UK to “mainline AI into it’s veins”, the next wave begins, and Compute boosts image generation

The UK government has unveiled a strategy for AI development, shaped by VC Matt Clifford. At its centre is a plan to increase public-sector AI computing power by 20 times before 2030, alongside new “AI growth zones” for data centres. Companies including Vantage Data Centres, Nscale, and Kyndryl have pledged £14 billion to build AI infrastructure, with plans to create 13,000 new jobs. This builds on £25 billion of tech investment commitments from the recent International Investment Summit.

We break down the key components and assess their impact:

Compute: The UK aims to multiply its AI computing power by 20 times before 2030, including a new national super cluster, expanding the AI Research Resource (AIRR). This is a clear, measurable goal that could help Britain remain competitive. Whilst the US and China dominate in the compute stakes, the UK currently sits at or near the top of a group of countries that include UAE, France, India, South Korea, Japan and Singapore. A 20x increase is welcome but is still likely to see the US stretch its lead, with planned clusters measured in gigawatts rather than the megawatt scale facilities being considered for the UK. Meanwhile, a new AI Energy Council will examine how to source generation capacity in the future including small modular nuclear reactors and renewables.

AI growth zones: New special regions for data centres, starting near Culham Science Centre in Oxfordshire are planned. While the first location is set, details about further zones and yet to be confirmed. The site has been chosen as it’s also home to the UK’s fusion reactor programme and has 100MW of generation capacity ready to go. The key criteria for these new zones will be where there is a surfeit of energy capacity such as in the North of the country and in Scotland. Peter Kyle the Secretary of State for Science, Innovation and Technology (DSIT) outlined on the Rest is Money podcast that in the coming weeks regions across the UK will be invited to apply for the programme, and the government will thus help connect infrastructure investment with the most suitable locations.

Regulation: The emphasis is innovation over restriction, clearly diverging from the EU’s comprehensive AI Act model. While new legislation is planned to address risks from frontier AI models and DSIT will shortly begin consultation, the government is pursuing a lighter-touch framework that maintains Britain’s competitive advantage. The AI Safety Institute will be established as a statutory body, but the broader regulatory landscape will rely on existing regulators adapting their approaches rather than creating new oversight mechanisms. A particularly thorny issue remains the balance between AI development and copyright protection, with the government considering pro-AI reforms to text and data mining regulations. Regulators will be required to actively promote AI innovation as part of their growth duty, with public reporting requirements on their AI-related activities. A £10 million funding package will boost regulators’ AI expertise, whilst they must publish their strategic approaches to AI by 30th April 2025.

Public datasets: Plans are also afoot to release at least five major government datasets, possibly including NHS data. A new National Data Library will manage this, though exact timelines aren’t set. Quality data access could speed up breakthroughs in healthcare, biotech, transport and many other fields.

Talent pipeline: The plan sets out targets to train tens of thousands of AI professionals by 2030. Enhanced Turing Fellowships and new graduate scholarships will help attract international talent. While headcount goals exist, the mix of university courses versus apprenticeships remains flexible. The Institute for Public Policy Research suggests this could lead to “£306 billion in yearly GDP gains, depending on AI deployment strategies.”

Increased government adoption: A new ‘Scan > Pilot > Scale’ approach will test AI solutions before wider rollout. Cabinet Office Minister Georgia Gould is leading changes to procurement rules to help British AI companies win government contracts. Success depends heavily on each department’s engagement. The strategy aims to keep promising UK startups from being acquired by US tech giants – a common pattern seen with companies like DeepMind. Major tech companies have welcomed the UK’s plan, “calling for government that is on their side.”

UK Sovereign AI: A new public-private organisation to build national AI leaders through funding and partnerships. While potentially crucial for the UK’s global AI position, the unit’s structure and accountability are still to be confirmed.

While AI could add hundreds of billions of annual GDP to the UK economy, some roles may change significantly, with rapid job destruction in high value service industries a looming possibility. The plan is also likely to put UK net zero targets under pressure, and the compute increase will need to be combined with a significant increase in power efficiency for future infrastructure. Other major concerns include the skills gap, with the flagship scholarship programme “a drop in the ocean” given the ambitious goal of training tens of thousands of AI professionals by 2030. There remains uncertainty about regulatory and legal details and in particular the ongoing consultation around IP restrictions and a potential text and data mining (TDM) exception for AI purposes, which should complete next month.

The UK’s position is a middle path between US market-driven innovation and the the EU’s rigid framework. While the US will likely maintain a regulation-light approach under Trump and China pursues state-backed development, Britain’s approach combines government coordinated infrastructure investment with reactive controls. The plan mirrors aspects of South Korea’s public-private partnership model, while its talk of responsible innovation aligns more closely with Japanese principles of agile governance. Unlike the EU’s regulation-first mindset, Britain’s strategy prioritises deployment and practical application, particularly in public services, falling back on targeted safeguards.

Takeaways: To summarise, the UK is focusing on infrastructure and pro-growth governance. Marc Andreessen, one of the leading Trump backers in Silicon Valley who was disappointed with Biden admins strategy, posted on X of the UK’s plan: “Progress and enlightenment!” Companies should look out for the regulator’s outputs at the end of April (we’ll cover those here) and when operating across multiple jurisdictions must prepare for divergence between the UK, EU, and US approaches. The UK initiatives are a welcome impetus, with global superpowers scaling up compute in gigawatts and a experimenting with a diverse mix of public-private collaborations. UK players must seize the moment for innovation and rapid adoption and be committed to the task of navigating through the ethical, economic and social disruption.

Q1 will see a slew of rumoured and highly anticipated ‘next generation’ models and agents emerge from the labs and big tech firms which will define the frontier capabilities for 2025.

xAI: Elon Musk’s AI outfit that put itself on the map in 2024 with a rapid super cluster build out:

• Grok 3: Rumours suggest that this model, trained on up to 100,000 Nvidia H100 GPUs, will surpass all currently available models in terms of pre-training compute. Grok could be available as soon as 1-2 weeks from now, but without reinforcement learning techniques and reasoning capabilities, we’re expecting it to be strong but not game-changing.

OpenAI: After a busy 12 days of shipping before Christmas several anticipated products are still to be made available to the world:

• Operator: Operator is designed to go beyond the scheduling or reminders seen in the new ChatGPT tasks feature. It aims to be a full-blown AI agent capable of performing complex tasks autonomously, such as scheduling meetings, writing documents, and automating multi-step processes like booking flights. Planned for release in January 2025.
• o3 full: Unveiled with benchmarks that were described as ground breaking, o3 set new records in reasoning, coding, and mathematical problem-solving, including an 87% score on the ARC-AGI benchmark. It’s slated for release shortly after o3-mini, likely in February or March, following safety testing and research access. Cost will be key, we can extrapolate that it likely to be a very pricey at around $30 per million input tokens and $120 for a million output tokens!
• o3-mini: The smaller model in the family, it’s designed to offer impressive performance while being adaptable with different reasoning effort levels. Public release is planned for February.
• OpenAI GPT-4.5/5/Orion: There has been no official announcement from OpenAI regarding the release of either GPT-4.5 or GPT-5. Reports and posts on X suggest that “Orion” is the internal codename for what might be publicly released as GPT-5. The Wall Street Journal and other tech publications have reported that the model is behind schedule due to technical challenges, high computing costs, and a scarcity of high-quality training data. But others have recently suggested that Orion/GPT-5 may be being used to feed high quality training data to the o-series reasoning models and supporting their rapid progress.

Google: After a big splash, Gemini 2.0’s roll-out in 2025 has been somewhat confused, with multiple experimental versions making it into the AI studio and elsewhere. 1.5 Pro, Flash, Pro with Deep Research, 2.0 Flash Experimental Preview and 2.0 Experimental Advanced Preview are all available in the Gemini web app? But what of the full family:

• Gemini 2.0 Pro / Pro Thinking: A more advanced version is now supposedly running Google Notebook LM research tool but has yet to be seen in the wider Google ecosystem. We can surely expect to see the fully functional 2.0 Pro model in the next few weeks, perhaps even enabled with the reasoning (thinking) mode that may rival or exceed OpenAI’s o1.
• Mariner, Astra, and Jules: These are Google’s agent offerings that were trailed in late 2024, showcasing the company’s push towards more advanced AI interactions with Project Mariner focusing on web navigation, Project Astra aiming to be a universal AI assistant, and Jules assisting with coding tasks. These products are likely several months away, although their release may be accelerated if OpenAI’s Operator offers significant capabilities.

Anthropic: They have been quiet of late but have tended to drop major new releases with little fanfare:

• Anthropic Claude 3.5 Opus: The release of Opus 3.5 by Anthropic has been subject to much speculation and delay. Initially, there were hints and teasers suggesting a release before the end of 2024, but it didn’t appear, although CEO Dario Amodei confirmed ongoing work, but remained vague about when it would be available. The most recent indications are that it was completed but ended up being used as a “teacher” model in a process akin to model distillation for training Sonnet 3.5, which ended up being the stand-out model in 2024.
• Anthropic Claude 4.0 Series: With perhaps models with more reasoning to come potentially for within the 3.5 range, it’s hard to guess when 4.0 will be available. Some suggest Q2, but Anthropic’s biggest headache is keeping up with demand, with Cursor the AI powered software engineering tool posting this week that they were easily the biggest Claude customer but were hitting capacity limits and having to restrict access. Amazon still seem to be behind the curve in deploying GPU infrastructure to match their competitors.

Meta: While Meta align with Trump’s vision of America, they continue for now to support the frontier of open weight models:

• Llama 4.0: Meta had aimed to have as many as 600,000 GPUs running by the end of 2024. Mark Zuckerberg confirmed during a Meta earnings call that Llama 4 is well into its training, with an anticipated release in early 2025. There are indications that Llama 4 will see multiple releases throughout the year, suggesting a strategy of iterative improvements similar to what we’ve seen with previous Llama versions. Llama 4 is expected to introduce or enhance multi-modal capabilities and advanced reasoning.

Takeaways: The AI landscape in early 2025 is set to be transformed by a wave of innovative model releases. These new products will define the state of the art and potential of AI in the first half of 2025. The successful models will need to balance three key elements: enhanced reasoning abilities, practical usefulness and instruction following, and infrastructure that can scale with demand.

We’ve covered the impact of increasing compute and adding branching search at the point of use in the context of text and OpenAI’s o1 model, and a new paper this week has explored what this can do for image generation. This picture shows the impact of increasing compute on AI image quality in several ways. The first row shows standard processing with more denoising steps, the lower set shows results from advanced search techniques. (Prompt: “Photo of an athlete cat explaining it’s latest scandal at a press conference to journalists.”)

Takeaways: Unlike LLMs which are typically fixed after training, image models can be adjusted during generation by changing the number of generation steps or using search methods to find better starting points. The research shows that additional compute at generation time leads to better quality images, with smaller models sometimes matching larger ones when given extra processing power. The team tested this with both class-based image generation and text-to-image generation. A key finding was that a model around one-tenth the size could match or outperform the much larger Flux model, suggesting efficient scaling is possible in the future for both image and video generation