Welcome to The Median, DataCamp’s newsletter for September 26, 2025.
In this edition: AI infrastructure arms race heats up globally, Elon Musk’s xAI joins the race to power government AI, DeepMind releases Gemini Robotics 1.5 for physical AI tasks, OpenAI introduces ChatGPT Pulse and new team features, and DeepSeek releases V3.1-Terminus while R2 anticipation builds.
This Week in 60 Seconds
AI Infrastructure Arms Race Heats Up Globally
OpenAI announced two massive infrastructure deals this week: a landmark partnership with Nvidia to deploy at least 10 gigawatts of GPU systems and an expansion of the Stargate initiative with Oracle and SoftBank to build five new AI datacenter sites. Together, these moves underscore the scale of investment pouring into AI compute, with Nvidia committing up to $100 billion as capacity comes online. At the same time, China’s recent ban on Nvidia’s AI chips highlights the growing fragmentation of supply chains and the intensifying global race to secure AI infrastructure. We’ll explore this arms race in more depth in our Deeper Look section.
Elon Musk’s xAI Joins the Race to Power Government AI
Musk’s startup xAI has struck a deal with the U.S. General Services Administration to sell its Grok chatbot to federal agencies for just 42 cents per user over 18 months. The bargain price, which comes with integration support from xAI engineers, positions Grok directly against its rivals in the contest for government adoption. OpenAI and Anthropic previously made similar $1 per user offers, and Perplexity has also joined the race with a $0.25 offer. As we explained in a past issue, these deeply discounted “loss leader” deals reflect a broader effort by AI firms to secure long-term government adoption.
DeepMind Releases Gemini Robotics 1.5 for Physical AI Tasks
Google DeepMind has introduced Gemini Robotics 1.5, a new system designed to bring its Gemini family of models into the physical world. The release includes two components: Gemini Robotics 1.5, a vision-language-action model that translates instructions into motor commands, and Gemini Robotics-ER 1.5, an embodied reasoning model that plans tasks, reasons about the environment, and can call external tools like Google Search. The two models work together in an agentic system, allowing robots to complete complex, multi-step tasks with natural language explanations and improved spatial understanding. According to DeepMind, this marks a step toward more general-purpose physical agents, with initial access rolling out via Google AI Studio.
OpenAI Introduces ChatGPT Pulse and New Team Features
OpenAI launched a preview of ChatGPT Pulse, a new daily update experience that proactively surfaces personalized content based on chat history, memory, feedback, and connected tools like Google Calendar. Pulse is currently available to Pro users on mobile and delivers short visual cards each morning that users can curate and adjust over time. Separately, OpenAI also announced shared projects for business users, allowing teams to collaborate in ChatGPT with persistent context, files, and instructions.
DeepSeek Releases V3.1-Terminus While R2 Anticipation Builds
While many are waiting for DeepSeek R2, the team has released V3.1-Terminus, a refinement of its existing V3.1 model that focuses on stability and user feedback. The update improves language consistency (reducing CN/EN mix-ups), removes random character issues, and brings more reliable results across benchmarks. Agentic tool use saw the biggest gains, with performance improvements in tasks like BrowseComp, SimpleQA, and Terminal-bench. The model is available now via app, web, and API, with open weights published on Hugging Face.
New Course: Intermediate SQL with AI
A Deeper Look at This Week’s News
The Great AI Infrastructure Arms Race
While the headlines often focus on AI models, a quieter but more consequential race is happening underneath: the scramble to build the hardware and data centers that power them.
Over the past three years, and especially in the last few months, the U.S., China, and Europe have each escalated their commitments to AI infrastructure.
The U.S.: Resilience of location but not of supply?
In the U.S., private tech giants are leading the charge with government support. Microsoft, Google, Meta, Amazon, and OpenAI invest tens of billions annually to expand AI datacenters, often with state-level incentives.
OpenAI’s Stargate project is already building sites across Texas, Ohio, and New Mexico, targeting 10 gigawatts of AI compute—enough power to rival the consumption of small nations.
Source: OpenAI
This week, OpenAI and Nvidia announced a letter of intent for a $100 billion partnership that will supply 10 GW of GPU systems, with the first phase set to go online in 2026 on Nvidia’s Vera Rubin platform (the successor to Blackwell, designed for more efficient large-scale training).
By anchoring compute capacity in U.S. territory, the country reduces its exposure to Taiwan, where most of Nvidia’s advanced chips are still manufactured.
However, America’s AI buildout is overwhelmingly dependent on a single supplier. If Nvidia stumbles—whether through supply chain disruptions, manufacturing issues, or competitive displacement—the entire U.S. ecosystem feels it. The U.S. has achieved resilience of location but not resilience of supply.
That’s why policymakers and industry voices have pushed for multi-sourcing. Supporting AMD, Intel, and AI chip startups like Cerebras or Groq is seen as the next step toward a more balanced ecosystem.
China: Self-reliance by necessity
China’s trajectory is shaped by U.S. export controls that cut it off from top-tier Nvidia chips. Rather than back down, Beijing has doubled down. This month, authorities went so far as to bar domestic tech giants from buying Nvidia GPUs, effectively forcing companies like Alibaba, Baidu, and Tencent to adopt local alternatives.
Huawei has stepped into that gap, putting forward a roadmap to release a new Ascend AI processor every year, with performance doubling each cycle. Its upcoming Atlas 950 and 960 “supernodes” will link thousands of Ascend chips into massive AI clusters, leaning on China’s vast power grid to compensate for lower efficiency compared to Nvidia GPUs.
Europe: Sovereignty and partnerships
Europe has long lagged behind in AI compute, but it’s catching up with a focus on “digital sovereignty” (the idea that Europe should host and control the infrastructure powering its AI systems, rather than depend entirely on U.S. cloud providers or Chinese hardware).
The EU’s €20 billion AI gigafactory program and EuroHPC’s AI Factories are creating regional hubs for researchers and businesses, while national projects in France and Germany are building sovereign supercomputing campuses.
Recent milestones show progress: Germany’s JUPITER supercomputer became Europe’s first exascale system this month, while the UK announced Stargate UK, a collaboration with OpenAI and Nvidia that will deploy 31,000 GPUs on British soil.
Compute becomes geopolitics
The deeper takeaway is that AI’s trajectory may depend less on who designs the smartest algorithm and more on who controls the energy, chips, and physical infrastructure to run them.
In the 20th century, nations projected power through oil pipelines, shipping lanes, and undersea cables. In the 21st, those chokepoints may look more like GPU clusters, transmission grids, and cooling systems.
Compute has become a form of economic leverage. If a country can train the largest models and rent capacity to others, it gains not just technical leadership but bargaining power in trade, defense, and diplomacy. That’s why governments are treating datacenters and fabs like strategic assets, subsidizing them the way they once did steel mills or aircraft plants.
It also hints at a new kind of vulnerability. Energy costs, supply chain bottlenecks, or even cyberattacks on datacenters could ripple far beyond the tech sector. Just as disruptions in oil once sent shockwaves through the global economy, disruptions in compute could one day do the same for industries that come to depend on AI.
Industry Use Cases
Hyundai’s Georgia Factory Built Around AI & Robotics
Hyundai’s new Metaplant America factory, a $7.6B facility spanning 278 football fields, was designed from the ground up as an AI-first manufacturing hub. Vehicles pass through 23 AI or robotic systems, with drones, robotic arms, and even Boston Dynamics’ Spot used for quality control and logistics. The factory runs a full digital twin simulation of operations, enabling predictive maintenance and real-time defect detection. Hyundai says the approach helps lower costs and better manage disruptions.
Citi Pilots Internal AI Agents for Client Research
Citigroup has begun a 5,000-person pilot of agentic AI inside its proprietary Stylus Workspaces platform. Employees can now prompt a single AI agent to carry out multi-step tasks—like researching a client across internal systems and public sources, building a profile, and translating it—all in one go. The system uses models from vendors like Google and Anthropic, and early goals include measuring impact, usage patterns, and cost-efficiency across research and profiling workflows.
AI-Enhanced Stethoscope Enables Rapid Heart Diagnosis
Researchers at Imperial College London and Eko Health have developed an AI-enabled stethoscope that can detect heart failure, valve disease, and arrhythmias in just 15 seconds. In a UK trial involving 12,000 patients, the tool doubled the heart failure detection rate compared to traditional exams and tripled diagnoses of atrial fibrillation. The device records ECG and audio data simultaneously, which is then analyzed in the cloud using AI models trained to detect subtle cardiac patterns. While not designed for routine screening, clinicians say it could bring faster diagnoses into primary care and reduce emergency admissions.
Tokens of Wisdom
The objective is not to teach a data literacy course. The objective is to change people’s behavior.
—Jordan Morrow, Godfather of Data Literacy
In our latest DataFramed podcast, Richie Cotton and Jordan Morrow explore progress and challenges in data literacy.
Super interesting thanks!