in Mistral AI.

The deal holds a massive symbolic weight in the era of sovereign AI and trade barriers. Although not big in the great scheme of things and especially compared with the eye-watering sums usually exchanged in the bubbly AI world, it brings together Europe’s AI superstar Mistral with the world’s only manufacturer of EUV lithography machines for AI accelerators. ASML may not be a well-known name outside the industry, but the company is a key player in global technology. Although not an acquisition, the deal reminds of the many alliances between AI accelerator companies and AI software companies, as Nvidia and AMD continue to buy startups such as Silo AI and others. Moreover, Mistral, which has never been short of US funding through VC activity, has received a financial boost at the right time when US bidders were rumored to be circling like sharks. Even Microsoft was said to be considering buying the company at some point. For GlobalData’s take on this, please see Three is a Crowd: Microsoft Strikes Sweetheart Deal with Mistral while OpenAI Trains GPT-5. ASML becomes now its main shareholder, helping keep at bay the threat of US ownership at a critical time to reinforce one of its unique selling points: its credentials in “sovereign AI” by remaining independent from US companies.

From a technological perspective, Mistral AI has also developed a unique modus operandi, leveraging open-source models and targeting only enterprise customers, setting it apart from US competitors. Last June, it launched its first reasoning model, Magistral, focused on domain-specific multilingual reasoning, code, and maths. Using open source from the outset has helped it build a large developer ecosystem, long before DeepSeek’s disruption in the landscape drove competitors such as OpenAI to adopt open-source alternatives.

The company’s use of innovative mixture of experts (MoE) architectures and other optimizations means that its models are efficient in terms of computational resources while maintaining high performance, a key competitive differentiator. This means its systems achieve high performance per compute cost, making them more cost effective. Techniques such as sparse MoE allow scaling capacity without proportional increases in resource usage.

In February 2024, Mistral AI launched Le Chat, a multilingual conversational assistant, positioning itself against OpenAI’s ChatGPT and Google Cloud’s Gemini but with more robust privacy credentials. The company has intensified efforts to expand its business platform and tools around Le Chat, recently releasing free enterprise features such as advanced memory capabilities and capacity, and extensive third-party integrations at no cost to users. The latter includes a connectors list, built on MCP, supporting platforms such as Databricks, Snowflake, GitHub, Atlassian, and Stripe, among many others. This move will help Mistral AI penetrate the enterprise market by democratizing access to advanced features, and signals an ambitious strategy to achieve market dominance through integrated suites, not just applications.

Of course, the challenges are plentiful, Mistral AI’s scale is really far behind its US counterparts, and estimates on LLM usage seem to indicate that it is not nibbling market share away from them yet. It has a mammoth task ahead. But this deal can carve a path for European ambitions in AI, and for the protection of European assets in an increasingly polarized world divided across geopolitical lines. Some of the largest European tech companies including SAP and Capgemini have tight links to Mistral AI. They could make a bid to expand their ecosystems with acquisitions of European AI labs, that have so often fallen in US hands, in the future. For ASML, which has so many Asian customers, and whose revenues are going through a rough patch, the geopolitical turmoil of late has not been good news: this partnership brings a much-needed push in the realm of software, a key competitive enabler. After the US launched America’s AI Action plan last July, to strengthen the US leadership in AI with a plan based on removing red tape and regulation, the stakes are undoubtedly higher than ever.

Summary Bullets:

• On August 2, 2025, the second stage of the EU AI Act came into force, including obligations for general purpose models.

• The AI Act first came into force in February 2025, with the first set of applications built into law; the legislation follows a staggered approach with the last wave expected for August 2, 2027.

August 2025 has been marked by the enforcement of a new set of rules as part of the AI Act, the world’s first comprehensive AI legislation, which is being implemented in gradual stages. Like GDPR was for data privacy in the 2010s, the AI Act will be the global blueprint for governance of the transformative technology of AI, for decades to come. Recent news of the latest case of legal action, this time against OpenAI, by the parents of 16-year-old Adam Raine, who ended his life after months of intensive use of ChatGPT, has thrown into stark relief the potential for harm and the need to regulate the technology.

The AI Act follows a risk management approach; it aims to regulate transparency and accountability for AI systems and their developers. Although it was enacted into law in 2024, the first wave of enforcement proper was implemented last February (please see GlobalData’s take on The AI Act: landmark regulation comes into force) covering “unacceptable risk,” including AI systems considered a clear threat to societal safety. The second wave, implemented this month, covers general purpose AI (GPAI) models and arguably is the most important one, at least in terms of scope. The next steps are expected to follow in August 2026 (“high-risk systems”) and August 2027 (final steps of implementation).

From August 2, 2025, GPAI providers must comply with transparency and copyright obligations when placing their models on the EU market. This applies not only to EU-based companies but any organization with operations in the EU. GPAI models already on the market before August 2, 2025, must ensure compliance by August 2, 2027. For the intents and purposes of the law, GPAI models include those trained with over 10^23 floating point operations (FLOP) and capable of generating language (whether text or audio), text-to-image, or text-to-video.

Providers of GPAI systems must keep technical documentation about the model, including a sufficiently detailed summary of its training corpus. In addition, they must implement a policy to comply with EU copyright law. Within the group of GPAI models there is a special tier considered to be of “systemic risk,” very advanced models that only a small handful of providers develop. Firms within this tier face additional obligations, for instance, notify the European Commission when developing a model deemed with systemic risk and take steps to ensure the model’s safety and security. The classification of which models pose systemic risks can change over time as the technology evolves. There are exceptions: AI used for national security, military, and defense purposes is exempted in the act. Some open-source systems are also outside the reach of the legislation, as are AI models developed using publicly available code.

The European Commission has published a template to help providers summarize the data used to train their models, the GPAI Code of Practice, developed by independent experts as a voluntary tool for AI providers to demonstrate compliance with the AI Act. Signatories include Amazon, Anthropic, Cohere, Google, IBM, Microsoft, Mistral AI, OpenAI and ServiceNow, but some glaring absences include Meta (at the time of print). The code covers transparency and copyright rules that apply to all GPAI models, with additional safety and security rules for the systemic risk tier.

The AI Act has drawn criticism because of its disproportionate impact on startups and SMBs, with some experts arguing that it should include exceptions for technologies that are yet to have some hold on the general public, and don’t have a wide impact or potential for harm. Others say it could slow down progress among European organizations in the process of training their AI models, and that the rules are confusing. Last July, several tech lobbies including CCIA Europe, urged the EU to pause implementation of the act, arguing that the roll-out had been too rushed, without weighing in on the potential consequences… Sounds familiar?

However, it has been developed with the collaboration of thousands of stakeholders in the private sector, at a time when businesses are craving regulatory guidance. It is also introducing standard security practices across the EU, in a critical period of adoption. It is setting a global benchmark for others to follow in a time of great upheaval. After the AI Act, the US and other countries will find it increasingly harder to continue ignoring the calls for more responsible AI, a commendable effort that will make history.

Summary Bullets:

• OpenAI released GPT-5 on August 7, 2025, a multimodal large language model (LLM) with agentic capabilities.
• This is the latest iteration of the famous chatbot, and the most important upgrade since the release of the previous generation, GPT-4, in 2023.

As it happens sometimes when a product is thrust with such force into the realm of popular culture, the release of GPT-5 sparked a veritable PR crisis, leading CEO Sam Altman to make a public apology and backtrack on the decision to remove access to all previous AI models in ChatGPT. Unlike enterprise customers, which received advanced warnings of such movements, consumer ChatGPT users did not know their preferred models would disappear so suddenly. The ensuing kerfuffle highlighted the strange co-dependency relationship that some people have developed with the technology, creating no end of background noise surrounding this momentous release.

In truth, OpenAI handled this launch rather clumsily. But GPT-5 remains an extraordinary achievement, in terms of writing, research, analysis, coding, and problem-solving capabilities. The bête noire of generative AI (GenAI), hallucination, has been addressed (to a limited degree, of course), and GPT-5 is significantly less likely to hallucinate than previous generations, according to OpenAI. With web search enabled on anonymized prompts representative of ChatGPT production traffic, GPT-5’s responses are around 45% less likely to contain a factual error than GPT-4o. The startup claims that across several benchmarks, GPT-5 shows a sharp drop in hallucinations, about six times fewer than o3.

However, safety remains a concern. OpenAI has a patchy record in this area: Altman famously lobbied against the US California Senate Bill SB 1047 (SB 1047), which aimed to hold AI developers liable for catastrophic harm caused by their models if appropriate safety measures weren’t taken. In 2024, members of OpenAI’s safety team quit after voicing concerns about the company’s record in this area.

Meanwhile, there has been talk in industry circles and trade media outlets of artificial general intelligence (AGI) and GPT-5’s position in this regard. However, the AI landscape remains so dynamic that this is missing the point. Google’s announcement on August 5, 2025 (in limited research preview) of Google DeepMind’s Genie 3 frontier world models, which help users train AI agents in simulation environments, positions the company against AI behemoth Nvidia in the realm of world AI. World AI in this context means technologies that integrate so-called “world models,” i.e., simulations of how the world works from a physics, causality, or behavior perspective. It could be argued that this is where true AGI resides: in real-world representations and in the trenches of the simulation realm.

On the other hand, Google’s latest salvo in the enterprise space has involved a fierce onslaught of partnerships, with several deals announced in the last 48 hours. Oracle will sell Google Gemini models via Oracle’s cloud computing services and business applications through Google’s developer platform Vertex AI, an important step to boost its capillarity in corporate accounts. With Wipro, Google Cloud is going to launch 200 agentic AI solutions in different verticals that are production-ready and accessible via Google Cloud Marketplace. And with NTT Data, Google is launching industry-specific cloud and AI solutions, with joint go-to-market investments to support this important launch.

The AI market is advancing at rapid speed, including applications of agentic AI in enterprise environments. This includes a variety of AI-driven applications and platforms that are transforming business processes and interactions. The release of GPT-5 is simply another tool in this direction.

Summary Bullets:

• The Google Cloud summit was held in London (England) on July 9-10, 2025. The company said there are now over seven million developers operating within Google Vertex AI Studio.

• The company has expanded into 42 cloud regions, adding that 90% of AI unicorns are running on Google Cloud.

Google Cloud highlighted a wealth of customer cases during the Google Cloud summit held in London on July 9-10, 2025. Google Cloud has a data center in Waltham Cross (England), which will be fully operational by end-2025 to provide British businesses with high-performance computing (HPC) services. Various new capabilities for agentic AI that are aligned with business needs were highlighted. This shows strong momentum, but there are a few considerations for Google Cloud to strengthen its position in the enterprise, such as the need for stronger skills in the area of consulting, for example.

In line with the market direction, most new services and features are around agentic AI. Google Cloud unveiled plans to introduce a portfolio of purpose-built agents created through its experience in working across different industries, following patterns that are consistent across different sectors. There are other capabilities in the platform that mean users can create their own agents with no-code or low-code just by clicking a button or having a voice conversation. For an in-depth analysis of the event, please see GlobalData’s report, Generative AI Watch: Google Cloud Announces Customer Cases Leveraging Multi Agent Tools, July 11, 2025.

Google Cloud emphasized that is not necessary to build the AI agents within the Google environment. Customers can bring their own models, for example, if the company’s developers have created agents on external platforms such as Salesforce Agentforce and/or Microsoft Copilot, they can still deploy them on the Google Agentspace platform. The company is committed to building an ecosystem that gives users space and has collaborated with professional partners such as Capgemini, Accenture, and Deloitte as well as large software providers such as ServiceNow and Salesforce to ensure interoperability.

As part of the drive for interoperability, Google Cloud has launched the agent-to-agent (A2A) protocol within Google Vertex AI, its ML platform. While the model context protocol (MCP) is an open framework created by Anthropic that standardizes how large language models (LLMs) integrate data with external tools, and has enjoyed widespread adoption in the industry, Google wanted to take this a step further. To help firms integrate its prebuilt AI agents into their workflows, the A2A protocol aims to help AI agents be able to collaborate in a multi-agent ecosystem across siloed data systems and applications. A2A has had contributions from more than 50 technology partners like Langchain, MongoDB, Salesforce, SAP, ServiceNow, Accenture, BCG, Capgemini, Cognizant, Deloitte, KPMG, McKinsey, and more. The A2A protocol helps AI agents communicate with each other securely.

Fittingly for a UK summit, one of the presentations by customers included the UK government. Google Cloud EMEA has become a strategic partner of the UK Government to drive an ambitious digital transformation project to modernize the British government’s AI strategy and assets. Peter Kyle, Secretary of State for Science, Innovation and Technology (DSIT), took to the stage during the keynote to explain how his department secured almost GBP2 billion from the UK Spending Review 2025 (SR25) to implement the plan, moving straight to a hands-on product with the new trial version of the gov.uk app. Google Cloud will help it deal with a huge technical debt in the form of antiquated legacy systems that will be gradually replaced with a cloud-based service. The current contracts are very expensive, and the project is supposed to save the British taxpayer millions.

To sum up, some of the capabilities demonstrated during the event show Google Cloud’s strong momentum in the rapidly evolving cloud and AI markets. The company also showcased many customer references across multiple industries. However, competitors are also moving at a similar pace. For example, Amazon Web Services (AWS) recently added multi-agent collaboration to its Amazon Bedrock platform and is increasingly promoting open-source by adding support for Strands Agent (i.e., an open-source agent software development kit), and Model Context Protocol.

Summary Bullets:

• AMD’s “Advancing AI 2025” event, held in San Jose, California (US) in June 2025 helped analysts delve deeper into the company’s strategy for the next few years.

• The chip designer aims to build a fully open ecosystem and stack, supported by a string of acquisitions, including Silo AI and Brium.

AMD continues executing upon its annual roadmap cadence since it launched the AMD Instinct MI300 GPUs in late-2023. The launch of the AMD Instinct MI350 series, with a quadruple jump in performance compared with the previous generation, was a highlight of the conference. As AI agents become conspicuous, compute requirements will grow, driving an exponential demand for infrastructure. AMD also focused on its software roadmap and highlighted the importance of an open ecosystem, something the company has invested in through acquisitions.

The chip designer announced the launch of the AMD Instinct MI350 series GPUs, the fourth generation within the AMD Instinct family, and the forthcoming rack servers based on these chips, slated for availability in late 2025. The company is also unveiling the AMD Instinct MI400 processors in 2026, which will run on AMD’s Helios rack, pitted against Nvidia’s Vera Rubin.

AI is moving beyond the data center to intelligent devices at the edge and PCs. AMD expects to see AI deployed in every single device, running on different architectures. From a portfolio standpoint the company offers a suite of computing elements spanning GPUS, DPUS, CPUs, NICS, FPGAs, and adaptive SCIs. Its strategy is based on delivering a broad portfolio of compute engines so customers can match the right compute to the right use case, and on investing in an open, developer-first ecosystem that supports every major framework, library, and model. The chip designer believes that an open ecosystem is central to the future of AI and claims to be the only company committed to openness across hardware, software, and solutions.

Openness shouldn’t be just a buzzword because it will be critical to scale adoption of AI over the coming years. AMD has invested heavily both organically and through acquisitions to promote its open software stack; in the last year, it made 25 strategic investments in this area, including the Finnish company Silo AI, and more recently, Brium. Other acquisitions across the entire AI value chain include ZT Systems, Pensando, Lamini, Enosemi, and Xilinx. However, there are always risks associated with inorganic growth that the company needs to actively address.

However powerful AMD’s hardware may be, it is a common criticism in the industry that the software cannot match up to Nvidia’s CUDA platform. AMD has pinpointed software as a key AI enabler and therefore a crucial focus, shaping M&A plans. The ROCm 7 software stack is designed to broaden the coverage of AI models by accelerating the pace of updates and foster a developer-first mentality, with integration with open-source frameworks top of mind. This lends capillarity to the AMD hardware and makes it easier to scale.

The company highlighted that demand for compute based on inference workloads will soon be equal to model training, although training will remain the foundation to develop AI systems. As AI undertakes complex tasks like reasoning, driving demand for more compute, inference will soon become the majority stake of the market. AMD is focusing on inferencing as a crucial differentiator, with a focus on “tokens-per-dollar” as a metric.

Looking ahead, the chip designer believes there is further opportunity in an environment where customers have not invested enough in the refresh cycle of the last couple of years. However, and with the industry still relatively immature in the AI stakes, it is difficult to predict how successful the agentic AI experiment will be. Many enterprises remain in the PoC phase with lots of projects still in their infancy, and it is difficult to project the real size of the opportunity within this market. For a deeper analysis of the event, please read GlobalData’s report Advancing AI 2025: AMD Announces MI350 GPUs and Targets the Inference Opportunity, June 30, 2025.

Summary Bullets:

• Salesforce is leveraging generative AI (GenAI) capabilities to address customer pain points such as processing unstructured data and unlocking the value in this data creating unified customer profiles.

• Salesforce Data Cloud will be available on Hyperforce in the UK in July 2024. Salesforce Hyperforce aims to address customers’ growing appetite for compliance, safety, and standardization in the public cloud.

The Salesforce World Tour took place on June 6, 2024, at the Excel Centre in East London (England), with sponsors such as Amazon Web Services (AWS), Cognizant, Deloitte, and PWC. The annual event included workshops, demos, and discussions with partners, the announcement of an AI center, and innovations in the Salesforce Data Cloud and Slack platforms. For GenAI observers, the most salient news was the general availability of Salesforce Data Cloud Vector Database, built into the Salesforce Einstein 1 Platform, which infuses GenAI into the vendor’s CRM platform, Salesforce Customer 360. The vector database collects, “ingests,” and combines structured and unstructured data regarding end users. This is of great importance to Salesforce’s customers’ customers. According to the vendor, around 80% of customer data is scattered across internal corporate departments in an unstructured configuration, “trapped” in PDFs, emails, chat conversations, transcripts, customer reviews, and so on. This data can be leveraged to create a closer overall relationship with the customer by creating a unified profile of the so-called customer journey. Being able to ground all types of data in Salesforce Data Cloud – where it is processed – unlocks a ton of valuable information and not just to engage with the customer in positive ways: It makes it possible to be agile as possible in case of problems including issues such as product recall, returns, and so on.

During the keynote, it was emphasized that personalization is a critical tenet of customer engagement and one of the advantages of deploying GenAI in customer-facing verticals. “Putting data to work” was one of the highlights of the speech as well as how enterprises can augment employee productivity through upskilling to increase use of GenAI tech internally. Overcoming the fear factor and general mistrust of GenAI is also essential. Although there were no new product launches per se, the vendor announced that Salesforce Data Cloud will be available on Salesforce Hyperforce in the UK. Salesforce Hyperforce is designed to help firms tackle data residency problems by creating a layer where all Salesforce applications are integrated across the same compliance, security, privacy, and scalability standards. The solution is built for the public cloud and is composed of code rather than hardware so that all applications can be safely delivered to locations worldwide. Salesforce Hyperforce provides a common layer to deploy all the application stacks, offering Salesforce’s version of similar solutions available in the market. These solutions allow companies to handle data compliance for an increasingly fragmented technology world. Enterprises serve their employees and customers globally while providing choice and control for residency and compliance.

The event was also a launchpad for the Salesforce AI Center, whose pilot will be inaugurated in the UK to encourage collaboration among AI experts, support Salesforce partners and customers, and facilitate training and upskilling programs. The company said the center, which is planned to be the first of many globally, has capacity for 300 people and is located in Blue Fin Building near Blackfriars, London (England). Recognizing the value of training in the nascent GenAI market, Salesforce has set itself the ambitious goal of upskilling 100,000 developers worldwide leveraging a string of similar centers globally. The London facility will open on June 18, 2024 and is part of a $4 billion investment drive in the UK and Ireland.

Salesforce continues to incorporate GenAI across its portfolio from its data visualization platform Tableau, to Einstein for analytics, and Slack for collaboration. The company claims the Salesforce Data Cloud tool leverages all the metadata in the Salesforce Einstein 1 Platform by connecting unstructured and structured data, reducing the need to fine-tune large language models (LLMs) and enhancing the accuracy of the results delivered by Salesforce Einstein Copilot, Salesforce’s conversational AI assistant. Vector databases are not new, but the GenAI “revolution” has brought them to the forefront as enterprises use them alongside retrieval-augmented generation (RAG) techniques to link their proprietary data with LLMs, like OpenAI’s GPT-4, enabling them to generate more accurate results. Vector databases are becoming widespread because they power the RAG technique and are used by enterprises to build chatbots for employees needing to access internal company information (e.g., researchers using an AI hub or salespeople pulling information from knowledge hubs). Rivals including Oracle, Amazon, Microsoft, and Google have their own vector databases; Salesforce demonstrates its early investments in GenAI are bearing fruit with the Salesforce Data Cloud Vector Database launch.