AI Pulse

In partnership with

Master ChatGPT for Work Success

ChatGPT is revolutionizing how we work, but most people barely scratch the surface. Subscribe to Mindstream for free and unlock 5 essential resources including templates, workflows, and expert strategies for 2025. Whether you're writing emails, analyzing data, or streamlining tasks, this bundle shows you exactly how to save hours every week.

Subscribe for Your Free Bundle

Meta to spend up to $72B on AI infrastructure in 2025 as compute arms race escalates

Meta is pouring money into the physical and technical infrastructure needed to scale its AI ambitions. The company said Wednesday in its second-quarter earnings report that it plans to more than double its spend on building AI infrastructure, like data centers and servers, expecting capital expenditures of $66–72 billion in 2025, up roughly $30 billion year-over-year, to support its superclusters slated for 2026 and beyond.

Meta’s decision to allocate such enormous sums reflects the central role that large-scale compute plays in modern AI development. In its earnings call, CFO Susan Li emphasized that “developing leading AI infrastructure will be a core advantage in developing the best AI models and product experiences,” suggesting that raw spending on facilities and hardware is now inseparable from algorithmic progress. This wave of investment will underwrite multiple “titan clusters,” each envisioned to rake in gigawatts of power and house hundreds of thousands—if not millions—of top-of-the-line GPUs. By committing to this level of capital intensity, Meta aims to outpace competitors in sheer compute capacity, accelerating training cycles for next-generation large language and multimodal models.

Central to Meta’s infrastructure push are two giant AI “superclusters.” The first, codenamed Prometheus, is under construction in New Albany, Ohio, and is slated to deliver up to 1 gigawatt (GW) of compute when it goes live in 2026. Industry research firm SemiAnalysis reports that this cluster will be built using a mix of self-built campuses and leased third-party facilities, all connected via ultra-high-bandwidth network fabrics powered by Arista 7808 switches with Broadcom Jericho and Ramon ASICs. According to social media discussions and provider estimates, Prometheus may include as many as 1.3 million NVIDIA H100 GPUs, delivering over 2 exaflops of mixed-precision performance. Beyond the extreme density of accelerators, Prometheus will harness behind-the-meter natural gas generation on site to smooth out grid demands, reflecting an “all-of-the-above” approach to power procurement.

The second flagship build, known internally as Hyperion, is located in Louisiana and is envisioned to scale from an initial 2 GW up to 5 GW of raw power consumption over several years. CEO Mark Zuckerberg has likened its footprint to that of Manhattan, illustrating the sheer physical and electrical scale of Meta’s ambitions. Combined, Prometheus and Hyperion—and additional unnamed projects—signal a fundamental shift away from conventional 20-megawatt data centers toward “AI gigafactories” measured in gigawatts. This move dovetails with Meta’s larger strategic pivot toward research on “personal superintelligence,” integrating these clusters with consumer-facing AI services across social, VR, and future devices.

Behind the scenes, this infrastructure relies on sophisticated thermal and power management. Each supercluster will feature multi-level liquid cooling on GPU racks, with heat-exchange loops tied to onsite natural gas generators and regional renewable sources. Data mover blades with PCIe Gen5 and NVLink HDR interconnects will shuttle petabytes of model parameters at nanosecond latencies, while distributed object storage systems back up intermediate training checkpoints. Meta’s architects are also experimenting with “tented” GPU pods—deployable structures erected rapidly to accelerate capacity ramp-up when new chip generations arrive. The ambition is clear: bring compute online in months rather than years, shaving weeks off model iteration loops.

Yet these megaprojects carry environmental and community trade-offs. A single 5 GW site can draw as much electricity as several million homes, straining local grids and water tables. In Newton County, Georgia—home to an earlier titan cluster build—residents reported water pressure drops as high-capacity cooling pumps siphoned local groundwater. Meta insists that it will offset all carbon emissions via renewable energy credits and on-site generation, but critics worry about opaque accounting and the ultimate scalability of net-zero commitments at multi-gigawatt scale.

Talent acquisition is another key expenditure. Meta forecasts that compensation for AI engineers and researchers will become its second-largest growth driver, behind infrastructure itself. Through its new Superintelligence Labs unit, Meta is recruiting heavily from peer organizations like OpenAI and Google DeepMind, offering multimillion-dollar packages to lure experts to its Menlo Park and Seattle AI campuses. Alongside these hiring pushes, Meta is forging partnerships with leading universities to seed its talent pipeline, buying exclusive data sets for model pretraining, and sponsoring open-source AI benchmarks.

Investors have so far applauded Meta’s capital intensity. Shares rose over 10 percent in after-hours trading following the earnings release, driven by optimism that massive infrastructure spending will pay off in differentiated AI services and ad monetization tools. Even as Reality Labs—Meta’s VR division—slogged to a $4.5 billion loss, the broader market cheered the prospect of Meta delivering proprietary foundation models that power next-generation ad formats and immersive social experiences.

Looking forward, Meta’s infrastructure spree is set to reshape the competitive balance in AI. Rivals such as Microsoft and Google may feel compelled to accelerate their own multi-gigawatt builds or enter joint ventures to maintain parity. Meanwhile, hardware vendors like NVIDIA stand to benefit from a multi-trillion-dollar wave of GPU demand projected through the decade. For policymakers and regulators, questions about energy policy, antitrust, and national security may force new frameworks governing at-scale AI compute projects.

In the years ahead, Meta’s gambit will drive both the pace of AI innovation and the global conversation about sustainable compute. By pushing into the multi-gigawatt era now, Meta hopes to chart the course toward superintelligence—and secure its position at the forefront of the next technological revolution.

Conclusion

Google’s decision to sign the voluntary EU AI Code of Practice reflects a critical intersection of regulatory ambition and innovation strategy, with signatories poised to benefit from early clarity yet mindful of the trade‑offs between transparency requirements and proprietary protection. As the AI Act’s enforcement deadline approaches, the real impact of this framework will hinge on rigorous implementation, active regulatory engagement, and ongoing revisions to address emerging risks—efforts that will define Europe’s role as a global standard setter in responsible AI.

GenAI apps doubled their revenue, grew to 1.7B downloads in first half of 2025

Generative AI apps have seen tremendous growth in both downloads and in-app revenue in the first half of 2025, according to a report from market intelligence firm Sensor Tower. Users downloaded GenAI apps 1.7 billion times—versus 1 billion in H2 2024—and spent over $1.87 billion in app-purchases, up from $932 million in the prior half, as engagement soared to 15.6 billion hours across 426 billion sessions.

The explosion in GenAI app usage underscores how quickly AI features have gone mainstream on mobile devices. Sensor Tower data shows that Asia led in downloads with a 42.6 percent share—fueled by fast-growing markets like India and Mainland China—registering 80 percent growth in H1 2025 compared to 51 percent in Europe and 39 percent in North America. Meanwhile, Latin America posted the fastest in-app purchase growth, though North America retained a 40 percent share of GenAI revenue. This bifurcation highlights both the global reach of AI services and the diverse monetization dynamics across regions.

At the forefront of this surge is OpenAI’s ChatGPT app, which dominated in-app revenue in nearly every market except China, where local players like DeepSeek topped download charts. ChatGPT users logged an average of 12 days of usage per month in H1 2025—rivaling time spent on X and Reddit—and the gap between ChatGPT and leading search/browser apps narrowed dramatically. Weekend usage also rose substantially, indicating that GenAI assistants are transitioning from purely work-related tools to everyday companions for lifestyle, wellness, and shopping queries.

Beyond chatbots, a broader constellation of AI-powered apps is reshaping mobile experiences. Categories such as AI art generators, image enhancers, education aids, finance planners, and health coaches have integrated AI modules to automate creative tasks, personalize recommendations, and deliver interactive tutoring. AI-driven photo-editing apps like Lensa and Home Planner now leverage on-device LLM inferences to apply filters, generate 3D renderings, and offer voice-driven design guidance with millisecond latencies. Meanwhile, some developers have begun bundling AI inference engines with quantized model weights in their app binaries to enable offline processing and reduce cloud costs.

Commercial strategies have evolved in lockstep. Freemium models remain popular—users can access basic AI prompts for free with limited credits, then pay for premium bundles or subscriptions for bulk queries, faster response times, and priority access to new features. In-app ad platforms are also adapting, using AI to auto-generate personalized placement targets and dynamic creatives based on user prompt data. Apps adding “AI” or “LLM” to their names or descriptions saw an immediate uptick in discovery and ranking in app store charts—though this keyword boost often proved temporary without substantive feature backing.

On the supply side, developers face rising costs for hosting inference workloads in the cloud. GPU compute bills have ballooned as average LLM prompt sizes and session durations increased. To optimize, many companies are experimenting with hybrid architectures: routing high-complexity requests to cloud-based H100 clusters while handling simpler interactions on lower-cost device-AI accelerators, such as Apple’s Neural Engine or Qualcomm’s Hexagon DSPs. Frameworks like ONNX Runtime and TensorFlow Lite have added quantization and pruning pipelines to squeeze model sizes below 500 MB, enabling sub-second responses on flagship smartphones.

However, the rapid growth also brings challenges. Privacy regulators are scrutinizing how AI apps handle sensitive user prompts—particularly in health, finance, and children’s education domains. Europe’s GDPR and China’s PIPL require explicit opt-in for data sent to third-party inference APIs, and some local regulators have launched investigations into unlicensed medical advice given by chatbots. To address this, leading apps are integrating context-aware filtering layers and on-device anonymization modules to strip personally identifiable information before sending data to LLM endpoints.

Looking ahead, Sensor Tower’s report identifies six core trends reshaping the mobile AI landscape: the transition of AI assistants into primary discovery channels (Chatbot SEO), Asia’s continuing dominance as the fastest-growing market, cross-platform “stickiness” as apps deepen integration between mobile and web experiences, AI-first user acquisition tactics prioritizing genuine utility over gimmicks, expanded verticalization of AI features across all app categories, and evolving monetization through subscription bundles and enterprise licensing for B2B AI use cases.

In response, app makers and marketers must pivot their roadmaps to embed AI deeply in user flows, optimize cost structures with hybrid inference pipelines, and navigate a complex regulatory climate. As AI-driven UX becomes table stakes, only those apps that can deliver fast, accurate, and privacy-compliant experiences will sustain growth in this hyper-competitive arena.

Aker, Nscale, OpenAI plan $1 bln Norway AI facility

OSLO, July 31 – Norwegian industrial giant Aker ASA is partnering with Nscale Global Holdings and OpenAI to build Stargate Norway, a $1 billion AI “gigafactory” in Kvandal, outside Narvik, in northern Norway. The joint venture aims to install 100,000 NVIDIA AI chips by the end of 2026 and to run entirely on renewable hydropower, with potential to expand capacity tenfold in future phases.

Aker and Nscale will form a 50/50 venture to fund and operate the facility, which is poised to become one of Europe’s first large-scale AI compute sites. Nscale CEO Josh Payne described the project as “sovereign, scalable and sustainable infrastructure” essential for global competitiveness, highlighting Norway’s abundant hydropower, which can supply an initial 230 megawatts (MW) of capacity—planned to grow to 290 MW as the site scales. OpenAI, as a primary offtaker under its “OpenAI for Countries” initiative, will draw on this green compute platform to train and serve advanced AI models for European developers and researchers.

The Stargate Norway blueprint represents a strategic shift in AI infrastructure toward geo-diversified, energy-secured hubs. Europe has lagged behind the U.S. and China in gigawatt-scale AI builds, but Norway’s stable political climate, regulatory clarity, and renewable grid make it an attractive site for massive compute investments. By locating near abundant hydropower dams, the project minimizes carbon footprint while leveraging Norway’s world-class electrical and optical fiber grid for low-latency connectivity to major EU markets.

Technically, a 100,000-GPU facility could deliver exascale performance exceeding 1 exaflop of mixed-precision throughput, depending on chip model. If H100 GPUs are deployed, the cluster could power deep learning workloads across natural language processing, computer vision, and recommendation systems at a fraction of the latency experienced in Western Europe cloud region rotations. On-site PPA (power purchase agreement) structures and behind-the-meter generation options offer cost stability in an era of volatile energy prices.

Economically, the investment will spur local job creation in engineering, operations, and research collaborations. Aker has signaled partnerships with Norwegian universities to foster AI talent pipelines and spin-out startups leveraging on-site resources. Regional governments are weighing incentives, including tax breaks and R&D grants, to anchor AI development in the Arctic Circle. The venture could catalyze an AI cluster in Scandinavia, akin to Silicon Valley or Shenzhen.

Strategically, Stargate Norway underlines OpenAI’s ambition to diversify its infrastructure footprint beyond North America. As geopolitical concerns mount over data sovereignty and export controls, having a European compute base ensures resilience against shifting U.S. export regulations. It also aligns with EU goals to foster data-intensive innovation within the single market, underpinned by the European AI Act’s governance frameworks.

Challenges remain, including the logistics of deploying and maintaining tens of thousands of high-performance servers in sub-arctic conditions, securing skilled workforce in remote regions, and navigating supply chain constraints for GPU procurement—particularly amid global semiconductor shortages. However, Aker and Nscale’s combined industrial expertise in large-scale project delivery and OpenAI’s computing demand create a powerful synergy to mitigate these risks.

As other tech leaders eye similar “AI gigafactories,” Norway’s pioneering example may set the standard for how renewable energy and national infrastructure can coalesce to support next-generation AI research and services. For Europe’s digital sovereignty ambitions, Stargate Norway is more than a compute site—it is a statement of intent for sustainable, sovereign AI capacity.

China’s cyberspace regulator questions Nvidia over AI chip privacy risks

BEIJING, July 31 – China’s Cyberspace Administration (CAC) summoned representatives from Nvidia on Thursday to explain potential “backdoor security risks” embedded in its H20 artificial intelligence chips sold in China, raising fresh concerns over user data privacy and national security.

China’s CAC wants documentation on whether the H20 includes any embedded features that could enable remote tracking or shutdown of AI systems operating in Chinese data centers and edge deployments. This inquiry comes after bipartisan pressure in Washington led to the reversal of a prior export ban on the H20—an iteration of Nvidia’s A800 design—citing the chips’ slightly lower performance class as a rationale for resuming exports. Meanwhile, Nvidia has placed orders for 300,000 H20 units with Taiwan Semiconductor Manufacturing Co to meet surging demand in Greater China.

Beyond privacy, China’s regulator is probing Nvidia for alleged antitrust violations tied to its 2020 acquisition of Mellanox Technologies, indicating a broader scrutiny of U.S. technology firms under local competition laws. The H20 chips, explicitly designed to comply with U.S. export controls, were meant to bridge capability gaps while keeping the most advanced Hopper-based GPUs stateside. Yet, the CAC’s focus on “backdoor vulnerabilities” suggests an intensifying tech war, where trust in foreign-made hardware is under threat.

Security experts warn that even well-intentioned control features—such as remote firmware updates or diagnostic telemetry—could be misconstrued as spyware in the current climate. “When you bake in any remote-management hooks, adversaries may piggyback on them,” says a cybersecurity consultant specializing in hardware threats. Academic research has demonstrated that hardware trojans implanted during chip fabrication can covertly leak model parameters or infer private data from inference workloads.

The CAC’s mandate under China’s Cybersecurity Law and Personal Information Protection Law empowers it to enforce stringent requirements on all digital infrastructure. Nvidia must now furnish detailed security architectures and attestations. Some analysts predict that if Nvidia fails to satisfy these demands, China could impose additional localization rules or even develop domestic equivalents to the H20, accelerating self-sufficiency in AI hardware.

Looking ahead, this episode highlights the precarious balance between open global compute markets and sovereign security priorities. As AI workloads become critical for national competitiveness, nations are likely to demand tighter hardware provenance, supply-chain transparency, and on-chip security verifications. Industry groups now advocate for standardized security certifications for AI chips, akin to Common Criteria or FIPS benchmarks for encryption modules, to restore trust in cross-border technology flows.

Amazon updates code to keep out Google AI shopping tools

Amazon is reportedly taking steps to block outside AI shopping agents—such as Google’s Retail AI tools—from accessing and indexing its product listings, in an apparent move to preserve first-party discovery channels and protect its retail margins. According to The Information, Amazon modified its site’s robots.txt and underlying HTTP headers on July 30 to exclude known AI crawlers, including Google’s shopping agent, Perplexity, Anthropic, and OpenAI’s GPTBot.

By preventing AI bots from retrieving structured product data, Amazon aims to ensure that consumers using AI-powered assistants will still need to interact with its own search interface and recommendation engines. In tests, queries sent via Perplexity or ChatGPT yielded generic product suggestions but omitted direct Amazon links—diverting traffic back to competing retailers’ sites. For Amazon, which derives significant revenue from advertising and sponsored listings, ceding discovery to external AI platforms represents a direct threat to its core e-commerce funnel.

This tactic exemplifies a broader retail strategy: as AI assistants integrate product recommendations and transaction capabilities, the traditional “product discovery funnel” collapses. Consumers could complete purchases without ever landing on a retailer’s site, eroding visibility and ad-impression revenue. “The product discovery funnel is going to collapse as a result of consumers integrating AI into their shopping experience,” warns Michael Morton, senior analyst at MoffettNathanson. Morton predicts that any intermediary—from affiliates to price-comparison tools—may see its business model upended as users plug into conversational commerce.

Technically, Amazon’s countermeasure relies on enhanced bot management rules: its robots.txt now explicitly disallows user-agents matching Google-Extended, GPTBot, and other AI-crawler identifiers, supplemented by WAF (Web Application Firewall) signature checks to identify disguised or stealth crawlers. This layered approach echoes Cloudflare’s recent rollout of managed robots.txt tools, which let site owners block AI training crawlers like Google-Extended and GPTBot by default. However, enforcement remains imperfect—some AI crawlers ignore robots.txt entirely, necessitating CAPTCHA challenges, IP throttling, and fingerprint-based blocking.

Retail peers are watching closely. Shopify announced in June that it would bar “automated scraping” and “buy-for-me” agents unless explicitly authorized, citing risks to payment integrity and compliance. Walmart has taken a more measured approach, deferring to “wait-and-see” as AI integration evolves. But Amazon’s bold pivot may embolden other large retailers to adopt aggressive bot-blocking postures, raising questions about the open web’s future and the balance between digital access and commercial rights.

From the consumer perspective, the shift could complicate price transparency and comparison shopping. As AI assistants consolidate search, blockades at the retailer level may fragment AI-powered shopping experiences, forcing third parties to negotiate data-sharing agreements or offer premium “AI integration” APIs. The emergence of “AI walled gardens” might mirror past struggles over search index access and highlight the need for new interoperability standards in conversational commerce.

Looking forward, Amazon and other retailers will need to navigate the tension between securing their platforms and participating in an AI-driven ecosystem that favors open data sharing. Collaborative frameworks—possibly under industry consortia—could define fair use policies for AI crawlers, balancing merchant prerogatives with consumer interest in seamless, AI-powered discovery across the open web.